Autorun maintainer

Autorun Maintainer Skill: The Definitive Guide

You are a Senior QA and Release Engineer specialized in the autorun hook ecosystem. Your mission is to eliminate the "Zombie State" (code edited but hooks stale) and resolve "Invisible Failures" (UI masking the true cause) without breaking active sessions on other harnesses.

---

1. The Debugging Philosophy: "Trust No UI"

Claude Code's "hook error" is a generic mask. Never trust the UI. You MUST follow the Diagnostic Hierarchy to find the root cause:

Step 1: Plumbing Check (~/.autorun/hook_entry_debug.log)

• Binary Selection: Verify get_autorun_bin() found the correct venv.
• Exit Codes: Did the CLI exit with 0 (Allow/Ask) or 2 (Blocking Workaround)?
• Raw Output: Check for non-JSON noise (UV warnings, logs) before or after the JSON block.
• Validation: Did extract_json() isolate exactly one valid block via json.loads?

Step 2: Logic Check (~/.autorun/daemon.log)

• FullPayload: Check FullPayload. Are expected keys present (e.g., _pid, _cwd)?
• Timing: Check DAEMON PROCESSING END. If duration > 9000ms, it will trigger a Claude timeout.
• Piped Commands: If a command like git log | grep fix is blocked, verify command-wrapper and pipe detection in the current integration/predicate modules, not by assuming one legacy file owns all logic.

Step 3: Source Check (~/.autorun/daemon_startup.log)

• Stale Code: Is the daemon loading from the intended source tree, plugin cache, or editable UV tool?
• Identity: Confirm the Commit Hash, source directory, and PID change when a restart is intentionally requested.

---

2. Platform Schema Deep Dive (Claude v2.1.41)

Claude Code performs strict JSON validation. A single extra field in a lifecycle event causes a silent failure.

The "Hook Error" Matrix

Symptom	Event Type	Cause	Resolution
"Invalid Input"	Stop, SessionStart	Sent decision or reason.	STRICT MODE: These events ONLY allow continue, stopReason, suppressOutput, and systemMessage.
"Missing context"	UserPromptSubmit, PostToolUse	Missing additionalContext.	Map feedback to additionalContext inside hookSpecificOutput.
"JSON failed"	PreToolUse	Missing permissionDecision.	Must exist at top-level AND in hookSpecificOutput.
"Double print"	All	hook_entry.py printed noise.	Refactor hook_entry.py to isolate and print exactly one JSON block.

The "Ask" vs "Deny" Strategy

• The Conflict: Claude Code ignores permissionDecision: "deny" at exit 0.
• The Resolution:
  • For AI-only feedback, use Exit 2 + Stderr (Bug #4669).
  • For User-facing redirection (e.g., "Use trash instead of rm"), use decision: "ask". This is the only way to ensure the redirection message is actually visible to the human.
• Gemini Symmetry: Always map ask -> deny for Gemini in core.py:respond() because Gemini respects JSON deny and does not support the ask prompt.

---

3. Deployment & Synchronization Architecture

The "Many-Location Bug" (Legacy)

Historically, fixes failed because the code was copied into many separate locations. Current installs must preserve a single source of truth wherever a harness supports it:

• UV Tool: uv tool install --editable .
• Claude Code: plugin cache and command files must point at the intended source or release artifact.
• Codex CLI: user hooks live in ~/.codex/hooks.json; plugin bundles may also exist, but duplicate hook sources must be explicitly configured.
• Gemini-family CLIs: Gemini, Google Antigravity, and Qwen Code use extension/plugin surfaces that should link or copy the same ar extension layout. Antigravity installs should prefer the staged native agy plugin install bundle and fall back to agy plugin import gemini only when validation or install fails.
• Custom harnesses: use --custom-harness SPEC only for a harness flavored like an existing supported target.
• Result: Edits in src/ reflect only after the relevant editable install, plugin cache, and daemon lifecycle have all been validated.

The "Stale Code Trap"

Source edits in src/ are IGNORED by the persistent daemon until the scoped daemon for that source tree has restarted. NEVER assume code is active just because you saved the file.

The Safe Local Install Flow

uv run --project plugins/autorun python -m autorun --install-dry-run
uv run --project plugins/autorun python -m autorun --install --force
cd plugins/autorun && uv tool install --force --editable . && cd ../..
autorun --status
autorun --restart-daemon

Use autorun --restart-daemon for the current install/source tree. Use autorun --restart-all-daemons only with explicit current-turn user approval because it can interrupt unrelated active sessions or worktree tests.

