opencode-yaml-hooks

opencode-yaml-hooks is an OpenCode plugin that loads hook definitions from hooks.yaml and runs command, tool, or bash actions on session and tool lifecycle events.

Use it to run tests after edits, lint changed files, block risky commands before they run, or trigger local automation without another LLM step.

Install

Install from npm with Bun:

bun add opencode-yaml-hooks

Or install directly from this repo:

bun add "https://github.com/KristjanPikhof/OpenCode-Hooks.git"

Then register it in your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-yaml-hooks"]
}

OpenCode resolves the package by name, so the plugin entry stays opencode-yaml-hooks even when you install it from GitHub.

Quick start

Create one of:

• ~/.config/opencode/hook/hooks.yaml
• <project>/.opencode/hook/hooks.yaml

Then add a minimal hook:

hooks:
  - event: file.changed
    conditions: [matchesCodeFiles]
    actions:
      - bash: "npm test"

This runs npm test after a supported file mutation tool changes at least one tracked code file.

Current config locations

Hooks are merged from global and project locations.

Platform	Global config	Project config
macOS / Linux	~/.config/opencode/hook/hooks.yaml	<project>/.opencode/hook/hooks.yaml
Windows	~/.config/opencode/hook/hooks.yaml preferred, otherwise %APPDATA%/opencode/hook/hooks.yaml	<project>/.opencode/hook/hooks.yaml

Start with these defaults

Unless you need something more specific:

• prefer file.changed for file-oriented automation
• leave scope unset unless you need main or child
• leave runIn unset unless you need actions to execute in the root session
• treat tool.after.* and tool.after.<name> as advanced hooks for observability or non-file workflows

Explicit defaults in the current runtime:

• scope defaults to all
• runIn defaults to current
• conditions are optional
• bash timeout defaults to 60000 milliseconds

Schema overview

hooks:
  - event: <hook-event>
    action: <stop>             # optional, only for tool.before.* hooks
    scope: <all|main|child>   # optional, defaults to all
    runIn: <current|main>     # optional, defaults to current
    async: <boolean>          # optional, fire-and-forget execution
    conditions:               # optional
      - matchesCodeFiles
      - matchesAnyPath: src/**/*.ts
      - matchesAllPaths:
          - package.json
          - apps/*/package.json
    actions:                  # required, non-empty
      - command: <string>
      - command:
          name: <string>
          args: <string>
      - tool:
          name: <string>
          args: <object>
      - bash: <string>
      - bash:
          command: <string>
          timeout: <positive integer milliseconds>

Validation rules enforced by the runtime:

• hooks must exist and be an array
• each hook must be an object with a supported event
• scope, if present, must be all, main, or child
• runIn, if present, must be current or main
• action, if present, must be stop and is only supported on tool.before.* and tool.before.<name> hooks
• async, if present, must be a boolean; cannot be true on tool.before or session.idle events; async hooks must use only bash actions
• conditions, if present, must be an array of supported condition entries
• actions must be a non-empty array
• each action must define exactly one of command, tool, or bash

Supported events

Session events

Event	When it fires
session.created	When OpenCode creates a session
session.deleted	When OpenCode deletes a session
session.idle	When a session becomes idle
file.changed	After a supported mutation tool reports file changes

Tool events

Event	When it fires
tool.before.*	Before every tool execution
tool.before.<name>	Before a specific tool, such as tool.before.write
tool.after.*	Advanced: after every tool execution
tool.after.<name>	Advanced: after a specific tool, such as tool.after.edit

Tool hook order for a tool named write:

1. tool.before.*
2. tool.before.write
3. tool executes
4. file.changed if the tool changed tracked files
5. tool.after.*
6. tool.after.write

Public API versus advanced hooks

Preferred public API: file.changed

Use file.changed when your automation depends on changed files.

Why it is preferred:

• it only fires for supported mutation tools: write, edit, multiedit, patch, and apply_patch
• it includes files and structured changes metadata
• it avoids catch-all after-hook ambiguity
• it is the recommended path for linting, formatting, indexing, and atomic commit workflows

Advanced escape hatches: tool.after.* and tool.after.<name>

Keep using low-level tool hooks only when you need:

• observability for every tool call, including non-file tools
• tool-specific post-processing unrelated to changed files
• compatibility with workflows that truly depend on raw tool arguments instead of normalized file changes

Conditions

All configured conditions must pass for a hook to run.

Condition	Meaning
matchesCodeFiles	Run only when tracked modified files include at least one supported code extension
matchesAnyPath	Run only when at least one final changed file path matches one or more glob patterns
matchesAllPaths	Run only when every final changed file path matches at least one glob pattern

matchesCodeFiles is extension-based. Extensionless files such as Dockerfile do not currently count as code changes.

matchesAnyPath and matchesAllPaths only work on file.changed and session.idle. Both accept either a single string or a string array, and both fail when there are no changed files to evaluate.

