TODO: Add pre-agent fetch/pull and post-merge worktree cleanup

todo-worktree-staleness-prevention Todo

todo done

Project

Problem

Dev agent spawned with isolation: "worktree" created a worktree from stale local main. The worktree was missing ~260 lines of infrastructure (DORA, CNPG, Postgres) that had been merged to remote but not fetched locally. Result: agent's PR would have destroyed production resources. 40K+ tokens wasted. QA also failed to catch the scope mismatch.

Root Cause

Two SOP gaps with no enforcement:

No fetch/pull before worktree creation. The worktree-workflow SOP says "ALWAYS fetch + pull main before creating a worktree" but Claude Code's built-in isolation: "worktree" bypasses this — it branches from whatever HEAD is, with no fetch.
No worktree cleanup after merge. SOP says clean up after merge, but no hook enforces it. 10 stale worktrees accumulated in .claude/worktrees/, some weeks old.

Fix Options

Option A: PreToolUse hook on Agent tool

When isolation: "worktree" is detected, run git fetch forgejo && git pull forgejo main before the worktree is created. Problem: the PreToolUse hook can block but can't modify behavior — it can only pass/fail.

Option B: Betty Sue runs fetch/pull before every agent spawn

Add to agent-spawn-conventions: "Before spawning a dev agent, run git fetch <remote> && git pull <remote> main." Relies on discipline, not automation.

Option C: Post-merge cleanup hook

After a PR merge, auto-remove the associated worktree. Could be a PostToolUse hook on mcp__forgejo__merge_approved_pr.

Option D: Periodic cleanup

A session-start hook that removes worktrees whose branches have been merged to main.

Recommendation

Option B (immediate, low-effort) + Option D (automated safety net). Option A would be ideal but PreToolUse hooks can't inject commands before tool execution.

Incident

2026-03-06: PR #19 on pal-e-platform. Dev agent worktree branched from stale main missing DORA+CNPG+Postgres resources. Caught during tofu plan review before merge. No production impact.