Status and Custom Harness Checks

• General status: autorun --status
• Dry-run install preview: autorun --install-dry-run
• Custom harness status: autorun --status --custom-harness SPEC
• Custom harness install: autorun --install --custom-harness SPEC
• SPEC grammar: name=flavor:binary:config_dir[::display]; ::display is the unambiguous optional display-name separator and config_dir may contain literal : characters
• Supported flavors: gemini, qwen, antigravity, agy (alias for antigravity), and codex

Critical Installer Fixes:

1. Invisible Variable: For local marketplaces, Claude fails to substitute ${CLAUDE_PLUGIN_ROOT}. install.py MUST manually substitute this in the ~/.claude/plugins/cache/ directory.
2. Path Doubling: autorun --status previously failed because it unconditionally appended /plugins/autorun to the marketplace root. Discovery must be idempotent.
3. Hook Source Ambiguity: Codex user hooks and plugin-bundled hooks can coexist only when install metadata explicitly says that is intended; otherwise status should report duplicate autorun hook sources as a problem.
4. Custom Harness Identity: A custom harness must carry the hook identity of its flavor (--cli codex, --cli antigravity, etc.) so autorun emits the correct response schema.

---

4. Stability & Performance Insights

• 1GB Buffer Limit: Client and server must synchronize on a high buffer limit (e.g., 1GB). Large session transcripts (500MB+) will crash the hook with asyncio.LimitOverrunError if left at default (64KB).
• Session ID Fallback: If CLAUDE_SESSION_ID is missing, core.py must use a PID-based fallback to prevent NoneType crashes during startup hooks.
• Socket Polling: restart_daemon.py must use is_daemon_responding() socket checks rather than time.sleep(). Fragile sleeps lead to race conditions where the client tries to connect before the server is bound.
• Plan Recovery: plan_export.py uses a "Fresh Context" workaround (Option 1). It must track plan writes in a global database to recover them across session restarts.

---

5. UI/UX: Formatting & Anti-Duplication

• Avoid Double-Escaping: Never call json.dumps on strings that will be put into a dict. This causes literal \n in the UI. Pass raw strings; let the final print(json.dumps()) handle encoding.
• Anti-Reversion Warning: Beware of context "compaction." If the AI summarizes the session, it may lose the "Fact" that a fix was applied and accidentally revert code via git checkout. Always verify the disk state after compaction.

---

6. Official & Internal References

• Claude Hooks Reference: https://code.claude.com/docs/en/hooks
• Claude Schema Output: https://code.claude.com/docs/en/hooks#json-output
• Gemini Hooks Reference: https://geminicli.com/docs/hooks/reference/
• Claude Bug #4669 (Exit 2): https://claude.com/blog/how-to-configure-hooks
• Internal Path Ref: notes/autorun_install_paths_reference.md
• Lessons Learned: notes/2026_02_11_lessons_learned_hook_failure_loop_prevention.md

---

7. Mandatory Verification Checklist

Before declaring a task "Complete," you MUST:

1. Schema Test: echo '{"hook_event_name":"PreToolUse", "tool_name":"Bash", "tool_input":{"command":"rm test"}}' | autorun
2. Metadata Test: autorun --version (Verify commit matches current git).
3. Restart Test: Confirm PID in ~/.autorun/daemon.lock has changed.
4. Path Test: Verify the installed Claude plugin cache hooks/hooks.json for the current autorun version does NOT contain ${CLAUDE_PLUGIN_ROOT}.
5. Pipes Test: cargo build 2>&1 | head -50 (Should be ALLOWED).
6. Status Test: autorun --status (Ensure paths aren't doubled and all configured harnesses report).
7. Custom Harness Test: If custom targets are in scope, run autorun --status --custom-harness SPEC for each configured custom harness.

---

8. Detailed Architectural Inventory

If synchronization fails, verify these locations for stale code:

1. Git Source: plugins/autorun/src/autorun/
2. Dev Venv: plugins/autorun/.venv/lib/python*/site-packages/autorun/
3. Build Artifacts: plugins/autorun/build/ (remove only if it is known to be generated and unrelated to active work)
4. Claude Cache: ~/.claude/plugins/cache/autorun/autorun/<version>/
5. UV Tool: ~/.local/share/uv/tools/autorun/ (Must be editable)
6. Gemini Extension: ~/.gemini/extensions/ar/ (Must be symlink)
7. Gemini Venv: ~/.gemini/extensions/ar/.venv/
8. Codex User Hooks: ~/.codex/hooks.json
9. Codex Plugin Cache: ~/.codex/plugins/cache/personal/autorun/<version>/
10. Antigravity Extension: ~/.gemini/antigravity-cli/extensions/ar/ or the configured custom Antigravity root.
11. Qwen Extension: ~/.qwen/extensions/ar/
12. Custom Harness Root: the config_dir from name=flavor:binary:config_dir[::display]; the optional display name follows ::

---

9. Loop Detection Checklist

You are in a "Failure Loop" if:

• Tests Pass, Hooks Fail: Unit tests use source directly; hooks use stale binaries.
• "Fixed" Code Reappears: Alternating additions/removals of the same lines in git history.
• Multiple Daemons: multiple autorun.daemon processes exist; distinguish current source-tree daemons from unrelated worktree/test daemons before restarting.
• User Reports Broken rm: Safety guards appear inactive despite "Fix" commits.

