Repository navigation
Releases
v0.6.2
v0.6.1
Fixed
- The loopback adapter is unref'd, so one-shot CLI commands that load the plugin can exit instead of hanging on the listener.
- The manifest catalog no longer pins
api.meta.ai: the host resolves the provider base URL from the runtime entry, which points at the adapter, so inference keeps going through aliasing and normalization. reasoning.effort: "none"(thinking off) maps tominimalinstead of being rejected by Meta.- The adapter floors
max_output_tokensat 1024 for OpenClaw, so connection probes are not consumed by the reasoning budget and return text. - The static catalog lists
muse-wonder-1.0andmuse-wonder-1.0-contributor, so those models are selectable before live discovery.
Documented
- Multi-agent gateways need one
models auth login --provider muse-code --method device-code --agent <id>per agent.
Fixes #17.
v0.6.0
What changed
- Local loopback wire adapter in front of
api.meta.ai: it aliases the client tool names that collide with the 40-name Muse builtin catalog, restores them in the response, and normalizes the Responses parameters. - The provider entry and every model publish the adapter's
/v1origin as the base URL, so subscription requests keep working after Meta's 2026-10-29 client restriction. - If Meta still refuses the request, the adapter retries once with every tool aliased.
ADAPTER_CONTRACT.mdv3 pins the rules;MUSE_CODE_ADAPTER=offkeeps the proxy but disables the reshaping.
Install
npm install @thestreamcode/openclaw-muse-codev0.5.1: onboarding version requirement
0.5.1
- Document that the onboarding browser-login choice requires OpenClaw 2026.9.5 or newer.
- Older gateways can use
openclaw models auth login --provider muse-code --method device-code; the provider and the login command remain supported from 2026.7.1.
v0.5.0: Muse Code browser login
0.5.0
- Onboarding now runs the Meta device-code login and saves an OpenClaw auth profile; the API-key prompt is gone.
- The login honours cancellation, shows the URL when the browser cannot open, and never prints secret material.
- Login errors no longer echo endpoint response bodies.
- The existing credential cache is still written for older installations; an explicit
MUSE_CODE_SUB_TOKENis never overwritten. - The README states that Meta can reject subscription requests from clients other than Muse Code, and that signing in again does not remove that restriction.
v0.4.0: credential env sync and keyed catalog cache
0.4.0
- Keep MUSE_CODE_SUB_TOKEN aligned with the login cache without overwriting an explicit value.
- Key the live catalog cache by credential hash.
- Extract and test login argument parsing.
- Add the auth contract; rebuild dist.
v0.3.0: dynamic model discovery
Serve every model id returned by the live account catalog except the endpoint families verified as non-chat (sam-, muse-image-, muse-voice-*). The static baseline stays as the offline and pre-login fallback only. Supersedes #6.
v0.2.0
Live model discovery: the catalog is fetched from GET /models (60s TTL, muse-spark-* only), so new models appear automatically. Static baseline stays for pre-credential discovery and fallback. No gateway restart needed after login (key re-resolved per call).
v0.1.1
Security hardening: the credential cache no longer persists the Meta OAuth access token (apiKey/accountId/email only); credential-source resolution moved to network-free src/config.ts; README documents the credential story. Also fixes the login script dist import path. Re-run npm run login once to refresh pre-0.1.1 caches.
v0.1.0
Initial release: Muse Spark in OpenClaw billed to the Muse Code monthly subscription (Meta device login, no API key). Static catalog muse-spark-1.1/1.2/1.3 incl. contributor tiers, Responses API transport, dynamic resolver, one-time login script. Verified end-to-end against the live Meta Model API.