docs(roadmap): add #448 — sandbox JSON has contradictory enabled/supported/active flags

Pinpoint: 'enabled:true, supported:false' is semantic nonsense.
'filesystem_active:true allowed_mounts:[]' contradicts 'workspace-only'.
'active:false filesystem_active:true' has no documented aggregation rule.
Renaming 'enabled' to 'requested' and exposing 'active_components:[]'
would surface real isolation state to automation.

ROADMAP.md

@@ -6292,3 +6292,133 @@ Original filing (2026-04-18): the session emitted `SessionStart hook (completed)
 350. **Top-level `plugins enable <missing-plugin> --output-format json` hangs with zero stdout/stderr instead of returning a typed plugin-not-found or unsupported-target response** — dogfooded 2026-04-29 for the 22:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `ee44ff98`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw plugins enable does-not-exist --output-format json` exited `124` with `stdout=0` and `stderr=0`; a third sample was still stuck until killed. In the same rebuilt binary, `plugins list --output-format json` returned promptly with the known plugin inventory payload, proving the plugin top-level surface is reachable and narrowing the hang to missing-plugin lifecycle mutation. This is distinct from #348's prose-only list inventory and #349's unsupported `plugins show` success-shaped JSON: #350 covers a supported lifecycle verb (`enable`) against an absent target, where the CLI should be able to fail fast before any plugin runtime work. **Required fix shape:** (a) validate the target plugin against the discovered/configured inventory before invoking enable-side effects; (b) return bounded stdout JSON such as `kind:"plugin"`, `action:"enable"`, `status:"not_found"` or `kind:"error"`, `code:"plugin_not_found"`, `plugin`, and optional `available_plugins[]`; (c) add internal timeout/diagnostic metadata for plugin lifecycle operations so registry or hook stalls do not produce silent zero-byte hangs; (d) add regression coverage proving `plugins enable does-not-exist --output-format json` returns a typed JSON outcome within a deterministic budget and does not mutate plugin state. **Why this matters:** enable/disable/update/uninstall are destructive control-plane actions. A missing or stale plugin name must fail safely and machine-readably; otherwise claws cannot preflight plugin lifecycle operations, distinguish typo from loader deadlock, or recover without killing a hung process. Source: gaebal-gajae dogfood follow-up for the 22:00 nudge on rebuilt `./rust/target/debug/claw` `ee44ff98`.
 351. **Top-level `plugins disable <missing-plugin> --output-format json` sends the JSON error envelope to stderr only, leaving stdout empty** — dogfooded 2026-04-29 for the 22:30 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `0f9e8915`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw plugins disable does-not-exist --output-format json` exited `1` with `stdout=0` and `stderr=113`; stderr contained JSON (`{"error":"plugin `does-not-exist` is not installed or discoverable","hint":null,"kind":"unknown","type":"error"}`), but stdout was empty. In the same rebuilt binary, `plugins list --output-format json` returned stdout JSON promptly with the known plugin inventory payload, proving the plugin command surface is reachable. This is distinct from #350's missing-target `plugins enable` zero-byte timeout: the disable path fails fast, but its JSON-mode error envelope is routed to stderr and uses generic `kind:"unknown"`/`type:"error"` instead of a plugin-specific stdout outcome. **Required fix shape:** (a) define and consistently document whether JSON mode emits machine-readable envelopes on stdout, stderr, or both for nonzero exits; (b) return a plugin-specific typed error with `kind:"plugin"` or `domain:"plugin"`, `action:"disable"`, `status:"not_found"` or `code:"plugin_not_found"`, `plugin`, and optional `available_plugins[]`; (c) keep stdout/stderr placement consistent across plugin lifecycle verbs so callers do not need per-action stream heuristics; (d) add regression coverage proving `plugins disable does-not-exist --output-format json` produces a typed plugin-not-found JSON contract on the documented stream. **Why this matters:** disable is a recovery/control-plane operation. A stale plugin name should be a structured, domain-specific not-found result on a predictable stream; otherwise claws that read stdout JSON for normal responses and stderr for human diagnostics must special-case this lifecycle failure. Source: gaebal-gajae dogfood follow-up for the 22:30 nudge on rebuilt `./rust/target/debug/claw` `0f9e8915`; invalid hang PR #2891 was closed after repeated bounded repros returned exit 1 with JSON on stderr.
 352. **Top-level `plugins update <missing-plugin> --output-format json` sends a generic JSON error envelope to stderr only, leaving stdout empty** — dogfooded 2026-04-29 for the 23:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `5eb1d7d8`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw plugins update does-not-exist --output-format json` exited `1` with `stdout=0` and `stderr=97`; stderr contained JSON (`{"error":"plugin `does-not-exist` is not installed","hint":null,"kind":"unknown","type":"error"}`), but stdout was empty. In the same rebuilt binary, `plugins list --output-format json` returned stdout JSON promptly with the known plugin inventory payload. This is distinct from #350's missing-target `plugins enable` zero-byte timeout and parallel to #351's `plugins disable` stderr-only JSON envelope: update fails fast, but the JSON-mode error lives on stderr only and uses generic `kind:"unknown"`/`type:"error"` instead of a plugin-specific not-found contract. **Required fix shape:** (a) define and consistently document stdout/stderr placement for JSON-mode lifecycle errors; (b) return a plugin-specific typed error with `kind:"plugin"` or `domain:"plugin"`, `action:"update"`, `status:"not_found"` or `code:"plugin_not_found"`, `plugin`, and optional `available_plugins[]`; (c) share missing-target error-envelope behavior across disable/update/uninstall and reconcile it with enable's timeout path; (d) add regression coverage proving `plugins update does-not-exist --output-format json` produces a typed plugin-not-found JSON contract on the documented stream. **Why this matters:** update is a maintenance/control-plane operation often run in automation. A stale plugin name should produce a predictable, domain-specific not-found result, not require callers to special-case stderr-only generic error envelopes after explicitly requesting JSON. Source: gaebal-gajae dogfood follow-up for the 23:00 nudge on rebuilt `./rust/target/debug/claw` `5eb1d7d8`; invalid hang PR #2894 was closed after repeated bounded repros returned exit 1 with JSON on stderr.
+353. **Top-level `plugins uninstall <missing-plugin> --output-format json` sends a generic JSON error envelope to stderr only, leaving stdout empty** — dogfooded 2026-04-29 for the 23:30 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `6f92e54d`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw plugins uninstall does-not-exist --output-format json` exited `1` with `stdout=0` and `stderr=97`; stderr contained JSON (`{"error":"plugin `does-not-exist` is not installed","hint":null,"kind":"unknown","type":"error"}`), but stdout was empty. In the same rebuilt binary, `plugins list --output-format json` returned stdout JSON promptly with the known plugin inventory payload. This is distinct from #350's missing-target `plugins enable` zero-byte timeout and parallel to #351/#352 for disable/update: uninstall fails fast, but the JSON-mode error lives on stderr only and uses generic `kind:"unknown"`/`type:"error"` instead of a plugin-specific not-found contract. **Required fix shape:** (a) define and consistently document stdout/stderr placement for JSON-mode lifecycle errors; (b) return a plugin-specific typed error with `kind:"plugin"` or `domain:"plugin"`, `action:"uninstall"`, `status:"not_found"` or `code:"plugin_not_found"`, `plugin`, and optional `available_plugins[]`; (c) share missing-target error-envelope behavior across disable/update/uninstall and reconcile it with enable's timeout path; (d) add regression coverage proving `plugins uninstall does-not-exist --output-format json` produces a typed plugin-not-found JSON contract on the documented stream. **Why this matters:** uninstall is the most destructive plugin lifecycle action. A stale plugin name should produce a predictable, domain-specific not-found result before cleanup hooks or loader work, not require callers to special-case stderr-only generic error envelopes after explicitly requesting JSON. Source: gaebal-gajae dogfood follow-up for the 23:30 nudge on rebuilt `./rust/target/debug/claw` `6f92e54d`; invalid hang PR #2897 was closed after repeated bounded repros returned exit 1 with JSON on stderr.
+354. **Top-level `memory list` and `memory help` with `--output-format json` hang with zero stdout/stderr instead of returning bounded memory inventory/help or a typed unavailable response** — dogfooded 2026-04-30 for the 00:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `19947545`. After rebuilding and verifying the binary provenance, bounded runs of `timeout 8 ./rust/target/debug/claw memory list --output-format json` produced `stdout=0` and `stderr=0`; the first sample exited `124` and the second sample was still stuck until killed. A follow-up sanity check of `timeout 8 ./rust/target/debug/claw memory help --output-format json` also exited `124` with `stdout=0` and `stderr=0`, so the issue is broader than list inventory: even the memory help path can hang silently in JSON mode. This is distinct from prior plugin lifecycle stream/status items: the affected surface is memory command introspection, where claws need safe local help/inventory before reading or mutating memory. **Required fix shape:** (a) make `memory help` and `memory list --output-format json` return bounded local JSON without requiring external/authenticated backing store availability; (b) return stdout JSON with `kind:"memory"`, `action:"help"|"list"`, `status`, usage or `entries[]`, source/provenance, counts, and truncation metadata; (c) if credentials/config/backing store are missing or slow, return a typed JSON unavailable/config/timeout error instead of hanging; (d) add regression coverage proving both `memory help --output-format json` and `memory list --output-format json` return machine-readable outcomes within a deterministic budget. **Why this matters:** memory is a core clawability surface. If even help/list can hang silently with no bytes, agents cannot tell whether memory is empty, unavailable, remote-auth blocked, or deadlocked, and any higher-level recall/debug flow stalls at the first introspection step. Source: gaebal-gajae dogfood follow-up for the 00:00 nudge on rebuilt `./rust/target/debug/claw` `19947545`.
+355. **Top-level `session list` and `session help` with `--output-format json` hang with zero stdout/stderr instead of returning bounded session inventory/help or a typed unavailable response** — dogfooded 2026-04-30 for the 00:30 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `8e24f304`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw session list --output-format json` exited `124` with `stdout=0` and `stderr=0`. A follow-up bounded `session help --output-format json` probe also produced no stdout/stderr before it had to be killed, so the issue is broader than inventory: even the session help path can silently hang in JSON mode. This is distinct from #354's memory help/list hang: the affected surface is session command introspection, where claws need a safe local way to enumerate resumable sessions or at least read usage before deciding whether to resume, inspect, or clean them up. **Required fix shape:** (a) make `session help` and `session list --output-format json` return bounded local JSON without waiting indefinitely on remote API/auth/session-store availability; (b) return stdout JSON with `kind:"session"`, `action:"help"|"list"`, `status`, usage or `sessions[]`, source/provenance, counts, and truncation metadata, or typed `status:"unavailable"`/`code` when backing state cannot be reached; (c) add explicit timeout diagnostics if a remote/authenticated session source is consulted; (d) add regression coverage proving both `session help --output-format json` and `session list --output-format json` return machine-readable outcomes within a deterministic budget. **Why this matters:** session inventory/help is a core recovery/control-plane path. If even help/list can hang silently with no bytes, claws cannot distinguish no sessions, missing credentials, remote API stall, corrupted local store, or dispatch deadlock, and resume/cleanup automation blocks before it can choose a safe next action. Source: gaebal-gajae dogfood follow-up for the 00:30 nudge on rebuilt `./rust/target/debug/claw` `8e24f304`.
+356. **Top-level `status --help --output-format json` exits successfully but emits plain text help instead of JSON** — dogfooded 2026-04-30 for the 01:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `74338dc6`. After rebuilding and verifying the binary provenance, repeated bounded runs of `./rust/target/debug/claw status --help --output-format json` exited `0` with `stdout=326` and `stderr=0`, but stdout was plain text (`Status`, `Usage`, `Purpose`, `Output`, `Formats`, `Related`) rather than a JSON object. In the same rebuilt binary, `version --output-format json` returned proper stdout JSON with version/build metadata, proving the JSON output path itself is reachable. This is distinct from #354/#355 memory/session JSON help/list hangs: the status help path returns promptly, but ignores the requested JSON format. **Required fix shape:** (a) make `status --help --output-format json` emit valid stdout JSON with `kind:"help"` or `kind:"status"`, `action:"help"`, usage, options, examples, supported output formats, and related slash/direct commands; (b) preserve text help for default/text mode only; (c) add a `format:"json"` or equivalent field so callers can assert the contract without parsing prose; (d) add regression coverage proving status help with JSON format parses as JSON and does not silently fall back to plain text. **Why this matters:** help is the discovery surface automation uses before invoking status. If `--output-format json` is accepted but help remains plain text, claws must scrape formatting-sensitive prose or special-case help output, defeating the point of machine-readable CLI contracts. Source: gaebal-gajae dogfood follow-up for the 01:00 nudge on rebuilt `./rust/target/debug/claw` `74338dc6`; invalid hang PR #2907 was closed after repeated bounded repros returned promptly.
+357. **Top-level `doctor --help --output-format json` exits successfully but emits plain text help instead of JSON** — dogfooded 2026-04-30 for the 01:30 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `52a909ce`. After rebuilding and verifying the binary provenance, repeated bounded runs of `./rust/target/debug/claw doctor --help --output-format json` exited `0` with `stdout=343` and `stderr=0`, but stdout was plain text (`Doctor`, `Usage`, `Purpose`, `Output`, `Formats`, `Related`) rather than a JSON object. In the same rebuilt binary, `status --help --output-format json` also returned promptly as plain text (#356), confirming a broader help-format fallback class while keeping this pinpoint on the doctor surface. This is distinct from #354/#355 memory/session JSON help/list hangs: doctor help returns promptly, but ignores the requested JSON format. **Required fix shape:** (a) make `doctor --help --output-format json` emit valid stdout JSON with `kind:"help"` or `kind:"doctor"`, `action:"help"`, usage, checks, options, examples, supported output formats, and related slash/direct commands; (b) preserve text help for default/text mode only; (c) add a `format:"json"` or equivalent field so callers can assert the contract without parsing prose; (d) add regression coverage proving doctor help with JSON format parses as JSON and does not silently fall back to plain text. **Why this matters:** doctor is the diagnostic entrypoint users reach for when things are broken. If JSON help falls back to prose, claws cannot discover diagnostic semantics or present structured recovery instructions without scraping formatting-sensitive text. Source: gaebal-gajae dogfood follow-up for the 01:30 nudge on rebuilt `./rust/target/debug/claw` `52a909ce`; invalid hang PR #2911 was closed after repeated bounded repros returned promptly.
+358. **Top-level `cost --help --output-format json` hangs with zero stdout/stderr instead of returning bounded command help JSON** — dogfooded 2026-04-30 for the 02:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `d95b230c`. After rebuilding and verifying the binary provenance, repeated bounded runs of `timeout 8 ./rust/target/debug/claw cost --help --output-format json` exited `124` with `stdout=0` and `stderr=0`. In the same rebuilt binary, `version --output-format json` returned promptly with version/build metadata, proving the binary itself and the JSON output path are reachable; the hang is specific to the cost help path, though other help surfaces have separate known JSON contract issues (#356/#357). **Required fix shape:** (a) make `cost --help --output-format json` return static/bounded stdout JSON with `kind:"help"` or `kind:"cost"`, `action:"help"`, usage, options, examples, supported output formats, and related slash/direct commands; (b) ensure help rendering does not initialize slow cost/session/accounting providers; (c) if any dynamic provider is accidentally consulted, return a typed JSON timeout/unavailable error instead of hanging; (d) add regression coverage proving cost help in JSON mode returns within a deterministic budget. **Why this matters:** cost/tokens surfaces are commonly consumed by automation for budgeting. If even cost help can hang silently, claws cannot discover cost command semantics or present safe budget diagnostics before running potentially slow accounting paths. Source: gaebal-gajae dogfood follow-up for the 02:00 nudge on rebuilt `./rust/target/debug/claw` `d95b230c`.
+380. **Top-level `tokens --help --output-format json` hangs with zero stdout/stderr instead of returning bounded command help JSON** — dogfooded 2026-04-30 for the 02:30 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `d95b230c`. After verifying #358 covered `cost --help`, a fresh adjacent probe on the token-budget surface showed the same silent failure class: repeated bounded runs of `timeout 8 ./rust/target/debug/claw tokens --help --output-format json` exited `124` with `stdout=0` and `stderr=0`. In the same rebuilt binary, `version --output-format json` returned promptly with version/build metadata, proving the binary itself and JSON output path are reachable. This is distinct from #358's cost help hang: the affected surface is the sibling `tokens` command help, which agents use before estimating prompt/session token budgets. **Required fix shape:** (a) make `tokens --help --output-format json` return static/bounded stdout JSON with `kind:"help"` or `kind:"tokens"`, `action:"help"`, usage, options, examples, supported output formats, and related slash/direct commands; (b) ensure help rendering does not initialize slow token accounting, session, or provider state; (c) if any dynamic provider is consulted, return a typed JSON timeout/unavailable error instead of hanging; (d) add regression coverage proving tokens help in JSON mode returns within a deterministic budget. **Why this matters:** token budgeting is a preflight clawability surface. If help hangs silently, automation cannot safely discover how to inspect or constrain token usage before running expensive prompts, and budget-aware wrappers stall at the discovery step. Source: gaebal-gajae dogfood follow-up for the 02:30 nudge on rebuilt `./rust/target/debug/claw` `d95b230c`.
+381. **Top-level `cache --help --output-format json` hangs with zero stdout/stderr instead of returning bounded command help JSON** — dogfooded 2026-04-30 for the 03:00 nudge on current `origin/main` / rebuilt `./rust/target/debug/claw` with embedded `git_sha` `d95b230c`. After #358 and #380 landed for the cost/tokens preflight help hangs, a fresh adjacent probe on the cache-control surface showed the same silent failure class: repeated bounded runs of `timeout --kill-after=1s 8s ./rust/target/debug/claw cache --help --output-format json` exited `124` with `stdout=0` and `stderr=0`. In the same rebuilt binary, `version --output-format json` returned promptly with version/build metadata, proving the binary itself and JSON output path are reachable. This is distinct from the separate `/cache` slash-command envelope mismatch class: the affected surface here is top-level `cache` command help, where agents need bounded local discovery before deciding whether to inspect, clear, or summarize cache state. **Required fix shape:** (a) make `cache --help --output-format json` return static/bounded stdout JSON with `kind:"help"` or `kind:"cache"`, `action:"help"`, usage, options, examples, supported output formats, and related slash/direct commands; (b) ensure help rendering does not initialize slow cache/session/provider state; (c) if any dynamic provider is consulted, return a typed JSON timeout/unavailable error instead of hanging; (d) add regression coverage proving cache help in JSON mode returns within a deterministic budget. **Why this matters:** cache inspection and cleanup are recovery/control-plane operations. If cache help hangs silently, claws cannot safely discover cache semantics before attempting cleanup, and automation stalls before it can choose a non-destructive cache action. Source: gaebal-gajae dogfood follow-up for the 03:00 nudge on rebuilt `./rust/target/debug/claw` `d95b230c`.
+
+422. **`export --output-format json` and `--resume latest` report the same "no managed sessions" scenario using two different `kind` codes — `no_managed_sessions` vs `session_load_failed` — making "no session found" undetectable by a single kind-code check** — dogfooded 2026-04-30 KST (UTC+9) by Jobdori on `e939777f`. Running `claw export --output-format json` with no session present returns (on stderr, exit 1): `{"error":"no managed sessions found in .claw/sessions/<fingerprint>/","hint":"Start \`claw\` to create a session, then rerun with \`--resume latest\`.\nNote: claw partitions sessions per workspace fingerprint; sessions from other CWDs are invisible.","kind":"no_managed_sessions","type":"error"}`. Running `claw --resume latest /status --output-format json` with no session present returns (on stderr, exit 1): `{"error":"failed to restore session: no managed sessions found in .claw/sessions/<fingerprint>/","hint":"Start \`claw\` to create a session, then rerun with \`--resume latest\`.\nNote: claw partitions sessions per workspace fingerprint; sessions from other CWDs are invisible.","kind":"session_load_failed","type":"error"}`. Both describe the same root condition — there are no sessions to operate on — but they expose it via different `kind` discriminants. Automation that checks `kind == "no_managed_sessions"` to detect a cold workspace will miss the `--resume` path's `session_load_failed`, and vice versa. A wrapper that guards "run with --resume only if a session exists" must special-case both codes. The hint text is identical between them, suggesting the messages are logically equivalent. Additionally neither code matches the proposed canonical names `session_not_found` / `session_load_failed` as stable `ErrorKind` discriminants described in ROADMAP #77's fix shape, which explicitly proposes typed error-kind codes for session lifecycle failures. **Required fix shape:** (a) unify "no sessions found for this workspace fingerprint" under a single canonical `kind` code — either `no_managed_sessions` or `session_not_found` — used consistently by every command path that encounters an empty session registry; (b) if `session_load_failed` is a more general category (covering e.g. corrupt session files, IO errors, schema version mismatches), it should nest a concrete `reason:"no_managed_sessions"` or `reason:"session_not_found"` sub-field so callers can distinguish "empty registry" from "found but unreadable"; (c) align with the canonical error-kind contract proposed in #77; (d) add regression coverage proving `export` and `--resume latest` in an empty workspace both return an error with the same top-level `kind` code. **Why this matters:** session guard-rails in orchestration need a single stable `kind` to detect cold workspaces without enumerating all possible no-session synonyms. Two divergent codes for the same condition make defensive automation brittle and contradict the promise of machine-readable error envelopes. Source: Jobdori live dogfood, `e939777f`, 2026-04-30 KST (UTC+9).
+
+407. **`config --output-format json` returns `files[].loaded:false` with no `load_error`, `not_found`, or `skip_reason` field — automation cannot distinguish "file does not exist", "file exists but parse failed", and "file exists but was skipped by policy" from the same `loaded:false` value; also `loaded_files` and `merged_keys` are bare integers with no per-file attribution** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `./claw --output-format json config` on a workspace with 5 discovered config files returns `{"kind":"config","cwd":"...","files":[{"loaded":false,"path":"/Users/yeongyu/.claw.json","source":"user"},{"loaded":true,"path":"/Users/yeongyu/.claw/settings.json","source":"user"},{"loaded":true,"path":"/Users/yeongyu/clawd/claw-code/.claw.json","source":"project"},{"loaded":false,"path":"/Users/yeongyu/clawd/claw-code/.claw/settings.json","source":"project"},{"loaded":false,"path":"/Users/yeongyu/clawd/claw-code/.claw/settings.local.json","source":"local"}],"loaded_files":2,"merged_keys":2}`. Three of five files have `loaded:false` with no accompanying `not_found:true`, `parse_error`, `io_error`, or `skip_reason`; automation must stat each path separately to guess why. Also `loaded_files:2` and `merged_keys:2` are bare counts — ambiguous whether `merged_keys:2` means 2 total top-level JSON keys across all files or 2 unique merged settings. **Required fix shape:** (a) add `not_found: bool` and optional `load_error: string` to each `files[]` entry so callers can distinguish missing, parse-broken, and policy-skipped files without filesystem probing; (b) document or rename `merged_keys` as `merged_setting_count` or `total_merged_keys` to remove the int-semantics ambiguity; (c) optionally add `merged_keys_by_file: [{path, keys}]` for attribution; (d) add regression coverage proving `files[]` entries with `loaded:false` carry at minimum `not_found` distinguishing non-existent paths from load failures. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+408. **`status --output-format json` `workspace.changed_files` is ambiguous — on a workspace with 5 untracked files, `changed_files:5`, `staged_files:0`, `unstaged_files:0`, `untracked_files:5`; it is unclear whether `changed_files` is the sum of all four git-status categories or only a subset; automation cannot tell if `changed_files:5` means "5 tracked modified" or "5 total non-clean files including untracked"** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `./claw --output-format json status` returns `{"workspace":{"changed_files":5,"staged_files":0,"unstaged_files":0,"untracked_files":5,...}}` — `changed_files==untracked_files==5` with staged and unstaged both zero. The field name `changed_files` implies "modified tracked files" but the value equals the untracked count, not `staged+unstaged`. Without a comment or documented definition, automation must probe whether `changed_files = staged + unstaged` (excludes untracked) or `changed_files = staged + unstaged + untracked + conflicted` (total dirty). Also `git_state:"dirty · 5 files · 5 untracked"` repeats the same data as a prose string alongside the structured integer fields — redundant human-readable string alongside machine-readable integers. **Required fix shape:** (a) document and stabilize `changed_files` as either `tracked_dirty_count` (staged+unstaged only) or `total_non-clean_count` (staged+unstaged+untracked+conflicted) and rename to remove the ambiguity; (b) ensure a machine consumer can compute `is_clean` as a single boolean field without interpreting `git_state` prose; (c) deprecate or remove `git_state` prose string now that all its constituent counts are available as integers; (d) add regression coverage proving `changed_files` semantics against a workspace with staged, unstaged, untracked, and conflicted files. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+409. **`init --output-format json` emits redundant parallel artifact schemas — `artifacts[].status` and flat `created[]`/`skipped[]`/`updated[]` arrays carry identical state, and `artifacts[].status:"skipped"` omits `skip_reason`** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw init --output-format json` on a fresh directory returns a JSON object with two parallel representations of the same artifact set: (1) `artifacts: [{name, status}]` — a structured per-artifact array; and (2) `created: [...]`, `skipped: [...]`, `updated: [...]` — flat string arrays partitioned by status. Both encode the same four artifact names and their outcomes with no additional information between them. On a subsequent run in an already-initialized directory, every artifact has `status:"skipped"`, but no `reason` field is present on any artifact entry — automation cannot distinguish `"already_exists"` (safe to ignore) from `"permission_denied"`, `"dry_run"`, or `"conflicting_contents"` (each requiring a different response). The `message` field also embeds `"skipped (already exists)"` prose that is absent from the structured payload. **Required fix shape:** (a) pick one canonical artifact representation — either `artifacts[{name, status, reason?, path?}]` or the flat status arrays — and deprecate the other; (b) add a `skip_reason` or `reason` field to `artifacts[]` entries with `status:"skipped"` and `status:"error"`, using an enum such as `already_exists`, `permission_denied`, `dry_run`, `conflict`, `unknown`; (c) add optional `path` (absolute) to each artifact entry so automation can act on the real on-disk location without re-joining with `project_path`; (d) add regression coverage proving `init --output-format json` on an existing directory includes machine-classifiable skip reasons for every skipped artifact and does not rely on the prose `message` field for structured state. **Why this matters:** init is the bootstrapping surface automation uses to ensure a project is claw-ready. If skip classification requires parsing human prose and the structured payload has two redundant formats, claws either over-provision re-inits or cannot distinguish safe skips from blocked writes without brittle message scraping. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+410. **`agents list`, `skills list`, and `mcp list` use three different count-field names and divergent envelope schemas despite being sibling list commands** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running all three list commands with `--output-format json` reveals incompatible envelope shapes: `agents list` emits `count:int` at the top level plus `summary:{active,shadowed,total}` and `working_directory`; `skills list` emits no top-level `count`, only `summary:{active,shadowed,total}`, and omits `working_directory`; `mcp list` uses a different count-field name `configured_servers:int`, has no `count`, no `summary`, and instead adds `status:"ok"` and `config_load_error:null` fields absent from the other two. The three sibling commands cannot be polymorphically consumed with the same count-extraction logic, requiring per-command special-casing at the cardinality check level. **Required fix shape:** (a) define one canonical top-level count field name (`count`, `total`, or `item_count`) and use it across `agents`, `skills`, and `mcp` list envelopes; (b) define one canonical `summary` object shape with at minimum `active`, `total`, and optionally `shadowed` and include it on all three; (c) expose `working_directory` consistently on all list commands or omit it from all; (d) add regression coverage proving the three list envelopes share the same count-field name and summary shape before each release. **Why this matters:** orchestration lanes that inventory agents, skills, and servers before delegation need one count-extraction pattern. Three different field names force per-command special-casing of the most basic cardinality check. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+411. **`plugins enable/disable --output-format json` always emits `reload_runtime:true` regardless of whether state actually changed, and omits `previous_status`, `changed`, `version`, and `source` fields — automation cannot tell if a reload is necessary or if the mutation was a no-op** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw plugins enable example-bundled --output-format json` on an already-enabled plugin returns `{"action":"enable","kind":"plugin","message":"…","reload_runtime":true,"target":"example-bundled"}` — `reload_runtime:true` every time, even on a no-op re-enable. The same applies to idempotent `disable`. Structured fields present: `action`, `kind`, `message`, `reload_runtime`, `target`. Structured fields absent: `previous_status`, `status`, `changed`, `version`, `source`. The actual plugin name, version, and new status are embedded only in the prose `message` field (`"Result  enabled example-bundled@bundled\n  Name  example-bundled\n  Version  0.1.0\n  Status  enabled"`), requiring callers to scrape column-aligned text to extract the post-mutation state. A no-op mutation emitting `reload_runtime:true` forces orchestration to trigger an expensive runtime reload even when no config change occurred. **Required fix shape:** (a) add `changed:bool` so callers can skip runtime reload when `changed:false`; (b) add `previous_status` and `status` fields (enums: `enabled`/`disabled`) so pre/post state is machine-readable without parsing `message`; (c) add `version` and `source` fields at the mutation response level, consistent with `plugins list` entry shape; (d) emit `reload_runtime:false` when `changed:false`; (e) add regression coverage proving idempotent enable/disable sets `changed:false` and `reload_runtime:false`. **Why this matters:** plugin lifecycle is a hot path for automation that conditionally enables plugins before running sessions. If every enable emits `reload_runtime:true` and no `changed` field exists, orchestration must reload unconditionally or maintain external state — both brittle patterns. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+412. **`bootstrap-plan --output-format json` returns `phases: string[]` of raw Rust enum variant names with no description, steps, duration, or dependency metadata — unusable by automation** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw bootstrap-plan --output-format json` returns `{"kind":"bootstrap-plan","phases":["CliEntry","FastPathVersion","StartupProfiler","SystemPromptFastPath","ChromeMcpFastPath","DaemonWorkerFastPath","BridgeFastPath","DaemonFastPath","BackgroundSessionFastPath","TemplateFastPath","EnvironmentRunnerFastPath","MainRuntime"]}`. The envelope has only two keys: `kind` and `phases`. The `phases` array contains 12 raw Rust enum variant name strings — opaque identifiers with no `description`, no `label`, no `steps[]`, no `estimated_ms`, no `dependencies[]`, no `optional:bool`, and no `status` (enabled/disabled/skipped). Automation that calls `bootstrap-plan` to understand startup costs or profile initialization paths receives 12 name strings that reveal nothing about what each phase does, how long it takes, whether it depends on credentials/network/MCP, or which ones can be skipped. **Required fix shape:** (a) replace `phases: string[]` with `phases: [{id, label, description, optional, estimated_ms?, dependencies?, status?