docs(roadmap): add #466 — provider *_BASE_URL env vars can be malformed while doctor/status stay green

ROADMAP.md

@@ -6429,4 +6429,44 @@ Original filing (2026-04-18): the session emitted `SessionStart hook (completed)
 
 450. **`prompt` emits `kind:"missing_credentials"` JSON on STDERR (not stdout), leaving stdout at 0 bytes — automation pattern `output=$(claw prompt hello --output-format json)` captures nothing on auth-absent failure; `doctor` correctly surfaces `auth.status:"warn"` with `api_key_present:false` but exposes no `prompt_ready:false` field that automation can check before invoking `prompt`** — dogfooded 2026-05-16 by Jobdori on `a35ee9a0` in response to Clawhip pinpoint nudge at `1505208225321062521`. Exact reproduction (isolated env, no creds, fresh git repo, HEAD `a35ee9a0`): `timeout 5 env -i HOME=$ISOLATED_HOME PATH=$PATH CLAW_CONFIG_HOME=$PROBE/.claw-cfg claw prompt hello --output-format json > stdout.txt 2> stderr.txt` → stdout = **0 bytes**, stderr = 195 bytes containing `{"error":"missing Anthropic credentials…","exit_code":1,"hint":null,"kind":"missing_credentials","type":"error"}`, exit code 1. Confirms Gaebal's `1505208553793781792` pinpoint that `prompt` timeout + zero bytes was the prior state — HEAD `a35ee9a0` now correctly exits 1 with `kind:"missing_credentials"` **but the envelope is still routed to stderr** (issue #447 class, same class as prior entries #422, #435). **Contrast with `doctor`:** `claw doctor --output-format json 2>/dev/null` succeeds to stdout with `checks[auth].status:"warn"`, `api_key_present:false`, `auth_token_present:false` — but the auth check has no `prompt_ready:false` field. Automation that gates on `doctor` before invoking `prompt` must re-derive readiness from `api_key_present && auth_token_present` — there is no single canonical boolean. **Three compound problems:** (a) **stdout-empty on `--output-format json` failure**: same class as #447; `prompt`'s error envelope goes to stderr, not stdout. The canonical automation idiom `if ! result=$(claw prompt "q" --output-format json); then echo "$result" | jq .kind; fi` sees `$result=""` on failure — the jq call gets nothing. All `--output-format json` error paths must route JSON to stdout per #447 contract; (b) **`doctor` missing `prompt_ready` field**: `doctor --output-format json` already knows auth is absent (`api_key_present:false`) but surfaces no derived `prompt_ready:bool` or `prompt_blocked_reason:string` field. Automation must infer readiness from `api_key_present || auth_token_present || legacy_*_present` — a 5-field OR across legacy fields that is fragile as auth mechanisms evolve. A single `prompt_ready:false` (with `prompt_blocked_reason:"auth_missing"`) inside the `auth` check would give downstream a stable contract; (c) **`claw prompt` with no auth does no preflight and fires straight at the API**: the preflight check that `doctor` runs (auth discovery) is not reused by `prompt` to emit a fast typed error before attempting the network call. Both Gaebal's pinpoint (prompt hanging silently on older HEAD) and the current behavior (prompt hitting auth gate after a brief API attempt) stem from the same root: prompt does not short-circuit at the point where `doctor` already knows auth is absent. If `doctor` can emit `kind:"doctor"` with `auth.status:"warn"` in ~20ms without a network call, `prompt` should emit `kind:"missing_credentials"` in the same window and output it to stdout. **Required fix shape:** (a) `prompt --output-format json` must write the `kind:"missing_credentials"` JSON envelope to **stdout**, not stderr — same fix as #447 for all error envelopes; (b) add `prompt_ready:bool` and `prompt_blocked_reason:string|null` to the `auth` check in `doctor --output-format json`; derive it as `api_key_present || auth_token_present || legacy_saved_oauth_present`; (c) `prompt` must run the credential preflight check (same codepath as doctor's auth check) before attempting any API call and emit `{"kind":"missing_credentials","prompt_blocked_reason":"auth_missing"}` on **stdout** with exit 1 if the check fails; (d) `--output-format json` stdout routing fix must cover: `prompt`, `session list` (cross-ref #449), `skills uninstall` (cross-ref #431), `resume` (cross-ref #435), `acp serve` (cross-ref #443) — the full `kind:"missing_credentials"` class; (e) regression test: `claw prompt hello --output-format json` with no creds writes JSON to stdout (0 bytes stderr), exits 1, `kind:"missing_credentials"`, in under 200ms (no network attempt). **Why this matters:** `prompt` is the primary consumer entry point. Auth-absent failure routing to stderr breaks every automation wrapper that captures `$(claw prompt ... --output-format json)`. The `doctor` preflight metadata gap means auth-readiness checks require parsing 5 legacy fields instead of reading one boolean. Cross-references #447 (all JSON error envelopes on stderr), #449 (session list hits auth gate), #431 (skills uninstall hits auth gate), #357 (auth gate on local ops cluster), #422 (exit-code parity). Source: Jobdori live dogfood, `a35ee9a0`, 2026-05-16.
