Muse Code Linux PUW — issue #231

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

Date: 2026-10-02 (America/New_York) Host: Linux workstation (`uname` below) Muse Code: 1.4.2 (`1.4.2-R4684.1`), authenticated for the Meta provider AIWG version: 2026.9.24 (`main` at `6ded8c849`); hook/MCP steps re-run on `chore/231-promote-muse-stable` Sandboxing: user-scope and MCP steps used a temporary `XDG_CONFIG_HOME`; the operator's real `~/.config/muse` was never written.

Environment

Linux 7.0.0-34-generic #34-Ubuntu SMP PREEMPT_DYNAMIC Wed Sep  2 14:29:37 UTC 2026 x86_64 GNU/Linux
node v24.12.0
Muse Code 1.4.2 (1.4.2-R4684.1)

Commands and exit codes

Step	Command	Exit
Fail-closed relative XDG	`XDG_CONFIG_HOME=relative/cfg the agent-owned use operation all the provider option muse the scope option user the dry-run option`	1 (blocked)
Fail-closed bare `~`	`XDG_CONFIG_HOME='~' the agent-owned use operation all the provider option muse the scope option user the dry-run option`	1 (blocked)
Project dry-run	the agent-owned use operation	0 (zero writes)
Project deploy	the agent-owned use operation	0 (26 skills in `.agents/skills`)
User dry-run	`XDG_CONFIG_HOME=<tmp> the agent-owned use operation all the provider option muse the scope option user the dry-run option`	0
User deploy	`XDG_CONFIG_HOME=<tmp> the agent-owned use operation all the provider option muse the scope option user`	0 (26 skills in `<tmp>/muse/skills`)
Doctor	the agent-owned doctor operation	0 (29 passed, 4 warnings, Muse native section)
Status	the agent-owned status operation	0
Steward	the agent-owned steward operation	0
Live smoke	`AIWG_MUSE_LIVE_SMOKE=1 npm run smoke:muse:live`	0 (`LIVE_CHECKS_PASSED`, 26/26 skills loaded)
Muse loads skills	`muse skills list the source option all the workspace option <project> the trust-workspace option the json option`	0 (26 project skills, 0 duplicate ids)
Trust gate, untrusted	`muse exec the provider option echo "…"`	0 (`rules.context_load … skipped_untrusted=1`)
Trust gate, trusted	`muse exec the provider option echo the trust-workspace option "…"`	0 (`project_sources=1`: `AGENTS.md` loaded)
Hooks + MCP via `use all`	`XDG_CONFIG_HOME=<tmp> the agent-owned use operation all the provider option muse the mcp option`	0 (`.muse/hooks.json` + sidecar; settings backup)
Doctor after hooks/MCP	the agent-owned doctor operation	0 (`Hooks: managed`, `MCP: configured`)
Hook runs in Muse	trusted `muse exec the provider option echo`	0 (`hook_run_terminal status=completed`)
MCP loads in Muse	same run	`mcp.config.resolve … server_count=1`; 16 `mcp__aiwg__*` tools
Export	`muse export the session option <session.jsonl> the redacted option the out option trajectory.json`	0
Import	the agent-owned sessions operation	0 (1 session, 57 events committed)
Discovery, no root	the agent-owned sessions operation	0 (muse `export-required`, 0 sources)
Discovery, the muse-root option	`… the muse-root option ~/.local/share/muse/sessions the json option`	0 (2 sources: exactly this project's sessions)

The echo provider exercises Muse's startup path (skills, rules, hooks, MCP) without a model call.

Model calls

The Muse API quota was exhausted on 2026-10-02 (resets 2026-10-05), so real-model `muse exec` runs come from the 2026-09-25 qualification on Muse Code 1.4.0 (#2726): the the json option envelope (`run.output.delta`, `run.terminal.<state>`), exit codes 0/1/2, the session-id option create-or-continue, `--` before a leading-dash prompt, and observed models `muse-spark-1.3` and `muse-spark-1.3-contributor`. The Ralph adapter drove real `muse exec` end to end in that run. Nothing in 1.4.2's `exec` help or export format changed relative to those findings.

Findings fixed in the promotion

• the agent-owned use operation never installed the managed hook or an opted-in the mcp option profile:

`use all` deploys in kernel-only mode, which the Muse writer treated as skills-only. Fixed in `tools/agents/providers/muse.mjs`; the hook and MCP rows above were re-run with the fix.

Containment checks

• No `.cursor/` trees in the project or the temporary XDG root.
• No `~/.muse` exists.
• User skills only under `<tmp>/muse/skills` (0 `SKILL.md` outside it); `<tmp>/muse` holds only `skills/`.
• `~/.agents/skills` pre-dates the PUW (2026-09-01, README only) and no entry in it changed during the run.

Platforms

• Linux: verified above.
• macOS and Windows/WSL: waived by the maintainer on 2026-10-02; Linux verification is sufficient for release.