Agents Hooks

Claude Code Hooks + Codex Notifications

Use this skill when hook behavior is the main concern: Claude lifecycle hooks, runtime preflight, guardrails, async verification, worktree lifecycle checks, subagent coordination, or Codex notification callbacks.

Claude and Codex are not equivalent here. Claude has a broad, stable hook system (31 events, re-verified against official docs 2026-08-15). Codex has two surfaces: the long-stable notification surface (notify + tui.notifications), and a newer lifecycle-hooks system (hooks.json with events including SessionStart, UserPromptSubmit, SubagentStop, Stop, PreCompact/PostCompact — see developers.openai.com/codex/hooks). The Codex hooks system is less mature than Claude's and has known firing-reliability gaps (e.g. codex#17532: repo-local .codex/config.toml hooks may not fire in interactive sessions). Prefer notify for anything that must be dependable; treat Codex hooks.json as usable-but-verify and confirm it fires on your runtime before relying on it.

Quick Reference

When To Use This Skill

Use this skill when the task is:

• building Claude hook automation
• adding command guardrails or approval logic
• wiring verification or audit hooks
• managing hook-scoped repo hygiene
• configuring Codex notifications or callback programs

Route elsewhere when the main concern is:

Capability Boundary

Typical Scenarios

Real situations mapped to the smallest event + recipe that solves them. Pick the narrowest row that matches.

Workflow

1. Confirm whether the task is Claude hooks, Codex notifications, or mixed setup.
2. Choose the minimum event surface that satisfies the requirement.
3. Prefer deterministic command hooks for enforcement.
4. Keep synchronous hooks fast; move heavy work into async or background paths.
5. Validate event support and payload assumptions against current docs before final advice.

Validate and install checklist

# 1. Lint every hook script before deployment
shellcheck ~/.claude/hooks/*.sh

# 2. Dry-run a hook by piping a sample payload
echo '{"tool_name":"Bash","tool_input":{"command":"rm -rf /"}}' \
  | bash ~/.claude/hooks/preflight-guard.sh

# 3. Confirm hook files are executable
chmod +x ~/.claude/hooks/*.sh

# 4. Check audit log after a test session
cat /tmp/claude-hook-audit.log

# 5. Disable all hooks for emergency bypass
# Set "disableAllHooks": true in ~/.claude/settings.json

ASCII Flow

Hook request
  -> Identify runtime
     +-- Claude Code -> choose lifecycle event -> keep sync hook fast -> async heavy checks
     +-- Codex       -> configure notify/tui.notifications -> avoid lifecycle parity claims
  -> Validate payload, paths, and secrets boundary
  -> Test on the target runtime
  -> Document event, command, failure mode, and rollback path

Event Surface (Claude Code, verified 2026-08-15)

Source: code.claude.com/docs/en/hooks

Session lifecycle

Per-turn

Tool execution

Subagent / team

Task

Config / filesystem

Compaction / display

Hook Configuration Snippet

Minimal settings.json wiring (copy into ~/.claude/settings.json or .claude/settings.json):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/preflight-guard.sh"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/post-audit.sh",
            "async": true
          }
        ]
      }
    ],
    "PreCompact": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/pre-compact-state.sh"
          }
        ]
      }
    ]
  }
}

Matcher rules: empty string or "*" = match all; |-separated words = exact list (e.g. "Edit|Write"); any string with other characters = JavaScript regex (e.g. "mcp__memory__.*").

Hook types: command (shell script), http (POST to URL), mcp_tool (call MCP server tool), prompt (single-turn LLM), agent (spawns subagent, experimental).

Exit codes: 0 = success, parse stdout for JSON; 2 = hard block, stderr fed to Claude; other non-zero = non-blocking error logged.

Payload mutation (exit 0 with JSON hookSpecificOutput): PreToolUse can set permissionDecision to allow / deny / ask / defer; PostToolUse can replace the tool result with updatedToolOutput; MessageDisplay can rewrite on-screen text with displayContent (screen only — transcript and Claude's view are unchanged). Always set hookEventName in the output to the firing event.

Exec vs. shell form: add "args": [...] to avoid shell and support cross-platform use. Omit args to use sh -c with pipes and &&.