 
-452. **`claw models`, `claw models list`, `claw models help`, and `claw models --help` are not wired as a `CliAction` at all — every spelling falls through to `CliAction::Prompt` and is sent verbatim to the Anthropic API as a user prompt; with credentials the CLI spins on the LLM "Thinking…" spinner forever, without credentials it errors with `missing_credentials` from the provider path. Direct sibling of #78 (`claw plugins` had the same prompt-misdelivery failure mode) for an additional discovery surface that operators and claws naturally try first when they need a model registry/alias/provider list before invoking `--model <alias> prompt …`** — dogfooded 2026-05-24 for the 05:00 Clawhip pinpoint nudge at message `1507971434704797716`, reproduced on local `./rust/target/debug/claw` `git_sha 003b739d` (origin/main `f8e1bb72`; `models` dispatch is grep-clean across `rust/crates/` — `git grep -nE 'CliAction::Models|"/models"|"models"' rust/` returns 0 hits, while `CliAction::Plugins` is wired at `rust/crates/rusty-claude-cli/src/main.rs:356,891,10153,10158,10167,10180,10193`, so `models` is the analogous unrouted command exactly the way `plugins` was before #78 landed). Repros in a fully clean isolated environment (`HOME=/tmp/iso2/home` with `{"}` settings, fresh `/tmp/iso2/proj` git-init'd workspace, `stdin=/dev/null`): `timeout 8 claw models list` exits `1` with `stderr=490` carrying the **Anthropic provider** `missing_credentials` envelope when `ANTHROPIC_*` env vars are unset, proving the command was dispatched to the LLM rather than handled locally; with `ANTHROPIC_API_KEY` set, every spelling (`models list`, `models`, `models help`, `models --help`) shows the spinner ANSI sequence (`\x1b[38;5;12m⠋ 🦀 Thinking…\x1b[0m`) on stdout and never returns inside the 6–8s bounded budget. **Why this is distinct from #78 / #145 / the help-JSON cluster:** #78 covered `claw plugins` only; #145 added the regression for `Plugins` parsing; #451 covers `models` in `--output-format json` mode where the failure surface is the silent zero-byte JSON deadlock. This pinpoint is the **plain-text prompt-misdelivery path** for `models`, with three behavioral consequences not covered above: (1) operators get the wrong-shaped error (`missing_credentials` for an Anthropic prompt) when they meant to inspect the model registry; (2) with credentials, expensive token burn on a meaningless `"models list"` LLM completion; (3) no slash-command-vs-direct-command parity — there is also no `/models` REPL command, so claws have no recovery path either. **Required fix shape:** (a) add `CliAction::Models { action: ModelsAction }` with `List`, `Show { name }`, `Help` variants wired in `parse_args` next to the existing `CliAction::Plugins` arm, never falling through to `CliAction::Prompt` for any `models*` spelling; (b) implement `models list` to return the resolved provider registry merged from built-ins (`anthropic`, `openai`, `xai`) plus any `modelProviders.*` profiles in settings, with per-model `name`, `provider`, `aliases[]`, `available`, `requires_credentials`, `source`; (c) implement `models --help` / `models help` as a static bounded help renderer (text + JSON envelopes) that does not touch provider runtime; (d) mirror the slash surface (`/models` REPL command) to match the existing `/agents`, `/mcp`, `/skills`, `/config` pattern; (e) add regression coverage in `parses_models_subcommand`-style tests proving every `models*` spelling resolves to `CliAction::Models` (no LLM dispatch), AND that the action returns within a deterministic budget without provider credentials. **Why this matters:** `models list` is the canonical model-registry