Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
docs(muse): record the MTL-189 findings and live verification
Jack Rudenkodfa65e8121e4
diff --git a/ai-docs/reports/mtl-189-muse-code-routing-20261022.md b/ai-docs/reports/mtl-189-muse-code-routing-20261022.mdnew file mode 100644index 00000000..14997f3f--- /dev/null+++ b/ai-docs/reports/mtl-189-muse-code-routing-20261022.md@@ -0,0 +1,66 @@+# MTL-189: Muse Code routing — findings and live verification (2026-10-22)++Released in 13.3.0. Rationale lives in `ai-docs/architecture/providers/meta-muse-code.md`,+`routing.md` and `context-window.md`; this report keeps the evidence those files do not.++## The defect++A bare `muse-wonder-full` failed with `No provider in the catalog serves "muse-wonder-full"`,+and the other Muse Code models (`muse-wonder`, `muse-spark-1.3`, `muse-spark-1.2`) routed to+metered gateways, although catalog generation `g-20261021063612126-17350b71` maps all four on+`meta/muse-code-subscription`.++Cause: `CATALOG_ROUTE_BINDINGS` had no `meta-muse-code` row. The row was parked with a comment+until models-index published the profile (backend task M1); the catalog published it and nothing+checked the trigger. The ticket's hypothesis — that claudish maps the route through the+route-level `consumerProviderId` — was wrong: no code in `packages/cli/src` reads that field.+The `export META_API_KEY` hint in the error came from `meta`'s `nativeModelPatterns: /^muse-/i`.++Guard added: `catalog-route-bindings.coverage.test.ts` fails when a published `kind: subscription`+profile is neither bound nor listed in `UNBOUND_SUBSCRIPTION_PROFILES`. 13 of 13 were bound on+that generation. It reads a pinned copy of the registry, so it catches a new profile when that+copy is refreshed.++## A host leak the binding opened++Once bound, `muse@` is built by the generic catalog wire handler, which honoured a user `baseUrl`.+A test caught it before release:+`Expected: "https://api.meta.ai/v1/responses" Received: "https://override.invalid/v1/responses"`.+The handler now pins `meta-muse-code` to `MUSE_CODE_BASE_URL`. The Muse login can be a copy of the+Muse CLI's Keychain token, so this is a credential boundary, not a preference.++## Live verification (built binary of this branch, 2026-10-22)++Run from the worktree as `bun packages/cli/dist/index.js`, never the installed claudish: the MCP+`preflight` tool runs the installed version and cannot test an unreleased fix.++- `--probe muse-wonder-full muse-spark-1.3 --json` (live): `muse-wonder-full` →+ `[email protected]`, tier subscription, `availability: serves`, `probe.state: live`,+ 81 tokens, first token 2157 ms, `signedWith: Muse Code borrowed copy · Muse Code Power Usage`,+ wire `https://api.meta.ai/v1/responses`. `muse-spark-1.3` first hop `[email protected]`, live.+- `--model muse-wonder-full -y -p "Reply with exactly the word PONG…"` → `PONG`, exit 0.++## A second defect only a live launch showed++The bare-name run above printed `Claude Code doesn't recognize the model "muse-wonder-full", so it+assumes a context window of 200k tokens.`; the same prompt with `--model [email protected]`+printed nothing. That one-variable comparison located it: `resolveLocalContextWindow` asked live+discovery for the canonical id while discovery lists the wire id (`muse-wonder-full-1.0`; `k3` for+bare `kimi-k3`). It now asks for the routed wire id; the slim catalog is still read by canonical id+first, because a wire id can equal another catalog row's id (`qwen3.8-flash-next` on OpenRouter is+`qwen/qwen3.8-flash`). Unit tests could not see this: routing was correct, and the window is+decided on the launch path by a different function. After the fix the bare-name run printed no+warning.++## contentUse coverage (generation g-20261021063612126-17350b71)++The catalog publishes `contentUse` (`standard` | `contributor`) on 735 of 1924 connections,+including all 11 `-contributor` connections (`meta/direct-api` 2, `meta/muse-code-subscription` 4,+`opencode/go-subscription` 3, `openrouter/gateway` 2). On every connection carrying the field, the+field and the suffix rule agree. Decision (Jack, 2026-10-22): the field decides; a connection+without it keeps the suffix rule. Remove the fallback when every connection carries the field.++## Open++MTL-200: a `team` reviewer slot forked a worktree-isolated subagent and drove it with+`SendMessage` and `TaskStop` before voting; `team status` showed a clean `COMPLETED`.