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. Codex has two surfaces: the long-stable notification surface (notify + tui.notifications), and a lifecycle-hooks system (hooks.json; see developers.openai.com/codex/hooks) whose events, handler types, and trust rules change between releases and have had firing-reliability gaps. Both event sets grow, so look up the current list in each runtime's hooks reference instead of trusting a count. This skill owns the Codex hook facts as lookup steps in references/hook-templates.md §Codex. Prefer notify for anything that must be dependable, and confirm a Codex hook fires on your runtime before relying on it.

Quick Reference

Need	Event / Approach
enforce pre-tool policy or guardrails	PreToolUse command hook
fast runtime checks at session start	SessionStart or Setup
react to user prompt before Claude sees it	UserPromptSubmit
run checks after edits (non-blocking)	PostToolUse (async/background)
catch tool failures separately	PostToolUseFailure
check after a full batch of parallel tool calls	PostToolBatch
inject context into or coordinate subagents	SubagentStart / SubagentStop
transform what the user sees (not the transcript)	MessageDisplay
persist compact durable state before compaction	PreCompact
react after compaction completes	PostCompact
handle worktree setup and teardown	WorktreeCreate / WorktreeRemove
react to cd or watched file changes	CwdChanged / FileChanged
keep agent team from going idle	TeammateIdle
handle API errors at turn end	StopFailure
reload plugin hooks safely	atomic clear-then-register swap
add Codex callback behavior	notify external program plus tui.notifications
enforce budgets, iteration caps, stagnation, kill-switches	references/budget-and-loop-hooks.md

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:

Need	Use Instead
durable project memory or compaction content	../agents-memory/SKILL.md
MCP design or server integration	../agents-mcp/SKILL.md
subagent design or delegation boundaries	../agents-subagents/SKILL.md
multi-agent orchestration	../agents-swarm-orchestration/SKILL.md
permission modes and approval routing for coding agents	../ai-coding-agents-safety-envelope/SKILL.md

Capability Boundary

Capability	Claude Code	Codex
broad lifecycle hooks	yes (stable; look up the current event table)	yes via hooks.json (smaller event set, reliability gaps; look up the matcher table)
command or decision control	yes	yes (PreToolUse deny; Stop/SubagentStop block-and-continue); check which handler types execute and which are parsed but skipped
payload mutation on supported events	yes (updatedInput, updatedToolOutput)	check whether PreToolUse accepts updatedInput on your version
external callback program	yes	notify (dependable) + hooks.json hooks (verify-first; a trust review can silently skip a changed hook)
best use	policy, verification, hygiene	dependable: notify alerts/logging; experimental: lifecycle automation via hooks.json

Typical Scenarios

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

Scenario	Event(s)	Handler	Recipe
Block rm -rf, git push --force, git add . before they run	PreToolUse (matcher Bash)	command	references/hook-templates.md, patterns §8/§11
Auto-format and smoke-check after every edit, without slowing the agent	PostToolUse (matcher Edit|Write, async: true)	command	patterns §1
Require approval for production/destructive tools instead of hard-denying	PreToolUse → permissionDecision: ask, PermissionRequest	command	patterns §8
Inject repo state (branch, task id, top commands) at session start	SessionStart / Setup	command	references/runtime-preflight-hooks.md, patterns §6
Preserve critical state across context compaction	PreCompact (checkpoint) + SessionStart (restore)	command	hook-templates.md, patterns §5
Gate the next model call after a parallel tool burst	PostToolBatch	command/agent	hook-templates.md, patterns §1
Pass context into subagents and validate their output before they report back	SubagentStart (inject) + SubagentStop (block)	command	hook-templates.md, SKILL.md §Subagent coordination
Keep an agent team from going idle prematurely	TeammateIdle	command	Quick Reference
Per-worktree setup/teardown (caches, env files, indexes)	WorktreeCreate / WorktreeRemove	command	hook-templates.md, patterns §7
Ship audit events off-box to a SIEM/webhook immediately	PostToolUse / Notification	http	patterns §2
Semantic "are acceptance criteria met?" gate, not a syntactic one	Stop / SubagentStop	agent	patterns §3
Enforce a token/iteration/stagnation budget or kill-switch on an autonomous loop	Stop / PostToolBatch / UserPromptSubmit	command	references/budget-and-loop-hooks.md
Audit edits to hook/approval/sandbox policy itself	ConfigChange	command	hook-templates.md, patterns §4
Reload .envrc/local config when the directory changes	CwdChanged / FileChanged	command	patterns §6
Measure which skills actually trigger across sessions	PreToolUse (matcher Skill) + UserPromptExpansion (/skillname path)	command	patterns §15
Desktop alert / hand off "turn complete" to a local process (Codex)	notify + tui.notifications	external program	patterns §12

