Agentic AI Atlas

Page

docs/user-guide/features/hooks.md

repo-docs

Hooks: Extensible Lifecycle Events

docs/user-guide/features/hooks

wiki/docs/user-guide/features/hooks.md

[Docs](../index.md) › [Features](./index.md) › Hooks # Hooks: Extensible Lifecycle Events **Version:** 6.0.0 (v6) **Last Updated:** 2026-06-22 **Category:** Feature Guide --- ## In Plain English **Hooks are automatic triggers that run your custom code at specific moments.** Like setting up an automatic notification: "When the workflow finishes, send me a Slack message." That's a [hook](../reference/glossary.md). **Common examples:** - 📧 Send an email when a run completes - 📊 Log metrics to your dashboard when quality is scored - 🔔 Get a desktop notification when approval is needed **Do you need hooks as a beginner?** No - they're an advanced feature for customizing behavior. You can use Babysitter perfectly well without ever writing a hook. **One thing to know for v6:** how Babysitter *keeps a run going turn-to-turn* differs by harness. Claude Code uses a `Stop` hook; Gemini and antigravity use `AfterAgent`; openclaw uses a daemon `agent_end`; opencode uses `session.idle`; Hermes uses ACP. The **Hooks Adapter** normalizes all of these into one consistent contract so your processes don't care which harness they run on. --- ## On this page - [Overview](#overview) - [Hook Lifecycle Overview](#hook-lifecycle-overview) - [Available Hook Types](#available-hook-types) - [Creating Custom Hooks](#creating-custom-hooks) - [Hook Payloads and Environment Variables](#hook-payloads-and-environment-variables) - [Configuration in hooks.json](#configuration-in-hooksjson) - [Troubleshooting](#troubleshooting) - [Best Practices](#best-practices) --- ## Overview Hooks are scripts and handlers that execute at specific lifecycle points during Babysitter orchestration. They enable custom behavior for task execution, notifications, logging, metrics collection, and third-party integrations without modifying core SDK code. In v6 the hook system is unified by the **Hooks Adapter**: a canonical session store plus a merge engine that takes each harness's distinct hook/continuation model and presents one consistent contract to the runtime and to your custom hooks. The SDK lifecycle hooks (`on-run-start`, `on-iteration-start`, ...) are harness-agnostic and behave the same everywhere; what varies is the **per-harness continuation model** that drives the loop forward each turn (see the table below). The Hooks Adapter is part of the [Adapters](adapters.md) runtime. > **Do not assume the Claude `Stop`-hook model on other harnesses.** Claude's "one orchestration phase per turn, decided by a synchronous `Stop` hook" is specific to Claude Code. The per-harness table below is authoritative. ### Why Use Hooks - **Notifications**: Send Slack, email, or desktop alerts when runs complete or fail - **Metrics Collection**: Capture timing, quality scores, and execution data for dashboards - **Custom Orchestration**: Implement specialized task execution logic (e.g., native Node.js execution) - **Audit Logging**: Write events to external systems for compliance and debugging - **Integration**: Connect Babysitter to CI/CD pipelines, monitoring systems, or team tools - **Session Control**: Manage Claude Code session behavior for continuous orchestration loops --- ## Hook Lifecycle Overview The following diagram shows when each hook type fires during a Babysitter run. The SESSION LIFECYCLE block below illustrates the **Claude Code** continuation model (SessionStart + Stop); other harnesses substitute their own continuation mechanism (see the [Per-Harness Continuation Models](#per-harness-continuation-models) table). The RUN and GIT lifecycle hooks are harness-agnostic. ``` SESSION LIFECYCLE +==========================================================================+ | | | SessionStart -----> [Claude Code Session Active] -----> Stop | | | | | | v v | | Session setup, Decision: allow | | environment vars exit or continue | | | +==========================================================================+ RUN LIFECYCLE +==========================================================================+ | | | on-run-start -----> [Run Created] | | | | | v | | post-planning -----> [Plan Generated] | | | | | v | | +------------------------------------------------------------------+ | | | ORCHESTRATION LOOP | | | | | | | | on-iteration-start ---+ | | | | | | | | | | v | | | | | on-step-dispatch | | | | | | | | | | | v | | | | | on-task-start ------> [Task Executes] ------> on-task-complete | | | | | | | | | v | | | | on-score (if quality task) | | | | | | | | | v | | | | on-breakpoint (if breakpoint) | | | | | | | | | v | | | | on-iteration-end ---+ | | | | | | | | | | +-----------+ (loop continues) | | | +------------------------------------------------------------------+ | | | | | v | | +--- on-run-complete <----[Success] | | | | | +--- on-run-fail <--------[Failure] | | | +==========================================================================+ GIT LIFECYCLE +==========================================================================+ | | | pre-branch -----> [Branch Operation] -----> pre-commit | | | +==========================================================================+ ``` --- ## Available Hook Types ### SDK Lifecycle Hooks These hooks fire during the orchestration lifecycle managed by the Babysitter SDK. | Hook | Trigger | Purpose | |------|---------|---------| | `on-run-start` | Run creation | Initialize resources, set up monitoring | | `on-run-complete` | Successful completion | Cleanup, send success notifications | | `on-run-fail` | Run failure | Error alerts, debugging information | | `on-iteration-start` | Before each iteration | **Core orchestration logic**, effect execution | | `on-iteration-end` | After each iteration | Finalization, iteration logging | | `on-task-start` | Before task execution | Preparation, timing metrics | | `on-task-complete` | After task execution | Result processing, cleanup | | `on-breakpoint` | Breakpoint created | Notifications to reviewers | | `on-score` | Quality score computed | Metrics collection, dashboards | | `on-step-dispatch` | Step dispatch decision | Custom routing logic | | `post-planning` | After plan generation | Plan validation, notifications | | `pre-branch` | Before git branch operation | Branch naming, validation | | `pre-commit` | Before git commit | Linting, formatting, validation | ### Per-Harness Continuation Models The orchestration loop is *kept alive turn-to-turn* by a harness-specific mechanism. The Hooks Adapter normalizes these into one contract; you only need this table when integrating or debugging a specific harness. **These models are distinct - do not generalize one onto another.** | Harness | Continuation model | How the loop advances | |---------|--------------------|------------------------| | Claude Code | SessionStart + **Stop** | One orchestration phase per turn; the synchronous `Stop` hook decides block-vs-approve-exit, finishing only when the completion proof returns as `<promise>...</promise>` | | Codex | SessionStart + **UserPromptSubmit + Stop** | Plugin-level lifecycle hooks; the proxied-stop hook advances the loop each turn (hooks auto-detected via `./hooks/hooks.json`) | | Antigravity | SessionStart + **AfterAgent** (no Stop) | `AfterAgent` fires after every turn and returns `{decision: block, ..., systemMessage}` to inject the next iteration until `<promise>COMPLETION_PROOF</promise>` | | Gemini | SessionStart + **AfterAgent** (no Stop) | `session-start.sh` inits state; `after-agent.sh` returns `{decision: block, ...