New: muse hooks — local pre-commit gate mechanism (MVP)

New: muse hooks — local pre-commit gate mechanism (MVP)

Background

Muse has no commit-lifecycle hook mechanism at all today. Confirmed by reading the source directly: muse commit --help's own documented algorithm has no hook step, there is no .muse/hooks/ concept, no --no-verify-equivalent flag, and a full-codebase search for hook only turns up muse hub webhook (server-side, HTTP, for external CI notifications) and the unrelated git-bridge export hooks (core/bridge/hooks.py). Nothing lets a repo declare "run this check before every commit, and refuse the commit if it fails."

This gap surfaced concretely, twice, in the same session (2026-09-10/11):

1. musehub#182 — a stale, un-guarded proposal merge silently reverted 47 files. Root-caused and fixed, but the class of failure — a real invariant nobody automatically checks — is the same shape as #2 below.
2. This session's own docs drift — .museagent.md (the canonical, trackable agent-config source) had drifted from its compiled adapter outputs (CLAUDE.md, AGENTS.md) in two separate repos (~/ecosystem/.museagent.md and ~/ecosystem/musehub/.museagent.md). In ~/ecosystem/musehub, the drift meant CLAUDE.md's own @include pointed at a path (.muse/agent.md) that can never be valid — .muse/ is unconditionally excluded from every snapshot (muse/core/paths.py), so a file there can never be committed, pushed, or survive a clone. That means every Claude Code session rooted at ~/ecosystem/musehub, including this one, was silently getting zero content from its own project rules file — HARD RULEs on data loss included — until caught by hand today. muse agent-config status --json would have caught this instantly; nothing ever ran it automatically.

Both incidents are the same underlying pattern: a real, checkable invariant existed, a command already existed to check it, and nothing forced that command to run at the moment it mattered. The workspace's own philosophy (see musehub#185 — "five ambiguous muse resolutions were found and pruned … belt-and-suspenders even though X never touches canonical") already treats this class of problem as worth building real guardrails for, not just documenting a convention and hoping agents remember. This is a natural continuation of that same philosophy applied to local commit hygiene.

Worth implementing: yes. The cost of the MVP below is small (one new tracked config file format, three new subcommands, one integration point in muse commit), and it closes a gap that has already caused real, live drift twice.

Goal

A repo can declare pre-commit checks in a tracked, shareable file (not .muse/config.toml — that lives inside the unconditionally-excluded .muse/ directory and, per a related finding below, can never propagate via clone/push/pull either). A developer or agent explicitly activates those checks once per clone. Once active, muse commit runs them and refuses the commit on failure, with an explicit, visible, logged bypass — mirroring the --no-verify concept this workspace's own .museagent.md already assumes exists ("Never skip hooks (--no-verify) … unless explicitly asked").

Explicitly not a general git-hooks clone: no arbitrary shell scripts, no auto-execution on clone (a cloned repo that silently runs code at clone time is a supply-chain risk — git never does this either, hooks are always an opt-in local install step). MVP is a declarative list of commands for exactly one hook point: pre-commit. Other hook points and remote/server-side enforcement are explicitly out of scope (see below) — this ticket is about proving the mechanism end-to-end on the one hook point that would have caught both incidents above.

Related finding (context, not a phase of this ticket)

muse agent-config set --adapters ... (without --global) writes to <repo>/.muse/config.toml — which, like .muse/agent.md, can never be committed or shared via clone. Per-repo adapter selection is therefore silently machine-local only today, the same class of bug as the incident above, just in a different command. Worth its own follow-up ticket once this one lands (the fix is likely the same shape: move repo-level config that's meant to be shared out of .muse/ and into a tracked file).

Phases

Each phase must be fully green (tests passing, verified against a real scratch repo — not mocks, per this workspace's own testing-tier rules) before the next begins.

Phase 1 — Hook definition format

Add a new tracked file convention: .musehooks.toml at repo root (same naming family as .museignore/.museattributes/.museagent.md — discoverable, consistent, and crucially not inside .muse/).

[pre-commit]
commands = [
  "muse agent-config status --json",
]

Each command's exit code determines pass/fail (fully declarative — no shell scripts, no hashbang files, keeps the MVP's attack surface small).