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.
6. Declare the timeout and error posture for the exact runtime, event, and handler type. In current Claude Code, an explicit PreToolUse deny or exit 2 blocks the tool call, while a timed-out command, http, or mcp_tool handler returns the call to normal permission evaluation; an Agent SDK callback timeout blocks it. A policy that must remain closed on handler failure needs a separate deny rule or enforcement in the execution substrate. Test timeout, malformed output, and duplicate delivery separately from the happy path.

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

Event Surface (Claude Code)

Source: code.claude.com/docs/en/hooks. The tables below group the events by purpose; they are a map, not the authoritative list. New events arrive between releases, so before wiring an event confirm it in the reference's event table and read its blocking column.

Two rules hold across the table:

• Context-only events cannot block. For SessionStart, SubagentStart, and any event the reference marks "no blocking", exit 2 only shows stderr to the user as a hook-error notice; Claude never sees it. Send anything the model must know as hookSpecificOutput.additionalContext on exit 0.
• Blocking stop hooks need a loop guard. Stop and SubagentStop payloads carry stop_hook_active; exit 0 without blocking when it is true. The hooks reference documents a cap on consecutive stop-hook continuations and the environment variable that raises it; a gate that needs its own limit should count in a session-keyed state file rather than lean on the runtime cap.

Session lifecycle

Event	Fires when	Can block?
SessionStart	session begins or resumes	no (context only; see rule above)
Setup	claude --init-only, or --init / --maintenance in -p mode; one-time preparation for CI or scripts (matcher init / maintenance)	no
SessionEnd	session terminates	no (runs under a short default timeout; look up the default, the maximum, and the override variable before putting cleanup work here)

Per-turn

Event	Fires when	Can block?
UserPromptSubmit	user submits prompt, before Claude sees it	yes
UserPromptExpansion	user-typed command expands into a prompt	yes
Stop	Claude finishes responding	yes (check stop_hook_active)
StopFailure	turn ends due to API error	no
MessageDisplay	assistant text is displayed (display-only, does not alter transcript)	no
PreModelSwitch	before the session switches model	yes
PostModelSwitch	after a model switch completes	no (context only)

Tool execution

Event	Fires when	Can block?
PreToolUse	before any tool call	yes
PermissionRequest	permission dialog appears	via hookSpecificOutput.decision only; exit 2 is not honored
PermissionDenied	tool denied by auto-mode classifier	no
PostToolUse	after tool call succeeds	no
PostToolUseFailure	after tool call fails	no
PostToolBatch	after full batch of parallel tool calls resolves	yes
Elicitation	MCP server requests user input	yes
ElicitationResult	user responds to MCP elicitation	yes

Subagent / team

Event	Fires when	Can block?
SubagentStart	subagent spawned	no (context only, via hookSpecificOutput.additionalContext)
SubagentStop	subagent finishes	yes (check stop_hook_active; read agent_transcript_path, not transcript_path)
TeammateIdle	agent-team teammate about to go idle	yes

Task

Event	Fires when	Can block?
TaskCreated	task being created via TaskCreate	yes
TaskCompleted	task being marked complete	yes

Config / filesystem

Event	Fires when	Can block?
ConfigChange	config file changes during session	yes (except policy_settings)
InstructionsLoaded	instruction files load (CLAUDE.md, includes, glob/path matches, on compact)	no (observability; exit code ignored)
CwdChanged	working directory changes (cd)	no
DirectoryAdded	a directory is added to the session workspace	no
FileChanged	watched file changes on disk	no
WorktreeCreate	worktree being created	yes (any non-zero exit); read the requested name from the payload
WorktreeRemove	worktree being removed	check the reference's blocking column before using it as a gate