}` to inject the next iteration | | Cursor | SessionStart + **Stop** | The `Stop` hook emits `{followup_message: ...}` (Cursor-specific, **not** `{decision: block}`) to auto-continue while the run is in progress | | GitHub Copilot | plugin-bundle **SessionStart + SessionEnd + UserPromptSubmitted** | `sessionEnd` is the loop driver (checks completion / re-injects next step); cloud-agent path is driven by repo instructions/skills | | openclaw | daemon **session_start + before_prompt_build + agent_end + session_end** (no sync Stop) | `before_prompt_build` injects run state into each turn's prompt; `agent_end` fires async (fire-and-forget) and does not block the next turn | | opencode | **session.created + session.idle** (non-blocking) + shell.env + tool.execute.before/after | No blocking Stop hook; the agent runs the full loop within a single turn by calling `babysitter run:iterate` until completion | | Hermes | **ACP** (JSON-RPC over stdio) | The babysitter adapter speaks ACP; the SDK owns orchestration; no filesystem Stop hook | | omp (oh-my-pi) | thin-skill-alias: **session_start** proxied hook + `/skill:*` forwarding | SDK-owned loop; commands forward via `/skill:<name>` | | Pi | thin-skill-alias: **session_start + proxied stop** hook + `/skill:*` forwarding | Active-run detection lives in the SDK stop hook | | genty | genty extension API: **session_start** proxied hook | SDK-owned continuation via the `babysit` skill; no marketplace CLI | For each harness's full command surface and install steps, see the [Install Matrix](../harnesses/install-matrix.md), [Claude Code](../harnesses/claude-code.md), and [Codex](../harnesses/codex.md) pages. ### Host-Side Hook Management The host-side [`adapters` CLI](../reference/adapters-cli.md) manages unified agent hooks via `adapters hooks` (`discover`, `list`, `add`, `remove`, `set`, `handle`) - for example registering a cross-harness trace hook. This is distinct from the per-repo/per-user SDK lifecycle hooks below. --- ## Hook Discovery and Priority Hooks are discovered and executed in a specific priority order. All matching hooks in each location are executed. **Discovery Order (highest to lowest priority):** 1. **Per-repo hooks:** `.a5c/hooks/<hook-type>/*.sh` 2. **Per-user hooks:** `~/.config/babysitter/hooks/<hook-type>/*.sh` 3. **Plugin hooks:** `plugins/babysitter-unified/hooks/<hook-type>.sh` **Execution Order:** Within each location, hooks are executed in lexicographic (alphabetical) order by filename. ``` .a5c/hooks/on-run-complete/ 01-metrics.sh # Executes first 02-notify.sh # Executes second 99-cleanup.sh # Executes last ``` --- ## Hook Execution Model ### Input/Output Protocol | Channel | Purpose | Notes | |---------|---------|-------| | **stdin** | JSON payload input | Contains event-specific data | | **stdout** | JSON result output | Must be valid JSON (or empty) | | **stderr** | Logging output | Not captured, visible in console | ### Exit Codes | Exit Code | Meaning | |-----------|---------| | `0` | Success - hook completed normally | | Non-zero | Failure - logged but does not stop other hooks | **Important:** Hook failures are logged but do not stop the dispatcher from executing remaining hooks. --- ## Creating Custom Hooks ### Step 1: Create Hook Directory Choose the appropriate location for your hook: ```bash # Per-repo hook (version controlled, project-specific) mkdir -p .a5c/hooks/on-run-complete # Per-user hook (applies to all your projects) mkdir -p ~/.config/babysitter/hooks/on-run-complete ``` ### Step 2: Create Hook Script Create a shell script with the `.sh` extension: ```bash #!