Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
docs(meta): add the models-index tasks for a Muse Code route
Jack Rudenkof173620bf00f
diff --git a/ai-docs/reports/BACKEND_TASKS_meta_muse_code_route-20261016.md b/ai-docs/reports/BACKEND_TASKS_meta_muse_code_route-20261016.mdnew file mode 100644index 00000000..07696c2e--- /dev/null+++ b/ai-docs/reports/BACKEND_TASKS_meta_muse_code_route-20261016.md@@ -0,0 +1,127 @@+# Tasks for models-index: a route for the Muse Code subscription (claudish, 2026-10-16)++Read against generation `g-20261015133900604-fb6199b5` (registry 3.7.0), after models-index+PR #356 ("Meta Model API and Muse Code from Meta's own docs", 2.0.21).++## Why++claudish is adding the claudish provider `meta-muse-code` (provider shortcut `muse@`), which signs+requests with the Muse Code subscription credential, and a metered claudish provider `meta` for+`META_API_KEY` on route `meta`/`direct-api`. Jack's decision: the subscription's models come from+**catalog membership**, the same as `qwen-token-plan`. claudish never hardcodes the list.++Today the catalog cannot give that membership a route:++| Fact (live) | Value |+|---|---|+| plan `meta-muse-code` `routeStatus` | `unsupported`, `routeReason: no_reviewed_consumer_route`, no `route` |+| registry route `meta` profiles | `direct-api` only |+| connections on a Muse Code subscription route | none |+| plan inclusion `Muse Wonder Full` | its `externalModelId` is a display string; no model record `muse-wonder-full` exists |++`resolveSubscriptionRouting` only gates on a `supported` plan with a bound route, and+`gatherFromConnections` only gathers `mapped` connections. So with this generation `muse@` gets no+candidates for a bare name, and membership gates nothing.++**Reviewed consumer route.** The reason `unsupported` holds, that the subscription is "for use+with Muse Code only", is now a decision Jack has taken for claudish, as with Devin, Antigravity and+the Grok subscription. claudish reads the credential that `muse login` stores; it does not ask the+backend to hold one.++## Where the data comes from++Do not use the docs pages for this route. The subscription has its own model endpoint, measured on+2026-10-16 with a Power Usage account:++```+GET https://api.meta.ai/muse-code/models+Authorization: Bearer <Muse Code login token> # opaque "LLM|<digits>|<chars>", 49 chars+```++Each entry looks like this:++```jsonc+{ "id": "muse-wonder-full-1.0", "object": "model", "owned_by": "meta",+ "metadata": { "muse-code": {+ "name": "...", "family": "avocado", "release_date": "...", "is_hidden": false,+ "attachment": true, "reasoning": true, "tool_call": true,+ "modalities": { "input": ["text", "image"], "output": ["text"] },+ "limit": { "context": 1007997, "output": 128000 },+ "options": { "reasoningEffort": "high", "forceReasoning": true },+ "variants": { "minimal": {...}, "low": {...}, "medium": {...}, "high": {...}, "xhigh": {...}, "max": {...} },+ "background_agent_model": "muse-spark-1.3" } } }+```++The measured list, which is the membership as one Power Usage account sees it:++| wire id | hidden | context / output | efforts (`variants` keys) | input |+|---|---|---|---|---|+| `muse-wonder-full-1.0` | no | 1007997 / 128000 | minimal, low, medium, high, xhigh, max | text, image |+| `muse-wonder-full-1.0-contributor` | no | 1007997 / 128000 | minimal..max | text, image |+| `muse-wonder-1.0` | no | 1007997 / 128000 | minimal..max | text, image |+| `muse-wonder-1.0-contributor` | no | 1007997 / 128000 | minimal..max | text, image |+| `muse-spark-1.3` | no | 1007997 / 128000 | minimal..max | text, image |+| `muse-spark-1.3-contributor` | no | 1007997 / 128000 | minimal..max | text, image |+| `muse-spark-1.2` | no | 1007997 / 128000 | minimal..xhigh | text, image |+| `muse-spark-1.2-contributor` | no | 1007997 / 128000 | minimal..xhigh | text, image |+| `muse-image-1.0` | **yes** | — | — | text, image (not a chat model for this route) |++The cached copy muse keeps is `~/.local/share/muse/model-catalog/6d657461__p746268.json`. The+capture is in claudish `ai-docs/sessions/dev-feature-muse-code-*/research/muse-wire.md`, §4–§6.