coily dispatch should not mutate .claude/* during startup - poisons its own reap #122

New issue

Closed

opened 2026-05-28 00:49:59 +00:00 by coilysiren · 1 comment

coilysiren commented

2026-05-28 00:49:59 +00:00

Owner

Problem

coily dispatch startup mutates .claude/lockdown-deny.sh and .claude/settings.json (and sometimes .claude/scheduled_tasks.lock) inside the dispatch worktree as part of the lockdown wrapper. coily dispatch reap refuses to clean worktrees with modified or untracked files. Net result: every dispatch session becomes permanently un-reapable after it exits, because the dispatch itself dirties the tree.

Evidence

Six stale dispatch worktrees observed 2026-05-27, all with identical 2-file dirty state, all on branches that have already been merged to origin/main:

.dispatch-worktrees/agent-guard/issue-32 (closed)
.dispatch-worktrees/backend/issue-91 (+ D .claude/scheduled_tasks.lock)
.dispatch-worktrees/coily/issue-258
.dispatch-worktrees/repo-recall/issue-254
.dispatch-worktrees/session-lattice/issue-37
.dispatch-worktrees/voice-flow-learning-loop/issue-20 (merged HEAD message: closes #20)

All six show only:

 M .claude/lockdown-deny.sh
 M .claude/settings.json

The work shipped. The dirt is bookkeeping the dispatch itself wrote.

Fix

coily dispatch should not mutate .claude/* inside the worktree during startup. Kai's call 2026-05-27.

Implementation options to evaluate, in rough preference order:

Don't touch .claude/* at all in the dispatch worktree - read settings from the parent checkout, or generate per-session config under /tmp instead of the working tree.
Reset the .claude/* paths back to HEAD as the dispatch tears down (atexit handler in the wrapper).
As a last resort, teach reap to ignore mutations to a known-bookkeeping-paths allowlist - but that's papering over the root cause.

Out of scope

Cleanup of the existing seven stale worktrees. Once the systemic fix is in, git worktree remove --force (or a reap --force) clears them. The work in each is already merged to main; no salvage needed.
The orphaned .dispatch-worktrees/gauntlet/issue-39 directory, which is a different failure mode (worktree registration deleted, dir survived). Separate cleanup, rm -rf is safe.
The agentic-os/issue-149 worktree, which has a third dirty pattern (.claude/, agentic_os.egg-info/, uv.lock untracked). Likely a related but distinct gitignore-coverage issue.

Context

Discovered while trying to re-dispatch the seven stale worktrees per Kai's instinct, after noticing them in a coily dispatch interactive run for an unrelated issue (agentic-os-kai#364, the icon issue). Most issue refs no longer resolve on GitHub (issues disabled) or Forgejo (404), so re-dispatch was a dead end - which is what surfaced the cleanup question and the underlying bug.

**Problem** `coily dispatch` startup mutates `.claude/lockdown-deny.sh` and `.claude/settings.json` (and sometimes `.claude/scheduled_tasks.lock`) inside the dispatch worktree as part of the lockdown wrapper. `coily dispatch reap` refuses to clean worktrees with modified or untracked files. Net result: every dispatch session becomes permanently un-reapable after it exits, because the dispatch itself dirties the tree. **Evidence** Six stale dispatch worktrees observed 2026-05-27, all with identical 2-file dirty state, all on branches that have already been merged to `origin/main`: - `.dispatch-worktrees/agent-guard/issue-32` (closed) - `.dispatch-worktrees/backend/issue-91` (+ `D .claude/scheduled_tasks.lock`) - `.dispatch-worktrees/coily/issue-258` - `.dispatch-worktrees/repo-recall/issue-254` - `.dispatch-worktrees/session-lattice/issue-37` - `.dispatch-worktrees/voice-flow-learning-loop/issue-20` (merged HEAD message: `closes #20`) All six show only: ``` M .claude/lockdown-deny.sh M .claude/settings.json ``` The work shipped. The dirt is bookkeeping the dispatch itself wrote. **Fix** `coily dispatch` should not mutate `.claude/*` inside the worktree during startup. Kai's call 2026-05-27. Implementation options to evaluate, in rough preference order: 1. Don't touch `.claude/*` at all in the dispatch worktree - read settings from the parent checkout, or generate per-session config under `/tmp` instead of the working tree. 2. Reset the `.claude/*` paths back to HEAD as the dispatch tears down (atexit handler in the wrapper). 3. As a last resort, teach reap to ignore mutations to a known-bookkeeping-paths allowlist - but that's papering over the root cause. **Out of scope** - Cleanup of the existing seven stale worktrees. Once the systemic fix is in, `git worktree remove --force` (or a `reap --force`) clears them. The work in each is already merged to main; no salvage needed. - The orphaned `.dispatch-worktrees/gauntlet/issue-39` directory, which is a different failure mode (worktree registration deleted, dir survived). Separate cleanup, `rm -rf` is safe. - The `agentic-os/issue-149` worktree, which has a third dirty pattern (`.claude/`, `agentic_os.egg-info/`, `uv.lock` untracked). Likely a related but distinct gitignore-coverage issue. **Context** Discovered while trying to re-dispatch the seven stale worktrees per Kai's instinct, after noticing them in a `coily dispatch interactive` run for an unrelated issue (agentic-os-kai#364, the icon issue). Most issue refs no longer resolve on GitHub (issues disabled) or Forgejo (404), so re-dispatch was a dead end - which is what surfaced the cleanup question and the underlying bug.

coilysiren referenced this issue from coilyco-flight-deck/agentic-os

2026-05-28 00:57:46 +00:00

Dispatch worktree leaves untracked agentic_os.egg-info, uv.lock, .claude - settle gitignore / uv.lock policy #79

coilysiren referenced this issue

2026-05-28 00:59:59 +00:00

coily dispatch reap should handle orphaned worktree dirs - registration gone, dir survived #123

coilysiren commented

2026-05-30 05:43:13 +00:00

Author

Owner

Merged into #12 in the 2026-05-29 backlog burn-down. Same dispatch .claude mutation poisons own reap Reopen if it should stand alone.

coilysiren closed this issue