/bin/bash set -euo pipefail # Read JSON payload from stdin PAYLOAD=$(cat) # Parse payload using jq RUN_ID=$(echo "$PAYLOAD" | jq -r '.runId') STATUS=$(echo "$PAYLOAD" | jq -r '.status // "unknown"') # Log to stderr (visible in console) echo "[my-hook] Processing run: $RUN_ID" >&2 # Your custom logic here # ... # Return JSON result via stdout (must be valid JSON) echo '{"ok": true, "action": "processed"}' ``` ### Step 3: Make Executable ```bash chmod +x .a5c/hooks/on-run-complete/my-hook.sh ``` ### Step 4: Test Hook Test your hook manually by piping sample JSON: ```bash echo '{"runId": "run-test-123", "status": "completed"}' | \ .a5c/hooks/on-run-complete/my-hook.sh ``` --- ## Hook Payloads and Environment Variables ### Common Payload Fields Most SDK lifecycle hooks receive these fields: ```json { "runId": "run-20260125-143012", "timestamp": "2026-01-25T14:30:12.123Z" } ``` ### Hook-Specific Payloads #### on-iteration-start / on-iteration-end ```json { "runId": "run-20260125-143012", "iteration": 3, "timestamp": "2026-01-25T14:35:00.000Z" } ``` #### on-task-start / on-task-complete ```json { "runId": "run-20260125-143012", "effectId": "effect-01HJKMNPQR3STUVWXYZ", "taskId": "build", "kind": "node", "label": "Build project" } ``` #### on-breakpoint ```json { "runId": "run-20260125-143012", "question": "Approve the deployment?", "title": "Production Deployment", "context": { "runId": "run-20260125-143012", "files": [ {"path": "artifacts/plan.md", "format": "markdown"} ] } } ``` #### on-run-complete / on-run-fail ```json { "runId": "run-20260125-143012", "status": "completed", "duration": 45000 } ``` #### on-score ```json { "runId": "run-20260125-143012", "score": 85, "target": 90, "iteration": 2 } ``` ### Environment Variables These environment variables are available to hooks: | Variable | Description | |----------|-------------| | `HOOK_PAYLOAD` | The JSON payload (also available via stdin) | | `HOOK_TYPE` | The hook type being executed | | `REPO_ROOT` | Repository root directory | | `AGENT_SESSION_ID` | Harness-agnostic session identifier (supersedes the deprecated `CLAUDE_SESSION_ID` / `BABYSITTER_SESSION_ID`) | | `BABYSITTER_PLUGIN_ROOT` | Plugin installation directory (injected by the runtime; the legacy `CLAUDE_PLUGIN_ROOT` / `--plugin-root` are gone) | | `BABYSITTER_ENV_FILE` | Path to session environment file | --- ## Example Use Cases ### Example 1: Slack Notification on Run Complete **File:** `.a5c/hooks/on-run-complete/slack-notify.sh` ```bash #!/bin/bash set -euo pipefail PAYLOAD=$(cat) RUN_ID=$(echo "$PAYLOAD" | jq -r '.runId') STATUS=$(echo "$PAYLOAD" | jq -r '.status') DURATION=$(echo "$PAYLOAD" | jq -r '.duration') # Calculate duration in human-readable format DURATION_SEC=$((DURATION / 1000)) # Send to Slack webhook if [[ -n "${SLACK_WEBHOOK_URL:-}" ]]; then curl -s -X POST "$SLACK_WEBHOOK_URL" \ -H "Content-Type: application/json" \ -d "{ \"text\": \"Babysitter Run Complete\", \"attachments\": [{ \"color\": \"$([ \"$STATUS\" = \"completed\" ] && echo 'good' || echo 'danger')\", \"fields\": [ {\"title\": \"Run ID\", \"value\": \"$RUN_ID\", \"short\": true}, {\"title\": \"Status\", \"value\": \"$STATUS\", \"short\": true}, {\"title\": \"Duration\", \"value\": \"${DURATION_SEC}s\", \"short\": true} ] }] }" >&2 fi echo '{"ok": true}' ``` ### Example 2: Desktop Notification on Breakpoint **File:** `.a5c/hooks/on-breakpoint/desktop-notify.sh` ```bash #!/bin/bash set -euo pipefail PAYLOAD=$(cat) TITLE=$(echo "$PAYLOAD" | jq -r '.title // "Breakpoint"') QUESTION=$(echo "$PAYLOAD" | jq -r '.question') # macOS notification if command -v osascript &>/dev/null; then osascript -e "display notification \"$QUESTION\" with title \"$TITLE\" sound name \"Glass\"" fi # Linux notification if command -v notify-send &>/dev/null; then notify-send "$TITLE" "$QUESTION" --urgency=critical fi echo '{"ok": true}' ``` ### Example 3: Metrics Collection **File:** `.a5c/hooks/on-score/metrics-collector.sh` ```bash #!