Compaction / display

Event	Fires when	Can block?
PreCompact	before context compaction	yes
PostCompact	after compaction completes	no
Notification	Claude Code sends a notification	no

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

Type	Latency	Failure mode	Use when
command	usually lowest latency; measure	script bug, wrong exit code, missing binary	deterministic checks — the default choice
http	network round-trip	endpoint down, timeout, no local fallback	audit must leave the box immediately (SIEM, webhook)
mcp_tool	depends on server	server not connected, tool schema drift	policy lives on a connected MCP server already
prompt	one extra LLM call	non-determinism, model-dependent token cost	a single-turn semantic judgment is enough (no tool use needed)
agent (experimental)	usually higher latency and usage; measure	cost/latency creep if used on hot paths	multi-step semantic verification (e.g. "did this satisfy acceptance criteria")

Escalate down this list only when the simpler type cannot express the check; compare observed latency and usage before making a cost claim — 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 use restrictive precedence. Current Claude Code resolves deny before defer, ask, and allow. Still avoid overlapping hooks that disagree: narrow matchers so the effective rule is reviewable, and test the merged configuration rather than relying on one hook in isolation.
• Settings precedence (per code.claude.com/docs/en/settings; re-check there when a hook's scope matters): 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 (a skill's hooks stay registered for the rest of the session once the skill is invoked; a subagent's only while it runs). 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 does not inject plain stdout. Context reaches the subagent only through hookSpecificOutput.additionalContext in the JSON output; systemMessage is a top-level field shown to the user. SubagentStop can block the subagent from reporting back (exit 0 with decision: "block"); check stop_hook_active first and read agent_transcript_path, which is the subagent's transcript.

Codex

• use notify for external callbacks (long-stable, dependable)
• use tui.notifications for terminal notification policy
• for lifecycle automation, use hooks.json after running the lookup steps in references/hook-templates.md §Codex (events, handler types, async, updatedInput, managed policy and trust review, open firing bugs). Verify it fires on the target runtime; fall back to notify if dependability matters.
• before substituting Stop for end-of-session capture, check the Codex matcher table for a session-end event; Stop may fire per turn, so dedupe if you need once-per-session semantics.
• do not assume Claude-style lifecycle parity for any event you have not found in the Codex reference

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) — distributed through the claude-code-templates npm package. Do not run it as npx claude-code-templates@latest: a moving @latest executes whatever the registry serves at that moment with your user permissions. Either review one published version and pin it exactly (npx claude-code-templates@<reviewed-version> --hook monitoring/context-timeline), or vendor the hook script into .claude/hooks/ after reading it. 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 forked subagents (whether fork mode needs an opt-in depends on version and surface; see ../agents-subagents/SKILL.md §"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. Codex has two separate surfaces, notify/tui.notifications and the hooks.json lifecycle system, and neither matches Claude's event set. Look up the Codex matcher table and test on the actual runtime before deploying.

• Putting slow network calls, broad test suites, or repo-wide scans in always-on hooks. Resolution: Measure hook latency on your own runtime (no official budget is published) and move anything that visibly stalls the loop into the documented async: true path. A bare background job (&) loses output delivery and timeout handling. 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. Resolution: Claude Code runs all matching hooks in parallel and resolves conflicting decisions by the restrictive precedence in Execution Model above (deny first; check the hooks reference for the exact order on events with their own decision set). Still narrow matchers so hooks on the same event rarely disagree; a reviewed merged config is easier to reason about than arbitration.

• Writing a Stop or SubagentStop gate that blocks without checking stop_hook_active. Resolution: Exit 0 when stop_hook_active is true. Without the guard, a persistent failure condition forces extra turns until the runtime's block cap stops it.

• Reading a payload field by the name you expect rather than the name the reference documents. Resolution: Build dry-run payloads from the documented field names and test the input the guard must block. A wrong key reads as empty, so the guard prints nothing and exits 0, which looks like a pass. scripts/test_hook_templates.py does this for the templates.

• 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
• scripts/test_hook_templates.py — runs each template against doc-shaped payloads; run after editing a template
• data/sources.json

Related Skills

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

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

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.