}]`; (b) add a top-level `total_phases` count; (c) mark network/credential-dependent phases with a `requires_auth:bool` or `deps:["network","credentials","mcp"]` field so automation can plan for unavailability; (d) add regression coverage proving each phase entry has at least `id`, `label`, and `description` fields and that the count matches the phases array length. **Why this matters:** bootstrap-plan is the startup-cost introspection surface. If its JSON output is 12 opaque variant name strings, automation cannot profile startup, identify slow phases, skip optional phases, or present meaningful startup diagnostics — the entire command serves only as a list of internal identifiers. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+413. **`acp --output-format json` leaks internal ROADMAP tracking numbers and implementation notes as top-level JSON fields — `discoverability_tracking:"ROADMAP #64a"` and `tracking:"ROADMAP #76"` are internal backlog references that should not appear in the public machine-readable contract** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw acp --output-format json` returns a ten-key envelope: `aliases`, `discoverability_tracking`, `kind`, `launch_command`, `message`, `recommended_workflows`, `serve_alias_only`, `status`, `supported`, `tracking`. Two fields are verbatim internal backlog cross-references: `"discoverability_tracking":"ROADMAP #64a"` and `"tracking":"ROADMAP #76"`. These were presumably used during initial scaffolding to track which backlog items the stub relates to, but they are now part of the public JSON contract that automation consumes. The `message` field also contains implementation-note prose (`"ACP/Zed editor integration is not implemented in claw-code yet. \`claw acp serve\`..."`) that describes the build state rather than the command's machine-readable status. **Required fix shape:** (a) remove `discoverability_tracking` and `tracking` from the public JSON envelope or move them to an optional `_debug` or `_meta` sub-object gated on a debug flag; (b) replace `message` prose with a structured `reason` enum (`"not_implemented"`, `"discoverability_only"`, `"serve_only"`) plus optional `detail` string; (c) rename `supported:false` + `status:"discoverability_only"` to a single typed `availability` object with `status`, `reason`, and `target_command` fields; (d) add regression coverage proving the public `acp --output-format json` envelope contains no internal tracking/backlog fields and that `message` is not the sole machine-classifiable signal. **Why this matters:** public JSON APIs should not leak internal ticket references. Automation that snapshots or validates the ACP JSON schema will embed these internal identifiers into external contracts and need to change every time backlog numbering shifts. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+415. **`config <section> --output-format json` returns `merged_keys:int` (a count) with no actual merged key-value pairs — automation cannot read the resolved configuration values from JSON** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw config env --output-format json`, `claw config model --output-format json`, or `claw config hooks --output-format json` all return an identical five-key envelope: `{"cwd":"...","files":[...],"kind":"config","loaded_files":2,"merged_keys":1}`. The `merged_keys` field is an integer count of how many keys were merged across the loaded files, not an object or array of the actual key names and resolved values. The `files` array shows which config files were loaded/missing but contains no per-file key-value content. The merged section content — the actual resolved `env`, `model`, or `hooks` configuration — is entirely absent from the JSON output. It only appears in the prose output as a "Merged section: env / <value>" block. **Required fix shape:** (a) add a `merged` or `resolved` object/array field to the JSON envelope containing the actual key-value pairs that resulted from merging the loaded config files for the requested section; (b) rename `merged_keys` from an integer count to either remove it (derivable from `len(merged)`) or keep it as a companion count field; (c) for each entry in `merged`, include `key`, `value`, and optionally `source_file` so automation can attribute which file contributed the value; (d) add regression coverage proving `config env --output-format json` with a non-empty env section populates `merged` (or equivalent) with the actual resolved key-value pairs. **Why this matters:** the entire purpose of `config env/model/hooks --output-format json` is to allow automation to read the resolved runtime configuration without screen-scraping prose. Returning only a count defeats the purpose and forces callers to either re-parse the prose output or re-read and merge the source config files themselves. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+416. **`plugins list --output-format json` returns the mutation response shape with a prose `message` table instead of a structured `plugins:[]` array — `name`, `version`, `status`, `source` are embedded in `message` prose only** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw plugins list --output-format json` returns `{"action":"list","kind":"plugin","message":"Plugins\n  example-bundled      v0.1.0      disabled\n  sample-hooks         v0.1.0      disabled","reload_runtime":false,"target":null}`. This is the same four-key response envelope used by `plugins enable` and `plugins disable` mutation commands, not a list envelope. The `message` field contains the full rendered prose table (plugin name, version, and status as whitespace-aligned columns), but no `plugins` array with structured per-entry objects. `target` is `null` because no specific plugin was targeted. The `reload_runtime:false` field is meaningless for a read-only list operation. **This is distinct from ROADMAP #411** which covers the mutation commands' own missing `changed`/`previous_status`/`version`/`source` fields — #416 targets the list command's structural mismatch: it uses the mutation envelope entirely instead of emitting a dedicated list schema. **Required fix shape:** (a) emit a distinct `{kind:"plugin_list", plugins:[{name, version, status, source, path?, description?}], count}` envelope for the `list` action; (b) omit `action`, `reload_runtime`, and `target` from list responses (mutation-only fields); (c) the `message` field should be absent or optional and must not be the sole machine-readable inventory surface; (d) add regression coverage proving `plugins list --output-format json` populates a `plugins` array with at least `name`, `version`, and `status` fields for each installed plugin. **Why this matters:** automation that calls `plugins list --output-format json` to discover installed plugin inventory receives only a whitespace-aligned prose table in a string field, with `reload_runtime:false` and `target:null` as the only other machine-readable signals — identical noise to what a failed enable command returns. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+418. **`system-prompt --output-format json` exposes `"__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__"` as a literal element in the `sections` array — an internal split delimiter leaked into the public structured output** — dogfooded 2026-04-30 by Jobdori on `e939777f`. Running `claw system-prompt --output-format json` returns `{"kind":"system-prompt","message":"<full prose>","sections":["You are an interactive agent...", "# System\n...", "# Doing tasks\n...", "# Executing actions with care\n...", "__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__", "# Environment context\n...", "# Project context\n...", "# Claude instructions\n...", "# Runtime config\n..."]}`. The `sections` array has 9 elements; element index 4 is the raw string `"__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__"`. This internal sentinel marks the boundary between the static and dynamic sections of the compiled system prompt, used during assembly to split the prompt at injection time. It appears in the public JSON output verbatim as a first-class section, indistinguishable from real sections by type alone. Automation that iterates `sections[]` must special-case this sentinel or it will process an internal implementation string as if it were a real system prompt section. **Required fix shape:** (a) strip `"__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__"` and any similar internal delimiters from the `sections` array before serializing to JSON; (b) if the static/dynamic boundary is semantically meaningful for callers, expose it as a structured metadata field such as `boundary_index:4` or as a `section_type:"static"|"dynamic"` field on each section entry, not as a raw sentinel string in the array; (c) rename the `sections` type from `string[]` to `[{id, type, content}]` to enable this without breaking the boundary signal; (d) add regression coverage proving the `system-prompt --output-format json` output's `sections` array contains no elements whose value equals `"__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__"` or matches `/__[A-Z_]+__/`. **Why this matters:** internal sentinel strings in public JSON are a contract liability — they couple the wire format to internal implementation details. Any refactor that renames or removes the sentinel breaks callers that don't special-case it, and automation that doesn't know to filter it will miscount, misparse, or misrender the system prompt. Source: Jobdori live dogfood, `e939777f`, 2026-04-30.
+
+
+419. **`mcp <unknown-subcommand> --output-format json` returns `action:"help"` + `unexpected:<arg>` with exit 0 instead of an error envelope — unrecognized MCP subcommands silently succeed** — dogfooded 2026-05-01 by Jobdori on `e939777f`. Running `claw mcp add --output-format json` or `claw mcp remove --output-format json` (subcommands that do not exist) returns exit 0 with stdout JSON `{"action":"help","kind":"mcp","unexpected":"add","usage":{"direct_cli":"claw mcp [list|show <server>|help]","slash_command":"/mcp [list|show <server>|help]","sources":[...]}}`. Exit code is 0. The `action` field is `"help"` — not `"error"` — even though the caller issued a recognized token (`add`/`remove`) that maps to a real but unimplemented feature. The `unexpected` field correctly identifies the unrecognized arg, but automation that checks `exit == 0` or `action != "error"` will treat this as a successful invocation. This is distinct from ROADMAP #108 which covers *unrecognized CLI subcommands* falling through to the LLM prompt path — #419 targets MCP-specific *known-but-unimplemented* subcommands that return `action:"help"` with exit 0 instead of an explicit `action:"error"` envelope. **Required fix shape:** (a) return a non-zero exit code (exit 1 or exit 2) when an unrecognized or unimplemented MCP subcommand is provided; (b) emit `action:"error"` (or `kind:"error"`) with a `code:"unknown_subcommand"` and `unknown:"add"` field instead of `action:"help"`; (c) optionally include the help/usage payload as a sibling field `suggestion:{usage:{...}}` for context; (d) add regression coverage proving `mcp <unknown> --output-format json` returns a non-zero exit code and a non-help action token. **Why this matters:** `add` and `remove` are common MCP lifecycle operations that users will attempt; returning `action:"help"` with exit 0 makes these look like successful no-ops to any automation that doesn't deep-inspect the `unexpected` field. A pipeline that runs `claw mcp add my-server ... && claw mcp show my-server` will silently proceed to the show step even though add silently no-oped. Source: Jobdori live dogfood, `e939777f`, 2026-05-01.
+
+
+420. **`plugins help --output-format json` returns the mutation response shape (`message`, `reload_runtime`, `target`) instead of the help envelope (`action:"help"`, `kind`, `unexpected`, `usage`) that `mcp help`, `agents help`, and `skills help` all use — schema drift within the same command family** — dogfooded 2026-05-01 by Jobdori on `e939777f`. Running `claw plugins help --output-format json` returns `{"action":"help","kind":"plugin","message":"Unknown /plugins action 'help'. Use list, install, enable, disable, uninstall, or update.","reload_runtime":false,"target":null}`. By contrast, `claw mcp help --output-format json`, `claw agents help --output-format json`, and `claw skills help --output-format json` all return a help envelope: `{"action":"help","kind":"<surface>","unexpected":null,"usage":{"direct_cli":"...","slash_command":"...","sources":[...]}}`. The `plugins` subgroup has not adopted the help envelope schema used by all sibling subgroups. Instead it uses the mutation response shape (`message`, `reload_runtime`, `target`) with an error string in `message` that calls `help` an "unknown action." Automation that checks `usage.direct_cli` to discover plugin commands gets a `TypeError` (key not found) on the plugins help path while succeeding on all sibling subgroups. **Required fix shape:** (a) make `plugins help` return the same help envelope as `mcp help`/`agents help`/`skills help`: `{action:"help", kind:"plugin", unexpected:null, usage:{direct_cli:"claw plugins [list|enable|disable|install|uninstall|update|help]", slash_command:"/plugins [...]", sources:[...]}`; (b) drop `reload_runtime` and `target` from help responses for all plugin subcommands; (c) add regression coverage proving `plugins help --output-format json` contains a `usage.direct_cli` field matching the same envelope shape as `mcp help`/`agents help`/`skills help`; (d) audit all subgroup `help` handlers for the same mutation-envelope contamination. **Why this matters:** help discovery is the bootstrap surface for automation. If `plugins help --output-format json` returns a mutation envelope with an error message instead of a usage envelope, automated schema discovery fails silently for the entire plugins subgroup while working for every other subgroup. Source: Jobdori live dogfood, `e939777f`, 2026-05-01.
+
+
+421. **`status`, `mcp list`, `doctor` JSON output leak macOS `/private` symlink-canonicalized cwd instead of user-invocation cwd — automation that string-matches on cwd breaks across symlinked filesystems** — dogfooded 2026-05-11 by Jobdori on `b98b9a71` in response to Clawhip pinpoint nudge at `1503207549447573574`. Reproduction on macOS: invoke from `/tmp/claw-dog-cwd` (where `/tmp` symlinks to `/private/tmp`), then `claw status --output-format json` returns `workspace.cwd: "/private/tmp/claw-dog-cwd"`, `claw mcp list --output-format json` returns `working_directory: "/private/tmp/claw-dog-cwd"`. The user's invocation cwd (`$PWD`, `pwd`) is `/tmp/claw-dog-cwd`. Source: `session_control.rs:34` calls `fs::canonicalize(cwd)` for #151 cross-worktree session-bleed prevention, then leaks the canonicalized path through every JSON envelope that reports cwd. **Required fix shape:** (a) keep canonicalized cwd for session keying internally, but report user-input cwd (the value passed by `env::current_dir()` or `--cwd` flag) in JSON output as `cwd`; (b) optionally expose canonical path as a separate field `cwd_canonical` for diagnostic purposes; (c) audit every `--output-format json` surface that emits `cwd` / `working_directory` / `workspace.cwd` for the same leak (status, mcp list, doctor, session list, init, etc.); (d) add regression coverage proving JSON cwd matches `$PWD` on macOS where `/tmp -> /private/tmp` symlink exists. **Why this matters:** automation pipelines that route work to lanes by cwd, or that compare cwd against a registry, break across macOS hosts because the canonicalized form differs from the form the user/orchestrator passed. The leak is silent — no documentation indicates the path will be rewritten. Source: Jobdori live dogfood, `b98b9a71`, 2026-05-11.
+
+
+422. **Unknown top-level subcommands fall through to chat prompt path instead of returning `unknown_subcommand` error — typos silently send the subcommand string as a chat message to the configured LLM** — dogfooded 2026-05-11 by Jobdori on `b98b9a71` in response to Clawhip pinpoint nudge at `1503215095088676956`. Reproduction: `unset ANTHROPIC_AUTH_TOKEN; export ANTHROPIC_API_KEY=fake-key-for-routing-test; claw completely-bogus-subcommand --output-format json` returns `{"error":"api returned 401 Unauthorized (authentication_error) [trace req_011...]: invalid x-api-key","kind":"api_http_error"}` — proving the unknown token reached the Anthropic API endpoint as a chat prompt. With valid credentials, the bogus subcommand string would be silently consumed as a chat message, billing the user for a typo and producing whatever continuation the LLM generates. **Pre-error path:** `claw <unknown> --output-format json` with no creds returns `kind:"missing_credentials"` (the auth gate fires first), masking the routing bug. Only with creds present does the fallthrough manifest as the actual prompt being sent. **Sibling exit-code bug:** when the chat-path 401 returns, the JSON envelope is `kind:"api_http_error"` but exit code is **0**, while `cli_parse` errors (e.g. `--no-such-flag`) and `missing_credentials` errors correctly exit **1**. Exit-code parity between error envelopes is broken — automation that gates on `$?` will treat the 401-as-chat as success. **Required fix shape:** (a) reserve unknown top-level tokens that match no registered subcommand and emit `kind:"unknown_subcommand"` with `unknown:<token>` field and exit code 1, BEFORE the chat fallback path; (b) when a token is intended as a chat prompt, require an explicit verb (`prompt`, `chat`, `ask`) or `--prompt` flag; (c) ensure exit codes are non-zero for all `kind:*_error` envelopes; (d) regression test: `claw <bogus> --output-format json` with valid auth returns `kind:"unknown_subcommand"` exit 1, never reaches the API. **Why this matters:** automation that calls `claw <subcommand>` with a programmatically constructed verb (typo, version drift, refactored command) silently bills tokens and produces hallucinated output instead of a typed error. Cross-cluster with #108 (CLI fallthrough discovered earlier) — #422 is the post-#108 audit confirming the routing bug still bites with valid credentials. Source: Jobdori live dogfood, `b98b9a71`, 2026-05-11.
+
+
+423. **`claw prompt` does not read prompt text from stdin when no positional prompt arg is provided — `echo "what is 2+2" | claw prompt --output-format json` returns `kind:"unknown" error:"prompt subcommand requires a prompt string"` instead of consuming stdin** — dogfooded 2026-05-11 by Jobdori on `3c563fa1` in response to Clawhip pinpoint nudge at `1503222644739276951`. Reproduction: `echo "what is 2+2" | claw prompt --output-format json` → `{"error":"prompt subcommand requires a prompt string","hint":null,"kind":"unknown","type":"error"}` exit 1. Same for `claw prompt --output-format json` with stdin redirected from a file. The most common Unix automation pattern (`cmd | claw prompt`) is broken because the prompt subcommand only reads the positional argument, never falls through to stdin. **Sibling envelope-kind bug:** the error `kind` is `"unknown"` instead of a typed `"missing_argument"` or `"validation_error"`. The `unknown` discriminator is the catch-all bucket — automation that switches on `kind` to differentiate input-validation errors from runtime errors gets no signal here. **Required fix shape:** (a) when `prompt` subcommand has no positional prompt arg AND stdin is not a TTY (i.e., piped or redirected), read stdin to EOF and use that as the prompt; (b) emit `kind:"missing_argument"` (not `"unknown"`) when both positional arg and stdin are absent; (c) add `--prompt-stdin` or `--stdin` opt-in flag for explicit control; (d) regression tests: `echo X | claw prompt --output-format json` reaches the runtime with prompt=X, AND `claw prompt < /dev/null` returns `kind:"missing_argument"` exit 1. **Why this matters:** Unix pipelines are the foundation of CLI automation. Every other major CLI (curl, jq, gh, kubectl) accepts stdin as the primary input when no positional arg is given. Breaking this convention forces automation to either inline the prompt as a shell-quoted string (escaping nightmare for multiline/code) or write to a temp file first. The `kind:"unknown"` error category compounds the problem by making the failure indistinguishable from a runtime crash. Source: Jobdori live dogfood, `3c563fa1`, 2026-05-11.
+
+
+424. **`--model` rejects bare canonical Anthropic model names (`claude-opus-4-7`, `claude-opus-4-6`, `claude-sonnet-4-6`) as `invalid_model_syntax` — only short aliases (`opus`, `sonnet`, `haiku`) and full prefixed form (`anthropic/claude-opus-4-7`) work; sibling: error message stale-suggests `claude-opus-4-6` not `4-7`** — dogfooded 2026-05-11 by Jobdori on `6c0c305a` in response to Clawhip pinpoint nudge at `1503230194889134103`. Reproduction: `claw --model claude-opus-4-7 status --output-format json` → `{"error":"invalid model syntax: 'claude-opus-4-7'. Expected provider/model (e.g., anthropic/claude-opus-4-6) or known alias (opus, sonnet, haiku)","kind":"invalid_model_syntax"}`. Same for `claude-opus-4-6`, `claude-sonnet-4-6`. Forcing `--model anthropic/claude-opus-4-7` works (`model:"anthropic/claude-opus-4-7"`, `model_source:"flag"`). Three problems compounded: (a) Anthropic-canonical model names without provider prefix are rejected even though the `claude-` prefix unambiguously identifies the provider; (b) the error suggests `anthropic/claude-opus-4-6` as the example — `4-7` shipped 2026-04-16 and is the current production Anthropic frontier model, the suggestion is one model behind; (c) the alias list `opus, sonnet, haiku` doesn't disambiguate version (which `opus` does the alias resolve to — `opus-4-6` or `opus-4-7`?). **Required fix shape:** (a) accept bare `claude-*` and `gpt-*` model names as canonical-named-without-prefix and route via name-prefix detection (already implemented for prefix-routed mode); (b) update the example in `invalid_model_syntax` error to current frontier (`anthropic/claude-opus-4-7`); (c) document or expose `opus` → exact-version mapping in the error message and in `claw doctor`/`status` output (`model_alias_resolved_to: "claude-opus-4-7"`); (d) regression test: `claw --model claude-opus-4-7 status --output-format json` returns `model_source:"flag"`, not `kind:"invalid_model_syntax"`. **Sibling bug observed in same probe:** `enabledPlugins` deprecation warning repeats 3 times in stderr for the same `~/.claw/settings.json` load — config file is being loaded/parsed 3 times during a single `status` invocation. **Why this matters:** every Anthropic doc, every CCAPI route, every internal tooling references models by their bare canonical name (`claude-opus-4-7`). Forcing the `anthropic/` prefix breaks copy-paste from Anthropic's own examples and adds a redundant token to every invocation. The stale `4-6` suggestion in the error message actively misdirects users away from the current model. Source: Jobdori live dogfood, `6c0c305a`, 2026-05-11.
+
+
+425. **Config file precedence (`.claw/settings.json` always wins over `.claw.json`) is undocumented in user-facing surfaces — `config --output-format json` reports both files as `loaded:true` with no `precedence_rank` or `wins_for_keys` attribution; sibling: deprecation warning fires 4× per status invocation (was 3× in #424, regression upward)** — dogfooded 2026-05-11 by Jobdori on `d7dbe951` in response to Clawhip pinpoint nudge at `1503237744451649537`. Reproduction: create `.claw.json` with `{"model":"anthropic/claude-sonnet-4-6"}` and `.claw/settings.json` with `{"model":"anthropic/claude-opus-4-7"}` in the same workspace. `claw status --output-format json` returns `model:"anthropic/claude-opus-4-7", model_source:"config"`. Reverse the files (.claw.json=opus, settings.json=sonnet) → `model:"anthropic/claude-sonnet-4-6"`. Confirmed: `.claw/settings.json` **always** wins over `.claw.json` for conflicting keys, regardless of file mtime or alphabetical order. `claw config --output-format json` reports both as `loaded:true` with no `precedence_rank`, `effective_for_keys`, or `shadowed_keys` attribution. The only signal of precedence is the final merged value in `status` — automation cannot programmatically discover which file contributed which key without re-implementing the merge logic. **Sibling bug (regression from #424):** the `enabledPlugins` deprecation warning now fires **4 times** in stderr per single `status` invocation (was 3× in #424's probe at HEAD `6c0c305a`; current HEAD `d7dbe951` shows 4×). Config load count went up by 1. **Sibling bug observed in config-section probe:** `claw config model --output-format json` with a `.claw.json` that contains a benign unknown key (e.g., `"alpha":"x"`) returns `{"error":"/path/.claw.json: unknown key \"alpha\" (line 1)","kind":"unknown"}` — the entire config command fails with a generic `unknown` kind instead of (a) tolerating unrecognized keys with a warning, or (b) emitting a typed `kind:"unknown_key"` error scoped to the offending file/key. **Required fix shape:** (a) document precedence order in `USAGE.md` (`.claw/settings.local.json > .claw/settings.json > .claw.json` for project scope; `user`/`system` scope at each layer); (b) add `precedence_rank:int` and optional `wins_for_keys:[string]` / `shadowed_keys:[string]` to each entry in `config --output-format json` `files[]`; (c) dedupe the deprecation warning to fire **once per discovered file** instead of N× per load pass; (d) make `config <section> --output-format json` tolerate unknown keys with warnings, OR emit `kind:"unknown_key"` with `path:` and `key:` fields scoped to the offending file. **Why this matters:** users mixing legacy `.claw.json` with new `.claw/settings.json` have no way to verify which file is actually controlling their runtime. The undocumented precedence + missing per-key attribution forces trial-and-error to debug config drift. Cross-references #407 (config files no load_error) and #415 (config section returns merged_keys count not values). Source: Jobdori live dogfood, `d7dbe951`, 2026-05-11.
+
+
+426. **`ANTHROPIC_MODEL` env var bypasses the `invalid_model_syntax` validator that `--model` enforces — bogus model strings are accepted with `status:"ok"`, deferred-failing only when the first API call is made** — dogfooded 2026-05-11 by Jobdori on `3730b459` in response to Clawhip pinpoint nudge at `1503245298800136296`. Reproduction (asymmetric validation): `claw --model bogus-model-xyz status --output-format json` returns `kind:"invalid_model_syntax"` exit 1; `ANTHROPIC_MODEL=bogus-model-xyz claw status --output-format json` returns `model:"bogus-model-xyz", model_raw:"bogus-model-xyz", model_source:"env", status:"ok"` — the doctor surface lies that the configured model is valid when it is not. The bogus model only manifests as a failure when the first prompt fires and the API rejects it with 404/400. Three sibling discoveries in the same probe: (a) **alias indirection invisible**: `ANTHROPIC_MODEL=opus claw status --output-format json` returns `model:"claude-opus-4-6", model_raw:"opus", model_source:"env"` — the `opus` alias resolves to `claude-opus-4-6` (the *previous* frontier, not the current `claude-opus-4-7` released 2026-04-16). Users typing `opus` get yesterday's model with no warning. (b) **`CLAW_MODEL` env var silently ignored**: `CLAW_MODEL=opus claw status` shows `model:"claude-opus-4-6" model_source:"default"` — the `CLAW_MODEL` env var (the project-namespaced equivalent that users expect) does not exist; only `ANTHROPIC_MODEL` is honored. No warning when a `CLAW_*` env var that looks like it should work is set. (c) **`ANTHROPIC_DEFAULT_MODEL` also silently ignored**: the longer-named env var that some Anthropic SDKs use is not recognized. **Required fix shape:** (a) symmetric validation: `ANTHROPIC_MODEL` env value must pass the same `invalid_model_syntax` check that `--model` does, and `claw status` must return `kind:"invalid_model"` / `status:"warn"` (not `status:"ok"`) when the resolved model is unrecognized; (b) expose alias resolution in `status`: add `model_alias_resolved_to:string|null` field so automation can see `opus → claude-opus-4-6`; (c) bump the `opus` alias to `claude-opus-4-7` (current frontier) or document the alias-to-version mapping policy explicitly; (d) accept `CLAW_MODEL` and `ANTHROPIC_DEFAULT_MODEL` env vars with parity to `ANTHROPIC_MODEL`, OR emit a warning when those env vars are set but unrecognized. **Why this matters:** the most common automation pattern is `export ANTHROPIC_MODEL=...` in a shell rc file. Bogus values pass silently, alias indirection hides the actual model in use, and `CLAW_MODEL` looking like a working name but doing nothing is a footgun. Cross-references #424 (bare canonical names rejected at validator level) — together #424 + #426 make model selection inconsistent across CLI flag, env var, and alias paths. Source: Jobdori live dogfood, `3730b459`, 2026-05-11.
+
+
+427. **Subcommand `--help` paths (`resume`, `session`, `compact`) hit the auth gate and trigger config validation before returning static help — `claw resume --help` with no credentials returns `missing_credentials` error instead of help text** — dogfooded 2026-05-11 by Jobdori on `1fecdf09` in response to Clawhip pinpoint nudge at `1503252843669491892`. Reproduction (no env vars, isolated `CLAW_CONFIG_HOME`): `claw resume --help` returns `{"error":"missing Anthropic credentials; export ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY..."}` instead of usage text. Same for `claw session --help`, `claw compact --help`. By contrast, `claw prompt --help` and `claw --help` (top-level) return proper usage text without auth. Even worse: with a broken `.claw.json` discovered up the parent directory tree (e.g., `mcpServers.missing-command: missing string field command`), the subcommand `--help` paths fail with `[error-kind: unknown]` from config validation — config load is happening before `--help` is parsed. **Sibling exit-code bug:** `claw resume --help --output-format json` returns `kind:"missing_credentials"` but exits **0** (the exit-code parity bug from #422 reproduces on this path too — only `cli_parse` exits 1 consistently). **Sibling: `claw resume <bogus-id>` should be local-only** but also hits `missing_credentials` — `resume` of a session that doesn't exist on disk should return `kind:"session_not_found"` from a local lookup, not require API credentials. Same class as ROADMAP #357 (session list requires creds) and #369 (session help/fork require credentials) — now confirmed for `resume`. **Required fix shape:** (a) `--help` MUST short-circuit before any auth check, config load, or session resolution — emit static usage text from a compiled-in string table, no I/O; (b) `resume <id>` must check the local session store first; if the id is absent on disk, emit `kind:"session_not_found"` with `sessions_dir` field; only require auth when resuming a known-on-disk session that requires re-establishing API context; (c) ensure exit code 1 for all error envelopes including `missing_credentials` returned from a `--help` path that should never have reached the auth gate; (d) regression test: with empty `CLAW_CONFIG_HOME` and no env vars, every `claw <subcommand> --help` returns usage text on stdout, exit 0, no `kind:*_error` envelope. **Why this matters:** `--help` is the universal CLI discovery primitive. Failing `--help` because of missing API credentials or broken config files makes claw undiscoverable to users debugging an already-broken setup. Cross-references #357 (session list), #369 (session help/fork), #422 (exit code parity), #108 (subcommand fallthrough). Source: Jobdori live dogfood, `1fecdf09`, 2026-05-11.
+
+
+428. **Default `permission_mode` is `danger-full-access` — claw runs with FULL filesystem + network + tool access out of the box, with no opt-in flag and no warning from `doctor`** — dogfooded 2026-05-11 by Jobdori on `72048449` in response to Clawhip pinpoint nudge at `1503260393622212628`. Reproduction (no env vars, isolated `CLAW_CONFIG_HOME`, no config files, no CLI flags): `claw status --output-format json` returns `permission_mode:"danger-full-access"` as the default. The three supported modes per the validator error message are `read-only`, `workspace-write`, `danger-full-access` — and `danger-full-access` is chosen with zero user opt-in. `claw doctor --output-format json` produces a `sandbox` check with `status:"warn", summary:"sandbox was requested but is not currently active"` (because macOS lacks Linux `unshare`), but **emits no warning, info, or summary about the permission_mode itself being danger-full-access**. There is no `permissions` check in `doctor` output at all. **Required fix shape:** (a) change default `permission_mode` to `workspace-write` (safe-by-default: filesystem write limited to cwd, network limited to LLM endpoints, no arbitrary command exec); (b) require explicit `--permission-mode danger-full-access` or `--dangerously-skip-permissions` to opt into full access; (c) add a `permissions` check to `doctor --output-format json` that emits `status:"warn"` when `permission_mode == "danger-full-access"` without explicit source (flag/env/config), with details like `mode:"danger-full-access", source:"default", message:"running with full access without explicit opt-in"`; (d) document the three modes and the default in USAGE.md with one-paragraph descriptions of what each mode allows. **Sibling typed-error bug:** `claw --permission-mode bogus-mode status --output-format json` returns `kind:"unknown"` instead of `kind:"invalid_permission_mode"` — same catch-all problem as #424, #426. **Sibling flag-name asymmetry:** `--dangerously-skip-permissions` works but `--skip-permissions` (Claude Code's flag) returns `kind:"cli_parse"` `unknown option`. Users migrating from Claude Code lose the short flag name. **Why this matters:** every other security-conscious CLI (Docker, kubectl, terraform) requires explicit opt-in for dangerous modes. Defaulting to `danger-full-access` is a footgun for first-time users who pipe `curl install.sh | sh` and immediately get a tool with full filesystem write and arbitrary command exec. The doctor surface is the only diagnostic users consult before trusting the tool, and it stays silent about the most permissive setting. Cross-references #50, #87, #91, #94, #97, #101, #106, #115, #123 (permission-audit sweep) — those all cover permission *rule* and *list* surfaces; #428 covers the *mode default* itself. Source: Jobdori live dogfood, `72048449`, 2026-05-11.
+
+