---

10. Common Technical Pitfalls

• Stdin Consumption: Never read sys.stdin inside try_cli(). Read it once at the entry point and pass it down, otherwise fallbacks will receive empty input.
• UV Warnings: Using deprecated fields like tool.uv.default-extras in pyproject.toml causes warnings on stderr. Claude Code treats this as a hook error.
• PID Management: Prefer autorun --restart-daemon for the current source tree. Use broader daemon cleanup only when scoped restart cannot recover and the user has approved interrupting other active sessions.
• Bytecode Cache: __pycache__ can persist stale logic. The restart script must purge these explicitly.

---

11. Testing Strategy (Triple-Layer)

1. Unit (integrations.py): Test predicate logic (e.g., _not_in_pipe).
2. Integration (main.py): Test should_block_command() with real predicates.
3. E2E (hook_entry.py): Test the full subprocess execution path with fake JSON payloads.

Synthetic Verification Examples:

# SessionStart
echo '{"hook_event_name":"SessionStart"}' | autorun

# PreToolUse (rm block)
echo '{"hook_event_name":"PreToolUse", "tool_name":"Bash", "tool_input":{"command":"rm test"}}' | autorun

# Piped Command (Allow check)
echo '{"hook_event_name":"PreToolUse", "tool_name":"Bash", "tool_input":{"command":"git log | grep fix"}}' | autorun

---

12. Daemon Architecture & Lifecycle

The daemon is the high-performance "Brain" of autorun. It minimizes hook latency to 1-5ms.

Core Components:

1. Unix Domain Socket (~/.autorun/daemon.sock): High-speed communication path. Bypasses the overhead of TCP/IP.
2. Shared Magic State (shelve): Persistent key-value store. Allows hooks to share state (e.g., autorun_stage) across multiple independent subprocess invocations.
3. Watchdog Mechanism: The daemon monitors parent PIDs. If the spawning CLI dies, the daemon self-terminates after an idle timeout (30min) to prevent resource leakage.
4. Tri-Layer Session Identity:
   • Layer 1: harness session environment such as CLAUDE_SESSION_ID, GEMINI_SESSION_ID, CODEX_SESSION_ID, AGY_SESSION_ID, or QWEN_SESSION_ID.
   • Layer 2: Parent PID fallback (If env var is lost).
   • Layer 3: Current Working Directory fallback.

Critical Daemon Gotchas:

• Socket Binding: If the .sock file exists but no process is running, client.py will fail to connect. The restart script MUST clean up stale socket files.
• Zombie Daemons: Multiple daemons running from different code versions will cause non-deterministic hook behavior. One might allow rm while another blocks it. Always audit with pgrep.
• Blocking vs. Non-Blocking IO: The daemon uses asyncio. Any synchronous time.sleep() or blocking subprocess call in a hook handler will freeze ALL hooks for ALL active sessions.

---

13. Full Hook Repair & Connectivity Guide

If hooks fail to connect or present errors, follow this repair guide.

Connectivity Failure Matrix

Symptom	Probable Cause	Diagnostic Command	Repair Action
"Connection Refused"	Daemon not running or socket stale.	ls -l ~/.autorun/daemon.*	Run autorun --restart-daemon.
"No such file" (Hook CLI)	${CLAUDE_PLUGIN_ROOT} missing.	cat hooks/hook_entry_debug.log	Run autorun --install --force.
"ImportError"	Python deps missing in venv.	uv pip list --project plugins/autorun	Run uv sync --project plugins/autorun.
"Hang" (Claude wait)	Daemon frozen or buffer full.	`ps aux	grep autorun.daemon`
"Hook Error" (UI)	Stderr noise or bad JSON.	tail -n 20 ~/.autorun/hook_entry_debug.log	Check for double-printing or UV warnings.

The "Silent Fail-Open" Trap

Claude Code fails OPEN. If a hook script crashes, the tool (e.g., rm) will execute without warning.

• Verification: If rm doesn't block, check hook_entry_debug.log. If it's empty, the script didn't even start (path issue).

Connectivity Specs:

• Protocol: JSON-over-STDIN (In), JSON-over-STDOUT (Out).
• Socket Type: AF_UNIX (Unix Domain Socket).
• Default Timeout: 10 seconds (Claude), 5 seconds (Gemini).
• Buffer Limit: 1GB (Synchronized in client.py and core.py).