Disable all hooks: set "disableAllHooks": true in settings for emergency bypass.

Choosing A Hook Type

Escalate down this list only when the cheaper type can't express the check — see references/hook-patterns.md §3 for the command-vs-agent decision in practice.

Execution Model And Precedence

• Parallel, not sequential. When multiple registered hooks match the same event, Claude Code runs all of them in parallel. Do not assume one hook's output is visible to another, and do not rely on registration order to break ties. Identical command hooks are deduplicated by command string + args; identical http hooks by URL — near-duplicate hooks (different flags, same intent) are not deduplicated and will double-fire.
• Conflicting decisions are not spec'd as "deny wins." The docs do not officially guarantee a resolution order when one matching hook returns allow and another returns deny on the same event. Design as if any single deny should be treated as authoritative (fail-closed), and avoid registering two hooks with overlapping matchers that can disagree — narrow the matchers instead.
• Settings precedence (per code.claude.com/docs/en/settings, re-verified 2026-07-11): managed (org) policy > CLI flags > project .claude/settings.local.json > project .claude/settings.json > user ~/.claude/settings.json. This corrects an earlier version of this skill, which put .claude/settings.local.json last — it actually overrides both project and user settings, not the reverse. Two more hook-bearing scopes exist beyond these four: plugin hooks/hooks.json (active whenever the plugin is enabled) and skill/agent frontmatter (active only while that component is active). Hooks from every scope merge and run together rather than override each other — a broader-scoped hook does not silently replace a narrower one — so this ordering mainly governs disableAllHooks and single-value settings conflicts, not whether a given hook fires.
• allowManagedHooksOnly: an enterprise admin can set this in managed settings to block all user/project/plugin hooks except those bundled with plugins force-enabled via managed enabledPlugins. If a hook you registered mysteriously stops firing in a managed environment, check this first before debugging the hook script.
• Scoping a hook without a shell condition: tool-event hooks (PreToolUse etc.) accept an if field — a permission-rule string like "if": "Bash(git *)" — to narrow when a handler fires beyond what matcher alone expresses. Prefer this over duplicating the same logic inside the script.

Recommended Patterns

Claude

• SessionStart or Setup for runtime preflight
• UserPromptSubmit to intercept or enrich prompts before Claude processes them
• PreToolUse for narrow allow, deny, or ask guardrails
• PostToolUse for formatting and smoke checks; PostToolUseFailure for failure-specific handling
• PostToolBatch to gate the next model call after a parallel tool batch
• SubagentStart to inject context into spawned subagents; SubagentStop to validate their output
• PreCompact for terse state reinjection; PostCompact for post-compaction orientation
• CwdChanged to reload .envrc or local configs when directory changes
• ConfigChange for auditing hook-policy edits
• WorktreeCreate and WorktreeRemove for worktree hygiene
• clear and re-register plugin hooks atomically during reloads so stale handlers never coexist with new ones

Subagent coordination

Inject context into spawned subagents via SubagentStart and validate their output via SubagentStop:

{
  "hooks": {
    "SubagentStart": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/subagent-context.sh"
          }
        ]
      }
    ],
    "SubagentStop": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/.claude/hooks/subagent-validate.sh"
          }
        ]
      }
    ]
  }
}

SubagentStart command stdout is injected as systemMessage context for the subagent. SubagentStop can block the subagent from reporting back (exit 0 with decision: "block").

Codex

• use notify for external callbacks (long-stable, dependable)
• use tui.notifications for terminal notification policy
• for lifecycle automation, Codex now has hooks.json (events incl. SessionStart, UserPromptSubmit, SubagentStop, Stop, PreCompact/PostCompact; docs) — but it is newer and has firing-reliability gaps (codex#17532). Verify it fires on the target runtime before depending on it; fall back to notify if dependability matters.
• Codex has no Claude-style SessionEnd; use Stop for end-of-turn capture, and note Stop may fire per turn (dedupe if you need once-per-session semantics).
• do not assume full Claude-style lifecycle parity

Community Recipes

Third-party hooks worth knowing about, treat as community-sourced (verify provenance and review code before installing on production sessions):

• monitoring/context-timeline (aitmpl.com, via Daniel San, 2026-04-26 thread) — installs with npx claude-code-templates@latest --hook monitoring/context-timeline. Shows a live timeline of the main agent's context window plus every subagent running in parallel, including the context each subagent returns when it finishes. Useful when debugging multi-worker fan-out or after enabling CLAUDE_CODE_FORK_SUBAGENT=1 (see agents-subagents §"Forking Parent Context Into Subagents"). Treat as observability, not policy enforcement — and audit the hook source before adding to a session that touches secrets.