discovery spelling across competing CLIs (`gh models list`, `openai api models.list`, `codex models`, the Anthropic Console UI). A claw or operator who reaches for it before deciding `--model <alias>` cannot discover what models exist, cannot validate an alias before paying for a prompt, and — worst case — burns provider tokens sending the string `"models list"` to Claude on a credentialed setup. The cost-of-doing-nothing here is real spend, not just opacity, which is why the prompt-misdelivery class deserves its own surface entry beyond #78/#145's `plugins` precedent. Source: gaebal-gajae dogfood follow-up for the 2026-05-24 05:00 Clawhip pinpoint nudge at message `1507971434704797716`.
+466. **Provider `*_BASE_URL` env vars are accepted as routing/transport configuration but `doctor` / `status` do zero validation and surface zero provenance: 24 malformed/unsupported values across `ANTHROPIC_BASE_URL`, `OPENAI_BASE_URL`, `XAI_BASE_URL`, and `DASHSCOPE_BASE_URL` all return `doctor_exit=0`, `has_failures=false`, `auth ok`, `config ok`, and `system ok`, even for `not-a-url`, `ftp://example.com`, `http://`, `http://localhost:99999`, `javascript:alert(1)`, and empty string. This makes the preflight surface say “green” while the next prompt call will fail later in the HTTP client / URL parser / provider edge, with no machine-readable clue which base URL env var poisoned the lane** — dogfooded 2026-05-24 for the 17:00–17:30 Clawhip nudge window (finalized for message `1508160182386167961`), reproduced on local `./rust/target/debug/claw` `git_sha 003b739d` (origin/main `f8e1bb72`) in a clean isolated env.
+
+    Reproduction matrix (credential-free except fake API key env vars so auth check passes):
+
+    ```bash
+    for var in ANTHROPIC_BASE_URL OPENAI_BASE_URL XAI_BASE_URL DASHSCOPE_BASE_URL; do
+      for url in 'not-a-url' 'ftp://example.com' 'http://' 'http://localhost:99999' 'javascript:alert(1)' ''; do
+        env -i HOME=/tmp/iso24/home ANTHROPIC_API_KEY=sk-ant-fake OPENAI_API_KEY=sk-openai-fake \
+          "$var=$url" PATH=/usr/bin:/bin TERM=dumb \
+          claw doctor --output-format json
+      done
+    done
+    ```
+
+    Every row produced the same health shape:
+
+    ```text
+    doctor_exit=0 stderr_bytes=0
+    has_failures false
+    auth ok 'supported auth env vars are configured'
+    config ok 'no config files present; defaults are active'
+    system ok 'captured local runtime metadata'
+    ```
+
+    No diagnostic says the lane is configured to talk to `javascript:alert(1)`, `ftp://example.com`, `http://`, or an invalid port `99999`.
+
+    **Root cause traced:** provider metadata knows which env var controls each provider's base URL (`rust/crates/api/src/providers/mod.rs:42-43`, with `base_url_env` / `default_base_url` fields), and the Anthropic client reads `ANTHROPIC_BASE_URL` directly in `anthropic.rs:765-766`:
+
+    ```rust
+    pub fn read_base_url() -> String {
+        std::env::var("ANTHROPIC_BASE_URL").unwrap_or_else(|_| DEFAULT_BASE_URL.to_string())
+    }
+    ```
+
+    OpenAI-compat providers carry the same metadata in `openai_compat.rs` (`XAI_BASE_URL`, `OPENAI_BASE_URL`, `DASHSCOPE_BASE_URL`). But `check_auth_health()` only checks auth booleans, `check_config_health()` only checks config files/MCP counts, and `check_system_health()` only reports local runtime metadata. There is no provider-endpoint check and no `base_url_source` field in `status`. The runtime will later concatenate/parse/use these env vars as actual URLs (`anthropic.rs:470` builds `{base_url}/v1/messages`), but the preflight surfaces do not validate scheme, host, parseability, or port range.