/bin/bash set -euo pipefail PAYLOAD=$(cat) RUN_ID=$(echo "$PAYLOAD" | jq -r '.runId') SCORE=$(echo "$PAYLOAD" | jq -r '.score') TARGET=$(echo "$PAYLOAD" | jq -r '.target') ITERATION=$(echo "$PAYLOAD" | jq -r '.iteration') # Log to metrics file METRICS_FILE="${HOME}/.babysitter/metrics.jsonl" mkdir -p "$(dirname "$METRICS_FILE")" jq -n --compact-output \ --arg runId "$RUN_ID" \ --argjson score "$SCORE" \ --argjson target "$TARGET" \ --argjson iteration "$ITERATION" \ --arg timestamp "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \ '{ timestamp: $timestamp, runId: $runId, score: $score, target: $target, iteration: $iteration, gap: ($target - $score) }' >> "$METRICS_FILE" echo "[metrics] Recorded score $SCORE/$TARGET for iteration $ITERATION" >&2 echo '{"ok": true}' ``` ### Example 4: Native Task Orchestration The plugin includes a `native-orchestrator.sh` hook that automatically executes Node.js tasks: **File:** generated harness-specific runtime bundle under `artifacts/generated-plugins/<target>/hooks/` This hook: 1. Queries run status via CLI 2. Identifies pending node tasks 3. Executes them externally (up to 3 in parallel) 4. Posts results back to the SDK ```bash # Key excerpt - executes node tasks (cd "$CWD_ABS" && node "$ENTRY_ABS" "${NODE_ARGS[@]}") >"$STDOUT_ABS" 2>"$STDERR_ABS" EXIT_CODE=$? if [ "$EXIT_CODE" -eq 0 ]; then "${CLI[@]}" task:post "$RUN_ID" "$EFFECT_ID" --status ok --value "$OUTPUT_REF" else "${CLI[@]}" task:post "$RUN_ID" "$EFFECT_ID" --status error --error - <<< '{"message":"Task failed"}' fi ``` ### Example 5: In-Session Loop Control (Stop Hook) The `babysitter-stop-hook.sh` implements continuous orchestration by intercepting exit attempts: ```bash # Returns JSON to block exit and continue loop jq -n \ --arg prompt "$PROMPT_TEXT" \ --arg msg "Babysitter iteration $NEXT_ITERATION | Continue orchestration" \ '{ "decision": "block", "reason": $prompt, "systemMessage": $msg }' ``` --- ## Hook Execution The SDK discovers per-repo and per-user runtime hooks directly. Harness entrypoints in the maintained plugin source live under `plugins/babysitter-unified/hooks/*.sh` and invoke `babysitter hook:run` for harness-specific lifecycle hooks such as `session-start` and `stop`. ### Example Dispatcher Output ``` [per-repo] Executing hooks from: .a5c/hooks/on-run-complete [per-repo] Running: 01-metrics.sh [per-repo] + 01-metrics.sh succeeded [per-user] Executing hooks from: /home/user/.config/babysitter/hooks/on-run-complete [per-user] Running: notify.sh [per-user] + notify.sh succeeded Hook execution summary: per-repo:01-metrics.sh:success per-user:notify.sh:success ``` --- ## Configuration in hooks.json The `hooks.json` file registers **Claude Code** hooks (SessionStart, Stop, PreToolUse, PostToolUse). Other harnesses register continuation hooks through their own manifests — see [Per-Harness Continuation Models](#per-harness-continuation-models) above. **Location:** generated from `plugins/babysitter-unified/plugin.json` ```json { "description": "Babysitter plugin hooks for orchestration loops", "hooks": { "SessionStart": [ { "hooks": [ { "type": "command", "command": "bash ${BABYSITTER_PLUGIN_ROOT}/hooks/session-start.sh" } ] } ], "Stop": [ { "hooks": [ { "type": "command", "command": "bash ${BABYSITTER_PLUGIN_ROOT}/hooks/stop.sh" } ] } ] } } ``` --- ## Troubleshooting ### Hook Not Executing **Symptom:** Custom hook is not being called. **Solutions:** 1. **Verify file is executable:** ```bash ls -la .a5c/hooks/on-run-complete/my-hook.sh # Should show: -rwxr-xr-x (x permission required) chmod +x .a5c/hooks/on-run-complete/my-hook.sh ``` 2. **Verify hook type directory name:** ```bash # Correct .a5c/hooks/on-run-complete/ # Wrong .a5c/hooks/on_run_complete/ # underscore instead of hyphen ``` 3. **Check for syntax errors:** ```bash bash -n .a5c/hooks/on-run-complete/my-hook.sh ``` ### Hook Failing Silently **Symptom:** Hook runs but produces no output or effect. **Solutions:** 1. **Add verbose logging to stderr:** ```bash echo "[my-hook] Starting..." >&2 echo "[my-hook] Payload: $PAYLOAD" >&2 ``` 2. **Check jq parsing:** ```bash # Test jq command echo '{"runId":"test"}' | jq -r '.runId' ``` 3. **Verify external services are accessible:** ```bash # Test Slack webhook curl -X POST "$SLACK_WEBHOOK_URL" -d '{"text":"test"}' ``` ### Hook Breaking JSON Output **Symptom:** Error messages about invalid JSON. **Solutions:** 1. **Ensure stdout only contains JSON:** ```bash # Wrong - prints to stdout echo "Processing..." echo '{"ok": true}' # Correct - logging to stderr echo "Processing..." >&2 echo '{"ok": true}' ``` 2. **Validate JSON output:** ```bash # Test your hook's output echo '{"runId":"test"}' | ./my-hook.sh | jq . ``` ### Stop Hook Not Blocking Exit **Symptom:** Claude Code exits instead of continuing the loop. **Solutions:** 1. **Verify state file exists:** ```bash ls -la ~/.a5c/state/ ``` 2. **Check stop hook output is valid JSON:** ```bash # Must include decision: "block" to prevent exit {"decision": "block", "reason": "...", "systemMessage": "..."} ``` 3. **Verify session ID is being passed:** ```bash babysitter session:whoami --json ``` --- ## Best Practices ### Do - **Log to stderr** - Keep stdout clean for JSON output - **Use `set -euo pipefail`** - Fail fast on errors - **Parse JSON with jq** - Robust JSON handling - **Make hooks idempotent** - Safe to run multiple times - **Use meaningful exit codes** - 0 for success, non-zero for failure - **Prefix log messages** - `[hook-name]` for easy identification ### Don't - **Don't block indefinitely** - Use timeouts for external calls - **Don't print non-JSON to stdout** - Breaks the output protocol - **Don't rely on working directory** - Use absolute paths - **Don't store secrets in scripts** - Use environment variables - **Don't skip error handling** - Validate inputs before processing --- ## Related Documentation - [Adapters](./adapters.md) - The Hooks Adapter and the harness-agnostic runtime - [Install Matrix](../harnesses/install-matrix.md) - Per-harness install and continuation models - [Adapters CLI Reference](../reference/adapters-cli.md) - Host-side `adapters hooks` management - [Configuration Reference](../reference/configuration.md) - Hook configuration options - [Glossary](../reference/glossary.md) - Hook terminology definitions - [Process Definitions](./process-definitions.md) - Using hooks in processes - [Breakpoints](./breakpoints.md) - on-breakpoint hook integration - [Run Resumption](./run-resumption.md) - How hooks interact with resumption - [Best Practices](./best-practices.md) - Patterns for workflow design and team collaboration --- ## Summary Hooks provide a powerful extension mechanism for customizing Babysitter behavior at every lifecycle stage. Use SDK lifecycle hooks for run orchestration, notifications, and metrics. Use Claude Code hooks for session management and continuous orchestration loops. Follow the input/output protocol (stdin JSON, stdout JSON, stderr logging) and ensure scripts are executable. Place hooks in per-repo, per-user, or plugin directories based on your needs. --- ## Next steps - **Next:** [Install Matrix](../harnesses/install-matrix.md) - **Related:** [Adapters](./adapters.md), [Breakpoints](./breakpoints.md), [Architecture Overview](./architecture-overview.md)

[]

Hooks: Extensible Lifecycle Events overview

Attributes

Outgoing edges

Incoming edges

Hooks: Extensible Lifecycle Events overview

Attributes

Outgoing edges

Incoming edges