Security Rules

• treat stdin JSON as untrusted input
• validate fields before use
• prefer canonical path checks over filename regex
• quote shell variables and avoid eval
• keep secrets out of logs and checked-in config
• keep blocking hooks narrow and auditable
• run ShellCheck on non-trivial shell hooks

Known Traps

• Assuming Claude lifecycle events and Codex notification surfaces are interchangeable. Resolution: Check the Capability Boundary table above before wiring any hook. Claude has broad lifecycle events; Codex exposes only notify and tui.notifications. Test on the actual runtime before deploying.

• Putting slow network calls, broad test suites, or repo-wide scans in always-on hooks. Resolution: Move anything over ~200ms into an async or background path. Use PostToolUse with a background job (&) rather than blocking the tool call. Reserve synchronous hooks for fast, narrow checks.

• Mutating files or config in a hook without leaving a reviewable diff or audit trail. Resolution: Write mutations through the normal git-tracked file path. Log every mutation to an append-only audit file (e.g. /tmp/claude-hook-audit.log). See references/scenario-preflight-chain.md for a working example.

• Reloading plugin hooks non-atomically and leaving stale handlers active beside new ones. Resolution: Clear all handlers first, then register the new set. Never add new handlers before removing old ones. Use a lock file or atomic swap (write to a temp path, then mv) if the registration sequence can be interrupted.

• Trusting payload shape, cwd, or path values without canonicalization and boundary checks. Resolution: Always call realpath -m (or equivalent) on any path from the payload before using it. Validate that the resolved path is within the expected root before acting. Treat stdin JSON as untrusted input regardless of hook type.

• Assuming two hooks on the same event run in order, or that "deny" is guaranteed to beat "allow" when they disagree. Resolution: Claude Code runs all matching hooks in parallel with no documented tie-break rule. Narrow matchers so hooks on the same event cannot disagree, and treat any single deny as authoritative in your own hook logic rather than depending on runtime arbitration.

• A hook silently stops firing after it worked fine in dev, and the script itself looks correct. Resolution: Check settings precedence and allowManagedHooksOnly before debugging the script — an org-level managed policy can suppress user/project/plugin hooks entirely in a way that looks identical to a broken hook.

Anti-Patterns

• heavy synchronous test suites in every hook
• undocumented assumptions about Codex event parity
• regex-only path validation
• raw payload or environment logging without redaction
• silent dangerous rewrites that are not reviewable

Navigation

Resources

• references/hook-templates.md
• references/hook-patterns.md
• references/hook-security.md
• references/runtime-preflight-hooks.md
• references/scenario-preflight-chain.md — end-to-end runnable scenario: PreToolUse + PostToolUse + PreCompact composed
• references/budget-and-loop-hooks.md — budget, iteration cap, stagnation, kill-switch hooks for autonomous loops, triggered runs, and always-on bots
• assets/template-preflight-runtime-hook.sh
• data/sources.json

Related Skills

• agents-subagents
• ../agents-mcp/SKILL.md
• ../agents-memory/SKILL.md
• ../agents-skills/SKILL.md
• ../agents-swarm-orchestration/SKILL.md
• ../ai-coding-agents-permissions/SKILL.md
• ../ops-devops-platform/SKILL.md

Fact-Checking

• Known bugs, regressions, framework/compiler/runtime footguns, and version-specific crash or workaround guidance must be verified against current primary web sources before being treated as current fact.
• Verify current hook behavior, event support, JSON schema, and Codex notification capabilities against official docs before operational guidance.
• Prefer official Claude and OpenAI sources over community posts.
• If live verification is unavailable, mark hook-surface claims as unverified.

Learnings Loop

Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.