Examples:

hooks:
  - event: file.changed
    conditions:
      - matchesAnyPath: src/**/*.ts
    actions:
      - bash: "npm run lint -- --fix"

  - event: session.idle
    scope: main
    conditions:
      - matchesAllPaths:
          - package.json
          - apps/*/package.json
    actions:
      - bash: "npm test"

Invalid usage example:

hooks:
  - event: tool.after.write
    conditions:
      - matchesAnyPath: src/**/*.ts # invalid: path conditions are file.changed/session.idle only
    actions:
      - bash: "echo nope"

Actions

Command action

Runs an OpenCode command in the same session, unless runIn: main redirects it to the root session.

actions:
  - command: simplify-changes
  - command:
      name: review-pr
      args: "main feature"

Tool action

Prompts the session to use a tool with specific arguments.

actions:
  - tool:
      name: bash
      args:
        command: "echo done"

Bash action

Runs a bash command directly without another LLM step.

actions:
  - bash: "npm run lint"
  - bash:
      command: "$OPENCODE_PROJECT_DIR/.opencode/hooks/init.sh"
      timeout: 30000

If timeout is omitted, bash actions use the runtime default of 60000 milliseconds.

OPENCODE_PROJECT_DIR remains the action cwd / project directory that triggered the hook.

Bash payloads

Every bash action receives:

• inherited process.env
• OPENCODE_PROJECT_DIR for the action cwd / project directory
• OPENCODE_SESSION_ID
• OPENCODE_GIT_COMMON_DIR when available
• JSON over stdin

Example file.changed payload:

{
  "session_id": "abc123",
  "event": "file.changed",
  "cwd": "/path/to/project",
  "files": ["src/index.ts", "src/renamed.ts"],
  "changes": [
    { "operation": "modify", "path": "src/index.ts" },
    { "operation": "rename", "fromPath": "src/old.ts", "toPath": "src/renamed.ts" }
  ],
  "tool_name": "apply_patch",
  "tool_args": {
    "patchText": "*** Begin Patch\\n...\\n*** End Patch"
  }
}

Blocking behavior

Only tool.before.* and tool.before.<name> hooks can block execution.

• a bash action that exits with 2 blocks the tool
• action: stop escalates a blocking pre-tool hook into a best-effort session.abort(...) for the active session
• tool.after.*, tool.after.<name>, file.changed, and session hooks do not block execution
• non-blocking failures are logged and later actions continue

Example:

hooks:
  - event: tool.before.bash
    action: stop
    actions:
      - bash: |
          payload=$(cat)
          cmd=$(printf '%s' "$payload" | jq -r '.tool_args.command // empty')

          case "$cmd" in
            "git push"|git\ push\ *)
              echo "Blocked and stopping session: git push is not allowed." >&2
              exit 2
              ;;
          esac

Execution behavior on this branch

• hooks for the same event run in declaration order
• global hooks load before project hooks
• the runtime reloads discovered hooks.yaml files at each hook entrypoint
• invalid reloads are rejected and the last known good config stays active
• session.idle clears tracked changes only after successful dispatch
• if idle dispatch fails, tracked changes are preserved for retry
• reentrant file.changed and tool.after.* dispatches are queued and replayed after the active dispatch finishes
• async: true hooks return immediately without blocking the tool pipeline; their actions run in the background as best-effort work
• async actions for the same event and source session are serialized to prevent overlapping executions; note that serialization is per source session, not per target — runIn: main hooks from different child sessions are not serialized against each other

Copy-paste examples

See examples/hooks.yaml for:

• main-session only examples
• child-to-main runIn: main routing
• recommended file.changed automation
• advanced tool.after.* observability
• conservative atomic commit wiring

See also

• docs/hooks-v2-reference.md for the current public config shape
• docs/comparison-with-claude-code-hooks.md for how this compares to Claude Code's hook system

Known limitations

• file tracking is limited to supported OpenCode mutation tools, not arbitrary filesystem changes
• matchesCodeFiles is extension-based and ignores extensionless code-like files
• tool hooks depend on actual emitted OpenCode tool names
• Windows discovery is supported, but bash actions still require a working shell runtime
• async: true is not allowed on tool.before.* or session.idle events; async hooks cannot block tool execution or idle dispatch
• async hooks must use only bash actions; command and tool actions have no timeout and can stall the queue
• async hook failures are logged but not retried; async execution is best-effort and not guaranteed to complete if the host process exits

Explicit non-goals for v1/v2 runtime scope

This package does not currently try to:

• define custom hook events beyond session, file, and tool lifecycle events
• provide config inheritance or override priority beyond global-then-project merging
• provide retries, scheduling, or concurrency controls per hook
• track arbitrary filesystem changes outside OpenCode mutation tools
• make command or tool actions blocking

Development

npm install
npm run build
npm test