+429. **No global `--cwd`/`-C`/`--directory` flag — `claw` cannot be invoked against an arbitrary working directory without first `cd`-ing into it; `--cwd` only exists as a subcommand option for `system-prompt`, and the `cli_parse` "Did you mean --acp?" suggestion is misleading (the `--acp` flag is unrelated to directory selection)** — dogfooded 2026-05-11 by Jobdori on `ec882f4c` in response to Clawhip pinpoint nudge at `1503267943285264394`. Reproduction: `claw --cwd /tmp/claw-dog-cwd status --output-format json` → `{"error":"unknown option: --cwd","hint":"Did you mean --acp?\nRun `claw --help` for usage.","kind":"cli_parse"}`. Same error for `--cwd <relative>`, `--cwd <nonexistent>`, `--cwd <file-not-dir>`, `--cwd ""`. Inspecting `claw --help`: `--cwd PATH` appears ONLY in the usage line `claw system-prompt [--cwd PATH] [--date YYYY-MM-DD]` — it is not a global flag and is not accepted by `status`, `doctor`, `mcp list`, `init`, or any other subcommand. Users programmatically running claw against multiple workspaces must `cd` into each one before invoking, breaking the `subprocess.run(['claw', 'status', '--cwd', ws], cwd=other_dir)` pattern that every other major CLI (cargo `-C`, git `-C`, npm `--prefix`, gh `--repo` semantically, kubectl `--kubeconfig`+`--context`) supports. **Sibling misleading-suggestion bug:** the `cli_parse` error's `hint` field suggests `Did you mean --acp?` for `--cwd`. `--acp` is the alias for ACP/Zed editor integration (entirely unrelated to working directory). The Levenshtein-distance auto-complete is matching on first-character similarity without considering semantic relatedness. Users following the hint get a totally orthogonal feature. **Required fix shape:** (a) add a global `--cwd PATH` / `-C PATH` flag accepted before any subcommand, parsed in the global flag pre-pass; (b) validate the path exists and is a directory; emit `kind:"invalid_cwd"` with `path:` and `reason:` (`"not_found"`/`"not_a_directory"`/`"empty"`) when validation fails; (c) document the precedence: `--cwd` flag > `$PWD` > `env::current_dir()`; (d) fix the "Did you mean" hint algorithm to filter suggestions by semantic category (don't suggest `--acp` for `--cwd`; suggest `claw system-prompt --cwd PATH` if the user clearly wants `cwd` override but used the wrong scope); (e) regression test: `claw --cwd /tmp status --output-format json` from any `$PWD` returns `workspace.cwd:"/private/tmp"` (or `cwd:"/tmp"` after #421 fix). **Why this matters:** every claw automation orchestrator runs claw against multiple workspaces from a single parent process. Forcing `cd` before each invocation breaks parallelism (can't use shared cwd across concurrent invocations), breaks subprocess wrappers that want to pass cwd explicitly, and breaks `xargs`/`parallel`-style pipelines. Cross-references #421 (cwd canonicalization leak — fix should canonicalize but report user-input via `--cwd`). Source: Jobdori live dogfood, `ec882f4c`, 2026-05-11.
+
+
+430. **`dump-manifests` is documented as "emit every skill/agent/tool manifest the resolver would load for the current cwd" but actually requires the upstream Claude Code TypeScript source files (`src/commands.ts`, `src/tools.ts`, `src/entrypoints/cli.tsx`) — the command is unusable for any user who installed claw without cloning the original Claude Code repo** — dogfooded 2026-05-11 by Jobdori on `075c2144` in response to Clawhip pinpoint nudge at `1503275502046023690`. Reproduction: `claw dump-manifests --output-format json` returns `{"error":"Manifest source files are missing.","hint":"repo root: /private/tmp/claw-dog-0530\n  missing: src/commands.ts, src/tools.ts, src/entrypoints/cli.tsx\n  Hint: set CLAUDE_CODE_UPSTREAM=/path/to/upstream or pass \`claw dump-manifests --manifests-dir /path/to/upstream\`.","kind":"missing_manifests"}`. The fresh-main worktree at `/private/tmp/claw-dog-0530` does not contain these TypeScript files because the Rust port doesn't include the upstream TS source. The `--help` text says the command works against "the current cwd" but in practice it requires `CLAUDE_CODE_UPSTREAM=` pointing at an unshipped TS source tree. **Three sibling problems compounded:** (a) **derivative-work disclosure leak**: the error message exposes that `claw-code` is a port of Claude Code (`CLAUDE_CODE_UPSTREAM` env var name) — even if true, surfacing this in a casual diagnostic message couples user-facing behavior to upstream provenance details. (b) **kind drift**: `claw dump-manifests --manifests-dir /tmp/nonexistent --output-format json` returns `kind:"unknown"`, while `claw dump-manifests` (no override) returns `kind:"missing_manifests"`. Same root cause (no usable upstream), two different `kind` discriminators — automation cannot switch on a single error type. (c) **export-positional-arg silently dropped**: probed in the same run — `claw export <bogus-positional>` ignores the path and returns `kind:"no_managed_sessions"` regardless of what positional arg was passed. The `--help` advertises `[PATH]` as the output-file destination but the path is discarded before validation, indistinguishable from invocation with no args. **Required fix shape:** (a) make `dump-manifests` emit the manifests claw-code itself ships with (Rust-resolver-discovered skills/agents/tools), independent of any upstream TS source — that matches the `--help` description; (b) if upstream-comparison is genuinely needed for parity work, move it to a separate command like `parity dump-upstream-manifests` and remove the upstream dependency from `dump-manifests`; (c) standardize on one error `kind` for the manifest-missing failure mode (`missing_manifests` is more descriptive than `unknown`); (d) `claw export <PATH>` must validate the path positional arg before the session-discovery check, so users see `kind:"invalid_output_path"` (or similar) when the path is malformed instead of always seeing `kind:"no_managed_sessions"`. **Why this matters:** `dump-manifests` is the inventory surface a downstream automation lane would call to learn what claw can do in the current workspace. If it's broken without upstream TS source, downstream lanes can't introspect — they have to fall back to `agents list`/`skills list`/`mcp list` separately and re-aggregate. Cross-references #422 (kind:unknown for unknown_subcommand), #423 (kind:unknown for missing_argument), #428 (kind:unknown for invalid_permission_mode) — `kind:"unknown"` keeps appearing as the catch-all for surfaces that should have typed kinds. Source: Jobdori live dogfood, `075c2144`, 2026-05-11.
+
+
+431. **`skills uninstall <name>` requires Anthropic credentials despite being a local filesystem operation — `claw skills uninstall nonexistent-skill-xyz --output-format json` returns `kind:"missing_credentials"` instead of resolving locally that the skill doesn't exist** — dogfooded 2026-05-11 by Jobdori on `328fd114` in response to Clawhip pinpoint nudge at `1503275502046023690` (sibling probe to #430). Reproduction (no creds, isolated `CLAW_CONFIG_HOME`): `claw skills uninstall nonexistent-skill-xyz --output-format json` returns `{"error":"missing Anthropic credentials; export ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY...","kind":"missing_credentials"}`. Uninstalling a skill is a pure local filesystem operation: read the skills directory, find the named skill, remove its files. There is no semantic reason to require API credentials. Same class of bug as #357 (`session list` requires creds), #369 (`session help/fork` require creds), and #427 (`resume <bogus-id>` requires creds). **Three sibling findings in same probe:** (a) `claw skills install <bogus-name>` returns `{"error":"No such file or directory (os error 2)","kind":"unknown"}` — leaks raw OS error string with no hint about expected install source format (path vs name vs URL?), and the catch-all `kind:"unknown"` again instead of typed `kind:"skill_install_source_not_found"`. (b) `claw skills install` (no args) returns `action:"help"` with `unexpected:"install"` — but `install` IS a documented subcommand. The handler treats it as "unknown action" instead of "missing required argument". Should emit `kind:"missing_argument"` with `argument:"install_source"`. (c) `claw agents create my-agent` returns `action:"help"` with `unexpected:"create my-agent"` — there is no agent-creation surface at all. Users must hand-craft `.claw/agents/<name>.md` files with no scaffolding command, while `claw init` only creates the top-level `.claw/` skeleton. **Required fix shape:** (a) `skills uninstall <name>` must be local-first: enumerate the local skills dir, return `kind:"skill_not_found"` (with `skills_dir:` and `available_names:[]` fields) for missing, or remove the files and return `kind:"skills"` with `action:"uninstall", removed:<name>` for present skills; (b) `skills install <source>` must distinguish source forms (`path:`, `name:`, `url:`) and emit `kind:"invalid_install_source"` with the parsed-and-failed reason; (c) `skills install` (no args) emits `kind:"missing_argument"` with `argument:"install_source"`; (d) add `claw agents create <name>` (or `claw init agent <name>`) that scaffolds `.claw/agents/<name>.md` with a stub frontmatter; or document explicitly that agents are user-authored only. **Why this matters:** lifecycle commands (`uninstall`, `install`, `create`) are the primary surface for managing claw's extension surface area. If `uninstall` requires API creds, an offline user who fat-fingered an install can't undo it. If `install` returns a raw OS error, automation can't programmatically recover. If `agents create` doesn't exist, agent authoring is undocumented file-touching only. Cross-references #357, #369, #427 (auth-gate-on-local-ops cluster), and #422/#423/#428/#430 (`kind:"unknown"` catch-all cluster). Source: Jobdori live dogfood, `328fd114`, 2026-05-11.
+
+
+432. **`--allowedTools` validator inconsistency: tool name list is half snake_case (`bash`, `read_file`, `write_file`, `edit_file`, `glob_search`, `grep_search`) and half PascalCase (`WebFetch`, `WebSearch`, `TodoWrite`, `Skill`, `Agent`, `Sleep`) with three UPPERCASE entries (`REPL`, `LSP`, `MCP`); accepts undocumented CamelCase aliases (`Read`, `Write`, `Edit`) and silently translates them to snake_case; argument parsing consumes the next positional when value is missing** — dogfooded 2026-05-11 by Jobdori on `fad53e2d` in response to Clawhip pinpoint nudge at `1503283046856655029`. Reproduction: `claw --allowedTools status --output-format json` → `{"error":"unsupported tool in --allowedTools: status (expected one of: bash, read_file, write_file, edit_file, glob_search, grep_search, WebFetch, WebSearch, TodoWrite, Skill, Agent, ToolSearch, NotebookEdit, Sleep, SendUserMessage, Config, EnterPlanMode, ExitPlanMode, StructuredOutput, REPL, PowerShell, AskUserQuestion, TaskCreate, RunTaskPacket, TaskGet, TaskList, TaskStop, TaskUpdate, TaskOutput, WorkerCreate, WorkerGet, WorkerObserve, WorkerResolveTrust, WorkerAwaitReady, WorkerSendPrompt, WorkerRestart, WorkerTerminate, WorkerObserveCompletion, TeamCreate, TeamDelete, CronCreate, CronDelete, CronList, LSP, ListMcpResources, ReadMcpResource, McpAuth, RemoteTrigger, MCP, TestingPermission)","kind":"unknown"}`. The `status` subcommand was consumed as the `--allowedTools` value because the flag parser doesn't distinguish missing-value from end-of-flag-args. The error reveals **the supported tool list mixes naming conventions inconsistently within a single error message**: snake_case (`bash`, `read_file`, `write_file`, `edit_file`, `glob_search`, `grep_search`), PascalCase (`WebFetch`, `WebSearch`, `TodoWrite`, `Skill`, `Agent`, `Sleep`, `Config`, `PowerShell`, `AskUserQuestion`, `TaskCreate`, `WorkerCreate`, `TeamCreate`, `CronCreate`), UPPERCASE (`REPL`, `LSP`, `MCP`), and CamelCase compounds (`McpAuth`, `RemoteTrigger`). **Hidden alias mapping**: `claw --allowedTools Read,Write,Edit status --output-format json` is accepted and returns `allowed_tools.entries:["edit_file","read_file","write_file"]` — proving the validator has an undocumented CamelCase→snake_case alias map (`Read`→`read_file`, `Write`→`write_file`, `Edit`→`edit_file`) that is not surfaced in the error message. Users who copy-paste tool names from Claude Code documentation work, users who copy from the validator error don't. **Sibling missing-value bug:** `claw --allowedTools status` with `status` as a positional subcommand is interpreted as `--allowedTools=status`, swallowing the subcommand. The flag parser must require a value for `--allowedTools` and emit `kind:"missing_argument"` when followed by a recognized subcommand or `--`-prefixed flag instead of silently treating the next arg as a tool name. **Sibling typed-kind bug:** both errors use `kind:"unknown"` instead of typed `kind:"invalid_tool_name"` / `kind:"missing_argument"` — the catch-all keeps appearing (#422/#423/#424/#428/#430/#431/#432). **Required fix shape:** (a) standardize the canonical tool-name registry on one casing convention (snake_case is most CLI-ergonomic) and update both the registry and all CamelCase aliases; (b) document and expose the alias map (`tool_aliases:{Read:"read_file",...}`) in `claw doctor`/`status` and in the validator error; (c) flag parser must require a value for `--allowedTools` and refuse to consume a recognized subcommand or `-`/`--`-prefixed token as the value, emit `kind:"missing_argument"` with `argument:"--allowedTools"`; (d) emit `kind:"invalid_tool_name"` with `tool_name:` and `available:[]` fields instead of `kind:"unknown"`; (e) regression test that `claw --allowedTools <subcommand>` rejects with `missing_argument`, and that the canonical name list in errors uses the same casing as the alias map. **Why this matters:** `--allowedTools` is the primary surface for restricting claw's tool surface area (security-relevant). Inconsistent naming between the validator error and the alias map means users following the error message guidance pick names that work in some places and fail in others. The missing-value bug silently swallows a subcommand, leading to confusing "unsupported tool: status" errors when the user actually wanted to run `claw status`. Cross-references #94/#97/#101/#106/#115/#123 (permission-rule audit), #428 (default permission_mode), #422/#423/#424/#428/#430/#431 (`kind:"unknown"` catch-all). Source: Jobdori live dogfood, `fad53e2d`, 2026-05-11.
+
+
+433. **Repeated `--output-format` flag silently takes the last value without warning — `claw --output-format json --output-format text status` produces text output, no signal that the prior `json` was overridden; sibling: `--output-format` value is case-sensitive (`JSON` rejected as `kind:"unknown"`); sibling: no `CLAW_OUTPUT_FORMAT` env var for default format override** — dogfooded 2026-05-11 by Jobdori on `ce39d5c5` in response to Clawhip pinpoint nudge at `1503290592556220488`. Reproduction: `claw --output-format json --output-format text status` returns the text-format `Status\n  Model claude-opus-4-6...` table — the first `--output-format json` was silently overridden. No warning, no `format_overridden:true` field, no stderr message. Scripts that compose flag arrays from multiple sources (`flags=("${BASE_FLAGS[@]}" --output-format json)` while `BASE_FLAGS` already contains `--output-format text`) silently get the wrong format. **Three sibling findings in same probe:** (a) **case-sensitivity drift**: `claw --output-format JSON status` returns `{"error":"unsupported value for --output-format: JSON (expected text or json)","kind":"unknown"}` — error message tells user to use lowercase `json` but doesn't accept the uppercase form that users often type from muscle memory. Most CLI flag-value validators (cargo, kubectl, gh) are case-insensitive for enum values or accept both forms with normalization. (b) **`kind:"unknown"` for invalid format value**: same catch-all bucket bug as #422/#423/#424/#428/#430/#431/#432 — should be `kind:"invalid_output_format"` with `value:` and `expected:["text","json"]` fields. (c) **no env-var default for output format**: `CLAW_OUTPUT_FORMAT=json claw status` silently ignored — no env override for the global default, forcing scripts to repeat `--output-format json` on every invocation. Other major CLIs honor `KUBECTL_OUTPUT=`, `AWS_DEFAULT_OUTPUT=`, `GH_NO_PROMPT=` etc. (d) **silently-ignored env vars `CLAW_LOG`/`RUST_LOG`**: no env-based log level control surfaced in `claw doctor` — debug logging requires undocumented `RUST_LOG=` (Rust convention) but `claw --help` doesn't mention either. **Required fix shape:** (a) repeated `--output-format` (or any flag that takes a value, not a count flag) emits a warning to stderr (`warning: --output-format specified multiple times; using last value 'text'`) and adds a `format_source:"flag", format_overridden:[]` field to the JSON envelope; (b) accept case-insensitive enum values for `--output-format` (`JSON`, `Json`, `json` all work), document the canonical lowercase form in `--help`; (c) emit `kind:"invalid_output_format"` (not `kind:"unknown"`) when value is invalid; (d) accept `CLAW_OUTPUT_FORMAT` env var as the default for `--output-format`, with flag-overrides-env precedence documented; (e) document `RUST_LOG` / `CLAW_LOG` in `--help` or doctor output as the log-level env vars; (f) regression test: repeated flag emits stderr warning + JSON metadata field; case-insensitive enum accepts all three casings; env-var default is honored when flag is absent. **Why this matters:** scripts that compose flag arrays from multiple sources (CI envs + per-invocation flags) silently get the wrong output format. Case-sensitive enum values trip up users typing from muscle memory. Missing env-var defaults force per-invocation flag repetition. Cross-references #422/#423/#424/#428/#430/#431/#432 (`kind:"unknown"` catch-all cluster). Source: Jobdori live dogfood, `ce39d5c5`, 2026-05-11.
+
+
+434. **POSIX `--` end-of-flags separator is not recognized — `claw -- "-prompt-with-dash"` returns `{"error":"unknown option: --","hint":"Did you mean -V?","kind":"cli_parse"}` instead of treating subsequent args as positional; shorthand prompt mode cannot accept dash-prefixed prompts at all** — dogfooded 2026-05-11 by Jobdori on `0e5f6958` in response to Clawhip pinpoint nudge at `1503298142286905484`. Reproduction: `claw -- "-prompt-with-dash" --output-format json` returns `{"error":"unknown option: --","hint":"Did you mean -V?\nRun \`claw --help\` for usage.","kind":"cli_parse"}`. The POSIX/GNU CLI convention — universally honored by cargo, git, npm, gh, kubectl, grep, ls, find, etc. — is that `--` terminates flag parsing and treats everything after it as positional arguments. claw rejects `--` itself as an unknown flag. **Sibling misleading-suggestion bug (recurring from #429):** the `cli_parse` hint suggests `Did you mean -V?` for `--`. `-V` is the version flag; `--` is the end-of-flags separator. They have no semantic relationship; the auto-complete is matching on prefix-character similarity only. **Sibling shorthand-prompt limitation:** `claw "-just a prompt" --output-format json` returns `{"error":"unknown option: -just a prompt","kind":"cli_parse"}` and `claw "--bogus-flag-like" --output-format json` returns the same. The shorthand non-interactive prompt mode (documented as `claw [--model MODEL] [--output-format text|json] TEXT`) cannot accept any TEXT that starts with `-` or `--`, even when the entire string is shell-quoted as a single token. Users must use the explicit `prompt` verb (`claw prompt "-prompt-with-dash"` works) to escape this, but the explicit verb is documented as alternative not required. **Required fix shape:** (a) accept POSIX `--` as the end-of-flags marker globally — every arg after `--` is positional; (b) shorthand prompt mode must distinguish "this looks like a flag" from "this is a quoted positional that happens to start with `-`" by looking at whether the token matches any registered flag name (`-h`, `-V`, `--help`, `--version`, etc.) — strings that don't match any flag should be treated as prompt text; (c) fix the "Did you mean" hint algorithm to filter by semantic category (don't suggest `-V` for `--`, suggest "use \`--\` to terminate flag parsing" if the user types just `--`); (d) regression test: `claw -- "-foo"` reaches the runtime with prompt=`-foo`; `claw "-not-a-flag"` is treated as shorthand prompt when no registered flag matches; canonical `--` is recognized. **Why this matters:** POSIX `--` is the universal mechanism for passing arbitrary text (filenames starting with `-`, prompts containing flag-like syntax, log lines, etc.) to a CLI. Failing on `--` makes claw fundamentally unergonomic in shell pipelines (`echo "-q for quiet" | xargs claw` fails). The shorthand-prompt limitation forces users to remember the `prompt` verb specifically when their prompt happens to start with `-`. Cross-references #422 (unknown subcommand fallthrough), #423 (stdin not consumed by prompt), #429 ("Did you mean --acp" misleading suggestion). Source: Jobdori live dogfood, `0e5f6958`, 2026-05-11.
+
+
+435. **`claw --resume latest` on a fresh workspace exit code is 0 in text mode but 1 in JSON mode (text mode lies about success); sibling: failed `--resume` creates the `.claw/sessions/<fingerprint>/` directory tree as a filesystem side effect of the failure** — dogfooded 2026-05-11 by Jobdori on `e29010ed` in response to Clawhip pinpoint nudge at `1503305692566655096`. Reproduction (fresh empty dir, no `.claw/`, no sessions): `claw --resume latest` (text mode) prints `failed to restore session: no managed sessions found in .claw/sessions/0ead448127a2de44/` and exits **0**. Same invocation with `--output-format json` correctly exits **1** with `kind:"session_load_failed"`. Exit-code parity broken on the same input depending on format flag. **Sibling filesystem-side-effect bug:** after the failed `--resume latest` on a fresh empty workspace, the directory `.claw/sessions/0ead448127a2de44/` (the workspace-fingerprint partition) is created on disk despite the operation failing. The user did not opt into creating workspace metadata — they asked to resume an existing session, the resume failed, and now there's a partition directory hanging around. The fingerprint directory ought to be created lazily on first successful session save, not as a side effect of every resume attempt. **Three sibling findings in the same probe:** (a) **`claw --compact` alone (no other args) drops into the interactive REPL with the ANSI welcome banner** — `--compact` is documented as a modifier that strips tool call details in text mode for piping (`--compact ... useful for piping`), not as a verb that activates the REPL. Running `claw --compact` with no positional should be a no-op or an error explaining the flag needs a subcommand or prompt; entering the REPL is the wrong default. (b) **`claw --compact "hello"` (shorthand prompt) returns `{"error":"unknown subcommand: hello.","hint":"Did you mean help","kind":"unknown"}` — `--compact` disables shorthand prompt mode entirely**, treating the positional as a subcommand instead of as prompt text. Users must use the explicit `prompt` verb (`claw --compact prompt "hello"`) which contradicts the `claw [flags] TEXT` usage line in `--help`. (c) `kind:"unknown"` again for the unknown-subcommand error in --compact path — same catch-all bucket bug appearing for the 11th time across pinpoints. **Required fix shape:** (a) exit code 1 for all `failed_to_restore` / `session_load_failed` text-mode failures; text mode should print to stderr and exit non-zero, not print to stdout and exit 0; (b) defer `.claw/sessions/<fingerprint>/` creation to first successful save; failed `--resume` must not leave filesystem droppings; (c) `claw --compact` alone (no positional, no subcommand, stdin is TTY) should emit `kind:"missing_argument"` with `argument:"prompt or subcommand"` rather than activating the REPL; (d) `--compact` must be transparent to shorthand prompt mode parsing — `claw --compact "hello"` is equivalent to `claw --compact prompt "hello"`, both should reach the prompt path; (e) emit typed `kind:"unknown_subcommand"` not `kind:"unknown"` for fallthrough cases. **Why this matters:** scripts that gate on `$?` after `claw --resume latest` see success on text mode and failure on JSON mode — the same operation, two outcomes. The filesystem side effect pollutes a user's worktree with workspace partitions they didn't ask for, and CI pipelines that snapshot `.claw/` size silently grow on every failed `--resume`. Cross-references #422 (exit-code parity across error envelopes), #423 (`kind:"unknown"` for `missing_argument`), #434 (shorthand prompt limitations). Source: Jobdori live dogfood, `e29010ed`, 2026-05-11.
+
+
+436. **`claw init` shipped `.claw.json` template explicitly sets `permissions.defaultMode:"dontAsk"` — every user who runs `claw init` gets a config file that disables permission prompts by default; sibling: `init` creates an empty `.claw/` directory with no settings.json template inside, and when `.claw/` already exists it skips the whole artifact (no settings template materialized)** — dogfooded 2026-05-11 by Jobdori on `b8f989b6` in response to Clawhip pinpoint nudge at `1503313241751949335`. Reproduction: `mkdir /tmp/probe && cd /tmp/probe && claw init --output-format json` returns `artifacts:[{name:".claw/",status:"created"},{name:".claw.json",status:"created"},...]`. Inspecting the created `.claw.json`: `{"permissions":{"defaultMode":"dontAsk"}}`. This is the polar opposite of safe-by-default: every user who follows the documented onboarding flow (`claw init` after `curl install.sh`) ships their workspace with permission prompts disabled. Compounds with **#428** (default runtime permission_mode is `danger-full-access`) — between the runtime default and the init template, a fresh claw setup has zero user-facing safety friction. **Sibling: `.claw/` artifact is an empty directory.** After `claw init`, `find .claw -type f` returns nothing. No `settings.json`, no template, no scaffolding — just `mkdir .claw`. The `--help` description implies init produces a usable workspace, but `.claw/settings.json` (the project-scope counterpart of `~/.claw/settings.json`) is never templated. **Sibling: `.claw/` skip-on-exists drops the entire artifact.** If `.claw/` already exists (e.g., from a partial setup, a `--resume` failure side effect per #435, or manual creation), `claw init` returns `.claw/: skipped` and does not materialize any expected sub-content. The other artifacts (`.claw.json`, `.gitignore`, `CLAUDE.md`) are still created, but a future `claw skills install` or `claw plugins enable` may expect `.claw/` to contain template files that are now missing. **Required fix shape:** (a) the shipped `.claw.json` template must default to `permissions.defaultMode:"acceptEdits"` or `"plan"` (safe-by-default modes per #428 spec) — `"dontAsk"` requires explicit opt-in; (b) `claw init` must materialize `.claw/settings.json` with documented schema defaults inside `.claw/` so the directory is useful on its own; (c) when `.claw/` already exists, `init` must report `partial` status (not `skipped`) and still try to create missing sub-files like `.claw/settings.json` without overwriting existing files; (d) emit per-sub-file artifact entries for `.claw/settings.json` and `.claw/sessions/` (skipped status if absent, deferred-to-first-save acceptable) so automation knows what's present; (e) regression test: `claw init` produces a `.claw.json` whose `permissions.defaultMode` is NOT `dontAsk`; `.claw/` contains at least one templated file. **Why this matters:** init is the primary onboarding surface. Every first-time user piping `curl install.sh | sh && claw init` gets a workspace pre-configured to skip permission prompts — and that workspace gets committed to the user's repo via the `init`-added entry. The `.claw/` empty-directory bug means feature discovery (skills, plugins) lacks the scaffolding it implies. Cross-references #428 (runtime default permission_mode), #50/#87/#91/#94/#97/#101/#106/#115/#123 (permission-rule audit), #435 (filesystem side effects on failed resume). Source: Jobdori live dogfood, `b8f989b6`, 2026-05-11.
+
+
+437. **`version --output-format json` omits build provenance fields — no `is_dirty`, `branch`, `commit_date`, `commit_timestamp`, `rustc_version`; `git_sha` is truncated to 7 chars instead of full 40-char hash; sibling: `executable_path` leaks the build host's path (`/tmp/claw-dog-0530/...`) into runtime output** — dogfooded 2026-05-11 by Jobdori on `8cf628a5` in response to Clawhip pinpoint nudge at `1503320791582900344`. Reproduction: `claw version --output-format json` returns `{"build_date":"2026-05-11","executable_path":"/tmp/claw-dog-0530/rust/target/release/claw","git_sha":"b98b9a7","kind":"version","message":"Claw Code\n  Version 0.1.0\n  Git SHA b98b9a7\n  Target aarch64-apple-darwin\n  Build date 2026-05-11","target":"aarch64-apple-darwin","version":"0.1.0"}`. Critical provenance fields missing: (a) **`is_dirty`** — was the working tree clean at build time? Automation that pins on build provenance cannot tell if the binary was built from a clean commit or includes uncommitted changes; (b) **`branch`** — was this built from `main`, `dev/rust`, a release tag, or a feature branch? The `git_sha` alone doesn't reveal the integration point; (c) **`commit_date` / `commit_timestamp`** — only `build_date` (when the binary was compiled) is exposed; the commit itself might be days/weeks older if the build happened later. Reproducibility audits need both; (d) **`rustc_version`** — what Rust compiler version produced this binary? Critical for security advisories (e.g., known regressions in specific rustc versions); (e) **`git_sha` truncated to 7 chars** ("b98b9a7" instead of full "b98b9a71..."): 7-char shas have known collision rates in large repos and prevent unambiguous git rev-parse round-trip. **Sibling: `executable_path` leaks build-host path.** The `executable_path` field returns `/tmp/claw-dog-0530/rust/target/release/claw` — the directory where the binary was compiled, embedded into the binary metadata. For a binary copied/installed/symlinked to a different location, this field still reports the build path, not the actual invocation path. Either the field should reflect the runtime path via `std::env::current_exe()` at runtime (not compile-time), or it should be dropped to avoid leaking compile-host filesystem layout. **Sibling: prose `message` field duplicates structured data.** The `message` field still contains the entire text-mode prose version block (`"Claw Code\n  Version 0.1.0\n  Git SHA b98b9a7\n..."`) — every field present as structured JSON (`version`, `git_sha`, `target`, `build_date`) is also embedded in the prose. Same issue as #391 (`version json includes prose message field`) which was closed as "fixed" — the prose remains. **Required fix shape:** (a) add `is_dirty:bool`, `branch:string|null`, `commit_date:string` (ISO-8601), `commit_timestamp:int` (Unix epoch), `rustc_version:string` to the JSON envelope; (b) preserve full 40-char `git_sha` and add `git_sha_short:string` as a derived field if 7-char form is needed for UX; (c) `executable_path` should be `std::env::current_exe()` at runtime, not the compile-time path; (d) drop the prose `message` field from JSON or rename it `human_readable:string` and make it explicitly secondary to the structured fields; (e) re-verify #391 closure — the prose `message` is still present, the fix didn't fully land. **Why this matters:** version surface is the canonical provenance probe for security audits, build reproducibility, and bug-report metadata. Missing `is_dirty` means automated triage cannot distinguish "issue against a clean main commit" from "issue against a developer's uncommitted hack". Truncated `git_sha` blocks unambiguous git lookup. Leaked `executable_path` exposes build-host layout. Cross-references #391 (version prose duplication — apparently not fully fixed), #334 (version json omits build_date — fixed, but partial scope), #100 (commit identity audit). Source: Jobdori live dogfood, `8cf628a5`, 2026-05-11.
+
+
+438. **Memory file discovery only recognizes `CLAUDE.md` — `AGENTS.md` (industry convention used by OpenCode/Codex/Aider/Cursor) and `CLAW.md` (project's own brand name) are silently ignored despite being present in the workspace** — dogfooded 2026-05-11 by Jobdori on `d3a982dd` in response to Clawhip pinpoint nudge at `1503328341422244012`. Reproduction (fresh empty dir, isolated `CLAW_CONFIG_HOME`): create three files in cwd — `CLAUDE.md` (marker `MARKER-FROM-CLAUDE-MD`), `AGENTS.md` (marker `MARKER-FROM-AGENTS-MD`), `CLAW.md` (marker `MARKER-FROM-CLAW-MD`). Run `claw status --output-format json` → `workspace.memory_file_count: 1`. Run `claw system-prompt --output-format json` and search the `message` field for each marker: only `MARKER-FROM-CLAUDE-MD` is found; `MARKER-FROM-AGENTS-MD` and `MARKER-FROM-CLAW-MD` are absent. `claw-code` exclusively recognizes the Claude-branded filename inherited from upstream Claude Code; the project's own `CLAW.md` brand name and the cross-tool industry convention `AGENTS.md` are both silently dropped. **Three sibling implications:** (a) **brand-consistency gap**: a project rebranded from Claude Code to Claw Code that introduces `CLAUDE.md` as its only memory file is internally inconsistent. Users naturally expect `claw <subcommand>` to read `CLAW.md`. (b) **industry-convention gap**: `AGENTS.md` is the convergent convention for OpenCode (oh-my-opencode/sisyphus), OpenAI Codex CLI, Aider, Cursor, Continue.dev, and most ACP harnesses. Users with mixed-tool workflows maintain a shared `AGENTS.md` and expect every AI coding tool to honor it. (c) **silent failure mode**: there is no warning when `AGENTS.md` or `CLAW.md` exist but are not loaded. Users who copy-paste `AGENTS.md` from another tool's docs see `memory_file_count` stay at 0 or 1 and have to guess why their instructions aren't applied. **Required fix shape:** (a) discover and load **`CLAUDE.md`, `CLAW.md`, `AGENTS.md`** in that priority order (existing config-precedence pattern); (b) all three contribute to `memory_file_count` with `memory_files:[{path, source:"claude_md"|"claw_md"|"agents_md", chars}]` array exposed in `status --output-format json`; (c) when multiple files exist, merge or document the precedence: project-specific `CLAUDE.md`/`CLAW.md` overrides industry-shared `AGENTS.md`; (d) `claw doctor --output-format json` adds a `memory` check that warns when `AGENTS.md` exists but is not the loaded variant (alerting users that they may be relying on the wrong file); (e) regression test: workspace with all three files results in `memory_file_count >= 1` and the system prompt contains markers from at least the highest-precedence file. **Why this matters:** `AGENTS.md` is the lingua-franca instruction file for cross-tool AI coding workflows. A team using OpenCode for one project and Claw Code for another keeps their conventions in a shared `AGENTS.md`. Forcing them to also maintain a `CLAUDE.md` for claw-code (with identical content) is friction that breaks the value proposition of a fork. Cross-references #438 itself (the multi-file convention), and AGENTS.md ecosystem references in oh-my-opencode/sisyphus docs. Source: Jobdori live dogfood, `d3a982dd`, 2026-05-11.
+
+