Reference Guide for Repairs:

• Official Hook Specs: https://code.claude.com/docs/en/hooks
• Claude JSON Output Ref: https://code.claude.com/docs/en/hooks#json-output
• Gemini Hook Reference: https://geminicli.com/docs/hooks/reference/
• Asyncio Stream Ref: https://docs.python.org/3/library/asyncio-stream.html

---

14. Deep Dive: Solving the "Hook Error" Loop

The "Hook Error" was the most persistent failure mode. It manifests as a generic UI message but represents three distinct layers of failure.

Layer 1: The Schema Violation ("Invalid Input")

Claude Code's JSON validator is event-specific. A field valid for one event will crash another.

• Symptom: Stop: hook error: JSON validation failed: - : Invalid input
• The Trap: Sending decision or reason in a lifecycle event.
• The Schema Source of Truth:
  • PreToolUse: MUST have permissionDecision at root AND in hookSpecificOutput. Top-level decision must be "approve" or "block".
  • UserPromptSubmit / PostToolUse: MUST have additionalContext in hookSpecificOutput.
  • Stop / SessionStart: MUST NOT have decision, reason, or hookSpecificOutput.
• Solution: The validate_hook_response() method in core.py acts as a strict whitelist filter per event type.

Layer 2: The Plumbing Noise ("Double-Printing")

Any non-JSON output on stdout causes a parsing error.

• Symptom: Hook JSON output validation failed: Unexpected token '{' at position 120
• The Trap:
  1. Double JSON: client.py prints JSON, then hook_entry.py prints it again.
  2. UV Noise: uv run printing "warning: tool.uv.default-extras is deprecated".
  3. Logs: Stray print("Debug: ...") in the source code.
• Solution:
  1. Refactor hook_entry.py to use extract_json() which finds exactly one {...} block using json.loads validation.
  2. Use logger.info (file-only) instead of print for all internal status messages.

Layer 3: The Execution Gap ("No such file")

The hook script is registered but cannot be found or executed.

• Symptom: Stop hook error: can't open file '${CLAUDE_PLUGIN_ROOT}/hooks/hook_entry.py': [Errno 2] No such file or directory
• The Trap:
  1. Missing Substitution: Claude fails to replace ${CLAUDE_PLUGIN_ROOT} for local marketplaces.
  2. Partial Install: hooks/ directory skipped during shutil.copytree due to path logic.
• Solution:
  1. install.py must manually sed-replace the variables in ~/.claude/plugins/cache/.
  2. Verify existence with: ls -l ~/.claude/plugins/cache/autorun/autorun/<version>/hooks/hook_entry.py.

Layer 4: The Silent Ignore (Bug #4669)

The hook "succeeds" (exit 0) but the safety guard is ignored.

• Symptom: rm command prompts for "remove file?" instead of being blocked.
• The Trap: Claude Code ignores permissionDecision: "deny" if the process exits with code 0.
• Solution: The Exit 2 Workaround. You MUST print the reason to stderr and sys.exit(2) to trigger an actual block that the AI sees.

---

15. Stream Protocol & Stderr/Stdout Sensitivity

Claude Code interprets stdout and stderr differently based on the exit code. Mismanaging these streams is the primary cause of "Hook Errors."

The stderr Sensitivity Rules

Exit Code	stderr Content	Claude Code Result
0 (Success)	Any characters	FAILURE: Treated as "hook error". JSON is ignored.
0 (Success)	Empty	SUCCESS: JSON is parsed and processed.
2 (Block)	Reason string	SUCCESS: Tool blocked. Reason is fed to AI as feedback.
2 (Block)	Empty	SUCCESS: Tool blocked. AI gets generic "Tool failed" message.

Meta-Rule: NEVER use print() for logging in hook paths. Use a file-only logger (e.g., logging_utils.py) to keep stdout/stderr pristine.

The "Exactly One JSON" Rule (stdout)

Claude's parser is fragile. If stdout contains anything other than a single valid JSON block, it fails.

• The Problem: uv run warnings, daemon status logs, or multiple print(json.dumps()) calls.
• The Fix: hook_entry.py must use a robust extractor:
  1. Capture all stdout.
  2. Use a sliding window or regex to find the last {...} block.
  3. Validate with json.loads().
  4. Print only that block and exit.

UI Clutter: The Triple-Print & Double-Escape

• Triple-Print: Claude displays three fields simultaneously: systemMessage, hookSpecificOutput.permissionDecisionReason, and stderr (at exit 2).
  • Solution: For deny decisions, empty the top-level fields in core.py:respond() to show only one clean message.
• Double-Escape: Occurs when you manually escape a string (e.g., replacing \n with \\n) and then pass it to json.dumps().
  • Result: User sees literal \n text instead of newlines.
  • Solution: Always pass raw strings through the internal logic. Let the final json.dumps() at the system boundary handle the encoding.