+
+    **Why distinct from existing items:** ROADMAP #28/#29 cover provider routing and auth-copy issues, including OpenRouter guidance and prefix routing, but not malformed base URL validation. ROADMAP #248 asks non-interactive prompt lifecycle events to include provider/model/base-url identity after a prompt starts; this entry is the *preflight* counterpart: `doctor` / `status` should surface bad base URLs before any prompt/API call. ROADMAP #465 covers effective auth source when both Anthropic auth env vars are set; this entry covers endpoint configuration across all provider env vars. ROADMAP #111 says `/providers` should list providers/base URLs/reachability, but `/providers` is a slash-command spec mismatch; this entry is specifically about the existing `doctor`/`status` green-light surfaces ignoring base URL env vars. No existing entry documents that every malformed `*_BASE_URL` value tested leaves doctor green.
+
+    **Why this matters:** (1) **Startup friction:** users configuring OpenRouter/Ollama/local proxy inevitably touch `OPENAI_BASE_URL`. A one-character typo (`http://`, trailing junk, unsupported scheme) currently passes doctor and fails only during a live prompt. (2) **Automation cannot preflight lanes.** Clawhip wants to know whether a lane is safe before spending tokens / starting a long prompt. Green doctor with poisoned base URL is false confidence. (3) **Security posture:** accepting `javascript:` / `ftp:` / empty string as silently “configured” is not just UX badness; provider clients should only ever use HTTP(S) URLs, and diagnostics should reject or warn on anything else. (4) **The metadata already exists.** Provider metadata names every env var and default URL; the missing layer is just validation and redaction-safe surfacing. (5) **Base URL is provider identity.** Status currently reports model and permission mode, but not where requests will be sent. For OpenAI-compatible providers especially, `OPENAI_BASE_URL` is the difference between OpenAI, OpenRouter, Ollama, local proxy, or malicious typo.
+
+    **Required fix shape:** (a) Add a provider endpoint diagnostics check to `doctor` that iterates provider metadata, reads each `*_BASE_URL` env var if present, trims it, parses it with `Url`, and validates `scheme in {http, https}`, non-empty host, valid port, and no unsupported schemes. Empty string should be treated as unset or explicit invalid with a dedicated warning, not silent ok. (b) Add redaction-safe fields to `status --output-format json`: active provider, `base_url_env`, `base_url_source: "default" | "env"`, `base_url_valid`, `base_url_scheme`, `base_url_host`, and `base_url_error` if invalid. Do not include credentials or path secrets; host/scheme are enough. (c) When the selected model/provider is affected by an invalid base URL, `doctor` should be `warn` or `fail` and `status.status` should be `degraded`, not `ok`. (d) Add tests for the 24-row matrix above plus a valid local URL (`http://127.0.0.1:11434/v1`) and valid HTTPS URL. (e) Optional: `/providers` (when fixed from #111) should reuse the same endpoint validation so base URL truth has one source. **Acceptance check:** `env OPENAI_API_KEY=sk-test OPENAI_BASE_URL=javascript:alert\(1\) claw doctor --output-format json | jq -e '.checks[] | select(.name=="providers" or .name=="provider_endpoints") | .status != "ok" and (.details[]? | test("OPENAI_BASE_URL"))'` should pass; currently there is no such check and doctor is green. Source: gaebal-gajae dogfood for the 2026-05-24 17:00/17:30 Clawhip nudges. Coordination note: still avoided F/CLAW_CONFIG_HOME because Jobdori publicly queued it; this endpoint-validation surface is orthogonal and credential-free.