+439. **Memory file discovery walks ALL ancestor directories up to `$HOME` boundary, silently loading any `CLAUDE.md` it finds — `/tmp/CLAUDE.md` left from a previous test silently bleeds into every project under `/tmp/*/`; no `--no-parent-memory` flag, no `.no-claude-md-boundary` marker file to limit discovery scope** — dogfooded 2026-05-11 by Jobdori on `f4a96740` in response to Clawhip pinpoint nudge at `1503335892461293675`. Reproduction: create three nested `CLAUDE.md` files with unique markers — `/tmp/claw-nested-probe/CLAUDE.md` (`PARENT_CLAUDE`), `subproj/CLAUDE.md` (`CHILD_CLAUDE`), `subproj/deep/CLAUDE.md` (`DEEP_CLAUDE`). Run `claw system-prompt --output-format json` from `subproj/deep/nest/` (note: `nest` has no `CLAUDE.md`). The `message` field contains **all three markers** (PARENT + CHILD + DEEP) and `status --output-format json` reports `memory_file_count: 3`. Boundary tests: (a) `$HOME/CLAUDE.md` is NOT picked up from `/tmp/no-claude-dir` (discovery stops at `$HOME` boundary, good); (b) From `/tmp/deep` (no nested CLAUDE.md), `/tmp/CLAUDE.md` IS picked up (count: 1); (c) git-root is NOT a discovery boundary — running from a git subdir still walks above the git root. **Ambient-context-bleed footgun:** any stale `/tmp/CLAUDE.md` (or `/home/<user>/projects/CLAUDE.md`, or any ancestor-path CLAUDE.md left over from a previous experiment, copy-paste, or AI-generated example) silently bleeds into every workspace nested below it. The user has no signal in `status --output-format json` indicating which ancestor file is contributing — only the aggregate `memory_file_count`. **Three required fixes:** (a) **expose discovery list**: `status --output-format json` and `system-prompt --output-format json` must include `memory_files:[{path, source:"workspace"|"ancestor"|"parent_dir"|"home", chars, contributes:bool}]` so users can see what's leaking in; (b) **add `--no-parent-memory` flag** to limit discovery to cwd only (no ancestor walk), or add a boundary marker (`.claude-no-walk`, `.claw-root`, or honor `.git` as the boundary by default — most users expect repo-root scope); (c) **`doctor` warns** when ancestor `CLAUDE.md` files are loaded from outside the current git repo (suggests they may be unintentional). **Sibling discovery scope question:** discovery walks up to `$HOME` — but for a user with a project at `/Users/foo/work/proj`, that's `/Users/foo/work/CLAUDE.md` + `/Users/foo/CLAUDE.md` (if it exists) both load. The home boundary is exclusive, but the entire `/Users/foo` tree under home is in scope. **Why this matters:** test workspaces, scratch dirs, AI-generated example projects, and shared `/tmp` workdirs are full of stale `CLAUDE.md` files. The current discovery rule means every claw invocation can silently inherit context from arbitrary ancestor paths. Cross-references #438 (memory discovery only finds CLAUDE.md, not AGENTS.md or CLAW.md), #421 (cwd canonicalization leak — the canonicalized form determines which ancestor walk path is used). Source: Jobdori live dogfood, `f4a96740`, 2026-05-11.
+
+
+440. **One invalid `mcpServers` entry blocks ALL OTHER valid MCP servers from loading — `mcp list --output-format json` returns `configured_servers: 0, servers: []` when even one server has a missing/invalid `command` field, despite other servers in the same config being well-formed; sibling: config parser halts on first invalid entry, never reports the remaining invalid entries** — dogfooded 2026-05-11 by Jobdori on `bd126905` in response to Clawhip pinpoint nudge at `1503343442904879156`. Reproduction: write `.claw.json` containing six `mcpServers` entries — one valid (`valid-server: {command:"/bin/echo", args:["hello"]}`) and five with progressive defects (missing-command, empty-command, null-command, wrong-type-command, extra-unknown-field). Run `claw mcp list --output-format json` → `{"action":"list","config_load_error":"/private/tmp/claw-mcp-probe/.claw.json: mcpServers.missing-command-server: missing string field command","configured_servers":0,"kind":"mcp","servers":[],"status":"degraded"}`. The error mentions only `missing-command-server` (the first invalid entry in JSON-object iteration order); the other four invalid entries are never surfaced. The valid `valid-server` entry is silently dropped because the parser bails on the first error. `status --output-format json` correctly propagates the same `config_load_error` and sets `status:"degraded"`, but no field tells automation which servers are valid vs broken — `servers:[]` is the only signal. **Three problems compounded:** (a) **all-or-nothing loading**: ROADMAP product principle #5 says "partial success is first-class," but mcp config loading is binary. One bad server kills the entire MCP plane; (b) **first-error-only reporting**: a `.claw.json` with five invalid entries surfaces only one error message — the user fixes that one and runs again, gets the next error, and so on. Five iterations needed to discover all errors; (c) **no per-server status**: even with the partial-success fix, the JSON envelope needs `servers:[{name, valid:bool, error?, command?, args?}]` so automation can see which entries are usable. **Required fix shape:** (a) the MCP config parser must collect ALL invalid entries into an `invalid_servers:[{name, error_field, reason}]` array and load all valid ones into `servers:[]`; do not abort on first error; (b) `configured_servers` reflects the count of *valid* loaded servers (not zero) when there are valid entries alongside invalid ones; (c) expose `total_configured:int` (count of entries in source `.claw.json`) AND `valid_count:int` (loaded), AND `invalid_count:int` (rejected) — three distinct counts; (d) `doctor --output-format json` adds an `mcp_validation` check that lists each invalid entry with its error message; (e) regression test: `.claw.json` with one valid + one invalid entry results in `configured_servers: 1, invalid_servers: [{name:"...", reason:"..."}]`. **Why this matters:** users iterate on MCP server lists during onboarding — one typo kills the entire plane, including servers they got working previously. The first-error-only reporting forces N iterations through N invalid entries instead of a single fix-everything-at-once pass. Cross-references #407 (config files no load_error per-file), #415 (config section merged_keys count only), #416 (plugins list prose), #428 (default permission mode), and Product Principle #5. Source: Jobdori live dogfood, `bd126905`, 2026-05-11.
+
+
+441. **`hooks` config schema diverges from Claude Code documented format — claw-code expects `{"hooks":{"PreToolUse":["command-string"]}}` (array of command strings) while Claude Code documentation specifies `{"hooks":{"PreToolUse":[{"matcher":"Read","hooks":[{"type":"command","command":"..."}]}]}}` (structured matcher objects); users copy-pasting from Claude Code docs see `field "hooks.PreToolUse" must be an array of strings`** — dogfooded 2026-05-11 by Jobdori on `86ff83c2` in response to Clawhip pinpoint nudge at `1503350990680887418`. Reproduction: write `.claw.json` with the Claude-Code-documented hook format `{"hooks":{"PreToolUse":[{"matcher":"Read","hooks":[{"type":"command","command":"/bin/echo pretool"}]}]}}`. Run `claw status --output-format json` → `config_load_error: "/private/tmp/claw-hook-probe/.claw.json: field \"hooks.PreToolUse\" must be an array of strings, got an array (line 3)"`, `status: "degraded"`. The error wording ("must be an array of strings, got an array") is confusingly tautological — the user did provide an array; the parser objects that the array contains objects instead of strings. Replacing with the claw-code-actual format `{"hooks":{"PreToolUse":["/bin/echo pretool"]}}` succeeds: `config_load_error: null, status: "ok"`. The two formats are fundamentally incompatible: claw-code drops the `matcher` field (no tool-specific filtering at the config layer), drops the `type:"command"` discriminator (no future expansion to other hook types), and treats each entry as a bare command string instead of a structured hook spec. **Sibling: PR #3000 (justcode049) was attempting to tolerate object-style hook entries** — that PR's title `fix: tolerate object-style hook entries in config parser` confirms this is a known user complaint, but the PR is still conflicting and unmerged. **Three sibling findings in same probe:** (a) **unknown event names reject entire hooks config**: `.claw.json` with `hooks.InvalidEvent` (not a real event name like `PreToolUse`/`PostToolUse`/`Stop`/`Notification`) triggers `config_load_error: "unknown key \"hooks.InvalidEvent\""` and rejects ALL hooks in the same file, even valid ones — same "one bad apple kills all" pattern as #440 (MCP servers). (b) **`kind:"unknown"` for the validation error** — should be `kind:"invalid_hooks_config"` or `kind:"unknown_hook_event"` (catch-all cluster #422/#423/#424/#428/#430/#431/#432/#433/#435 — 13th occurrence). (c) **first-error-only halting**: a `.claw.json` with `hooks.Stop:"not-an-array"` (type mismatch) AND `hooks.InvalidEvent` (unknown name) AND `hooks.Notification:[{}]` (empty entry) surfaces only the FIRST error in iteration order — user must fix one at a time across 3 iterations. **Required fix shape:** (a) **adopt Claude Code's structured hook format as the canonical**: support `{matcher, hooks:[{type, command}]}` natively, with `matcher` for tool-filtering, `type` for hook-type discriminator (future-proof for `inline`/`webhook`/etc beyond just `command`); (b) **keep backward compat for bare command strings**: legacy `["command-string"]` arrays still load, but emit a deprecation warning suggesting migration to the structured form; (c) **partial-success loading**: invalid hook entries surface in `invalid_hooks:[{event, index, reason}]` while valid ones load — same fix as #440 for MCP; (d) **typed `kind:"invalid_hooks_config"` envelope** instead of `kind:"unknown"`; (e) **rebase and merge PR #3000** which addresses this directly; (f) regression test: Claude-Code-documented hook config loads without error on claw-code. **Why this matters:** users migrating from Claude Code to Claw Code hit this on their first `.claw.json` write. The error message ("array of strings, got an array") is unhelpful; the documentation doesn't surface the schema divergence; and Claude Code's structured format is strictly more expressive (matchers, types) than claw-code's bare-string format. Cross-references #407 (config files no load_error), #410 (list-envelope schema drift), #428 (default permission mode), #440 (one invalid MCP entry blocks all), PR #3000 (justcode049's pending fix). Source: Jobdori live dogfood, `86ff83c2`, 2026-05-11.
+
+
+442. **`agents` discovery requires TOML format (`.toml` files) while Claude Code documents agents as Markdown with YAML frontmatter (`.md`) — claw-code silently ignores `.md` files in `.claw/agents/` without any warning; the help text lists `.claw/agents, ~/.claw/agents, $CLAW_CONFIG_HOME/agents` as sources but does not mention the `.toml` file format requirement** — dogfooded 2026-05-11 by Jobdori on `8499599b` in response to Clawhip pinpoint nudge at `1503358540230692876`. Reproduction: write `.claw/agents/valid-agent.md` with Claude-Code-format YAML frontmatter `---\nname: valid-agent\ndescription: A simple test agent\ntools: [bash, read_file]\n---\nYou are a helpful agent.` Run `claw agents list --output-format json` → `{"agents":[], "count":0, "summary":{"active":0,"shadowed":0,"total":0}}`. The valid `.md` agent is silently dropped. Replace with `.claw/agents/toml-agent.toml` containing TOML format `name = "toml-agent"\ndescription = "..."` → loads correctly with `count:1`. Source code confirms (`rust/crates/commands/src/lib.rs:3378`): `if entry.path().extension().is_none_or(|ext| ext != "toml") { continue; }` — only `.toml` extension is recognized, all others (including `.md`) skipped without warning. The help text `claw agents --help` documents the source paths but **omits the file-format requirement**. **Five sibling problems compounded:** (a) **schema divergence from Claude Code**: Claude Code's `agents` are documented as `.md` files with YAML frontmatter (matching the `CLAUDE.md`/`.claude/agents/` convention upstream). claw-code chose TOML for no documented reason. Users migrating from Claude Code or copy-pasting community agent definitions hit silent failure. (b) **silent file drop**: invalid agent files (wrong extension, broken frontmatter, missing required fields, file-name vs frontmatter-name mismatch) are all silently ignored with `count:0`. No `invalid_agents:[]` array, no warning, no `kind:"agent_load_failed"` envelope. Same all-or-nothing pattern as #440 (MCP servers) and #441 (hooks). (c) **no documentation of the schema**: `claw agents --help --output-format json` (per #427, this hits the auth gate; without auth it doesn't return the schema either). The required TOML fields (`name`, `description`, `model`, `model_reasoning_effort` per source code) aren't documented in any user-facing surface. (d) **missing `.claude/agents/` discovery**: many existing projects have `.claude/agents/` from Claude Code installs. claw-code only looks at `.claw/agents/` — users have to copy/move their existing agents. (e) **no agent-scaffolding command**: cross-reference #431 — there's no `claw agents create <name>` to generate a valid `.toml` skeleton; users must hand-craft. **Required fix shape:** (a) accept BOTH `.md` (with YAML frontmatter) AND `.toml` formats in `.claw/agents/`; prefer YAML frontmatter for Claude Code parity, keep TOML for back-compat; (b) include `.claude/agents/` in the discovery sources alongside `.claw/agents/` with documented precedence; (c) expose `invalid_agents:[{path, reason}]` array in `agents list --output-format json` so users can see what was skipped and why; (d) document the agent schema (required + optional fields) in `claw agents --help` and in USAGE.md; (e) add `claw agents create <name>` scaffolding command per #431; (f) regression test: `.claw/agents/foo.md` with YAML frontmatter loads correctly. **Why this matters:** agents are the primary extension surface for custom workflows. A silent-drop on the wrong file format breaks the discoverability promise of CLI agents. Claude Code's `.md`-with-YAML convention is the lingua franca across AI coding tools; deviating to TOML breaks copy-paste compatibility. Cross-references #430 (dump-manifests needs upstream), #431 (skills/agents lifecycle), #440 (MCP all-or-nothing), #441 (hooks all-or-nothing), #438 (memory file discovery only CLAUDE.md). Source: Jobdori live dogfood, `8499599b`, 2026-05-11.
+
+
+443. **`claw acp serve` exits 0 with `status:"discoverability_only", supported:false` instead of failing — automation pipelines see "success" from a command that explicitly says "not implemented"; ROADMAP #413's internal-tracking leak (`discoverability_tracking:"ROADMAP #64a"`, `tracking:"ROADMAP #76"`) still present despite being filed 2026-04-30** — dogfooded 2026-05-11 by Jobdori on `19aaf9d0` in response to Clawhip pinpoint nudge at `1503366101533200435`. Reproduction: `claw acp serve --output-format json` returns exit code **0** with envelope `{aliases:["acp","--acp","-acp"], discoverability_tracking:"ROADMAP #64a", kind:"acp", launch_command:null, message:"ACP/Zed editor integration is not implemented in claw-code yet. \`claw acp serve\` is only a discoverability alias today; it does not launch a daemon or Zed-specific protocol endpoint. Use the normal terminal surfaces for now and track ROADMAP #76 for real ACP support.", recommended_workflows:["claw prompt TEXT","claw","claw doctor"], serve_alias_only:true, status:"discoverability_only", supported:false, tracking:"ROADMAP #76"}`. The exit code is 0 (success) but the command explicitly states it is not implemented. Pipeline like `claw acp serve && zed --connect localhost:12345` will proceed to the zed connect step despite `acp serve` being a no-op. The only signal of no-op is `supported:false` in the JSON body — easy to miss for automation gating on `$?`. **ROADMAP #413 reproduction confirmed unfixed:** #413 (filed 2026-04-30) called out `discoverability_tracking:"ROADMAP #64a"` and `tracking:"ROADMAP #76"` as internal ticket references leaked into public JSON. **11 days later, both fields are still present in the envelope.** The fix was prescribed but never landed. Also `recommended_workflows:["claw prompt TEXT","claw","claw doctor"]` is internal scaffolding (curated suggestion list) exposed as a top-level public field — not normally part of an "ACP status" public contract. **Sibling unknown-subcommand bug:** `claw acp status --output-format json` (a reasonable next-thing-to-try) returns `{"error":"unsupported ACP invocation. Use \`claw acp\`, \`claw acp serve\`, \`claw --acp\`, or \`claw -acp\`.","kind":"unknown"}` exit 0 — the `kind:"unknown"` catch-all yet again (#422/#423/#424/#428/#430/#431/#432/#433/#435/#440/#441/#442 — **14th occurrence**), should be `kind:"unsupported_acp_invocation"`. **Required fix shape:** (a) `claw acp serve` exits **non-zero** (exit code 2 = "not implemented" is conventional) so automation `$?`-gating detects the no-op; (b) deliver #413's fix: remove `discoverability_tracking` and `tracking` top-level fields, OR move them under an optional `_meta` sub-object gated on a debug flag; (c) replace `message` prose with a typed `reason:"not_implemented"` enum + optional `detail` string for downstream pipelines that need a stable signal; (d) drop `recommended_workflows` from the ACP envelope OR move it under `_meta`; (e) the `status:"discoverability_only"` value is non-standard — replace with `status:"not_implemented"` (matching the `supported:false` boolean); (f) typed `kind:"unsupported_acp_invocation"` for the bad-arg path. **Why this matters:** ACP/Zed integration is the integration point for IDE-based AI workflows. A "success" exit code on a "not implemented" stub breaks the contract for any wrapper script that tries to detect ACP availability via `claw acp serve && ...`. The internal-tracking-ID leak (#413) being unfixed for 11 days suggests the JSON envelope audit isn't being executed against the ROADMAP backlog. Cross-references #413 (internal tracking leak — unfixed), #422 (exit-code parity), `kind:"unknown"` catch-all cluster. Source: Jobdori live dogfood, `19aaf9d0`, 2026-05-11.
+
+
+444. **No broad-cwd safety guard for `--resume` — `claw --resume latest` from `/` attempts to `mkdir /.claw/sessions/<fingerprint>/` and is only stopped by the read-only filesystem at root; from any writable system directory (`/tmp`, `/var/tmp`, `$HOME` itself) it silently creates `.claw/sessions/<fingerprint>/` droppings; exit code is 0 (success) on the read-only filesystem error path** — dogfooded 2026-05-11 by Jobdori on `b2048856` in response to Clawhip pinpoint nudge at `1503373639884607629`. Reproduction: `cd / && claw --resume latest --output-format json` returns `{"error":"failed to restore session: Read-only file system (os error 30)","hint":null,"kind":"session_load_failed","type":"error"}` exit **0**. The OS permission denial is the only thing preventing claw from creating `/.claw/sessions/<fingerprint>/` in the root filesystem. Compare with `cd /tmp && claw --resume latest --output-format json`: silently creates `/tmp/.claw/sessions/<fingerprint>/` partition (confirmed by `ls /tmp/.claw` showing a directory from a prior dogfood session at `13:31` — the May 11 11:00 pinpoint #435 dropping is still there 10+ hours later, despite documented cleanup). Same dogfood session: `cd $HOME && claw --resume latest` would silently create `~/.claw/sessions/<fingerprint>/` (the user's home claw config dir). The shorthand prompt path has a broad-cwd guard (`claw is running from a very broad directory (/). The agent can read and search everything under this path. Use --allow-broad-cwd to proceed anyway`) — but the guard does NOT fire on `--resume`, `--status`, or `claw status` invocations. Inconsistent safety surface: the dangerous path (LLM prompt with full tool access) has a guard, but session-management paths that create filesystem artifacts in broad locations have none. **Three sibling findings in same probe:** (a) **exit-code 0 on filesystem error** (`session_load_failed` envelope returns exit code 0): the read-only-filesystem error from `/.claw` creation path is an unrecoverable failure but the process exits 0 — same exit-parity bug as #422/#435; (b) **stale filesystem droppings**: `/tmp/.claw/` from a 13:31 dogfood session at HEAD `6c0c305a` is still present at 21:30 (10 hours later, 6+ HEADs later). The "deferred cleanup" or "lazy creation" fix prescribed in #435 hasn't landed; (c) **broad-cwd guard misfires on resume**: the existing guard from `run` path (visible in `claw --help` as "Use --allow-broad-cwd to proceed anyway") never fires on `--resume`. Either both paths should guard, or the guard should be promoted to a global pre-check. **Required fix shape:** (a) extend the broad-cwd guard to `--resume`, `claw status`, `claw doctor`, and every command that may create filesystem artifacts; `cd / && claw --resume latest` must fail fast with `kind:"broad_cwd_blocked"` before any filesystem operation; (b) `cd $HOME && claw` should warn that the workspace is your home directory and ask for `--allow-broad-cwd` (the LLM with full filesystem access in `$HOME` is the same blast radius as in `/`); (c) exit code 1 for `session_load_failed` regardless of underlying cause; (d) deliver #435's "defer fingerprint directory creation to first successful save" fix — failed `--resume` must not leave filesystem droppings; (e) cleanup `/tmp/.claw/` style scratch-dir artifacts via a `claw doctor --cleanup` or similar opt-in mechanism; (f) regression test: failed `--resume` does not create any directories under cwd. **Why this matters:** users running claw as part of CI/cron from system directories silently accumulate `.claw/sessions/<fingerprint>/` artifacts in /tmp, /var, /opt, $HOME, etc. Running as root from / would (with a writable root) silently pollute the root filesystem. The broad-cwd guard exists but only covers one entry point. Cross-references #427 (broad-cwd guard fires on resume too — actually it doesn't, that note in #427 was inaccurate), #428 (default permission_mode danger-full-access — compounds with this: full access + no broad-cwd guard = serious blast radius), #435 (filesystem side effects on failed resume), #422 (exit-code parity). Source: Jobdori live dogfood, `b2048856`, 2026-05-11.
+
+
+445. **Skill name-vs-directory mismatch is silently accepted — `.claw/skills/wrong-name/SKILL.md` with frontmatter `name: actually-different-name` loads as "actually-different-name" without any warning; users who reference the skill by directory name (`claw skills run wrong-name`) get `skill_not_found` while `skills list` shows it under the frontmatter name; sibling: loose `.md` files at the skills-dir root and subdirs without `SKILL.md` are silently dropped** — dogfooded 2026-05-11 by Jobdori on `9e1eafd0` in response to Clawhip pinpoint nudge at `1503381189539528897`. Reproduction: create `.claw/skills/wrong-name/SKILL.md` with frontmatter `---\nname: actually-different-name\ndescription: Skill where dir name and frontmatter name disagree\n---`. Run `claw skills list --output-format json` → the skill is listed with `name: "actually-different-name"` (the frontmatter value), no warning about the dir-vs-name mismatch. Users who type `claw skills run wrong-name` (the dirname they know from `ls`) get a `skill_not_found` error; `claw skills run actually-different-name` works. The two names are decoupled with no surfaced relationship. **Three sibling silent-drop bugs in same probe:** (a) **subdir without SKILL.md silently skipped**: `.claw/skills/no-skill-md/` containing only `README.md` (no `SKILL.md`) is silently skipped from `skills list`. No `invalid_skills:[{path, reason:"missing_SKILL.md"}]` array, no warning, just absent from output. (b) **Loose `.md` at skills dir root silently dropped**: `.claw/skills/loose-skill.md` (not inside a per-skill subdirectory) is silently ignored. Discovery only walks `.claw/skills/*/SKILL.md` — no support for flat `.claw/skills/<name>.md`. (c) **Workspace + user skills merged without per-source filter**: `skills list` returns 74 entries including all `~/.claw/skills/*` user-home skills alongside the project skills. There's no `--scope workspace` flag to limit output to just project-local skills; automation has to filter by `source.id == "project_claw"` post-hoc. **Required fix shape:** (a) when SKILL.md frontmatter `name` differs from the parent directory name, emit a `skills_metadata_drift:[{dir_name, frontmatter_name, path}]` array OR enforce `name = dir_name` as a hard rule; if neither, at minimum a stderr warning on each invocation; (b) skill subdirectories without `SKILL.md` should surface as `invalid_skills:[{path, reason}]` in `skills list --output-format json` (same pattern as #440 MCP servers, #441 hooks, #442 agents); (c) support loose `.md` files at skills-dir root OR document explicitly that only subdirectories with `SKILL.md` are discovered; (d) add `--scope workspace|user|all` flag to `skills list` for filtering; (e) regression test: dir/frontmatter mismatch triggers a deterministic warning or error; subdirs without SKILL.md show in invalid array. **Why this matters:** skill discovery is a security-relevant surface — a user's `claw skills run X` could end up running a different skill than they thought if dir-name and frontmatter-name diverge. The silent drops mean users can't tell why their skill files aren't recognized, leading to "I copied the example and it doesn't work" forum questions. Cross-references #440 (MCP all-or-nothing), #441 (hooks all-or-nothing), #442 (agents need TOML, .md dropped), #431 (skills install raw OS error). Source: Jobdori live dogfood, `9e1eafd0`, 2026-05-11.
+
+
+446. **Config is loaded 2-3 times per command invocation; each load re-emits identical deprecation warnings without deduplication — `status` triggers 3× `enabledPlugins` warning, `doctor`/`mcp` trigger 2× each, only `version` (config-free) emits 0** — dogfooded 2026-05-11 by Jobdori on `5a4cc506` in response to Clawhip pinpoint nudge at `1503388740595224717`. Reproduction: with a `~/.claw/settings.json` containing the deprecated `enabledPlugins` key, run each command from a fresh empty cwd and count `warning: ... is deprecated` lines on stderr — `claw status 2>&1 >/dev/null | grep -c deprecated` returns **3**, `claw doctor` returns **2**, `claw mcp` returns **2**, `claw version` returns **0**. Each duplicate is byte-identical (same file path, same line number, same field name). The pattern proves the config-load pipeline is invoked 2-3 times within a single command process; warnings are emitted at each load without checking a `warned_files: HashSet<PathBuf>` deduplication set. **Three sibling implications:** (a) **load-count varies by command** — status:3, doctor:2, mcp:2, version:0 — suggesting each command implements its own config-load call rather than going through a shared cached loader; (b) **noise pollution**: users running `claw status` once see the same 64-character warning 3 times in their terminal scrollback, making real warnings (other config errors, real deprecations) lost in the duplicate noise; (c) **performance signal**: 3× config load means 3× JSON parsing of `~/.claw/settings.json`, `~/.claw.json`, `$CLAW_CONFIG_HOME/settings.json`, and the project-local `.claw.json` / `.claw/settings.json` / `.claw/settings.local.json`. For a workspace with 5 config files, that's 15 redundant disk reads per status invocation. Earlier roadmap entries observed 3× (#424) and 4× (#425) warning counts at different HEADs; the count keeps fluctuating, suggesting the underlying issue is config-load fan-out that nobody has refactored. **Required fix shape:** (a) introduce a `ConfigLoader` cache scoped to the command-process lifetime: first load reads files and emits warnings; subsequent calls hit the cache and emit zero warnings; (b) move config validation/warnings to a single canonical entry point (`ConfigLoader::load_with_diagnostics()` returns `(RuntimeConfig, Vec<Warning>)` exactly once); (c) every command that needs config goes through the cached loader instead of re-reading from disk; (d) `doctor --output-format json` exposes `config_load_count:int` field so we can regression-test that loads are deduplicated; (e) regression test: any single command invocation emits each deprecation warning at most once. **Why this matters:** repeated identical warnings train users to ignore stderr noise. Real warnings (a new deprecation, a config error from a different file, an MCP server failure) get drowned out by 3-4 copies of the same notice. The 15-disk-read worst case is wasted I/O that adds startup latency. The fact that count fluctuates between HEADs (3 at `6c0c305a`, 4 at `d7dbe951`, back to 3 at `5a4cc506`) suggests dev velocity is moving config loads around without an architectural fix. Cross-references #424 (deprecation warning 3×), #425 (deprecation warning 4×), #421 (cwd canonicalization — possibly tied to per-load symlink resolution), #428 (default permission_mode loaded from same config files). Source: Jobdori live dogfood, `5a4cc506`, 2026-05-11.
+
+
+447. **All JSON error envelopes go to STDERR not STDOUT; stdout is empty (0 bytes) on every `--output-format json` failure — breaks the standard automation pattern `output=$(claw cmd --output-format json)` which captures nothing on error and forces ugly `2>&1` redirects to even see the JSON** — dogfooded 2026-05-11 by Jobdori on `5ab969e7` in response to Clawhip pinpoint nudge at `1503396289071808523`. Reproduction (stderr-vs-stdout discipline audit): `claw --no-such-flag --output-format json >stdout.txt 2>stderr.txt` → stdout = **0 bytes**, stderr = 115 bytes containing `{"error":"unknown option: --no-such-flag","hint":"Run \`claw --help\` for usage.","kind":"cli_parse","type":"error"}`. Same pattern across four error envelopes probed: (a) `cli_parse` → stdout 0 / stderr 115; (b) `missing_credentials` → stdout 0 / stderr 853 (includes deprecation warnings ahead of envelope); (c) `session_load_failed` → stdout 0 / stderr 322; (d) `invalid_model_syntax` → stdout 0 / stderr 199. Success paths route correctly: `claw status --output-format json` → stdout 1496 / stderr 0. **The asymmetry is wrong on two axes:** (a) **JSON-format outputs should always go to stdout regardless of success/failure**: every major CLI in this class (kubectl, gh, aws, jq, terraform `-json`, `npm --json`) emits JSON on stdout for both ok and error paths; consumers parse `stdout | jq .kind` and switch on the kind to detect errors. claw's split forces consumers to capture both streams or use `2>&1` which then includes deprecation prose alongside the JSON envelope and breaks parsing. (b) **Deprecation/info warnings leak into the JSON error envelope on stderr**: when stderr is the only path to get the JSON, the deprecation warning prefix (`warning: ... enabledPlugins ... is deprecated`) precedes the JSON, making `tail -1 stderr.txt | jq .` fragile. **Three sibling problems:** (i) **breaks the canonical Bash idiom** `if ! output=$(cmd --output-format json); then echo "$output" | jq .error; fi` — `$output` is empty on error so the `jq` call sees nothing. (ii) **forces N-line stderr parsing**: to get the JSON envelope from stderr, automation must read until EOF, then skip leading `warning:` lines, then parse only the last `{...}` JSON. This is a brittle heuristic that breaks if more warnings are added. (iii) **inconsistent with text mode**: text-mode error output ALSO goes to stderr (e.g., `claw --no-such-flag` → stderr `[error-kind: cli_parse]\nerror: ...`) — that's correct for text mode (stderr is the diagnostic channel). The bug is JSON mode inheriting the same routing. **Required fix shape:** (a) JSON error envelopes go to STDOUT when `--output-format json` is active; (b) keep text-mode error output on stderr (no change for text path); (c) deprecation/info warnings should ALSO go to stderr in JSON mode (they're diagnostic prose, not part of the JSON contract) — separate channels: JSON envelope on stdout, prose warnings on stderr; (d) add `--quiet` / `--no-warn` flag to fully suppress stderr warnings for clean automation; (e) regression test: every `--output-format json` failure path emits the JSON envelope on stdout, exit non-zero, no JSON ever on stderr. **Why this matters:** the entire point of `--output-format json` is enabling automation. Splitting JSON success vs error across stdout vs stderr defeats the purpose — automation must capture both, dedupe sources, and parse mixed streams. Cross-references #422 (exit-code parity across error envelopes), #424 (deprecation warnings noise), #428 (envelope vs prose tension), #446 (multi-load deprecation duplication). Source: Jobdori live dogfood, `5ab969e7`, 2026-05-11.
+
+
+448. **`sandbox --output-format json` has contradictory state flags — `enabled:true, supported:false, active:false, filesystem_active:true, allowed_mounts:[]`: claim that sandbox is "enabled" while OS doesn't support namespace isolation and `allowed_mounts:[]` is empty contradicts `filesystem_active:true filesystem_mode:"workspace-only"`** — dogfooded 2026-05-11 by Jobdori on `7244a82b` in response to Clawhip pinpoint nudge at `1503403842920779917` (using fresh-current-main runner at `/tmp/claw-dog-1430` per gajae's 14:00 protocol switch). Reproduction: `claw sandbox --output-format json` on macOS (where `unshare` is unavailable) returns `{"active":false,"active_namespace":false,"active_network":false,"allowed_mounts":[],"enabled":true,"fallback_reason":"namespace isolation unavailable (requires Linux with \`unshare\`)","filesystem_active":true,"filesystem_mode":"workspace-only","in_container":false,"kind":"sandbox","markers":[],"requested_namespace":true,"requested_network":false,"supported":false}`. **Three contradictions in the same envelope:** (a) `enabled:true` AND `supported:false`: what does "enabled" mean if the OS doesn't support sandboxing? Read literally, sandbox is *enabled but unsupported* — semantic nonsense. The likely intent is "user requested sandbox in config" but the field name `enabled` says "is ON". A better name would be `requested:true` or `config_intent:true`, with `enabled` reserved for the actually-active state. (b) `filesystem_active:true, filesystem_mode:"workspace-only"` AND `allowed_mounts:[]`: if the filesystem fence is active in workspace-only mode, the workspace directory itself MUST be an allowed mount. An empty `allowed_mounts:[]` array combined with `filesystem_active:true` means either (i) the fence is being misreported (it's not really active), (ii) the workspace is implicit and `allowed_mounts` only lists *additional* mounts, or (iii) the fence has no allowed paths and nothing is readable — all three are inconsistent with the user-facing summary. (c) `active:false` AND `filesystem_active:true`: the top-level `active` field is a single boolean summary, but it disagrees with `filesystem_active:true` (one component is active). Either `active` is "all components active" (then it should be `false` when any component is off) or "any component active" (then it should be `true` when filesystem is). The current value is `false` despite filesystem being active. **Sibling: no `claw sandbox --help`**: `claw sandbox status` and `claw sandbox --help` go to LLM-prompt fallback or hang (gajae confirmed at 13:00 that `sandbox status` returns typed `cli_parse` but `sandbox --help` is bounded — schema is non-uniform across help paths). **Required fix shape:** (a) rename `enabled` to `requested` or `config_intent` to disambiguate from "currently active"; (b) make `allowed_mounts` explicitly include the workspace when filesystem_mode is "workspace-only" (`allowed_mounts:[{path:"<cwd>",writable:true,reason:"workspace_root"}]`); (c) document the `active` aggregate semantics: pick either "all" or "any" composition rule and document the choice; (d) add `active_components:["filesystem"]` array as a richer alternative to the single boolean — surfaces exactly which sandbox subsystems are live; (e) regression test: when `filesystem_mode == "workspace-only"`, `allowed_mounts` MUST contain the cwd and `active` must agree with the documented composition rule. **Why this matters:** sandbox is the trust surface — automation that checks `sandbox.active == true` before running a risky LLM prompt sees `false` (no namespace, no network) and assumes no isolation, but `filesystem_active:true` means there IS partial isolation. The mixed signal forces consumers to OR all `*_active` fields together. Cross-references #428 (default permission_mode=danger-full-access — paired with sandbox-not-active means zero isolation), #444 (no broad-cwd guard — sandbox is the only safety net and its status is unclear). Source: Jobdori live dogfood, `7244a82b`, 2026-05-11.
+

