Conversation
…prevention plan User came back at 13:21 UTC (15:21 CEST) reporting CODEC was STILL firing every 5min — 5 windows / 5 same Notes. Investigation found a SECOND leak source distinct from the reminders one in the original incident doc. Root cause #2: Step 3 AskUserQuestion test-fixture leak. When my test runs (test_ask_user.py, test_destructive_consent.py) executed today between 12:22 and 13:22 UTC, the `temp_askuser_paths` monkeypatch fixture did not stick in some test orderings (likely module-cache reentry on the full suite with worktree-aware path resolution). Result: 11 entries written to ~/.codec/pending_questions.json (7 pending + 4 timed_out) AND 11 type="question" notifications in ~/.codec/notifications.json. Dashboard PWA polls these every 8s and renders an inline answer panel for each → user saw "5 same windows" and "TestAgent is asking a question". Root cause #3: same window saw 24 skill_proposal_staged emits because test_mcp_all_tools.py iterates EVERY MCP-exposed skill including self_improve. self_improve's run_once() calls Qwen and writes a .md proposal per gap — explains the cascade. Cleanup performed at 13:21 UTC (already done, documented here): - pending_questions.json: 11 → 0 (backup preserved) - notifications.json: 11 type=question removed (179 → 168, backup preserved) - Quit auto-opened Notes / Reminders / TextEdit - Killed NotificationCenter to clear stuck banners - Updated user's runtime ~/.codec/skills/reminders.py to FIXED version (read-mode for "list reminders" — prevents future leaks creating real Apple Reminders if any test or LLM ever calls reminders again) Permanent prevention plan added (6 items): 1. THIS hotfix already blocks reminders/notes/tts_say/qr in tests ✅ 2. Tighten Step 3 fixture monkeypatch BEFORE merge 3. Add self_improve to SKIP_SKILLS 4. Stop using Apple Reminders for monitoring checkpoints (decide format after Step 3 lands, per user's existing instruction) 5. Optional CI/pre-commit gate: fail if any test writes to ~/.codec/* 6. Document test-isolation contract in AGENTS.md §10 What was NOT done (per user contract): - No PM2 restart - No killing Claude.app's codec_mcp.py children - _HTTP_BLOCKED untouched - Backups preserved for forensic record
…clipboard, self_improve)
Second incident escalation at 15:35 CEST. User came back showing screenshot
of THREE auto-opening Terminal windows displaying:
- codec_clipboard_aek3vj4c.txt (CODEC CLIPBOARD HISTORY ...)
- codec_clipboard_j5dehi5i.txt (same content, second copy)
- codec_memory_36ypw9hl.txt (CODEC MEMORY SEARCH: 'test' ...)
Investigation:
- /tmp/codec_memory_*.txt files: 10 of them, mtimes 14:01 → 15:27 CEST
matching every full pytest suite I ran during this conversation
- Each file is a memory_search results dump, opened via:
subprocess.Popen(["osascript","-e",
'tell app "Terminal" to do script "cat <tmp> && ... && read && rm"'])
- CANONICAL_PROMPTS["memory_search"] = "test" → mod.run("test") on every
test suite invocation → opens 1 fresh Terminal window per run
- Same pattern in skills/clipboard.py (codec_clipboard_*.txt)
- self_improve writes Qwen-drafted .md proposals to ~/.codec/skill_proposals/
on every test run — slow, expensive, audit-log noise
I ran the full pytest suite ~5 times today doing the Step 3 rebase + the
pre-merge audit. Each run queued 1 memory_search Terminal popup. macOS
delivered them slowly, hence the user perceiving them "popping up out of
nowhere" several minutes AFTER my last test run.
This commit:
- Adds memory_search, clipboard, self_improve to SKIP_SKILLS
- Result: 36 → 39 skipped, 27 → 24 fired
- Verified: NONE of the 24 remaining-fired skills open Terminal windows,
write temp files via osascript, or open browser tabs (grep clean)
Cleanup performed at 15:36 CEST:
- find /tmp -name "codec_*.txt" -delete (12 files removed)
- closed any Terminal windows still showing those files
- (no PM2 restart, no _HTTP_BLOCKED change)
This is the third hotfix layer for the same incident:
PR #6 (merged): reminders/notes/tts_say/qr_generator/generate_qr_code
Step 3 PR #5: ask_user/stuck (interactive blockers)
THIS: memory_search/clipboard/self_improve (Terminal popups)
Permanent fix already documented in
docs/INCIDENT-2026-05-01-spurious-skill-fires.md prevention plan item #5:
add a CI/pre-commit gate that fails if any test writes to ~/.codec/* OR
spawns a `Terminal "do script"` subprocess.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Second incident escalation. User reported (with screenshot) at 15:35 CEST that 3 Terminal windows auto-opened showing
codec_clipboard_*.txtandcodec_memory_*.txtfiles. Investigation traced the cascade to my own pytest runs during the Step 3 rebase + audit work earlier this session — each full suite run queued a Terminal popup viamemory_search.run("test")→osascript "tell Terminal to do script ...".Root cause
tests/test_mcp_all_tools.py::CANONICAL_PROMPTShad:"memory_search": "test"→mod.run("test")→ writes/tmp/codec_memory_<hash>.txtand opens it in a fresh Terminal window via osascriptSame pattern (Terminal popup via
subprocess.Popen(["osascript", ...])) in:clipboard.py— writes/tmp/codec_clipboard_<hash>.txtand opensself_improve.py— writes Qwen-drafted.mdproposals to~/.codec/skill_proposals/, slow + audit noiseI ran the full pytest suite ~5× during the Step 3 rebase + pre-merge audit work earlier in this conversation. Each run = 1 queued Terminal window. macOS delivered them slowly — hence the user's perception that windows "kept appearing out of nowhere" several minutes after my last test run.
Verification
grepover the 24 skills that WILL still fire after this commit confirms NONE of them:Terminal.*do script)/tmp/codec_*.txtfileswebbrowser)subprocess.Popen.*openDistribution after this PR: 39 skipped (was 36), 24 fired (was 27).
Cleanup performed
find /tmp -name "codec_*.txt" -delete(12 leftover files removed)_HTTP_BLOCKEDchangeOut of scope
docs/INCIDENT-2026-05-01-spurious-skill-fires.md) covers the analysis; permanent prevention plan item feat(askuser): pause-and-ask + stuck detection + step budget (Phase 1 Step 3) #5 (CI/pre-commit gate that fails if any test writes to~/.codec/*OR spawns Terminal popups) becomes higher priority after this third hotfix layer.Hotfix layering for the same incident
fcbef2f)reminders, notes, tts_say, generate_qr_code, qr_generatorask_user, stuckmemory_search, clipboard, self_improve🤖 Generated with Claude Code