++## Tasks++**M1. Publish a subscription route profile.** For example `meta`/`muse-code-subscription` (the+name is yours, and claudish copies it verbatim into `CATALOG_ROUTE_BINDINGS`):+- `kind: subscription`, base URL `https://api.meta.ai/v1`.+- `apis`: `responses` only. It is what muse itself sends (`POST /v1/responses`), and it is the+ only stream that reports the subscription quota (a Meta-only `response.subscription_usage`+ event). `/v1/chat/completions` and `/v1/messages` also answer 200 to this token, but whether they+ count against the subscription or are billed per token is unverified, so do not publish them+ for this profile.+- Its own credential profile, distinct from `meta-api-key`. These are two silos, and+ `META_API_KEY` must never satisfy this route.++**M2. Flip plan `meta-muse-code` to `supported`** with that `route`, keeping+`modelDiscovery: "catalog"`.++**M3. Map one connection per member** on the new profile (`routeStatus: "mapped"`), with the+`externalModelId` exactly as `/muse-code/models` spells it:+- `muse-wonder-1.0` → `muse-wonder`+- `muse-spark-1.3` → `muse-spark-1.3`, and `muse-spark-1.2` → `muse-spark-1.2`+- `muse-wonder-full-1.0` → `muse-wonder-full` (see M4)+- Contributor ids (`*-contributor`) are second connections of the same model, as you already do+ on `direct-api`. Contributor mode means "content may be used for product improvement".+ claudish will not pick it unless the user asks, so please mark it in a way a reader can tell+ apart (for example a route variant or a flag), not only by the suffix.+- Leave out `is_hidden` entries (`muse-image-1.0`).++**M4. Create the model record `muse-wonder-full`.** Its wire id is `muse-wonder-full-1.0`, not+the display string `Muse Wonder Full`. It has no Model API id, so its only connection is the M3+subscription one. Meta lists it only on the Power Usage and Maximum Usage tiers, and the Power+account above does serve it.++**M5. Limits and effort per connection, from `/muse-code/models`, not from the docs:**+- context 1,007,997 and output 128,000 on this route. The `direct-api` records publish 1,048,576+ context. Both numbers can be true, but they belong to different routes, so put the subscription+ one on the connection or a route variant, not on the model.+- effort values = the `variants` keys. On the wire, `none` returns 400 (`reasoning_effort 'none'+ is not supported for model 'muse-spark-1.3'. Supported values: [minimal, low, medium, high,+ xhigh, max]`), and so does `ultra` (`unknown variant`). The default is `high`, with+ `forceReasoning: true`.+- `wire.api: responses`, with `max_output_tokens`, as on `direct-api`.++**M6. Membership varies by tier.** Wonder Full is Power and Maximum only. If one plan id keeps all+four tiers, publish the per-tier membership (or split the plans) so a lower tier does not get+`muse-wonder-full` as a candidate that answers 4xx.++## What claudish does meanwhile++- An explicit `muse@<wire id>` (for example `[email protected]`) works with no route:+ claudish passes the wire id through unchanged.+- claudish adds the binding row for the M1 profile name once you publish it. Until then a bare+ `muse-*` name routes to the metered `meta` provider (if `META_API_KEY` is set) and then to a+ gateway.++Please name the generation that carries M1–M5. claudish verifies it the same way as the+2026-09-21 delivery: live through its own reader, with no gating until that generation.