rust/crates/api/src/error.rs

@@ -14,6 +14,11 @@ const CONTEXT_WINDOW_ERROR_MARKERS: &[&str] = &[
     "too many tokens",
     "prompt is too long",
     "input is too long",
+    "input tokens exceed",
+    "configured limit",
+    "messages resulted in",
+    "completion tokens",
+    "prompt tokens",
     "request is too large",
 ];
 
@@ -542,6 +547,26 @@ mod tests {
         assert_eq!(error.request_id(), Some("req_ctx_123"));
     }
 
+    #[test]
+    fn classifies_openai_configured_limit_errors_as_context_window_failures() {
+        let error = ApiError::Api {
+            status: reqwest::StatusCode::BAD_REQUEST,
+            error_type: Some("invalid_request_error".to_string()),
+            message: Some(
+                "Input tokens exceed the configured limit of 922000 tokens. Your messages resulted in 1860900 tokens. Please reduce the length of the messages."
+                    .to_string(),
+            ),
+            request_id: Some("req_ctx_openai_123".to_string()),
+            body: String::new(),
+            retryable: false,
+            suggested_action: None,
+        };
+
+        assert!(error.is_context_window_failure());
+        assert_eq!(error.safe_failure_class(), "context_window");
+        assert_eq!(error.request_id(), Some("req_ctx_openai_123"));
+    }
+
     #[test]
     fn missing_credentials_without_hint_renders_the_canonical_message() {
         // given

rust/crates/api/src/lib.rs

@@ -21,11 +21,12 @@ pub use prompt_cache::{
 pub use providers::anthropic::{AnthropicClient, AnthropicClient as ApiClient, AuthSource};
 pub use providers::openai_compat::{
     build_chat_completion_request, flatten_tool_result_content, is_reasoning_model,
-    model_rejects_is_error_field, translate_message, OpenAiCompatClient, OpenAiCompatConfig,
+    model_rejects_is_error_field, model_requires_reasoning_content_in_history, translate_message,
+    OpenAiCompatClient, OpenAiCompatConfig,
 };
 pub use providers::{
     detect_provider_kind, max_tokens_for_model, max_tokens_for_model_with_override,
-    resolve_model_alias, ProviderKind,
+    model_family_identity_for, model_family_identity_for_kind, resolve_model_alias, ProviderKind,
 };
 pub use sse::{parse_frame, SseParser};
 pub use types::{

rust/crates/api/src/providers/mod.rs

@@ -250,19 +250,31 @@ pub fn detect_provider_kind(model: &str) -> ProviderKind {
     ProviderKind::Anthropic
 }
 
+#[must_use]
+pub const fn model_family_identity_for_kind(kind: ProviderKind) -> runtime::ModelFamilyIdentity {
+    match kind {
+        ProviderKind::Anthropic => runtime::ModelFamilyIdentity::Claude,
+        ProviderKind::Xai | ProviderKind::OpenAi => runtime::ModelFamilyIdentity::Generic,
+    }
+}
+
+#[must_use]
+pub fn model_family_identity_for(model: &str) -> runtime::ModelFamilyIdentity {
+    model_family_identity_for_kind(detect_provider_kind(model))
+}
+
 #[must_use]
 pub fn max_tokens_for_model(model: &str) -> u32 {
-    model_token_limit(model).map_or_else(
-        || {
-            let canonical = resolve_model_alias(model);
-            if canonical.contains("opus") {
-                32_000
-            } else {
-                64_000
-            }
-        },
-        |limit| limit.max_output_tokens,
-    )
+    let canonical = resolve_model_alias(model);
+    let heuristic = if canonical.contains("opus") {
+        32_000
+    } else {
+        64_000
+    };
+
+    model_token_limit(model)
+        .map(|limit| heuristic.min(limit.max_output_tokens))
+        .unwrap_or(heuristic)
 }
 
 /// Returns the effective max output tokens for a model, preferring a plugin
@@ -276,7 +288,8 @@ pub fn max_tokens_for_model_with_override(model: &str, plugin_override: Option<u
 #[must_use]
 pub fn model_token_limit(model: &str) -> Option<ModelTokenLimit> {
     let canonical = resolve_model_alias(model);
-    match canonical.as_str() {
+    let base_model = canonical.rsplit('/').next().unwrap_or(canonical.as_str());
+    match base_model {
         "claude-opus-4-6" => Some(ModelTokenLimit {
             max_output_tokens: 32_000,
             context_window_tokens: 200_000,
@@ -289,6 +302,20 @@ pub fn model_token_limit(model: &str) -> Option<ModelTokenLimit> {
             max_output_tokens: 64_000,
             context_window_tokens: 131_072,
         }),
+        // GPT-4.1 family via the OpenAI API.
+        "gpt-4.1" | "gpt-4.1-mini" | "gpt-4.1-nano" => Some(ModelTokenLimit {
+            max_output_tokens: 32_768,
+            context_window_tokens: 1_047_576,
+        }),
+        // GPT-5.4 family via the OpenAI API.
+        "gpt-5.4" => Some(ModelTokenLimit {
+            max_output_tokens: 128_000,
+            context_window_tokens: 1_000_000,
+        }),
+        "gpt-5.4-mini" | "gpt-5.4-nano" => Some(ModelTokenLimit {
+            max_output_tokens: 128_000,
+            context_window_tokens: 400_000,
+        }),
         // Kimi models via DashScope (Moonshot AI)
         // Source: https://platform.moonshot.cn/docs/intro
         "kimi-k2.5" | "kimi-k1.5" => Some(ModelTokenLimit {
@@ -470,8 +497,8 @@ mod tests {
     use super::{
         anthropic_missing_credentials, anthropic_missing_credentials_hint, detect_provider_kind,
         load_dotenv_file, max_tokens_for_model, max_tokens_for_model_with_override,
-        model_token_limit, parse_dotenv, preflight_message_request, resolve_model_alias,
-        ProviderKind,
+        model_family_identity_for, model_family_identity_for_kind, model_token_limit, parse_dotenv,
+        preflight_message_request, resolve_model_alias, ProviderKind,
     };
 
     /// Serializes every test in this module that mutates process-wide
@@ -530,6 +557,42 @@ mod tests {
         );
     }
 
+    #[test]
+    fn maps_provider_kind_to_model_family_identity() {
+        // given: each supported provider kind
+        let anthropic = ProviderKind::Anthropic;
+        let openai = ProviderKind::OpenAi;
+        let xai = ProviderKind::Xai;
+
+        // when: converting provider kinds to prompt model family identities
+        let anthropic_identity = model_family_identity_for_kind(anthropic);
+        let openai_identity = model_family_identity_for_kind(openai);
+        let xai_identity = model_family_identity_for_kind(xai);
+
+        // then: Anthropic stays Claude and OpenAI-compatible providers are generic
+        assert_eq!(anthropic_identity, runtime::ModelFamilyIdentity::Claude);
+        assert_eq!(openai_identity, runtime::ModelFamilyIdentity::Generic);
+        assert_eq!(xai_identity, runtime::ModelFamilyIdentity::Generic);
+    }
+
+    #[test]
+    fn maps_model_name_to_model_family_identity() {
+        // given: Anthropic, OpenAI-compatible, and xAI model names
+        let claude_model = "claude-opus-4-6";
+        let openai_model = "openai/gpt-4.1-mini";
+        let xai_model = "grok-3";
+
+        // when: detecting prompt model family identities from model names
+        let claude_identity = model_family_identity_for(claude_model);
+        let openai_identity = model_family_identity_for(openai_model);
+        let xai_identity = model_family_identity_for(xai_model);
+
+        // then: Anthropic stays Claude and OpenAI-compatible providers are generic
+        assert_eq!(claude_identity, runtime::ModelFamilyIdentity::Claude);
+        assert_eq!(openai_identity, runtime::ModelFamilyIdentity::Generic);
+        assert_eq!(xai_identity, runtime::ModelFamilyIdentity::Generic);
+    }
+
     #[test]
     fn openai_namespaced_model_routes_to_openai_not_anthropic() {
         // Regression: "openai/gpt-4.1-mini" was misrouted to Anthropic when
@@ -614,6 +677,15 @@ mod tests {
     fn keeps_existing_max_token_heuristic() {
         assert_eq!(max_tokens_for_model("opus"), 32_000);
         assert_eq!(max_tokens_for_model("grok-3"), 64_000);
+        assert_eq!(max_tokens_for_model("gpt-5.4"), 64_000);
+    }
+
+    #[test]
+    fn caps_default_max_tokens_to_openai_model_limits() {
+        assert_eq!(max_tokens_for_model("gpt-4.1-mini"), 32_768);
+        assert_eq!(max_tokens_for_model("openai/gpt-4.1-mini"), 32_768);
+        assert_eq!(max_tokens_for_model("gpt-5.4"), 64_000);
+        assert_eq!(max_tokens_for_model("openai/gpt-5.4"), 64_000);
     }
 
     #[test]
@@ -680,6 +752,18 @@ mod tests {
                 .context_window_tokens,
             131_072
         );
+        assert_eq!(
+            model_token_limit("openai/gpt-4.1-mini")
+                .expect("openai/gpt-4.1-mini should be registered")
+                .context_window_tokens,
+            1_047_576
+        );
+        assert_eq!(
+            model_token_limit("gpt-5.4")
+                .expect("gpt-5.4 should be registered")
+                .context_window_tokens,
+            1_000_000
+        );
     }
 
     #[test]
@@ -728,6 +812,42 @@ mod tests {
         }
     }
 
+    #[test]
+    fn preflight_blocks_oversized_requests_for_gpt_5_4() {
+        let request = MessageRequest {
+            model: "gpt-5.4".to_string(),
+            max_tokens: 64_000,
+            messages: vec![InputMessage {
+                role: "user".to_string(),
+                content: vec![InputContentBlock::Text {
+                    text: "x".repeat(3_900_000),
+                }],
+            }],
+            system: Some("Keep the answer short.".to_string()),
+            tools: None,
+            tool_choice: None,
+            stream: true,
+            ..Default::default()
+        };
+
+        let error = preflight_message_request(&request)
+            .expect_err("oversized gpt-5.4 request should be rejected before the provider call");
+
+        match error {
+            ApiError::ContextWindowExceeded {
+                model,
+                requested_output_tokens,
+                context_window_tokens,
+                ..
+            } => {
+                assert_eq!(model, "gpt-5.4");
+                assert_eq!(requested_output_tokens, 64_000);
+                assert_eq!(context_window_tokens, 1_000_000);
+            }
+            other => panic!("expected context-window preflight failure, got {other:?}"),
+        }
+    }
+
     #[test]
     fn preflight_skips_unknown_models() {
         let request = MessageRequest {

rust/crates/api/src/providers/openai_compat.rs

@@ -443,6 +443,8 @@ struct StreamState {
     stop_reason: Option<String>,
     usage: Option<Usage>,
     tool_calls: BTreeMap<u32, ToolCallState>,
+    thinking_started: bool,
+    thinking_finished: bool,
 }
 
 impl StreamState {
@@ -456,6 +458,8 @@ impl StreamState {
             stop_reason: None,
             usage: None,
             tool_calls: BTreeMap::new(),
+            thinking_started: false,
+            thinking_finished: false,
         }
     }
 
@@ -493,35 +497,61 @@ impl StreamState {
         }
 
         for choice in chunk.choices {
+            if let Some(reasoning) = choice
+                .delta
+                .reasoning_content
+                .filter(|value| !value.is_empty())
+            {
+                if !self.thinking_started {
+                    self.thinking_started = true;
+                    events.push(StreamEvent::ContentBlockStart(ContentBlockStartEvent {
+                        index: 0,
+                        content_block: OutputContentBlock::Thinking {
+                            thinking: String::new(),
+                            signature: None,
+                        },
+                    }));
+                }
+                events.push(StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
+                    index: 0,
+                    delta: ContentBlockDelta::ThinkingDelta {
+                        thinking: reasoning,
+                    },
+                }));
+            }
+
             if let Some(content) = choice.delta.content.filter(|value| !value.is_empty()) {
+                self.close_thinking(&mut events);
                 if !self.text_started {
                     self.text_started = true;
                     events.push(StreamEvent::ContentBlockStart(ContentBlockStartEvent {
-                        index: 0,
+                        index: self.text_block_index(),
                         content_block: OutputContentBlock::Text {
                             text: String::new(),
                         },
                     }));
                 }
                 events.push(StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
-                    index: 0,
+                    index: self.text_block_index(),
                     delta: ContentBlockDelta::TextDelta { text: content },
                 }));
             }
 
             for tool_call in choice.delta.tool_calls {
+                self.close_thinking(&mut events);
+                let tool_index_offset = self.tool_index_offset();
                 let state = self.tool_calls.entry(tool_call.index).or_default();
                 state.apply(tool_call);
-                let block_index = state.block_index();
+                let block_index = state.block_index(tool_index_offset);
                 if !state.started {
-                    if let Some(start_event) = state.start_event()? {
+                    if let Some(start_event) = state.start_event(tool_index_offset)? {
                         state.started = true;
                         events.push(StreamEvent::ContentBlockStart(start_event));
                     } else {
                         continue;
                     }
                 }
-                if let Some(delta_event) = state.delta_event() {
+                if let Some(delta_event) = state.delta_event(tool_index_offset) {
                     events.push(StreamEvent::ContentBlockDelta(delta_event));
                 }
                 if choice.finish_reason.as_deref() == Some("tool_calls") && !state.stopped {
@@ -535,11 +565,12 @@ impl StreamState {
             if let Some(finish_reason) = choice.finish_reason {
                 self.stop_reason = Some(normalize_finish_reason(&finish_reason));
                 if finish_reason == "tool_calls" {
+                    let tool_index_offset = self.tool_index_offset();
                     for state in self.tool_calls.values_mut() {
                         if state.started && !state.stopped {
                             state.stopped = true;
                             events.push(StreamEvent::ContentBlockStop(ContentBlockStopEvent {
-                                index: state.block_index(),
+                                index: state.block_index(tool_index_offset),
                             }));
                         }
                     }
@@ -557,19 +588,21 @@ impl StreamState {
         self.finished = true;
 
         let mut events = Vec::new();
+        self.close_thinking(&mut events);
         if self.text_started && !self.text_finished {
             self.text_finished = true;
             events.push(StreamEvent::ContentBlockStop(ContentBlockStopEvent {
-                index: 0,
+                index: self.text_block_index(),
             }));
         }
 
+        let tool_index_offset = self.tool_index_offset();
         for state in self.tool_calls.values_mut() {
             if !state.started {
-                if let Some(start_event) = state.start_event()? {
+                if let Some(start_event) = state.start_event(tool_index_offset)? {
                     state.started = true;
                     events.push(StreamEvent::ContentBlockStart(start_event));
-                    if let Some(delta_event) = state.delta_event() {
+                    if let Some(delta_event) = state.delta_event(tool_index_offset) {
                         events.push(StreamEvent::ContentBlockDelta(delta_event));
                     }
                 }
@@ -577,7 +610,7 @@ impl StreamState {
             if state.started && !state.stopped {
                 state.stopped = true;
                 events.push(StreamEvent::ContentBlockStop(ContentBlockStopEvent {
-                    index: state.block_index(),
+                    index: state.block_index(tool_index_offset),
                 }));
             }
         }
@@ -603,6 +636,31 @@ impl StreamState {
         }
         Ok(events)
     }
+
+    fn close_thinking(&mut self, events: &mut Vec<StreamEvent>) {
+        if self.thinking_started && !self.thinking_finished {
+            self.thinking_finished = true;
+            events.push(StreamEvent::ContentBlockStop(ContentBlockStopEvent {
+                index: 0,
+            }));
+        }
+    }
+
+    const fn text_block_index(&self) -> u32 {
+        if self.thinking_started {
+            1
+        } else {
+            0
+        }
+    }
+
+    const fn tool_index_offset(&self) -> u32 {
+        if self.thinking_started {
+            2
+        } else {
+            1
+        }
+    }
 }
 
 #[derive(Debug, Default)]
@@ -630,12 +688,12 @@ impl ToolCallState {
         }
     }
 
-    const fn block_index(&self) -> u32 {
-        self.openai_index + 1
+    const fn block_index(&self, offset: u32) -> u32 {
+        self.openai_index + offset
     }
 
     #[allow(clippy::unnecessary_wraps)]
-    fn start_event(&self) -> Result<Option<ContentBlockStartEvent>, ApiError> {
+    fn start_event(&self, offset: u32) -> Result<Option<ContentBlockStartEvent>, ApiError> {
         let Some(name) = self.name.clone() else {
             return Ok(None);
         };
@@ -644,7 +702,7 @@ impl ToolCallState {
             .clone()
             .unwrap_or_else(|| format!("tool_call_{}", self.openai_index));
         Ok(Some(ContentBlockStartEvent {
-            index: self.block_index(),
+            index: self.block_index(offset),
             content_block: OutputContentBlock::ToolUse {
                 id,
                 name,
@@ -653,14 +711,14 @@ impl ToolCallState {
         }))
     }
 
-    fn delta_event(&mut self) -> Option<ContentBlockDeltaEvent> {
+    fn delta_event(&mut self, offset: u32) -> Option<ContentBlockDeltaEvent> {
         if self.emitted_len >= self.arguments.len() {
             return None;
         }
         let delta = self.arguments[self.emitted_len..].to_string();
         self.emitted_len = self.arguments.len();
         Some(ContentBlockDeltaEvent {
-            index: self.block_index(),
+            index: self.block_index(offset),
             delta: ContentBlockDelta::InputJsonDelta {
                 partial_json: delta,
             },
@@ -690,6 +748,8 @@ struct ChatMessage {
     #[serde(default)]
     content: Option<String>,
     #[serde(default)]
+    reasoning_content: Option<String>,
+    #[serde(default)]
     tool_calls: Vec<ResponseToolCall>,
 }
 
@@ -735,6 +795,8 @@ struct ChunkChoice {
 struct ChunkDelta {
     #[serde(default)]
     content: Option<String>,
+    #[serde(default)]
+    reasoning_content: Option<String>,
     #[serde(default, deserialize_with = "deserialize_null_as_empty_vec")]
     tool_calls: Vec<DeltaToolCall>,
 }
@@ -793,6 +855,15 @@ pub fn is_reasoning_model(model: &str) -> bool {
         || canonical.contains("thinking")
 }
 
+/// Returns true for OpenAI-compatible DeepSeek V4 models that require prior
+/// assistant reasoning to be echoed back as `reasoning_content` in history.
+#[must_use]
+pub fn model_requires_reasoning_content_in_history(model: &str) -> bool {
+    let lowered = model.to_ascii_lowercase();
+    let canonical = lowered.rsplit('/').next().unwrap_or(lowered.as_str());
+    canonical.starts_with("deepseek-v4")
+}
+
 /// Strip routing prefix (e.g., "openai/gpt-4" → "gpt-4") for the wire.
 /// The prefix is used only to select transport; the backend expects the
 /// bare model id.
@@ -948,10 +1019,14 @@ pub fn translate_message(message: &InputMessage, model: &str) -> Vec<Value> {
     match message.role.as_str() {
         "assistant" => {
             let mut text = String::new();
+            let mut reasoning = String::new();
             let mut tool_calls = Vec::new();
             for block in &message.content {
                 match block {
                     InputContentBlock::Text { text: value } => text.push_str(value),
+                    InputContentBlock::Thinking {
+                        thinking: value, ..
+                    } => reasoning.push_str(value),
                     InputContentBlock::ToolUse { id, name, input } => tool_calls.push(json!({
                         "id": id,
                         "type": "function",
@@ -963,13 +1038,18 @@ pub fn translate_message(message: &InputMessage, model: &str) -> Vec<Value> {
                     InputContentBlock::ToolResult { .. } => {}
                 }
             }
-            if text.is_empty() && tool_calls.is_empty() {
+            let include_reasoning =
+                model_requires_reasoning_content_in_history(model) && !reasoning.is_empty();
+            if text.is_empty() && tool_calls.is_empty() && !include_reasoning {
                 Vec::new()
             } else {
                 let mut msg = serde_json::json!({
                     "role": "assistant",
                     "content": (!text.is_empty()).then_some(text),
                 });
+                if include_reasoning {
+                    msg["reasoning_content"] = json!(reasoning);
+                }
                 // Only include tool_calls when non-empty: some providers reject
                 // assistant messages with an explicit empty tool_calls array.
                 if !tool_calls.is_empty() {
@@ -1003,6 +1083,7 @@ pub fn translate_message(message: &InputMessage, model: &str) -> Vec<Value> {
                     }
                     Some(msg)
                 }
+                InputContentBlock::Thinking { .. } => None,
                 InputContentBlock::ToolUse { .. } => None,
             })
             .collect(),
@@ -1182,6 +1263,16 @@ fn normalize_response(
             "chat completion response missing choices",
         ))?;
     let mut content = Vec::new();
+    if let Some(thinking) = choice
+        .message
+        .reasoning_content
+        .filter(|value| !value.is_empty())
+    {
+        content.push(OutputContentBlock::Thinking {
+            thinking,
+            signature: None,
+        });
+    }
     if let Some(text) = choice.message.content.filter(|value| !value.is_empty()) {
         content.push(OutputContentBlock::Text { text });
     }
@@ -1413,13 +1504,15 @@ impl StringExt for String {
 mod tests {
     use super::{
         build_chat_completion_request, chat_completions_endpoint, is_reasoning_model,
-        normalize_finish_reason, openai_tool_choice, parse_tool_arguments, OpenAiCompatClient,
-        OpenAiCompatConfig,
+        model_requires_reasoning_content_in_history, normalize_finish_reason, normalize_response,
+        openai_tool_choice, parse_tool_arguments, OpenAiCompatClient, OpenAiCompatConfig,
+        StreamState,
     };
     use crate::error::ApiError;
     use crate::types::{
-        InputContentBlock, InputMessage, MessageRequest, ToolChoice, ToolDefinition,
-        ToolResultContentBlock,
+        ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
+        InputContentBlock, InputMessage, MessageRequest, OutputContentBlock, StreamEvent,
+        ToolChoice, ToolDefinition, ToolResultContentBlock,
     };
     use serde_json::json;
     use std::sync::{Mutex, OnceLock};
@@ -1465,6 +1558,188 @@ mod tests {
         assert_eq!(payload["tool_choice"], json!("auto"));
     }
 
+    #[test]
+    fn model_requires_reasoning_content_in_history_detects_deepseek_v4_models() {
+        // Given DeepSeek V4 and non-V4 model names.
+        let positive = [
+            "deepseek-v4-flash",
+            "deepseek-v4-pro",
+            "openai/deepseek-v4-pro",
+            "deepseek/deepseek-v4-flash",
+        ];
+        let negative = [
+            "deepseek-reasoner",
+            "deepseek-chat",
+            "gpt-4o",
+            "claude-sonnet-4-6",
+        ];
+
+        // When checking whether history reasoning_content is required.
+        // Then only DeepSeek V4 variants require it.
+        for model in positive {
+            assert!(model_requires_reasoning_content_in_history(model));
+        }
+        for model in negative {
+            assert!(!model_requires_reasoning_content_in_history(model));
+        }
+    }
+
+    #[test]
+    fn legacy_deepseek_reasoner_request_omits_reasoning_content_for_assistant_history() {
+        // Given an assistant history turn containing thinking.
+        let request = assistant_history_with_thinking_request("deepseek-reasoner");
+
+        // When serializing for legacy deepseek-reasoner.
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+
+        // Then reasoning_content is omitted.
+        let assistant = &payload["messages"][0];
+        assert_eq!(assistant["role"], json!("assistant"));
+        assert!(assistant.get("reasoning_content").is_none());
+    }
+
+    #[test]
+    fn deepseek_v4_pro_request_includes_reasoning_content_for_assistant_history() {
+        // Given an assistant history turn containing thinking.
+        let request = assistant_history_with_thinking_request("openai/deepseek-v4-pro");
+
+        // When serializing for DeepSeek V4 Pro.
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+
+        // Then reasoning_content is included on the assistant message.
+        let assistant = &payload["messages"][0];
+        assert_eq!(assistant["reasoning_content"], json!("prior reasoning"));
+        assert_eq!(assistant["content"], json!("answer"));
+    }
+
+    #[test]
+    fn deepseek_v4_flash_request_includes_reasoning_content_for_assistant_history() {
+        // Given an assistant history turn containing thinking.
+        let request = assistant_history_with_thinking_request("deepseek-v4-flash");
+
+        // When serializing for DeepSeek V4 Flash.
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+
+        // Then reasoning_content is included on the assistant message.
+        let assistant = &payload["messages"][0];
+        assert_eq!(assistant["reasoning_content"], json!("prior reasoning"));
+    }
+
+    #[test]
+    fn non_streaming_response_with_reasoning_content_emits_thinking_block_first() {
+        // Given a non-streaming OpenAI-compatible response with reasoning_content.
+        let response = super::ChatCompletionResponse {
+            id: "chatcmpl_reasoning".to_string(),
+            model: "deepseek-v4-pro".to_string(),
+            choices: vec![super::ChatChoice {
+                message: super::ChatMessage {
+                    role: "assistant".to_string(),
+                    content: Some("final answer".to_string()),
+                    reasoning_content: Some("hidden thought".to_string()),
+                    tool_calls: Vec::new(),
+                },
+                finish_reason: Some("stop".to_string()),
+            }],
+            usage: None,
+        };
+
+        // When normalizing the provider response.
+        let normalized = normalize_response("deepseek-v4-pro", response).expect("normalized");
+
+        // Then Thinking is the first content block, before text.
+        assert_eq!(
+            normalized.content,
+            vec![
+                OutputContentBlock::Thinking {
+                    thinking: "hidden thought".to_string(),
+                    signature: None,
+                },
+                OutputContentBlock::Text {
+                    text: "final answer".to_string(),
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn streaming_chunks_with_reasoning_content_emit_thinking_block_events_before_text() {
+        // Given streaming chunks with reasoning_content followed by text.
+        let mut state = StreamState::new("deepseek-v4-pro".to_string());
+        let mut events = state
+            .ingest_chunk(super::ChatCompletionChunk {
+                id: "chatcmpl_stream_reasoning".to_string(),
+                model: Some("deepseek-v4-pro".to_string()),
+                choices: vec![super::ChunkChoice {
+                    delta: super::ChunkDelta {
+                        content: None,
+                        reasoning_content: Some("think".to_string()),
+                        tool_calls: Vec::new(),
+                    },
+                    finish_reason: None,
+                }],
+                usage: None,
+            })
+            .expect("reasoning chunk");
+        events.extend(
+            state
+                .ingest_chunk(super::ChatCompletionChunk {
+                    id: "chatcmpl_stream_reasoning".to_string(),
+                    model: None,
+                    choices: vec![super::ChunkChoice {
+                        delta: super::ChunkDelta {
+                            content: Some(" answer".to_string()),
+                            reasoning_content: None,
+                            tool_calls: Vec::new(),
+                        },
+                        finish_reason: Some("stop".to_string()),
+                    }],
+                    usage: None,
+                })
+                .expect("text chunk"),
+        );
+        events.extend(state.finish().expect("finish"));
+
+        // When reading normalized stream events.
+        // Then Thinking starts at index 0, text is offset to index 1.
+        assert!(matches!(events[0], StreamEvent::MessageStart(_)));
+        assert!(matches!(
+            events[1],
+            StreamEvent::ContentBlockStart(ContentBlockStartEvent {
+                index: 0,
+                content_block: OutputContentBlock::Thinking { .. },
+            })
+        ));
+        assert!(matches!(
+            events[2],
+            StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
+                index: 0,
+                delta: ContentBlockDelta::ThinkingDelta { .. },
+            })
+        ));
+        assert!(matches!(
+            events[3],
+            StreamEvent::ContentBlockStop(ContentBlockStopEvent { index: 0 })
+        ));
+        assert!(matches!(
+            events[4],
+            StreamEvent::ContentBlockStart(ContentBlockStartEvent {
+                index: 1,
+                content_block: OutputContentBlock::Text { .. },
+            })
+        ));
+        assert!(matches!(
+            events[5],
+            StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
+                index: 1,
+                delta: ContentBlockDelta::TextDelta { .. },
+            })
+        ));
+        assert!(matches!(
+            events[6],
+            StreamEvent::ContentBlockStop(ContentBlockStopEvent { index: 1 })
+        ));
+    }
+
     #[test]
     fn tool_schema_object_gets_strict_fields_for_responses_endpoint() {
         // OpenAI /responses endpoint rejects object schemas missing
@@ -1624,6 +1899,27 @@ mod tests {
         );
     }
 
