Add Muse harness support (skills + always-on rules) - #1211
Conversation
- install.sh: install catstack skills into ${MUSE_SKILLS_DIR:-$HOME/workspace/skills}
and merge always-on fragments into ~/AGENTS.md as managed blocks
- install_codex_agents_md.py: generalized to an AGENTS-style merger with
--agents-path (defaults to ~/.codex/AGENTS.md)
- link_skill_three_harnesses.sh: link project skills into Muse's skill dir
- check_skills_three_harnesses.py: four-harness parity (claude-only skills
stay out of cursor, codex, and muse)
- Docs: README, CONTRIBUTING, create-skill, always-on, cursor rule,
provenance, CLAUDE.core updated for four harnesses
- Tests: muse skill install, AGENTS.md managed blocks, content preservation,
no hooks/commands dirs, reflect-enforcement for the muse block
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_2b27afb2-72cf-4cd7-b410-9bc05aa43c59) |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_f4451ddd-2e56-49ad-a1c1-edaa7eb8fcd8) |
Muse has no hook pipeline, so the Claude-only cat-mode-default hook can never fire there: the installed skill sat inert behind disable-model-invocation with nothing to override it. install.sh now merges an equivalent one-line cat-mode-default fragment into the muse ~/AGENTS.md block when CATSTACK_CAT_MODE_DEFAULT=on (and removes it otherwise), mirroring the reflect-enforcement --fragment/--without pattern. Codex behavior intentionally unchanged. Tests: block present when on, removed when off, always-on blocks otherwise untouched, and no such block in Codex's AGENTS.md.
|
Queued — the merge queue status continues in this comment ↓. |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_86434298-2904-40c6-9bda-37cae60b1432) |
|
@Mergifyio queue |
Merge Queue Status
This pull request spent 44 minutes 11 seconds in the queue, including 43 minutes 48 seconds running CI. Required conditions to merge
|
1 participant
Summary
Muse now receives the same reusable agent packages and standing guidance as the other coding agents. A setup run places them in Muse’s workspace and refreshes its global instructions file.
The problem was that Muse had no target in the setup path. The helper, parity checker, and instruction merger all assumed the older agent set, so Muse could drift.
The fix adds Muse as a first-class target, sends global instruction blocks to the file Muse reads, and enforces the four-agent invariant automatically.
Muse has no hook pipeline, so the Claude-only cat-mode default path cannot run there. When that flag is on, setup writes the matching directive into Muse’s global instructions.
Review Claim
Muse receives the same reusable agent packages and standing guidance as the other agents, with four-agent parity enforced automatically.
Review Lane
behavior
Review Unit
engine-runtime
Safety Invariant
This slice narrows the branch to runtime setup files plus direct tests. Existing Claude, Cursor, and Codex setup paths still use their same roots, and the four-agent parity check passes. (Proposed without user confirmation — repair is running headlessly; see Assumptions.)
Slice Rationale
This keeps the runtime Muse support in one review unit and leaves broad handbook wording for a docs-only follow-up.
Non-goals
Test Plan
Test Plan
node engine/skills/draft-pr/scripts/validate-pr-body.mjs --body-file /tmp/pr-body.md --changed-files-file /tmp/pr-files.txtpython3 scripts/ci/check_skill_test_coverage.py --base origin/main --head HEADpython3 scripts/ci/check_skills_three_harnesses.pypython3 scripts/ci/check_ecosystem_boundaries.pypython3 -m unittest discover -s tests -p "test_skills_three_harnesses.py"python3 -m unittest discover -s tests -p "test_validate_pr_body_local.py"python3 -m unittest discover -s tests -p "test_reflect_enforcement_install.py"python3 -m unittest tests.test_install.TestSkillSymlinks.test_cursor_always_on_create_skill_rule_symlinked tests.test_install.TestSkillSymlinks.test_codex_agents_md_gets_create_skill_blockRevert Plan
Revert Plan
57ccac2a./install.shto drop Muse links and managed blocksAssumptions
Note
Medium Risk
Changes
./install.shbehavior for every rerun (new symlinks,~/AGENTS.mdmerges, link sweep) but leaves existing Claude/Cursor/Codex roots and Codex AGENTS behavior unchanged aside from shared installer refactors.Overview
Adds Muse as a fourth agent harness alongside Claude, Cursor, and Codex for portable skills and standing “always-on” guidance.
./install.shnow symlinks skills intoMUSE_SKILLS_DIR(default~/workspace/skills), sweeps stale catstack links there, and merges the same marked always-on fragments into~/AGENTS.mdvia a generalizedinstall_codex_agents_md.py(--agents-path). Reflect-enforcement and standard always-on blocks follow Codex’s args; whenCATSTACK_CAT_MODE_DEFAULT=on, a Muse-onlycat-mode-defaultfragment is injected into~/AGENTS.md(not into~/.codex/AGENTS.md), since Muse has no hook pipeline.Project-skill linking (
link_skill_three_harnesses.sh) and CI/home parity checking (check_skills_three_harnesses.py) now require four roots; policy text increate-skill, always-on fragments, and the Cursor rule was updated to “Claude, Cursor, Codex, and Muse.” The repo check no longer gatesCONTRIBUTING.mdon that phrase.drafter.config.jsontreatsalways-on/**andcursor/rules/**as engine-runtime for review classification.Reviewed by Cursor Bugbot for commit 58012e0. Bugbot is set up for automated code reviews on this repo. Configure here.