Deliverables:

• .musehooks.toml parser + validator in muse/core/hooks.py — test HK_01
• Malformed/missing-file handling: absent file → no hooks defined (not an error); malformed TOML → clear parse error, never a silent no-op — test HK_02
• muse hooks list — prints configured hook points and commands from the tracked file (read-only, no local state) — test HK_03

Phase 2 — Local activation (muse hooks install)

Hooks are never auto-executed on clone or checkout. A human or agent must explicitly run muse hooks install once per clone to activate them locally. Installed state itself is legitimately local-machine state (whether this clone runs the hooks) — that's the one place .muse/ is the correct home, this is not a repeat of the earlier mistake since installed-state was never meant to be shared.

Deliverables:

• muse hooks install — reads .musehooks.toml, writes an "active" marker to .muse/hooks-installed.toml (local only) — test HK_10
• muse hooks uninstall — clears the local marker — test HK_11
• muse hooks status — reports both the tracked config and whether this clone has hooks installed, three distinct states (not_defined, defined_not_installed, installed) — test HK_12
• Idempotency: installing twice, uninstalling when never installed — both no-ops, not errors — test HK_13

Phase 3 — muse commit integration

Deliverables:

• muse commit runs all pre-commit commands (in order, stopping at the first failure) after step 4 (nothing-to-commit check) and before step 5 (snapshot/commit_id computation) of its documented algorithm, only when muse hooks status reports installed — test HK_20
• On failure: commit aborted before any object is written, exit code 1, JSON envelope names which command failed and its stderr — test HK_21
• muse commit --no-verify: explicit bypass, still logged (stdout/stderr notes hooks were skipped and why) — never silent — test HK_22
• defined_not_installed state: commit proceeds normally (no silent enforcement of an un-activated gate) but muse commit's JSON output includes a warnings entry noting hooks are defined but not installed — test HK_23 (this is the discoverability nudge for phase 4's scenario, proven at the point that actually matters: the moment someone commits)

Phase 4 — Discoverability at clone time

Deliverables:

• muse clone prints one line to stderr when the cloned repo has a tracked .musehooks.toml and hooks are (necessarily, since this is a fresh clone) not yet installed: This repo defines pre-commit hooks — run 'muse hooks install' to enable them. — test HK_30
• Same notice from muse status for repos that were cloned before this feature existed, or where install was skipped — test HK_31

Phase 5 — Dogfooding: close the actual incident

Deliverables:

• muse agent-config status gains --fail-if-out-of-sync (exit 1 if any configured adapter reports in_sync: false; exit 0 otherwise, including when no adapters are configured) — test HK_40
• Add .musehooks.toml to ~/ecosystem (workspace root) and ~/ecosystem/musehub with:

  [pre-commit]
  commands = ["muse agent-config status --fail-if-out-of-sync"]
  

• muse hooks install run in both repos as part of landing this ticket
• Acceptance test (HK_ACCEPT_01): reproduces this session's exact incident end-to-end against a real scratch repo — edit .museagent.md, do not run sync, attempt muse commit with hooks installed → commit is refused, message names muse agent-config status as the failing check. Then muse commit --no-verify on the same dirty state → commit succeeds, output notes the bypass.

Acceptance Criteria

• All HK_* tests pass against a real scratch repo (subprocess calls to the real muse binary — no mocks, matching this workspace's existing tests/test_backup.py/test_sandbox.py pattern from #185).
• HK_ACCEPT_01 demonstrates the mechanism would have caught today's real incident.
• No hook ever executes without an explicit, prior muse hooks install in that specific clone.
• --no-verify always produces visible output — no silent bypass path.

Out of Scope (this ticket)

• Additional hook points (post-commit, pre-push, post-checkout, etc.) — pre-commit alone is sufficient to prove the mechanism and close the incident that motivated it. Follow-up ticket once this lands.
• Server-side / CI enforcement (e.g., musehub rejecting a push whose HEAD commit would have failed a hook) — local-only for this MVP.
• Arbitrary shell scripts or hashbang hook files (git's model) — deliberately excluded to keep the attack surface to a declarative command list.
• Fixing muse agent-config set (non---global) writing to unshareable .muse/config.toml — noted above as a related but separate bug.

Mirrored from staging#192.