+    fn assistant_history_with_thinking_request(model: &str) -> MessageRequest {
+        MessageRequest {
+            model: model.to_string(),
+            max_tokens: 100,
+            messages: vec![InputMessage {
+                role: "assistant".to_string(),
+                content: vec![
+                    InputContentBlock::Thinking {
+                        thinking: "prior reasoning".to_string(),
+                        signature: None,
+                    },
+                    InputContentBlock::Text {
+                        text: "answer".to_string(),
+                    },
+                ],
+            }],
+            stream: false,
+            ..Default::default()
+        }
+    }
+
     fn env_lock() -> std::sync::MutexGuard<'static, ()> {
         static LOCK: OnceLock<Mutex<()>> = OnceLock::new();
         LOCK.get_or_init(|| Mutex::new(()))

rust/crates/api/src/types.rs

@@ -81,6 +81,11 @@ pub enum InputContentBlock {
     Text {
         text: String,
     },
+    Thinking {
+        thinking: String,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        signature: Option<String>,
+    },
     ToolUse {
         id: String,
         name: String,
@@ -268,8 +273,9 @@ pub enum StreamEvent {
 #[cfg(test)]
 mod tests {
     use runtime::format_usd;
+    use serde_json::json;
 
-    use super::{MessageResponse, Usage};
+    use super::{InputContentBlock, MessageResponse, Usage};
 
     #[test]
     fn usage_total_tokens_includes_cache_tokens() {
@@ -307,4 +313,33 @@ mod tests {
         assert_eq!(format_usd(cost.total_cost_usd()), "$54.6750");
         assert_eq!(response.total_tokens(), 1_800_000);
     }
+
+    #[test]
+    fn input_content_block_thinking_serializes_with_snake_case_type() {
+        // given
+        let block = InputContentBlock::Thinking {
+            thinking: "pondering".to_string(),
+            signature: Some("sig_123".to_string()),
+        };
+
+        // when
+        let serialized = serde_json::to_value(&block).unwrap();
+        let deserialized: InputContentBlock = serde_json::from_value(json!({
+            "type": "thinking",
+            "thinking": "pondering",
+            "signature": "sig_123"
+        }))
+        .unwrap();
+
+        // then
+        assert_eq!(
+            serialized,
+            json!({
+                "type": "thinking",
+                "thinking": "pondering",
+                "signature": "sig_123"
+            })
+        );
+        assert_eq!(deserialized, block);
+    }
 }

rust/crates/api/tests/openai_compat_integration.rs

@@ -63,6 +63,50 @@ async fn send_message_uses_openai_compatible_endpoint_and_auth() {
     assert_eq!(body["tools"][0]["type"], json!("function"));
 }
 
+#[tokio::test]
+async fn send_message_preserves_deepseek_reasoning_content_before_text() {
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let body = concat!(
+        "{",
+        "\"id\":\"chatcmpl_deepseek_reasoning\",",
+        "\"model\":\"deepseek-v4-pro\",",
+        "\"choices\":[{",
+        "\"message\":{\"role\":\"assistant\",\"reasoning_content\":\"Think first\",\"content\":\"Answer second\",\"tool_calls\":[]},",
+        "\"finish_reason\":\"stop\"",
+        "}],",
+        "\"usage\":{\"prompt_tokens\":11,\"completion_tokens\":5}",
+        "}"
+    );
+    let server = spawn_server(
+        state.clone(),
+        vec![http_response("200 OK", "application/json", body)],
+    )
+    .await;
+
+    let client = OpenAiCompatClient::new("openai-test-key", OpenAiCompatConfig::openai())
+        .with_base_url(server.base_url());
+    let response = client
+        .send_message(&MessageRequest {
+            model: "openai/deepseek-v4-pro".to_string(),
+            ..sample_request(false)
+        })
+        .await
+        .expect("request should succeed");
+
+    assert_eq!(
+        response.content,
+        vec![
+            OutputContentBlock::Thinking {
+                thinking: "Think first".to_string(),
+                signature: None,
+            },
+            OutputContentBlock::Text {
+                text: "Answer second".to_string(),
+            },
+        ]
+    );
+}
+
 #[tokio::test]
 async fn send_message_blocks_oversized_xai_requests_before_the_http_call() {
     let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));

rust/crates/commands/src/lib.rs

@@ -2371,6 +2371,40 @@ pub fn handle_skills_slash_command(args: Option<&str>, cwd: &Path) -> std::io::R
             let skills = load_skills_from_roots(&roots)?;
             Ok(render_skills_report(&skills))
         }
+        Some(args) if args.starts_with("list ") => {
+            let filter = args["list ".len()..].trim().to_lowercase();
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            let filtered: Vec<_> = skills
+                .into_iter()
+                .filter(|s| s.name.to_lowercase().contains(&filter))
+                .collect();
+            Ok(render_skills_report(&filtered))
+        }
+        Some("show" | "info" | "describe") => {
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            Ok(render_skills_report(&skills))
+        }
+        Some(args)
+            if args.starts_with("show ")
+                || args.starts_with("info ")
+                || args.starts_with("describe ") =>
+        {
+            let name = args
+                .splitn(2, ' ')
+                .nth(1)
+                .unwrap_or_default()
+                .trim()
+                .to_lowercase();
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            let matched: Vec<_> = skills
+                .into_iter()
+                .filter(|s| s.name.to_lowercase() == name)
+                .collect();
+            Ok(render_skills_report(&matched))
+        }
         Some("install") => Ok(render_skills_usage(Some("install"))),
         Some(args) if args.starts_with("install ") => {
             let target = args["install ".len()..].trim();
@@ -2402,6 +2436,40 @@ pub fn handle_skills_slash_command_json(args: Option<&str>, cwd: &Path) -> std::
             let skills = load_skills_from_roots(&roots)?;
             Ok(render_skills_report_json(&skills))
         }
+        Some(args) if args.starts_with("list ") => {
+            let filter = args["list ".len()..].trim().to_lowercase();
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            let filtered: Vec<_> = skills
+                .into_iter()
+                .filter(|s| s.name.to_lowercase().contains(&filter))
+                .collect();
+            Ok(render_skills_report_json(&filtered))
+        }
+        Some("show" | "info" | "describe") => {
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            Ok(render_skills_report_json(&skills))
+        }
+        Some(args)
+            if args.starts_with("show ")
+                || args.starts_with("info ")
+                || args.starts_with("describe ") =>
+        {
+            let name = args
+                .splitn(2, ' ')
+                .nth(1)
+                .unwrap_or_default()
+                .trim()
+                .to_lowercase();
+            let roots = discover_skill_roots(cwd);
+            let skills = load_skills_from_roots(&roots)?;
+            let matched: Vec<_> = skills
+                .into_iter()
+                .filter(|s| s.name.to_lowercase() == name)
+                .collect();
+            Ok(render_skills_report_json(&matched))
+        }
         Some("install") => Ok(render_skills_usage_json(Some("install"))),
         Some(args) if args.starts_with("install ") => {
             let target = args["install ".len()..].trim();
@@ -2419,10 +2487,27 @@ pub fn handle_skills_slash_command_json(args: Option<&str>, cwd: &Path) -> std::
 #[must_use]
 pub fn classify_skills_slash_command(args: Option<&str>) -> SkillSlashDispatch {
     match normalize_optional_args(args) {
-        None | Some("list" | "help" | "-h" | "--help") => SkillSlashDispatch::Local,
+        None | Some("list" | "help" | "-h" | "--help" | "show" | "info" | "describe") => {
+            SkillSlashDispatch::Local
+        }
+        Some(args)
+            if args
+                .split_whitespace()
+                .any(|part| matches!(part, "-h" | "--help")) =>
+        {
+            SkillSlashDispatch::Local
+        }
         Some(args) if args == "install" || args.starts_with("install ") => {
             SkillSlashDispatch::Local
         }
+        Some(args)
+            if args.starts_with("list ")
+                || args.starts_with("show ")
+                || args.starts_with("info ")
+                || args.starts_with("describe ") =>
+        {
+            SkillSlashDispatch::Local
+        }
         Some(args) => SkillSlashDispatch::Invoke(format!("${}", args.trim_start_matches('/'))),
     }
 }
@@ -2596,10 +2681,44 @@ fn render_mcp_report_for(
                 )),
             }
         }
+        Some(args) if args.split_whitespace().next() == Some("list") && args.contains(' ') => {
+            // `mcp list <filter>` — list does not accept arguments; treat as unsupported action.
+            Ok(render_mcp_unsupported_action_text(
+                args,
+                "list accepts no filter argument; use `claw mcp list`",
+            ))
+        }
+        Some(args) if matches!(args.split_whitespace().next(), Some("info" | "describe")) => {
+            Ok(render_mcp_unsupported_action_text(
+                args,
+                "use `claw mcp show <server>` to inspect a server",
+            ))
+        }
         Some(args) => Ok(render_mcp_usage(Some(args))),
     }
 }
 
+fn render_mcp_unsupported_action_text(action: &str, hint: &str) -> String {
+    format!(
+        "MCP\n  Error            unsupported action '{action}'\n  Hint             {hint}\n  Usage            /mcp [list|show <server>|help]"
+    )
+}
+
+fn render_mcp_unsupported_action_json(action: &str, hint: &str) -> Value {
+    json!({
+        "kind": "mcp",
+        "action": "error",
+        "ok": false,
+        "error_kind": "unsupported_action",
+        "requested_action": action,
+        "hint": hint,
+        "usage": {
+            "slash_command": "/mcp [list|show <server>|help]",
+            "direct_cli": "claw mcp [list|show <server>|help]",
+        },
+    })
+}
+
 fn render_mcp_report_json_for(
     loader: &ConfigLoader,
     cwd: &Path,
@@ -2680,6 +2799,18 @@ fn render_mcp_report_json_for(
                 })),
             }
         }
+        Some(args) if args.split_whitespace().next() == Some("list") && args.contains(' ') => {
+            Ok(render_mcp_unsupported_action_json(
+                args,
+                "list accepts no filter argument; use `claw mcp list`",
+            ))
+        }
+        Some(args) if matches!(args.split_whitespace().next(), Some("info" | "describe")) => {
+            Ok(render_mcp_unsupported_action_json(
+                args,
+                "use `claw mcp show <server>` to inspect a server",
+            ))
+        }
         Some(args) => Ok(render_mcp_usage_json(Some(args))),
     }
 }
@@ -4619,6 +4750,32 @@ mod tests {
         assert!(agents_error.contains("  Usage            /agents [list|help]"));
     }
 
+    #[test]
+    fn skills_show_and_list_filter_do_not_invoke_model() {
+        // `show`, `info`, `list <filter>` must route to Local, not Invoke.
+        // Regression for: `claw skills show plan` unexpectedly spawned a model session.
+        for token in &["show", "info", "describe"] {
+            assert_eq!(
+                classify_skills_slash_command(Some(token)),
+                SkillSlashDispatch::Local,
+                "`skills {token}` alone must be Local"
+            );
+        }
+        for prefix in &["show ", "info ", "list ", "describe "] {
+            let arg = format!("{prefix}plan");
+            assert_eq!(
+                classify_skills_slash_command(Some(&arg)),
+                SkillSlashDispatch::Local,
+                "`skills {arg}` must be Local, not Invoke"
+            );
+        }
+        // Bare invocable tokens still dispatch to Invoke.
+        assert_eq!(
+            classify_skills_slash_command(Some("plan")),
+            SkillSlashDispatch::Invoke("$plan".to_string()),
+        );
+    }
+
     #[test]
     fn accepts_skills_invocation_arguments_for_prompt_dispatch() {
         assert_eq!(
@@ -4641,6 +4798,38 @@ mod tests {
         );
     }
 
+    #[test]
+    fn mcp_unsupported_actions_return_typed_error_not_generic_help() {
+        // `mcp info <name>` and `mcp list <filter>` must return typed errors, not raw help.
+        // Regression for #504: these previously fell through to render_mcp_usage with
+        // unexpected=arg, giving no machine-readable error_kind.
+        use crate::handle_mcp_slash_command_json;
+        use std::path::PathBuf;
+        let cwd = PathBuf::from("/tmp");
+
+        let info_json = handle_mcp_slash_command_json(Some("info nonexistent"), &cwd)
+            .expect("info nonexistent should not error at IO level");
+        assert_eq!(info_json["kind"], "mcp");
+        assert_eq!(info_json["ok"], false);
+        assert_eq!(info_json["error_kind"], "unsupported_action");
+        assert!(info_json["hint"]
+            .as_str()
+            .unwrap_or_default()
+            .contains("show"));
+
+        let list_filter_json = handle_mcp_slash_command_json(Some("list nonexistent"), &cwd)
+            .expect("list nonexistent should not error at IO level");
+        assert_eq!(list_filter_json["kind"], "mcp");
+        assert_eq!(list_filter_json["ok"], false);
+        assert_eq!(list_filter_json["error_kind"], "unsupported_action");
+
+        let describe_json = handle_mcp_slash_command_json(Some("describe myserver"), &cwd)
+            .expect("describe myserver should not error at IO level");
+        assert_eq!(describe_json["kind"], "mcp");
+        assert_eq!(describe_json["ok"], false);
+        assert_eq!(describe_json["error_kind"], "unsupported_action");
+    }
+
     #[test]
     fn rejects_invalid_mcp_arguments() {
         let show_error = parse_error_message("/mcp show alpha beta");

rust/crates/mock-anthropic-service/src/lib.rs

@@ -248,6 +248,7 @@ fn detect_scenario(request: &MessageRequest) -> Option<Scenario> {
                 .split_whitespace()
                 .find_map(|token| token.strip_prefix(SCENARIO_PREFIX))
                 .and_then(Scenario::parse),
+            InputContentBlock::Thinking { .. } => None,
             _ => None,
         })
     })

rust/crates/runtime/src/compact.rs

@@ -213,6 +213,7 @@ fn summarize_messages(messages: &[ConversationMessage]) -> String {
             ContentBlock::ToolUse { name, .. } => Some(name.as_str()),
             ContentBlock::ToolResult { tool_name, .. } => Some(tool_name.as_str()),
             ContentBlock::Text { .. } => None,
+            ContentBlock::Thinking { .. } => None,
         })
         .collect::<Vec<_>>();
     tool_names.sort_unstable();
@@ -317,6 +318,9 @@ fn merge_compact_summaries(existing_summary: Option<&str>, new_summary: &str) ->
 fn summarize_block(block: &ContentBlock) -> String {
     let raw = match block {
         ContentBlock::Text { text } => text.clone(),
+        ContentBlock::Thinking { thinking, .. } => {
+            format!("thinking ({} chars)", thinking.chars().count())
+        }
         ContentBlock::ToolUse { name, input, .. } => format!("tool_use {name}({input})"),
         ContentBlock::ToolResult {
             tool_name,
@@ -378,6 +382,7 @@ fn collect_key_files(messages: &[ConversationMessage]) -> Vec<String> {
             ContentBlock::Text { text } => text.as_str(),
             ContentBlock::ToolUse { input, .. } => input.as_str(),
             ContentBlock::ToolResult { output, .. } => output.as_str(),
+            ContentBlock::Thinking { thinking, .. } => thinking.as_str(),
         })
         .flat_map(extract_file_candidates)
         .collect::<Vec<_>>();
@@ -400,6 +405,7 @@ fn first_text_block(message: &ConversationMessage) -> Option<&str> {
         ContentBlock::Text { text } if !text.trim().is_empty() => Some(text.as_str()),
         ContentBlock::ToolUse { .. }
         | ContentBlock::ToolResult { .. }
+        | ContentBlock::Thinking { .. }
         | ContentBlock::Text { .. } => None,
     })
 }
@@ -450,6 +456,10 @@ fn estimate_message_tokens(message: &ConversationMessage) -> usize {
             ContentBlock::ToolResult {
                 tool_name, output, ..
             } => (tool_name.len() + output.len()) / 4 + 1,
+            ContentBlock::Thinking {
+                thinking,
+                signature,
+            } => thinking.len() / 4 + signature.as_ref().map_or(0, |value| value.len() / 4 + 1),
         })
         .sum()
 }

rust/crates/runtime/src/conversation.rs

@@ -28,6 +28,10 @@ pub struct ApiRequest {
 /// Streamed events emitted while processing a single assistant turn.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AssistantEvent {
+    Thinking {
+        thinking: String,
+        signature: Option<String>,
+    },
     TextDelta(String),
     ToolUse {
         id: String,
@@ -721,6 +725,16 @@ fn build_assistant_message(
 
     for event in events {
         match event {
+            AssistantEvent::Thinking {
+                thinking,
+                signature,
+            } => {
+                flush_text_block(&mut text, &mut blocks);
+                blocks.push(ContentBlock::Thinking {
+                    thinking,
+                    signature,
+                });
+            }
             AssistantEvent::TextDelta(delta) => text.push_str(&delta),
             AssistantEvent::ToolUse { id, name, input } => {
                 flush_text_block(&mut text, &mut blocks);
@@ -1723,6 +1737,47 @@ mod tests {
             .contains("assistant stream produced no content"));
     }
 
+    #[test]
+    fn build_assistant_message_places_thinking_block_before_text_and_tool_use() {
+        // given
+        let events = vec![
+            AssistantEvent::Thinking {
+                thinking: "pondering".to_string(),
+                signature: Some("sig".to_string()),
+            },
+            AssistantEvent::TextDelta("hello".to_string()),
+            AssistantEvent::ToolUse {
+                id: "tool-1".to_string(),
+                name: "echo".to_string(),
+                input: "payload".to_string(),
+            },
+            AssistantEvent::MessageStop,
+        ];
+
+        // when
+        let (message, _, _) = build_assistant_message(events)
+            .expect("assistant message should preserve thinking, text, and tool blocks");
+
+        // then
+        assert_eq!(
+            message.blocks,
+            vec![
+                ContentBlock::Thinking {
+                    thinking: "pondering".to_string(),
+                    signature: Some("sig".to_string()),
+                },
+                ContentBlock::Text {
+                    text: "hello".to_string(),
+                },
+                ContentBlock::ToolUse {
+                    id: "tool-1".to_string(),
+                    name: "echo".to_string(),
+                    input: "payload".to_string(),
+                },
+            ]
+        );
+    }
+
     #[test]
     fn static_tool_executor_rejects_unknown_tools() {
         // given

rust/crates/runtime/src/file_ops.rs

@@ -1,4 +1,5 @@
 use std::cmp::Reverse;
+use std::collections::HashSet;
 use std::fs;
 use std::io;
 use std::path::{Path, PathBuf};
@@ -7,7 +8,7 @@ use std::time::Instant;
 use glob::Pattern;
 use regex::RegexBuilder;
 use serde::{Deserialize, Serialize};
-use walkdir::WalkDir;
+use walkdir::{DirEntry, WalkDir};
 
 /// Maximum file size that can be read (10 MB).
 const MAX_READ_SIZE: u64 = 10 * 1024 * 1024;
@@ -15,6 +16,15 @@ const MAX_READ_SIZE: u64 = 10 * 1024 * 1024;
 /// Maximum file size that can be written (10 MB).
 const MAX_WRITE_SIZE: usize = 10 * 1024 * 1024;
 
+const GLOB_SEARCH_IGNORED_DIRS: &[&str] = &[
+    ".git",
+    "node_modules",
+    ".build",
+    "target",
+    "dist",
+    "coverage",
+];
+
 /// Check whether a file appears to contain binary content by examining
 /// the first chunk for NUL bytes.
 fn is_binary_file(path: &Path) -> io::Result<bool> {
@@ -313,14 +323,22 @@ pub fn glob_search(pattern: &str, path: Option<&str>) -> io::Result<GlobSearchOu
     // `Assets/**/*.{cs,uxml,uss}` work correctly.
     let expanded = expand_braces(&search_pattern);
 
-    let mut seen = std::collections::HashSet::new();
+    let mut seen = HashSet::new();
     let mut matches = Vec::new();
     for pat in &expanded {
-        let entries = glob::glob(pat)
+        let compiled = Pattern::new(pat)
             .map_err(|error| io::Error::new(io::ErrorKind::InvalidInput, error.to_string()))?;
+        let walk_root = derive_glob_walk_root(pat);
+        let entries = WalkDir::new(&walk_root)
+            .into_iter()
+            .filter_entry(|entry| !should_skip_glob_dir(entry));
         for entry in entries.flatten() {
-            if entry.is_file() && seen.insert(entry.clone()) {
-                matches.push(entry);
+            let candidate = entry.path();
+            if entry.file_type().is_file()
+                && compiled.matches_path(candidate)
+                && seen.insert(candidate.to_path_buf())
+            {
+                matches.push(candidate.to_path_buf());
             }
         }
     }
@@ -457,6 +475,39 @@ pub fn grep_search(input: &GrepSearchInput) -> io::Result<GrepSearchOutput> {
     })
 }
 
+fn should_skip_glob_dir(entry: &DirEntry) -> bool {
+    entry.file_type().is_dir()
+        && entry
+            .file_name()
+            .to_str()
+            .is_some_and(|name| GLOB_SEARCH_IGNORED_DIRS.contains(&name))
+}
+
+fn derive_glob_walk_root(pattern: &str) -> PathBuf {
+    let path = Path::new(pattern);
+    let mut prefix = PathBuf::new();
+    let mut saw_component = false;
+
+    for component in path.components() {
+        let text = component.as_os_str().to_string_lossy();
+        if component_contains_glob(&text) {
+            break;
+        }
+        prefix.push(component.as_os_str());
+        saw_component = true;
+    }
+
+    if saw_component {
+        prefix
+    } else {
+        std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."))
+    }
+}
+
+fn component_contains_glob(component: &str) -> bool {
+    component.contains('*') || component.contains('?') || component.contains('[')
+}
+
 fn collect_search_files(base_path: &Path) -> io::Result<Vec<PathBuf>> {
     if base_path.is_file() {
         return Ok(vec![base_path.to_path_buf()]);
@@ -651,11 +702,13 @@ fn expand_braces(pattern: &str) -> Vec<String> {
 
 #[cfg(test)]
 mod tests {
+    use std::path::PathBuf;
     use std::time::{SystemTime, UNIX_EPOCH};
 
     use super::{
-        edit_file, expand_braces, glob_search, grep_search, is_symlink_escape, read_file,
-        read_file_in_workspace, write_file, GrepSearchInput, MAX_WRITE_SIZE,
+        component_contains_glob, derive_glob_walk_root, edit_file, expand_braces, glob_search,
+        grep_search, is_symlink_escape, read_file, read_file_in_workspace, write_file,
+        GrepSearchInput, MAX_WRITE_SIZE,
     };
 
     fn temp_path(name: &str) -> std::path::PathBuf {
@@ -836,4 +889,50 @@ mod tests {
         );
         let _ = std::fs::remove_dir_all(&dir);
     }
+
+    #[test]
+    fn glob_search_skips_common_heavy_directories() {
+        let dir = temp_path("glob-ignored-dirs");
+        std::fs::create_dir_all(dir.join("src")).unwrap();
+        std::fs::create_dir_all(dir.join("docs")).unwrap();
+        std::fs::create_dir_all(dir.join("node_modules/pkg")).unwrap();
+        std::fs::create_dir_all(dir.join(".build/checkouts/pkg")).unwrap();
+        std::fs::create_dir_all(dir.join("target/debug/deps")).unwrap();
+
+        std::fs::write(dir.join("src/AGENTS.md"), "src").unwrap();
+        std::fs::write(dir.join("docs/AGENTS.md"), "docs").unwrap();
+        std::fs::write(dir.join("node_modules/pkg/AGENTS.md"), "node_modules").unwrap();
+        std::fs::write(dir.join(".build/checkouts/pkg/AGENTS.md"), ".build").unwrap();
+        std::fs::write(dir.join("target/debug/deps/AGENTS.md"), "target").unwrap();
+
+        let result =
+            glob_search("**/AGENTS.md", Some(dir.to_str().unwrap())).expect("glob should succeed");
+
+        assert_eq!(result.num_files, 2, "ignored dirs should be pruned");
+        assert!(result
+            .filenames
+            .iter()
+            .any(|path| path.ends_with("src/AGENTS.md")));
+        assert!(result
+            .filenames
+            .iter()
+            .any(|path| path.ends_with("docs/AGENTS.md")));
+        assert!(!result
+            .filenames
+            .iter()
+            .any(|path| path.contains("node_modules")
+                || path.contains(".build")
+                || path.contains("/target/")));
+
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[test]
+    fn derive_glob_walk_root_stops_at_first_glob_component() {
+        let root = derive_glob_walk_root("/tmp/demo/**/AGENTS.md");
+        assert_eq!(root, PathBuf::from("/tmp/demo"));
+        assert!(component_contains_glob("**"));
+        assert!(component_contains_glob("*.rs"));
+        assert!(!component_contains_glob("src"));
+    }
 }

rust/crates/runtime/src/lib.rs

@@ -131,8 +131,8 @@ pub use policy_engine::{
     PolicyEngine, PolicyRule, ReconcileReason, ReviewStatus,
 };
 pub use prompt::{
-    load_system_prompt, prepend_bullets, ContextFile, ProjectContext, PromptBuildError,
-    SystemPromptBuilder, FRONTIER_MODEL_NAME, SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
+    load_system_prompt, prepend_bullets, ContextFile, ModelFamilyIdentity, ProjectContext,
+    PromptBuildError, SystemPromptBuilder, FRONTIER_MODEL_NAME, SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
 };
 pub use recovery_recipes::{
     attempt_recovery, recipe_for, EscalationPolicy, FailureScenario, RecoveryContext,

rust/crates/runtime/src/prompt.rs

@@ -43,6 +43,24 @@ pub const FRONTIER_MODEL_NAME: &str = "Claude Opus 4.6";
 const MAX_INSTRUCTION_FILE_CHARS: usize = 4_000;
 const MAX_TOTAL_INSTRUCTION_CHARS: usize = 12_000;
 
+/// Neutral identity for the model family line in generated prompts.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub enum ModelFamilyIdentity {
+    #[default]
+    Claude,
+    Generic,
+}
+
+impl ModelFamilyIdentity {
+    #[must_use]
+    pub const fn family_label(self) -> &'static str {
+        match self {
+            Self::Claude => FRONTIER_MODEL_NAME,
+            Self::Generic => "an AI assistant",
+        }
+    }
+}
+
 /// Contents of an instruction file included in prompt construction.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ContextFile {
@@ -97,6 +115,7 @@ pub struct SystemPromptBuilder {
     output_style_prompt: Option<String>,
     os_name: Option<String>,
     os_version: Option<String>,
+    model_family: Option<ModelFamilyIdentity>,
     append_sections: Vec<String>,
     project_context: Option<ProjectContext>,
     config: Option<RuntimeConfig>,
@@ -122,6 +141,12 @@ impl SystemPromptBuilder {
         self
     }
 
+    #[must_use]
+    pub fn with_model_family(mut self, model_family: ModelFamilyIdentity) -> Self {
+        self.model_family = Some(model_family);
+        self
+    }
+
     #[must_use]
     pub fn with_project_context(mut self, project_context: ProjectContext) -> Self {
         self.project_context = Some(project_context);
@@ -179,9 +204,10 @@ impl SystemPromptBuilder {
             || "unknown".to_string(),
             |context| context.current_date.clone(),
         );
+        let identity = self.model_family.unwrap_or_default();
         let mut lines = vec!["# Environment context".to_string()];
         lines.extend(prepend_bullets(vec![
-            format!("Model family: {FRONTIER_MODEL_NAME}"),
+            format!("Model family: {}", identity.family_label()),
             format!("Working directory: {cwd}"),
             format!("Date: {date}"),
             format!(
@@ -434,12 +460,14 @@ pub fn load_system_prompt(
     current_date: impl Into<String>,
     os_name: impl Into<String>,
     os_version: impl Into<String>,
+    model_family: ModelFamilyIdentity,
 ) -> Result<Vec<String>, PromptBuildError> {
     let cwd = cwd.into();
     let project_context = ProjectContext::discover_with_git(&cwd, current_date.into())?;
     let config = ConfigLoader::default_for(&cwd).load()?;
     Ok(SystemPromptBuilder::new()
         .with_os(os_name, os_version)
+        .with_model_family(model_family)
         .with_project_context(project_context)
         .with_runtime_config(config)
         .build())
@@ -522,7 +550,8 @@ mod tests {
     use super::{
         collapse_blank_lines, display_context_path, normalize_instruction_content,
         render_instruction_content, render_instruction_files, truncate_instruction_content,
-        ContextFile, ProjectContext, SystemPromptBuilder, SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
+        ContextFile, ModelFamilyIdentity, ProjectContext, SystemPromptBuilder,
+        SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
     };
     use crate::config::ConfigLoader;
     use std::fs;
@@ -804,13 +833,19 @@ mod tests {
         std::env::set_var("HOME", &root);
         std::env::set_var("CLAW_CONFIG_HOME", root.join("missing-home"));
         std::env::set_current_dir(&root).expect("change cwd");
-        let prompt = super::load_system_prompt(&root, "2026-03-31", "linux", "6.8")
-            .expect("system prompt should load")
-            .join(
-                "
+        let prompt = super::load_system_prompt(
+            &root,
+            "2026-03-31",
+            "linux",
+            "6.8",
+            ModelFamilyIdentity::Claude,
+        )
+        .expect("system prompt should load")
+        .join(
+            "
 
 ",
-            );
+        );
         std::env::set_current_dir(previous).expect("restore cwd");
         if let Some(value) = original_home {
             std::env::set_var("HOME", value);
@@ -828,6 +863,50 @@ mod tests {
         fs::remove_dir_all(root).expect("cleanup temp dir");
     }
 
+    #[test]
+    fn renders_default_claude_model_family_identity() {
+        // given: a prompt builder without an explicit model family override
+        let project_context = ProjectContext {
+            cwd: PathBuf::from("/tmp/project"),
+            current_date: "2026-03-31".to_string(),
+            ..ProjectContext::default()
+        };
+
+        // when: rendering the system prompt environment section
+        let prompt = SystemPromptBuilder::new()
+            .with_os("linux", "6.8")
+            .with_project_context(project_context)
+            .render();
+
+        // then: the Claude model family label is preserved by default
+        assert!(prompt.contains("Model family: Claude Opus 4.6"));
+    }
+
+    #[test]
+    fn renders_generic_model_family_identity_without_claude_label() {
+        // given: a prompt builder with generic model family identity
+        let project_context = ProjectContext {
+            cwd: PathBuf::from("/tmp/project"),
+            current_date: "2026-03-31".to_string(),
+            ..ProjectContext::default()
+        };
+
+        // when: rendering the system prompt environment section
+        let prompt = SystemPromptBuilder::new()
+            .with_os("linux", "6.8")
+            .with_model_family(ModelFamilyIdentity::Generic)
+            .with_project_context(project_context)
+            .render();
+        let model_family_line = prompt
+            .lines()
+            .find(|line| line.contains("Model family:"))
+            .expect("model family line should render");
+
+        // then: the model family line is neutral and excludes Claude Opus 4.6
+        assert_eq!(model_family_line, " - Model family: an AI assistant");
+        assert!(!model_family_line.contains("Claude Opus 4.6"));
+    }
+
     #[test]
     fn renders_claude_code_style_sections_with_project_context() {
         let root = temp_dir();

rust/crates/runtime/src/session.rs

@@ -30,6 +30,10 @@ pub enum ContentBlock {
     Text {
         text: String,
     },
+    Thinking {
+        thinking: String,
+        signature: Option<String>,
+    },
     ToolUse {
         id: String,
         name: String,
@@ -737,6 +741,22 @@ impl ContentBlock {
                 object.insert("type".to_string(), JsonValue::String("text".to_string()));
                 object.insert("text".to_string(), JsonValue::String(text.clone()));
             }
+            Self::Thinking {
+                thinking,
+                signature,
+            } => {
+                object.insert(
+                    "type".to_string(),
+                    JsonValue::String("thinking".to_string()),
+                );
+                object.insert("thinking".to_string(), JsonValue::String(thinking.clone()));
+                if let Some(signature) = signature {
+                    object.insert(
+                        "signature".to_string(),
+                        JsonValue::String(signature.clone()),
+                    );
+                }
+            }
             Self::ToolUse { id, name, input } => {
                 object.insert(
                     "type".to_string(),
@@ -783,6 +803,13 @@ impl ContentBlock {
             "text" => Ok(Self::Text {
                 text: required_string(object, "text")?,
             }),
+            "thinking" => Ok(Self::Thinking {
+                thinking: required_string(object, "thinking")?,
+                signature: object
+                    .get("signature")
+                    .and_then(JsonValue::as_str)
+                    .map(String::from),
+            }),
             "tool_use" => Ok(Self::ToolUse {
                 id: required_string(object, "id")?,
                 name: required_string(object, "name")?,
@@ -1208,6 +1235,36 @@ mod tests {
         assert_eq!(restored.session_id, session.session_id);
     }
 
