Meta Muse (Meta Model API) provider support — validated branch ready, PR creation restricted

[hongnoul]

Summary

Adds Meta's new Model API (Muse Spark family) as an OpenAI-compatible provider, following the same pattern as Celeris/Xiaomi MiMo.

• New META_MUSE_PROFILE / META_MUSE_LOGIN_PROVIDER catalog entries
  • base URL https://api.meta.ai/v1, key env META_MUSE_API_KEY, default model muse-spark-1.2
• jcode login meta-muse / --provider meta-muse with aliases meta, muse, muse-spark, meta-model-api, meta-ai
• Static muse-spark-1.2 / muse-spark-1.1 fallback so models are selectable immediately after login, before the live /models refresh
• 1,048,576-token context limit for the muse-spark family
• README provider list updated

Bonus fixes found during validation

• provider-doctor tool-smoke cap: the live tool-call probe capped output at 256 tokens; Muse Spark burns ~220 hidden reasoning tokens before emitting the tool call, so the completion ended at finish_reason: length with no tool_calls, failing a model that handles tools fine. Raised to 1,024 (still a positive multiple of 256, so Celeris-compatible).
• copilot auth test isolation: save_github_token_makes_future_loads_available fails on any machine with GITHUB_TOKEN/GH_TOKEN exported, because env tokens win over the persisted fixture in load_github_token(). The test now clears and restores those vars.

Tests

• bare_meta_muse_model_id_routes_to_meta_muse_profile: bare muse-spark-1.2 routes to the meta-muse profile from any active provider (mirrors the Celeris regression test)
• muse_spark_family_resolves_to_one_million_context
• All suites green on this branch: jcode-base 1253 passed, provider-core 121 passed, provider-metadata 14 passed, 0 failures

Verification

jcode provider-doctor meta-muse --tier full with a real Model API key: READY, every strict checkpoint passes:

• Credential loaded, live catalog (3 models), catalog hot reload, picker, model-switch route
• Non-streaming + streaming chat completions
• Tool-call parse / execution loop / tool-result followup / real jcode tool smoke

Offline tier also passes with the static-model fallback (no key).

[hongnoul]

Branch: https://github.com/hongnoul/jcode/tree/feat/meta-muse-provider (4 commits on current master, compare: master...hongnoul:jcode:feat/meta-muse-provider). PR creation from forks appears restricted on this repo (createPullRequest permission denied for READ-level users); happy to open the PR directly if access is granted, or feel free to pull the branch.

[hongnoul]

For future cherry-picking, the four commits are pinned under tag meta-muse-v1 on the fork:

git remote add hongnoul https://github.com/hongnoul/jcode
git fetch hongnoul feat/meta-muse-provider

# take the whole feature (recommended, they cherry-pick cleanly as a unit):
git cherry-pick 32d276c88 7a5ff6105 293025447 0c66c1ff3

Per-commit breakdown, in case only part is wanted:

• 32d276c88 — core provider support (catalog profile + login provider, static model fallback, 1M context limit, CLI --provider meta-muse) — the only commit strictly needed for Muse to work
• 7a5ff6105 — README provider list mention
• 293025447 — regression tests (bare-id routing, context limit) + copilot auth test env isolation fix (independently useful; that test fails for anyone with GITHUB_TOKEN exported)
• 0c66c1ff3 — provider-doctor tool-smoke cap 256 → 1,024 so reasoning models (Muse burns ~220 hidden reasoning tokens) don't false-fail the tool checkpoints

The only expected conflict against a moving master is scripts/code_size_budget.json (single-line bump for src/cli/provider_init.rs); re-run scripts/check_code_size_budget.py and take the new number.

[harrisonmichael809]

发版节奏稳不稳，跟不跟得上 ✨

[wellsashley00]

和现有脚手架集成顺畅吗 🚀🔥🧠

[1jehuang]

Thanks for implementing this in your fork. Adding a provider requires maintainer review of naming, credentials, model discovery, and test isolation. Please open a normal pull request from the fork so the implementation can be reviewed without importing an external branch directly.

--- — Jcode agent (automated triage), on behalf of @1jehuang