+    #[test]
+    fn persists_assistant_thinking_block_round_trip_through_jsonl() {
+        // given
+        let mut session = Session::new();
+        session
+            .push_message(ConversationMessage::assistant(vec![
+                ContentBlock::Thinking {
+                    thinking: "trace the path through session persistence".to_string(),
+                    signature: Some("sig-123".to_string()),
+                },
+            ]))
+            .expect("thinking block should append");
+        let path = temp_session_path("thinking-jsonl");
+
+        // when
+        session.save_to_path(&path).expect("session should save");
+        let restored = Session::load_from_path(&path).expect("session should load");
+        fs::remove_file(&path).expect("temp file should be removable");
+
+        // then
+        assert_eq!(restored, session);
+        assert_eq!(
+            restored.messages[0].blocks[0],
+            ContentBlock::Thinking {
+                thinking: "trace the path through session persistence".to_string(),
+                signature: Some("sig-123".to_string()),
+            }
+        );
+    }
+
     #[test]
     fn loads_legacy_session_json_object() {
         let path = temp_session_path("legacy");

rust/crates/rusty-claude-cli/src/main.rs

@@ -24,10 +24,11 @@ use std::thread::{self, JoinHandle};
 use std::time::{Duration, Instant, UNIX_EPOCH};
 
 use api::{
-    detect_provider_kind, resolve_startup_auth_source, AnthropicClient, AuthSource,
-    ContentBlockDelta, InputContentBlock, InputMessage, MessageRequest, MessageResponse,
-    OutputContentBlock, PromptCache, ProviderClient as ApiProviderClient, ProviderKind,
-    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
+    detect_provider_kind, model_family_identity_for, resolve_startup_auth_source, AnthropicClient,
+    AuthSource, ContentBlockDelta, InputContentBlock, InputMessage, MessageRequest,
+    MessageResponse, OutputContentBlock, PromptCache, ProviderClient as ApiProviderClient,
+    ProviderKind, StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition,
+    ToolResultContentBlock,
 };
 
 use commands::{
@@ -148,11 +149,7 @@ impl ModelProvenance {
 }
 
 fn max_tokens_for_model(model: &str) -> u32 {
-    if model.contains("opus") {
-        32_000
-    } else {
-        64_000
-    }
+    api::max_tokens_for_model(model)
 }
 // Build-time constants injected by build.rs (fall back to static values when
 // build.rs hasn't run, e.g. in doc-test or unusual toolchain environments).
@@ -361,8 +358,9 @@ fn run() -> Result<(), Box<dyn std::error::Error>> {
         CliAction::PrintSystemPrompt {
             cwd,
             date,
+            model,
             output_format,
-        } => print_system_prompt(cwd, date, output_format)?,
+        } => print_system_prompt(cwd, date, &model, output_format)?,
         CliAction::Version { output_format } => print_version(output_format)?,
         CliAction::ResumeSession {
             session_path,
@@ -464,7 +462,10 @@ fn run() -> Result<(), Box<dyn std::error::Error>> {
             reasoning_effort,
             allow_broad_cwd,
         )?,
-        CliAction::HelpTopic(topic) => print_help_topic(topic),
+        CliAction::HelpTopic {
+            topic,
+            output_format,
+        } => print_help_topic(topic, output_format)?,
         CliAction::Help { output_format } => print_help(output_format)?,
     }
     Ok(())
@@ -499,6 +500,7 @@ enum CliAction {
     PrintSystemPrompt {
         cwd: PathBuf,
         date: String,
+        model: String,
         output_format: CliOutputFormat,
     },
     Version {
@@ -567,7 +569,10 @@ enum CliAction {
         reasoning_effort: Option<String>,
         allow_broad_cwd: bool,
     },
-    HelpTopic(LocalHelpTopic),
+    HelpTopic {
+        topic: LocalHelpTopic,
+        output_format: CliOutputFormat,
+    },
     // prompt-mode formatting is only supported for non-interactive runs
     Help {
         output_format: CliOutputFormat,
@@ -843,7 +848,7 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
     if rest.first().map(String::as_str) == Some("--resume") {
         return parse_resume_args(&rest[1..], output_format);
     }
-    if let Some(action) = parse_local_help_action(&rest) {
+    if let Some(action) = parse_local_help_action(&rest, output_format) {
         return action;
     }
     if let Some(action) = parse_single_word_command_alias(
@@ -877,13 +882,17 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
         // `missing Anthropic credentials` even though the command is purely
         // local introspection. Mirror `agents`/`mcp`/`skills`: action is the
         // first positional arg, target is the second.
-        "plugins" => {
+        // `plugin` (singular) and `marketplace` are aliases for `plugins`.
+        // All three must route to the same local handler so that no form
+        // falls through to the LLM/prompt path.
+        "plugins" | "plugin" | "marketplace" => {
             let tail = &rest[1..];
             let action = tail.first().cloned();
             let target = tail.get(1).cloned();
             if tail.len() > 2 {
                 return Err(format!(
-                    "unexpected extra arguments after `claw plugins {}`: {}",
+                    "unexpected extra arguments after `claw {} {}`: {}",
+                    rest[0],
                     tail[..2].join(" "),
                     tail[2..].join(" ")
                 ));
@@ -926,6 +935,14 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
             }
             Ok(CliAction::Diff { output_format })
         }
+        // `claw permissions <mode>` falls through to the LLM when called
+        // with a subcommand argument because parse_single_word_command_alias
+        // only intercepts the bare single-word form. Catch all multi-word
+        // forms here and return a structured guidance error so no network
+        // call or session is created.
+        "permissions" => Err(format!(
+            "`claw permissions` is a slash command. Start `claw` and run `/permissions` inside the REPL.\n  Usage  /permissions [read-only|workspace-write|danger-full-access]"
+        )),
         "skills" => {
             let args = join_optional_args(&rest[1..]);
             match classify_skills_slash_command(args.as_deref()) {
@@ -946,7 +963,7 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
                 }),
             }
         }
-        "system-prompt" => parse_system_prompt_args(&rest[1..], output_format),
+        "system-prompt" => parse_system_prompt_args(&rest[1..], model, output_format),
         "acp" => parse_acp_args(&rest[1..], output_format),
         "login" | "logout" => Err(removed_auth_surface_error(rest[0].as_str())),
         "init" => Ok(CliAction::Init { output_format }),
@@ -1021,7 +1038,10 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
     }
 }
 
-fn parse_local_help_action(rest: &[String]) -> Option<Result<CliAction, String>> {
+fn parse_local_help_action(
+    rest: &[String],
+    output_format: CliOutputFormat,
+) -> Option<Result<CliAction, String>> {
     if rest.len() != 2 || !is_help_flag(&rest[1]) {
         return None;
     }
@@ -1044,7 +1064,10 @@ fn parse_local_help_action(rest: &[String]) -> Option<Result<CliAction, String>>
         "bootstrap-plan" => LocalHelpTopic::BootstrapPlan,
         _ => return None,
     };
-    Some(Ok(CliAction::HelpTopic(topic)))
+    Some(Ok(CliAction::HelpTopic {
+        topic,
+        output_format,
+    }))
 }
 
 fn is_help_flag(value: &str) -> bool {
@@ -1618,6 +1641,7 @@ fn filter_tool_specs(
 
 fn parse_system_prompt_args(
     args: &[String],
+    model: String,
     output_format: CliOutputFormat,
 ) -> Result<CliAction, String> {
     let mut cwd = env::current_dir().map_err(|error| error.to_string())?;
@@ -1654,6 +1678,7 @@ fn parse_system_prompt_args(
     Ok(CliAction::PrintSystemPrompt {
         cwd,
         date,
+        model,
         output_format,
     })
 }
@@ -2594,9 +2619,16 @@ fn print_bootstrap_plan(output_format: CliOutputFormat) -> Result<(), Box<dyn st
 fn print_system_prompt(
     cwd: PathBuf,
     date: String,
+    model: &str,
     output_format: CliOutputFormat,
 ) -> Result<(), Box<dyn std::error::Error>> {
-    let sections = load_system_prompt(cwd, date, env::consts::OS, "unknown")?;
+    let sections = load_system_prompt(
+        cwd,
+        date,
+        env::consts::OS,
+        "unknown",
+        model_family_identity_for(model),
+    )?;
     let message = sections.join(
         "
 
@@ -2627,12 +2659,15 @@ fn print_version(output_format: CliOutputFormat) -> Result<(), Box<dyn std::erro
 }
 
 fn version_json_value() -> serde_json::Value {
+    let executable_path = env::current_exe().ok().map(|p| p.display().to_string());
     json!({
         "kind": "version",
         "message": render_version_report(),
         "version": VERSION,
         "git_sha": GIT_SHA,
         "target": BUILD_TARGET,
+        "build_date": DEFAULT_DATE,
+        "executable_path": executable_path,
     })
 }
 
@@ -3523,10 +3558,10 @@ fn run_resume_command(
             Ok(ResumeCommandOutcome {
                 session: session.clone(),
                 message: Some(handle_agents_slash_command(args.as_deref(), &cwd)?),
-                json: Some(serde_json::json!({
-                    "kind": "agents",
-                    "text": handle_agents_slash_command(args.as_deref(), &cwd)?,
-                })),
+                json: Some(
+                    serde_json::to_value(handle_agents_slash_command_json(args.as_deref(), &cwd)?)
+                        .unwrap_or_else(|_| serde_json::json!(null)),
+                ),
             })
         }
         SlashCommand::Skills { args } => {
@@ -3542,6 +3577,37 @@ fn run_resume_command(
                 json: Some(handle_skills_slash_command_json(args.as_deref(), &cwd)?),
             })
         }
+        SlashCommand::Plugins { action, target } => {
+            // Only list is supported in resume mode (no runtime to reload)
+            match action.as_deref() {
+                Some("install") | Some("uninstall") | Some("enable") | Some("disable")
+                | Some("update") => {
+                    return Err(
+                        "resumed /plugins mutations are interactive-only; start `claw` and run `/plugins` in the REPL".into(),
+                    );
+                }
+                _ => {}
+            }
+            let cwd = env::current_dir()?;
+            let loader = ConfigLoader::default_for(&cwd);
+            let runtime_config = loader.load()?;
+            let mut manager = build_plugin_manager(&cwd, &loader, &runtime_config);
+            let result =
+                handle_plugins_slash_command(action.as_deref(), target.as_deref(), &mut manager)?;
+            let action_str = action.as_deref().unwrap_or("list");
+            let json = serde_json::json!({
+                "kind": "plugin",
+                "action": action_str,
+                "target": target,
+                "message": &result.message,
+                "reload_runtime": result.reload_runtime,
+            });
+            Ok(ResumeCommandOutcome {
+                session: session.clone(),
+                message: Some(result.message),
+                json: Some(json),
+            })
+        }
         SlashCommand::Doctor => {
             let report = render_doctor_report()?;
             Ok(ResumeCommandOutcome {
@@ -3628,7 +3694,6 @@ fn run_resume_command(
         | SlashCommand::Model { .. }
         | SlashCommand::Permissions { .. }
         | SlashCommand::Session { .. }
-        | SlashCommand::Plugins { .. }
         | SlashCommand::Login
         | SlashCommand::Logout
         | SlashCommand::Vim
@@ -4341,7 +4406,7 @@ impl LiveCli {
         allowed_tools: Option<AllowedToolSet>,
         permission_mode: PermissionMode,
     ) -> Result<Self, Box<dyn std::error::Error>> {
-        let system_prompt = build_system_prompt()?;
+        let system_prompt = build_system_prompt(&model)?;
         let session_state = new_cli_session()?;
         let session = create_managed_session_handle(&session_state.session_id)?;
         let runtime = build_runtime(
@@ -4477,6 +4542,10 @@ impl LiveCli {
                     TerminalRenderer::new().color_theme(),
                     &mut stdout,
                 )?;
+                let final_text = final_assistant_text(&summary);
+                if !final_text.is_empty() {
+                    println!("{final_text}");
+                }
                 println!();
                 if let Some(event) = summary.auto_compaction {
                     println!(
@@ -5062,10 +5131,17 @@ impl LiveCli {
         let cwd = env::current_dir()?;
         match output_format {
             CliOutputFormat::Text => println!("{}", handle_mcp_slash_command(args, &cwd)?),
-            CliOutputFormat::Json => println!(
-                "{}",
-                serde_json::to_string_pretty(&handle_mcp_slash_command_json(args, &cwd)?)?
-            ),
+            CliOutputFormat::Json => {
+                let value = handle_mcp_slash_command_json(args, &cwd)?;
+                // Propagate ok:false → non-zero exit so automation callers
+                // can rely on exit code instead of inspecting the envelope.
+                // (#68: mcp error envelopes previously always exited 0.)
+                let is_error = value.get("ok").and_then(|v| v.as_bool()) == Some(false);
+                println!("{}", serde_json::to_string_pretty(&value)?);
+                if is_error {
+                    std::process::exit(1);
+                }
+            }
         }
         Ok(())
     }
@@ -6090,8 +6166,90 @@ fn render_help_topic(topic: LocalHelpTopic) -> String {
     }
 }
 
-fn print_help_topic(topic: LocalHelpTopic) {
-    println!("{}", render_help_topic(topic));
+fn local_help_topic_command(topic: LocalHelpTopic) -> &'static str {
+    match topic {
+        LocalHelpTopic::Status => "status",
+        LocalHelpTopic::Sandbox => "sandbox",
+        LocalHelpTopic::Doctor => "doctor",
+        LocalHelpTopic::Acp => "acp",
+        LocalHelpTopic::Init => "init",
+        LocalHelpTopic::State => "state",
+        LocalHelpTopic::Export => "export",
+        LocalHelpTopic::Version => "version",
+        LocalHelpTopic::SystemPrompt => "system-prompt",
+        LocalHelpTopic::DumpManifests => "dump-manifests",
+        LocalHelpTopic::BootstrapPlan => "bootstrap-plan",
+    }
+}
+
+fn render_export_help_json() -> serde_json::Value {
+    json!({
+        "kind": "help",
+        "topic": "export",
+        "command": "export",
+        "usage": "claw export [--session <id|latest>] [--output <path>] [--output-format <format>]",
+        "purpose": "serialize a managed session to JSON for review, transfer, or archival",
+        "defaults": {
+            "session": LATEST_SESSION_REFERENCE,
+            "session_source": ".claw/sessions/",
+            "output": "derived from the selected session when omitted"
+        },
+        "formats": ["text", "json"],
+        "options": [
+            {
+                "name": "--session",
+                "value": "<id|latest>",
+                "default": LATEST_SESSION_REFERENCE,
+                "description": "managed session to export"
+            },
+            {
+                "name": "--output",
+                "aliases": ["-o"],
+                "value": "<path>",
+                "description": "write the exported transcript to this path"
+            },
+            {
+                "name": "--output-format",
+                "value": "<format>",
+                "values": ["text", "json"],
+                "default": "text",
+                "description": "format for the command result envelope"
+            },
+            {
+                "name": "--help",
+                "aliases": ["-h"],
+                "description": "show help for the export command"
+            }
+        ],
+        "related": ["/session list", "claw --resume latest"]
+    })
+}
+
+fn render_help_topic_json(topic: LocalHelpTopic) -> serde_json::Value {
+    if topic == LocalHelpTopic::Export {
+        return render_export_help_json();
+    }
+
+    json!({
+        "kind": "help",
+        "topic": local_help_topic_command(topic),
+        "command": local_help_topic_command(topic),
+        "message": render_help_topic(topic),
+    })
+}
+
+fn print_help_topic(
+    topic: LocalHelpTopic,
+    output_format: CliOutputFormat,
+) -> Result<(), Box<dyn std::error::Error>> {
+    match output_format {
+        CliOutputFormat::Text => println!("{}", render_help_topic(topic)),
+        CliOutputFormat::Json => println!(
+            "{}",
+            serde_json::to_string_pretty(&render_help_topic_json(topic))?
+        ),
+    }
+    Ok(())
 }
 
 fn print_acp_status(output_format: CliOutputFormat) -> Result<(), Box<dyn std::error::Error>> {
@@ -6207,7 +6365,7 @@ fn render_config_report(section: Option<&str>) -> Result<String, Box<dyn std::er
 }
 
 fn render_config_json(
-    _section: Option<&str>,
+    section: Option<&str>,
 ) -> Result<serde_json::Value, Box<dyn std::error::Error>> {
     let cwd = env::current_dir()?;
     let loader = ConfigLoader::default_for(&cwd);
@@ -6240,13 +6398,52 @@ fn render_config_json(
         })
         .collect();
 
-    Ok(serde_json::json!({
+    let base = serde_json::json!({
         "kind": "config",
         "cwd": cwd.display().to_string(),
         "loaded_files": loaded_paths.len(),
         "merged_keys": runtime_config.merged().len(),
         "files": files,
-    }))
+    });
+
+    if let Some(section) = section {
+        let section_rendered: Option<String> = match section {
+            "env" => runtime_config.get("env").map(|v| v.render()),
+            "hooks" => runtime_config.get("hooks").map(|v| v.render()),
+            "model" => runtime_config.get("model").map(|v| v.render()),
+            "plugins" => runtime_config
+                .get("plugins")
+                .or_else(|| runtime_config.get("enabledPlugins"))
+                .map(|v| v.render()),
+            other => {
+                return Ok(serde_json::json!({
+                    "kind": "config",
+                    "section": other,
+                    "ok": false,
+                    "error": format!("Unsupported config section '{other}'. Use env, hooks, model, or plugins."),
+                    "cwd": cwd.display().to_string(),
+                    "loaded_files": loaded_paths.len(),
+                    "files": files,
+                }));
+            }
+        };
+        // Parse the rendered JSON string back into serde_json::Value so that
+        // section_value is a real JSON object/array in the envelope, not a quoted string.
+        let section_value: serde_json::Value = section_rendered
+            .as_deref()
+            .and_then(|s| serde_json::from_str(s).ok())
+            .unwrap_or(serde_json::Value::Null);
+        let mut obj = base;
+        let map = obj.as_object_mut().expect("base is object");
+        map.insert(
+            "section".to_string(),
+            serde_json::Value::String(section.to_string()),
+        );
+        map.insert("section_value".to_string(), section_value);
+        return Ok(obj);
+    }
+
+    Ok(base)
 }
 
 fn render_memory_report() -> Result<String, Box<dyn std::error::Error>> {
@@ -6824,6 +7021,7 @@ fn render_export_text(session: &Session) -> String {
         for block in &message.blocks {
             match block {
                 ContentBlock::Text { text } => lines.push(text.clone()),
+                ContentBlock::Thinking { .. } => {}
                 ContentBlock::ToolUse { id, name, input } => {
                     lines.push(format!("[tool_use id={id} name={name}] {input}"));
                 }
@@ -7010,6 +7208,7 @@ fn render_session_markdown(session: &Session, session_id: &str, session_path: &P
                         lines.push(String::new());
                     }
                 }
+                ContentBlock::Thinking { .. } => {}
                 ContentBlock::ToolUse { id, name, input } => {
                     lines.push(format!(
                         "**Tool call** `{name}` _(id `{}`)_",
@@ -7063,12 +7262,13 @@ fn short_tool_id(id: &str) -> String {
     format!("{prefix}…")
 }
 
-fn build_system_prompt() -> Result<Vec<String>, Box<dyn std::error::Error>> {
+fn build_system_prompt(model: &str) -> Result<Vec<String>, Box<dyn std::error::Error>> {
     Ok(load_system_prompt(
         env::current_dir()?,
         DEFAULT_DATE,
         env::consts::OS,
         "unknown",
+        model_family_identity_for(model),
     )?)
 }
 
@@ -9030,26 +9230,29 @@ fn convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
             let content = message
                 .blocks
                 .iter()
-                .map(|block| match block {
-                    ContentBlock::Text { text } => InputContentBlock::Text { text: text.clone() },
-                    ContentBlock::ToolUse { id, name, input } => InputContentBlock::ToolUse {
+                .filter_map(|block| match block {
+                    ContentBlock::Text { text } => {
+                        Some(InputContentBlock::Text { text: text.clone() })
+                    }
+                    ContentBlock::Thinking { .. } => None,
+                    ContentBlock::ToolUse { id, name, input } => Some(InputContentBlock::ToolUse {
                         id: id.clone(),
                         name: name.clone(),
                         input: serde_json::from_str(input)
                             .unwrap_or_else(|_| serde_json::json!({ "raw": input })),
-                    },
+                    }),
                     ContentBlock::ToolResult {
                         tool_use_id,
                         output,
                         is_error,
                         ..
-                    } => InputContentBlock::ToolResult {
+                    } => Some(InputContentBlock::ToolResult {
                         tool_use_id: tool_use_id.clone(),
                         content: vec![ToolResultContentBlock::Text {
                             text: output.clone(),
                         }],
                         is_error: *is_error,
-                    },
+                    }),
                 })
                 .collect::<Vec<_>>();
             (!content.is_empty()).then(|| InputMessage {
@@ -9249,17 +9452,17 @@ mod tests {
         parse_git_status_branch, parse_git_status_metadata_for, parse_git_workspace_summary,
         parse_history_count, permission_policy, print_help_to, push_output_block,
         render_config_report, render_diff_report, render_diff_report_for, render_help_topic,
-        render_memory_report, render_prompt_history_report, render_repl_help, render_resume_usage,
-        render_session_list, render_session_markdown, resolve_model_alias,
-        resolve_model_alias_with_config, resolve_repl_model, resolve_session_reference,
-        response_to_events, resume_supported_slash_commands, run_resume_command, short_tool_id,
-        slash_command_completion_candidates_with_sessions, split_error_hint, status_context,
-        status_json_value, summarize_tool_payload_for_markdown, try_resolve_bare_skill_prompt,
-        validate_no_args, write_mcp_server_fixture, CliAction, CliOutputFormat, CliToolExecutor,
-        GitWorkspaceSummary, InternalPromptProgressEvent, InternalPromptProgressState, LiveCli,
-        LocalHelpTopic, PromptHistoryEntry, SessionLifecycleKind, SessionLifecycleSummary,
-        SlashCommand, StatusUsage, TmuxPaneSnapshot, DEFAULT_MODEL, LATEST_SESSION_REFERENCE,
-        STUB_COMMANDS,
+        render_help_topic_json, render_memory_report, render_prompt_history_report,
+        render_repl_help, render_resume_usage, render_session_list, render_session_markdown,
+        resolve_model_alias, resolve_model_alias_with_config, resolve_repl_model,
+        resolve_session_reference, response_to_events, resume_supported_slash_commands,
+        run_resume_command, short_tool_id, slash_command_completion_candidates_with_sessions,
+        split_error_hint, status_context, status_json_value, summarize_tool_payload_for_markdown,
+        try_resolve_bare_skill_prompt, validate_no_args, write_mcp_server_fixture, CliAction,
+        CliOutputFormat, CliToolExecutor, GitWorkspaceSummary, InternalPromptProgressEvent,
+        InternalPromptProgressState, LiveCli, LocalHelpTopic, PromptHistoryEntry,
+        SessionLifecycleKind, SessionLifecycleSummary, SlashCommand, StatusUsage, TmuxPaneSnapshot,
+        DEFAULT_MODEL, LATEST_SESSION_REFERENCE, STUB_COMMANDS,
     };
     use api::{ApiError, MessageResponse, OutputContentBlock, Usage};
     use plugins::{
@@ -9424,6 +9627,41 @@ mod tests {
         );
     }
 
+    #[test]
+    fn openai_configured_limit_errors_are_rendered_as_context_window_guidance() {
+        let error = ApiError::Api {
+            status: "400".parse().expect("status"),
+            error_type: Some("invalid_request_error".to_string()),
+            message: Some(
+                "Input tokens exceed the configured limit of 922000 tokens. Your messages resulted in 1860900 tokens. Please reduce the length of the messages."
+                    .to_string(),
+            ),
+            request_id: Some("req_ctx_openai_456".to_string()),
+            body: String::new(),
+            retryable: false,
+            suggested_action: None,
+        };
+
+        let rendered = format_user_visible_api_error("session-issue-32", &error);
+        assert!(rendered.contains("Context window blocked"), "{rendered}");
+        assert!(rendered.contains("context_window_blocked"), "{rendered}");
+        assert!(
+            rendered.contains("Trace            req_ctx_openai_456"),
+            "{rendered}"
+        );
+        assert!(
+            rendered.contains(
+                "Detail           Input tokens exceed the configured limit of 922000 tokens."
+            ),
+            "{rendered}"
+        );
+        assert!(rendered.contains("Compact          /compact"), "{rendered}");
+        assert!(
+            rendered.contains("Fresh session    /clear --confirm"),
+            "{rendered}"
+        );
+    }
+
     #[test]
     fn retry_wrapped_context_window_errors_keep_recovery_guidance() {
         let error = ApiError::RetriesExhausted {
@@ -10050,6 +10288,7 @@ mod tests {
 
     #[test]
     fn parses_system_prompt_options() {
+        // given: system-prompt options for cwd and date
         let args = vec![
             "system-prompt".to_string(),
             "--cwd".to_string(),
@@ -10057,16 +10296,43 @@ mod tests {
             "--date".to_string(),
             "2026-04-01".to_string(),
         ];
+
+        // when: parsing the direct system-prompt command
+        let action = parse_args(&args).expect("args should parse");
+
+        // then: the action carries prompt options and default model
         assert_eq!(
-            parse_args(&args).expect("args should parse"),
+            action,
             CliAction::PrintSystemPrompt {
                 cwd: PathBuf::from("/tmp/project"),
                 date: "2026-04-01".to_string(),
+                model: DEFAULT_MODEL.to_string(),
                 output_format: CliOutputFormat::Text,
             }
         );
     }
 
+    #[test]
+    fn parses_global_model_for_system_prompt() {
+        // given: a global OpenAI-compatible model before system-prompt
+        let args = vec![
+            "--model".to_string(),
+            "openai/gpt-4.1-mini".to_string(),
+            "system-prompt".to_string(),
+        ];
+
+        // when: parsing the CLI arguments
+        let action = parse_args(&args).expect("args should parse");
+
+        // then: the system-prompt action carries the selected model
+        match action {
+            CliAction::PrintSystemPrompt { model, .. } => {
+                assert_eq!(model, "openai/gpt-4.1-mini");
+            }
+            other => panic!("expected PrintSystemPrompt, got {other:?}"),
+        }
+    }
+
     #[test]
     fn removed_login_and_logout_subcommands_error_helpfully() {
         let login = parse_args(&["login".to_string()]).expect_err("login should be removed");
@@ -10379,21 +10645,33 @@ mod tests {
         assert_eq!(
             parse_args(&["status".to_string(), "--help".to_string()])
                 .expect("status help should parse"),
-            CliAction::HelpTopic(LocalHelpTopic::Status)
+            CliAction::HelpTopic {
+                topic: LocalHelpTopic::Status,
+                output_format: CliOutputFormat::Text,
+            }
         );
         assert_eq!(
             parse_args(&["sandbox".to_string(), "-h".to_string()])
                 .expect("sandbox help should parse"),
-            CliAction::HelpTopic(LocalHelpTopic::Sandbox)
+            CliAction::HelpTopic {
+                topic: LocalHelpTopic::Sandbox,
+                output_format: CliOutputFormat::Text,
+            }
         );
         assert_eq!(
             parse_args(&["doctor".to_string(), "--help".to_string()])
                 .expect("doctor help should parse"),
-            CliAction::HelpTopic(LocalHelpTopic::Doctor)
+            CliAction::HelpTopic {
+                topic: LocalHelpTopic::Doctor,
+                output_format: CliOutputFormat::Text,
+            }
         );
         assert_eq!(
             parse_args(&["acp".to_string(), "--help".to_string()]).expect("acp help should parse"),
-            CliAction::HelpTopic(LocalHelpTopic::Acp)
+            CliAction::HelpTopic {
+                topic: LocalHelpTopic::Acp,
+                output_format: CliOutputFormat::Text,
+            }
         );
     }
 
@@ -10423,10 +10701,30 @@ mod tests {
                     });
                 assert_eq!(
                     parsed,
-                    CliAction::HelpTopic(*expected_topic),
+                    CliAction::HelpTopic {
+                        topic: *expected_topic,
+                        output_format: CliOutputFormat::Text,
+                    },
                     "`{subcommand} {flag}` should resolve to HelpTopic({expected_topic:?})"
                 );
             }
+            let json_parsed = parse_args(&[
+                subcommand.to_string(),
+                "--help".to_string(),
+                "--output-format".to_string(),
+                "json".to_string(),
+            ])
+            .unwrap_or_else(|error| {
+                panic!("`{subcommand} --help --output-format json` should parse: {error}")
+            });
+            assert_eq!(
+                json_parsed,
+                CliAction::HelpTopic {
+                    topic: *expected_topic,
+                    output_format: CliOutputFormat::Json,
+                },
+                "`{subcommand} --help --output-format json` should preserve json output format"
+            );
             // And the rendered help must actually mention the subcommand name
             // (or its canonical title) so users know they got the right help.
             let rendered = render_help_topic(*expected_topic);
@@ -10441,6 +10739,24 @@ mod tests {
         }
     }
 
+    #[test]
+    fn export_help_json_is_bounded_and_parseable_384() {
+        let value = render_help_topic_json(LocalHelpTopic::Export);
+        assert_eq!(value["kind"], "help");
+        assert_eq!(value["topic"], "export");
+        assert_eq!(value["command"], "export");
+        assert_eq!(
+            value["usage"],
+            "claw export [--session <id|latest>] [--output <path>] [--output-format <format>]"
+        );
+        assert_eq!(value["defaults"]["session"], LATEST_SESSION_REFERENCE);
+        assert!(value["options"].as_array().expect("options array").len() >= 4);
+        assert!(
+            value.get("message").is_none(),
+            "export help json should be a bounded envelope, not plaintext help wrapped in json"
+        );
+    }
+
     #[test]
     fn status_degrades_gracefully_on_malformed_mcp_config_143() {
         // #143: previously `claw status` hard-failed on any config parse error,

rust/crates/rusty-claude-cli/tests/compact_output.rs

@@ -126,6 +126,66 @@ fn compact_flag_streaming_text_only_emits_final_message_text() {
     fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
 }
 
+#[test]
+fn text_prompt_mode_prints_final_assistant_text_after_spinner() {
+    // given a workspace pointed at the mock Anthropic service running the
+    // streaming_text scenario which only emits a single assistant text block
+    let runtime = tokio::runtime::Runtime::new().expect("tokio runtime should build");
+    let server = runtime
+        .block_on(MockAnthropicService::spawn())
+        .expect("mock service should start");
+    let base_url = server.base_url();
+
+    let workspace = unique_temp_dir("text-prompt-mode");
+    let config_home = workspace.join("config-home");
+    let home = workspace.join("home");
+    fs::create_dir_all(&workspace).expect("workspace should exist");
+    fs::create_dir_all(&config_home).expect("config home should exist");
+    fs::create_dir_all(&home).expect("home should exist");
+
+    // when we invoke claw in normal text prompt mode for the streaming text scenario
+    let prompt = format!("{SCENARIO_PREFIX}streaming_text");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            &prompt,
+        ],
+    );
+
+    // then stdout should contain the final assistant text, not just spinner output
+    assert!(
+        output.status.success(),
+        "text prompt run should succeed\nstdout:\n{}\n\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr),
+    );
+    let stdout = String::from_utf8(output.stdout).expect("stdout should be utf8");
+    let plain_stdout = strip_ansi_codes(&stdout);
+    assert!(
+        plain_stdout.contains("Mock streaming says hello from the parity harness."),
+        "text prompt stdout should include the assistant text ({stdout:?})"
+    );
+    assert!(
+        plain_stdout.contains("✔ ✨ Done"),
+        "text prompt stdout should still include spinner completion ({stdout:?})"
+    );
+    assert!(
+        plain_stdout
+            .lines()
+            .any(|line| line == "Mock streaming says hello from the parity harness."),
+        "text prompt stdout should print the assistant text as its own line ({stdout:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
 #[test]
 fn compact_flag_with_json_output_emits_structured_json() {
     let runtime = tokio::runtime::Runtime::new().expect("tokio runtime should build");
@@ -215,3 +275,21 @@ fn unique_temp_dir(label: &str) -> PathBuf {
         std::process::id()
     ))
 }
+
+fn strip_ansi_codes(input: &str) -> String {
+    let mut output = String::with_capacity(input.len());
+    let mut chars = input.chars().peekable();
+    while let Some(ch) = chars.next() {
+        if ch == '\u{1b}' && matches!(chars.peek(), Some('[')) {
+            chars.next();
+            while let Some(next) = chars.next() {
+                if ('@'..='~').contains(&next) {
+                    break;
+                }
+            }
+            continue;
+        }
+        output.push(ch);
+    }
+    output
+}

rust/crates/rusty-claude-cli/tests/compact_repl_panic.rs

@@ -0,0 +1,138 @@
+use std::fs;
+use std::io::Write;
+use std::path::PathBuf;
+use std::process::{Command, Output, Stdio};
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+static TEMP_COUNTER: AtomicU64 = AtomicU64::new(0);
+
+#[test]
+fn compact_slash_command_in_repl_does_not_start_nested_tokio_runtime() {
+    // given
+    let workspace = unique_temp_dir("compact-repl-panic");
+    let config_home = workspace.join("config-home");
+    let home = workspace.join("home");
+    fs::create_dir_all(&workspace).expect("workspace should exist");
+    fs::create_dir_all(&config_home).expect("config home should exist");
+    fs::create_dir_all(&home).expect("home should exist");
+
+    // when
+    let output = run_claw_repl(&workspace, &config_home, &home, "/compact\n/exit\n");
+
+    // then
+    assert!(
+        output.status.success(),
+        "compact repl run should succeed\nstdout:\n{}\n\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr),
+    );
+    let stderr = String::from_utf8(output.stderr).expect("stderr should be utf8");
+    assert!(
+        !stderr.contains("Cannot start a runtime"),
+        "stderr must not contain nested runtime panic: {stderr:?}"
+    );
+    assert!(
+        !stderr.contains("panicked at"),
+        "stderr must not contain panic output: {stderr:?}"
+    );
+
+    let stdout = String::from_utf8(output.stdout).expect("stdout should be utf8");
+    let plain_stdout = strip_ansi_codes(&stdout);
+    assert!(
+        plain_stdout.contains("Compaction skipped")
+            || plain_stdout.contains("Result           skipped")
+            || plain_stdout.contains("Result           compacted"),
+        "stdout should contain compact report output ({stdout:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
+fn run_claw_repl(
+    cwd: &std::path::Path,
+    config_home: &std::path::Path,
+    home: &std::path::Path,
+    stdin: &str,
+) -> Output {
+    let mut command = python_pty_command(env!("CARGO_BIN_EXE_claw"));
+    let mut child = command
+        .current_dir(cwd)
+        .env_clear()
+        .env("ANTHROPIC_API_KEY", "test-compact-repl-key")
+        .env("CLAW_CONFIG_HOME", config_home)
+        .env("HOME", home)
+        .env("NO_COLOR", "1")
+        .env("PATH", "/usr/bin:/bin")
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped())
+        .spawn()
+        .expect("claw should launch");
+
+    child
+        .stdin
+        .as_mut()
+        .expect("stdin should be piped")
+        .write_all(stdin.as_bytes())
+        .expect("stdin should write");
+
+    child.wait_with_output().expect("claw should finish")
+}
+
+fn python_pty_command(claw: &str) -> Command {
+    let mut command = Command::new("python3");
+    command.args([
+        "-c",
+        r#"
+import os
+import pty
+import subprocess
+import sys
+
+claw = sys.argv[1]
+payload = sys.stdin.buffer.read()
+master, slave = pty.openpty()
+child = subprocess.Popen([claw], stdin=slave, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+os.close(slave)
+os.write(master, payload)
+stdout, stderr = child.communicate(timeout=30)
+os.close(master)
+sys.stdout.buffer.write(stdout)
+sys.stderr.buffer.write(stderr)
+raise SystemExit(child.returncode)
+"#,
+        claw,
+    ]);
+    command
+}
+
+fn unique_temp_dir(label: &str) -> PathBuf {
+    let millis = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .expect("clock should be after epoch")
+        .as_millis();
+    let counter = TEMP_COUNTER.fetch_add(1, Ordering::Relaxed);
+    std::env::temp_dir().join(format!(
+        "claw-{label}-{}-{millis}-{counter}",
+        std::process::id()
+    ))
+}
+
+fn strip_ansi_codes(input: &str) -> String {
+    let mut output = String::with_capacity(input.len());
+    let mut chars = input.chars().peekable();
+    while let Some(ch) = chars.next() {
+        if ch == '\u{1b}' && matches!(chars.peek(), Some('[')) {
+            chars.next();
+            for next in chars.by_ref() {
+                if ('@'..='~').contains(&next) {
+                    break;
+                }
+            }
+            continue;
+        }
+        output.push(ch);
+    }
+    output
+}

rust/crates/rusty-claude-cli/tests/output_format_contract.rs

@@ -22,6 +22,42 @@ fn help_emits_json_when_requested() {
         .contains("Usage:"));
 }
 
+#[test]
+fn export_help_emits_bounded_json_when_requested_384() {
+    let root = unique_temp_dir("export-help-json");
+    fs::create_dir_all(&root).expect("temp dir should exist");
+
+    let parsed = assert_json_command(&root, &["export", "--help", "--output-format", "json"]);
+    assert_eq!(parsed["kind"], "help");
+    assert_eq!(parsed["topic"], "export");
+    assert_eq!(parsed["command"], "export");
+    assert_eq!(
+        parsed["usage"],
+        "claw export [--session <id|latest>] [--output <path>] [--output-format <format>]"
+    );
+    assert_eq!(parsed["defaults"]["session"], "latest");
+    assert!(parsed["options"].as_array().expect("options").len() >= 4);
+    assert!(parsed.get("message").is_none());
+}
+
+#[test]
+fn export_help_preserves_plaintext_in_text_mode_384() {
+    let root = unique_temp_dir("export-help-text");
+    fs::create_dir_all(&root).expect("temp dir should exist");
+
+    let output = run_claw(&root, &["export", "--help"], &[]);
+    assert!(
+        output.status.success(),
+        "stdout:\n{}\n\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr)
+    );
+    let stdout = String::from_utf8(output.stdout).expect("stdout utf8");
+    assert!(stdout.starts_with("Export\n"));
+    assert!(stdout.contains("Usage            claw export"));
+    serde_json::from_str::<Value>(&stdout).expect_err("text help should remain plaintext");
+}
+
 #[test]
 fn version_emits_json_when_requested() {
     let root = unique_temp_dir("version-json");
@@ -30,6 +66,15 @@ fn version_emits_json_when_requested() {
     let parsed = assert_json_command(&root, &["--output-format", "json", "version"]);
     assert_eq!(parsed["kind"], "version");
     assert_eq!(parsed["version"], env!("CARGO_PKG_VERSION"));
+    // Provenance fields must be present for binary identification (#507).
+    assert!(
+        parsed["build_date"].is_string(),
+        "build_date must be a string in version JSON"
+    );
+    assert!(
+        parsed["executable_path"].is_string(),
+        "executable_path must be a string in version JSON so callers can identify which binary is running"
+    );
 }
 
 #[test]
@@ -105,6 +150,18 @@ fn inventory_commands_emit_structured_json_when_requested() {
     let skills = assert_json_command(&root, &["--output-format", "json", "skills"]);
     assert_eq!(skills["kind"], "skills");
     assert_eq!(skills["action"], "list");
+
+    let plugins = assert_json_command(&root, &["--output-format", "json", "plugins"]);
+    assert_eq!(plugins["kind"], "plugin");
+    assert_eq!(plugins["action"], "list");
+    assert!(
+        plugins["reload_runtime"].is_boolean(),
+        "plugins reload_runtime should be a boolean"
+    );
+    assert!(
+        plugins["target"].is_null(),
+        "plugins target should be null when no plugin is targeted"
+    );
 }
 
 #[test]
@@ -348,6 +405,62 @@ fn resumed_inventory_commands_emit_structured_json_when_requested() {
     assert_eq!(skills["action"], "list");
     assert!(skills["summary"]["total"].is_number());
     assert!(skills["skills"].is_array());
+
+    let agents = assert_json_command_with_env(
+        &root,
+        &[
+            "--output-format",
+            "json",
+            "--resume",
+            session_path.to_str().expect("utf8 session path"),
+            "/agents",
+        ],
+        &[
+            (
+                "CLAW_CONFIG_HOME",
+                config_home.to_str().expect("utf8 config home"),
+            ),
+            ("HOME", home.to_str().expect("utf8 home")),
+        ],
+    );
+    assert_eq!(agents["kind"], "agents");
+    assert_eq!(agents["action"], "list");
+    assert!(
+        agents["agents"].is_array(),
+        "agents field must be a JSON array"
+    );
+    assert!(
+        agents["count"].is_number(),
+        "count must be a number, not a text render"
+    );
+
+    let plugins = assert_json_command_with_env(
+        &root,
+        &[
+            "--output-format",
+            "json",
+            "--resume",
+            session_path.to_str().expect("utf8 session path"),
+            "/plugins",
+        ],
+        &[
+            (
+                "CLAW_CONFIG_HOME",
+                config_home.to_str().expect("utf8 config home"),
+            ),
+            ("HOME", home.to_str().expect("utf8 home")),
+        ],
+    );
+    assert_eq!(plugins["kind"], "plugin");
+    assert_eq!(plugins["action"], "list");
+    assert!(
+        plugins["reload_runtime"].is_boolean(),
+        "plugins reload_runtime should be a boolean"
+    );
+    assert!(
+        plugins["target"].is_null(),
+        "plugins target should be null when no plugin is targeted"
+    );
 }
 
 #[test]
@@ -384,6 +497,44 @@ fn resumed_version_and_init_emit_structured_json_when_requested() {
     assert!(root.join("CLAUDE.md").exists());
 }
 
+#[test]
+fn config_section_json_emits_section_and_value() {
+    let root = unique_temp_dir("config-section-json");
+    fs::create_dir_all(&root).expect("temp dir should exist");
+
+    // Without a section: should return base envelope (no section field).
+    let base = assert_json_command(&root, &["--output-format", "json", "config"]);
+    assert_eq!(base["kind"], "config");
+    assert!(base["loaded_files"].is_number());
+    assert!(base["merged_keys"].is_number());
+    assert!(
+        base.get("section").is_none(),
+        "no section field without section arg"
+    );
+
+    // With a known section: should add section + section_value fields.
+    for section in &["model", "env", "hooks", "plugins"] {
+        let result = assert_json_command(&root, &["--output-format", "json", "config", section]);
+        assert_eq!(result["kind"], "config", "section={section}");
+        assert_eq!(
+            result["section"].as_str(),
+            Some(*section),
+            "section field must match requested section, got {result:?}"
+        );
+        assert!(
+            result.get("section_value").is_some(),
+            "section_value field must be present for section={section}"
+        );
+    }
+
+    // With an unsupported section: should return ok:false + error field.
+    let bad = assert_json_command(&root, &["--output-format", "json", "config", "unknown"]);
+    assert_eq!(bad["kind"], "config");
+    assert_eq!(bad["ok"], false);
+    assert!(bad["error"].as_str().is_some());
+    assert!(bad["section"].as_str().is_some());
+}
+
 fn assert_json_command(current_dir: &Path, args: &[&str]) -> Value {
     assert_json_command_with_env(current_dir, args, &[])
 }

rust/crates/rusty-claude-cli/tests/resume_slash_commands.rs

@@ -227,6 +227,8 @@ fn resumed_status_command_emits_structured_json_when_requested() {
     // given
     let temp_dir = unique_temp_dir("resume-status-json");
     fs::create_dir_all(&temp_dir).expect("temp dir should exist");
+    let config_home = temp_dir.join("config-home");
+    fs::create_dir_all(&config_home).expect("isolated config home should exist");
     let session_path = temp_dir.join("session.jsonl");
 
     let mut session = workspace_session(&temp_dir);
@@ -238,7 +240,9 @@ fn resumed_status_command_emits_structured_json_when_requested() {
         .expect("session should persist");
 
     // when
-    let output = run_claw(
+    // Use an isolated CLAW_CONFIG_HOME so ~/.claw/settings.json is not loaded,
+    // which would cause loaded_config_files to be non-zero (#65).
+    let output = run_claw_with_env(
         &temp_dir,
         &[
             "--output-format",
@@ -247,6 +251,7 @@ fn resumed_status_command_emits_structured_json_when_requested() {
             session_path.to_str().expect("utf8 path"),
             "/status",
         ],
+        &[("CLAW_CONFIG_HOME", config_home.to_str().expect("utf8 path"))],
     );
 
     // then

rust/crates/tools/src/lib.rs

@@ -4,9 +4,10 @@ use std::process::Command;
 use std::time::{Duration, Instant};
 
 use api::{
-    max_tokens_for_model, resolve_model_alias, ApiError, ContentBlockDelta, InputContentBlock,
-    InputMessage, MessageRequest, MessageResponse, OutputContentBlock, ProviderClient,
-    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
+    max_tokens_for_model, model_family_identity_for, resolve_model_alias, ApiError,
+    ContentBlockDelta, InputContentBlock, InputMessage, MessageRequest, MessageResponse,
+    OutputContentBlock, ProviderClient, StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition,
+    ToolResultContentBlock,
 };
 use plugins::PluginTool;
 use reqwest::blocking::Client;
@@ -3075,27 +3076,33 @@ fn extract_quoted_value(input: &str) -> Option<(String, &str)> {
 }
 
 fn decode_duckduckgo_redirect(url: &str) -> Option<String> {
-    if url.starts_with("http://") || url.starts_with("https://") {
-        return Some(html_entity_decode_url(url));
-    }
-
-    let joined = if url.starts_with("//") {
-        format!("https:{url}")
-    } else if url.starts_with('/') {
-        format!("https://duckduckgo.com{url}")
+    let decoded = html_entity_decode_url(url);
+    let parsed = if decoded.starts_with("http://") || decoded.starts_with("https://") {
+        reqwest::Url::parse(&decoded).ok()
+    } else if decoded.starts_with("//") {
+        reqwest::Url::parse(&format!("https:{decoded}")).ok()
+    } else if decoded.starts_with('/') {
+        reqwest::Url::parse(&format!("https://duckduckgo.com{decoded}")).ok()
     } else {
         return None;
-    };
+    }?;
 
-    let parsed = reqwest::Url::parse(&joined).ok()?;
-    if parsed.path() == "/l/" || parsed.path() == "/l" {
+    let host = parsed.host_str().unwrap_or_default().to_ascii_lowercase();
+    if (host == "duckduckgo.com" || host.ends_with(".duckduckgo.com"))
+        && (parsed.path() == "/l/" || parsed.path() == "/l")
+    {
         for (key, value) in parsed.query_pairs() {
             if key == "uddg" {
                 return Some(html_entity_decode_url(value.as_ref()));
             }
         }
     }
-    Some(joined)
+
+    if decoded.starts_with("http://") || decoded.starts_with("https://") {
+        Some(decoded)
+    } else {
+        Some(parsed.to_string())
+    }
 }
 
 fn html_entity_decode_url(url: &str) -> String {
@@ -3510,7 +3517,7 @@ where
         .filter(|name| !name.is_empty())
         .unwrap_or_else(|| slugify_agent_name(&input.description));
     let created_at = iso8601_now();
-    let system_prompt = build_agent_system_prompt(&normalized_subagent_type)?;
+    let system_prompt = build_agent_system_prompt(&normalized_subagent_type, &model)?;
     let allowed_tools = allowed_tools_for_subagent(&normalized_subagent_type);
 
     let output_contents = format!(
@@ -3623,13 +3630,14 @@ fn build_agent_runtime(
     ))
 }
 
-fn build_agent_system_prompt(subagent_type: &str) -> Result<Vec<String>, String> {
+fn build_agent_system_prompt(subagent_type: &str, model: &str) -> Result<Vec<String>, String> {
     let cwd = std::env::current_dir().map_err(|error| error.to_string())?;
     let mut prompt = load_system_prompt(
         cwd,
         DEFAULT_AGENT_SYSTEM_DATE.to_string(),
         std::env::consts::OS,
         "unknown",
+        model_family_identity_for(model),
     )
     .map_err(|error| error.to_string())?;
     prompt.push(format!(
@@ -4630,13 +4638,21 @@ async fn stream_with_provider(
     let mut stream = client.stream_message(message_request).await?;
     let mut events = Vec::new();
     let mut pending_tools: BTreeMap<u32, (String, String, String)> = BTreeMap::new();
+    let mut pending_thinking: BTreeMap<u32, (String, Option<String>)> = BTreeMap::new();
     let mut saw_stop = false;
 
     while let Some(event) = stream.next_event().await? {
         match event {
             ApiStreamEvent::MessageStart(start) => {
                 for block in start.message.content {
-                    push_output_block(block, 0, &mut events, &mut pending_tools, true);
+                    push_output_block(
+                        block,
+                        0,
+                        &mut events,
+                        &mut pending_tools,
+                        &mut pending_thinking,
+                        true,
+                    );
                 }
             }
             ApiStreamEvent::ContentBlockStart(start) => {
@@ -4645,6 +4661,7 @@ async fn stream_with_provider(
                     start.index,
                     &mut events,
                     &mut pending_tools,
+                    &mut pending_thinking,
                     true,
                 );
             }
@@ -4659,10 +4676,26 @@ async fn stream_with_provider(
                         input.push_str(&partial_json);
                     }
                 }
-                ContentBlockDelta::ThinkingDelta { .. }
-                | ContentBlockDelta::SignatureDelta { .. } => {}
+                ContentBlockDelta::ThinkingDelta { thinking } => {
+                    if let Some((pending, _)) = pending_thinking.get_mut(&delta.index) {
+                        pending.push_str(&thinking);
+                    }
+                }
+                ContentBlockDelta::SignatureDelta { signature } => {
+                    if let Some((_, pending_signature)) = pending_thinking.get_mut(&delta.index) {
+                        pending_signature
+                            .get_or_insert_with(String::new)
+                            .push_str(&signature);
+                    }
+                }
             },
             ApiStreamEvent::ContentBlockStop(stop) => {
+                if let Some((thinking, signature)) = pending_thinking.remove(&stop.index) {
+                    events.push(AssistantEvent::Thinking {
+                        thinking,
+                        signature,
+                    });
+                }
                 if let Some((id, name, input)) = pending_tools.remove(&stop.index) {
                     events.push(AssistantEvent::ToolUse { id, name, input });
                 }
@@ -4759,6 +4792,13 @@ fn convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
                 .iter()
                 .map(|block| match block {
                     ContentBlock::Text { text } => InputContentBlock::Text { text: text.clone() },
+                    ContentBlock::Thinking {
+                        thinking,
+                        signature,
+                    } => InputContentBlock::Thinking {
+                        thinking: thinking.clone(),
+                        signature: signature.clone(),
+                    },
                     ContentBlock::ToolUse { id, name, input } => InputContentBlock::ToolUse {
                         id: id.clone(),
                         name: name.clone(),
@@ -4778,6 +4818,9 @@ fn convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
                         is_error: *is_error,
                     },
                 })
+                .filter(
+                    |block| !matches!(block, InputContentBlock::Text { text } if text.is_empty()),
+                )
                 .collect::<Vec<_>>();
             (!content.is_empty()).then(|| InputMessage {
                 role: role.to_string(),
@@ -4792,6 +4835,7 @@ fn push_output_block(
     block_index: u32,
     events: &mut Vec<AssistantEvent>,
     pending_tools: &mut BTreeMap<u32, (String, String, String)>,
+    pending_thinking: &mut BTreeMap<u32, (String, Option<String>)>,
     streaming_tool_input: bool,
 ) {
     match block {
@@ -4811,17 +4855,38 @@ fn push_output_block(
             };
             pending_tools.insert(block_index, (id, name, initial_input));
         }
-        OutputContentBlock::Thinking { .. } | OutputContentBlock::RedactedThinking { .. } => {}
+        OutputContentBlock::Thinking {
+            thinking,
+            signature,
+        } => {
+            if streaming_tool_input {
+                pending_thinking.insert(block_index, (thinking, signature));
+            } else {
+                events.push(AssistantEvent::Thinking {
+                    thinking,
+                    signature,
+                });
+            }
+        }
+        OutputContentBlock::RedactedThinking { .. } => {}
     }
 }
 
 fn response_to_events(response: MessageResponse) -> Vec<AssistantEvent> {
     let mut events = Vec::new();
     let mut pending_tools = BTreeMap::new();
+    let mut pending_thinking = BTreeMap::new();
 
     for (index, block) in response.content.into_iter().enumerate() {
         let index = u32::try_from(index).expect("response block index overflow");
-        push_output_block(block, index, &mut events, &mut pending_tools, false);
+        push_output_block(
+            block,
+            index,
+            &mut events,
+            &mut pending_tools,
+            &mut pending_thinking,
+            false,
+        );
         if let Some((id, name, input)) = pending_tools.remove(&index) {
             events.push(AssistantEvent::ToolUse { id, name, input });
         }
@@ -6134,12 +6199,13 @@ mod tests {
     use std::time::Duration;
 
     use super::{
-        agent_permission_policy, allowed_tools_for_subagent, classify_lane_failure,
-        derive_agent_state, execute_agent_with_spawn, execute_tool, extract_recovery_outcome,
-        final_assistant_text, global_cron_registry, maybe_commit_provenance, mvp_tool_specs,
-        permission_mode_from_plugin, persist_agent_terminal_state, push_output_block,
-        run_task_packet, AgentInput, AgentJob, GlobalToolRegistry, LaneEventName, LaneFailureClass,
-        ProviderRuntimeClient, SubagentToolExecutor,
+        agent_permission_policy, allowed_tools_for_subagent, build_agent_system_prompt,
+        classify_lane_failure, derive_agent_state, execute_agent_with_spawn, execute_tool,
+        extract_recovery_outcome, final_assistant_text, global_cron_registry,
+        maybe_commit_provenance, mvp_tool_specs, permission_mode_from_plugin,
+        persist_agent_terminal_state, push_output_block, run_task_packet, AgentInput, AgentJob,
+        GlobalToolRegistry, LaneEventName, LaneFailureClass, ProviderRuntimeClient,
+        SubagentToolExecutor,
     };
     use api::OutputContentBlock;
     use runtime::ProviderFallbackConfig;
@@ -7148,10 +7214,103 @@ mod tests {
         assert!(error.contains("relative URL without a base") || error.contains("empty host"));
     }
 
+    #[test]
+    fn web_search_decodes_absolute_duckduckgo_redirect_urls() {
+        // given
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        let server = TestServer::spawn(Arc::new(|request_line: &str| {
+            assert!(request_line.contains("GET /search?q=duckduckgo+redirects "));
+            HttpResponse::html(
+                200,
+                "OK",
+                r#"
+                <html><body>
+                  <a rel="nofollow" class="result__a" href="https://duckduckgo.com/l/?uddg=https%3A%2F%2Fdocs.rs%2Freqwest&amp;rut=abc">Reqwest docs</a>
+                </body></html>
+                "#,
+            )
+        }));
+
+        // when
+        std::env::set_var(
+            "CLAWD_WEB_SEARCH_BASE_URL",
+            format!("http://{}/search", server.addr()),
+        );
+        let result = execute_tool(
+            "WebSearch",
+            &json!({
+                "query": "duckduckgo redirects"
+            }),
+        )
+        .expect("WebSearch should succeed");
+        std::env::remove_var("CLAWD_WEB_SEARCH_BASE_URL");
+
+        // then
+        let output: serde_json::Value = serde_json::from_str(&result).expect("valid json");
+        let results = output["results"].as_array().expect("results array");
+        let search_result = results
+            .iter()
+            .find(|item| item.get("content").is_some())
+            .expect("search result block present");
+        let content = search_result["content"].as_array().expect("content array");
+        assert_eq!(content.len(), 1);
+        assert_eq!(content[0]["title"], "Reqwest docs");
+        assert_eq!(content[0]["url"], "https://docs.rs/reqwest");
+    }
+
+    #[test]
+    fn web_search_decodes_protocol_relative_duckduckgo_redirect_urls() {
+        // given
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        let server = TestServer::spawn(Arc::new(|request_line: &str| {
+            assert!(request_line.contains("GET /search?q=duckduckgo+protocol+relative "));
+            HttpResponse::html(
+                200,
+                "OK",
+                r#"
+                <html><body>
+                  <a rel="nofollow" class="result__a" href="//duckduckgo.com/l/?uddg=https%3A%2F%2Fdocs.rs%2Ftokio&amp;rut=xyz">Tokio Docs</a>
+                </body></html>
+                "#,
+            )
+        }));
+
+        // when
+        std::env::set_var(
+            "CLAWD_WEB_SEARCH_BASE_URL",
+            format!("http://{}/search", server.addr()),
+        );
+        let result = execute_tool(
+            "WebSearch",
+            &json!({
+                "query": "duckduckgo protocol relative"
+            }),
+        )
+        .expect("WebSearch should succeed");
+        std::env::remove_var("CLAWD_WEB_SEARCH_BASE_URL");
+
+        // then
+        let output: serde_json::Value = serde_json::from_str(&result).expect("valid json");
+        let results = output["results"].as_array().expect("results array");
+        let search_result = results
+            .iter()
+            .find(|item| item.get("content").is_some())
+            .expect("search result block present");
+        let content = search_result["content"].as_array().expect("content array");
+        assert_eq!(content.len(), 1);
+        assert_eq!(content[0]["title"], "Tokio Docs");
+        assert_eq!(content[0]["url"], "https://docs.rs/tokio");
+    }
+
     #[test]
     fn pending_tools_preserve_multiple_streaming_tool_calls_by_index() {
         let mut events = Vec::new();
         let mut pending_tools = BTreeMap::new();
+        let mut pending_thinking = BTreeMap::new();
 
         push_output_block(
             OutputContentBlock::ToolUse {
@@ -7162,6 +7321,7 @@ mod tests {
             1,
             &mut events,
             &mut pending_tools,
+            &mut pending_thinking,
             true,
         );
         push_output_block(
@@ -7173,6 +7333,7 @@ mod tests {
             2,
             &mut events,
             &mut pending_tools,
+            &mut pending_thinking,
             true,
         );
 
@@ -8409,6 +8570,28 @@ mod tests {
         assert!(!verification.contains("write_file"));
     }
 
+    #[test]
+    fn subagent_system_prompt_uses_resolved_model_identity() {
+        // given: a temporary workspace and an OpenAI-compatible subagent model
+        let _guard = env_guard();
+        let root = temp_path("subagent-prompt-identity");
+        fs::create_dir_all(&root).expect("create temp workspace");
+        let previous = std::env::current_dir().expect("current dir");
+        std::env::set_current_dir(&root).expect("enter temp workspace");
+
+        // when: building the subagent system prompt
+        let prompt = build_agent_system_prompt("Explore", "openai/gpt-4.1-mini")
+            .expect("subagent system prompt should build")
+            .join("\n");
+        std::env::set_current_dir(previous).expect("restore current dir");
+
+        // then: the prompt renders a generic model family identity
+        assert!(prompt.contains("Model family: an AI assistant"));
+        assert!(!prompt.contains("Model family: Claude Opus 4.6"));
+
+        fs::remove_dir_all(root).expect("cleanup temp workspace");
+    }
+
     #[derive(Debug)]
     struct MockSubagentApiClient {
         calls: usize,

scripts/dogfood-build.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# dogfood-build.sh — Build claw from current checkout and verify provenance.
+#
+# Injects GIT_SHA at build time so version JSON is non-null.
+# Suppresses Cargo compile noise on stderr.
+# Prints the verified binary path on success. Use as:
+#
+#   CLAW=$(bash scripts/dogfood-build.sh)
+#
+# Then dogfood with config isolation (avoids real user config bleeding in):
+#
+#   CLAW_CONFIG_HOME=$(mktemp -d) $CLAW plugins list --output-format json
+#
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+RUST_DIR="$REPO_ROOT/rust"
+BINARY="$RUST_DIR/target/debug/claw"
+EXPECTED_SHA="$(git -C "$REPO_ROOT" rev-parse --short HEAD)"
+
+echo "▶ Building claw from $REPO_ROOT" >&2
+echo "  Commit: $(git -C "$REPO_ROOT" log --oneline -1)" >&2
+
+# Inject GIT_SHA so version JSON returns a non-null sha.
+# Redirect cargo stderr to /dev/null to suppress compile noise;
+# on build failure cargo exits non-zero and set -e aborts.
+if ! GIT_SHA="$EXPECTED_SHA" cargo build \
+        --manifest-path "$RUST_DIR/Cargo.toml" \
+        -p rusty-claude-cli -q 2>/dev/null; then
+    # Re-run with visible output so the user sees the error
+    echo "✗ Build failed — rerunning with output:" >&2
+    GIT_SHA="$EXPECTED_SHA" cargo build \
+        --manifest-path "$RUST_DIR/Cargo.toml" \
+        -p rusty-claude-cli 2>&1 | sed 's/^/  /' >&2
+    exit 1
+fi
+
+if [[ ! -x "$BINARY" ]]; then
+    echo "✗ Binary not found at $BINARY" >&2
+    exit 1
+fi
+
+BINARY_SHA=$("$BINARY" version --output-format json 2>/dev/null \
+    | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('git_sha') or 'null')" 2>/dev/null \
+    || echo "null")
+
+if [[ "$BINARY_SHA" == "null" || -z "$BINARY_SHA" ]]; then
+    echo "✗ Provenance check failed: binary reports git_sha: null" >&2
+    exit 1
+fi
+
+if [[ "$BINARY_SHA" != "$EXPECTED_SHA" ]]; then
+    echo "✗ Provenance mismatch: binary=$BINARY_SHA, HEAD=$EXPECTED_SHA" >&2
+    exit 1
+fi
+
+echo "✓ Binary verified: $BINARY_SHA == HEAD" >&2
+echo "" >&2
+echo "  export CLAW=$BINARY" >&2
+echo "" >&2
+echo "  Dogfood with isolated config (no real user config on stderr):" >&2
+echo "    CLAW_ISOLATED=\$(mktemp -d)" >&2
+echo "    CLAW_CONFIG_HOME=\$CLAW_ISOLATED \$CLAW plugins list --output-format json" >&2
+echo "    rm -rf \$CLAW_ISOLATED" >&2
+echo "" >&2
+echo "  cargo run overhead: ~1s/invocation vs 7ms for pre-built binary." >&2
+echo "  Prefer pre-built binary (\$CLAW) for dogfood loops." >&2
+echo "$BINARY"