seil is a tool for managing lifecycle hooks across AI-native development workflows.

It lets you define setup, teardown, and post-edit automation once in .seil.yml and reuse it across integrations.

go install github.com/sushichan044/seil/cmd/seil@latest

mise install github:sushichan044/seil

Download the latest release binaries from Releases.

• One config file for AI agent lifecycle hooks.
• Human-readable output by default, JSON output when automation needs structured results
• post-edit hooks can be filtered by glob and skipped for .gitignored files

You can route file edit events through seil post-edit instead of embedding formatter or lint commands directly in Claude Code hooks.

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path' | xargs -I {} env AI_AGENT=claude seil post-edit {} || exit 2"
          }
        ]
      }
    ]
  }
}

This keeps Claude Code responsible for detecting edits, while .seil.yml owns what should happen after the edit.

You can also connect workspace lifecycle events to setup and teardown.

[wt]
basedir = ".git/wt"
hook = "seil setup"
deletehook = "seil teardown"

This works well when a worktree should run initialization on create and cleanup on delete.

Create .seil.yml in your repository:

setup:
  jobs:
    - name: tidy
      run: go mod tidy

post_edit:
  jobs:
    - name: go-format
      glob: "**/*.go"
      run: mise run lint && mise run fmt

teardown:
  jobs:
    - name: cleanup
      run: echo "done"

# Run setup hooks
seil setup

# Run post-edit hooks for a file
seil post-edit internal/config/config.go

# Run teardown hooks
seil teardown

Use --reporter json when another tool needs structured results.

seil --reporter json post-edit internal/config/config.go

The JSON result is grouped into failure, success, and skipped.

seil uses .seil.yml.

• If -c, --config <path> is provided, that file is used.
• Otherwise, seil searches from the current directory upward until the filesystem root.
• If .seil.yml is not found during auto-discovery, seil treats it as a no-op and exits successfully.
• To use a different filename such as seil.yml, pass it with -c, --config.

setup:
  jobs:
    - name: optional-name
      run: required shell command

post_edit:
  jobs:
    - name: optional-name
      glob: optional doublestar pattern
      run: required shell command

teardown:
  jobs:
    - name: optional-name
      run: required shell command

• run is executed through sh -c.
• Hook commands run with the directory containing .seil.yml as their working directory.
• name is optional. If omitted, seil derives a normalized name from run.
• post_edit.jobs[].glob uses doublestar matching such as **/*.go.
• post-edit receives the edited file path and makes it available to command templating.

• setup and teardown run all configured jobs and preserve result order from the config.
• post-edit skips jobs when the glob does not match the file path.
• post-edit also skips jobs when the file is matched by .gitignore.
• If no config file is found during auto-discovery, all commands return an empty result set and exit with status code 0.
• The default human-readable output includes grouped results, status, log path, and a short summary.
• If any hook fails, seil exits with status code 1 for the default and JSON reporters, and 2 for the Claude reporter.

• --reporter auto is the default.
• AI_AGENT overrides automatic agent detection when --reporter auto is used.
• seil can auto-detect claude, cursor, devin, replit, gemini, codex, auggie, opencode, kiro, goose, and pi.
• --reporter auto selects an agent-specific reporter when one exists.
• --reporter json always keeps the JSON reporter, even when an agent is detected.
• --reporter default forces the default human-readable reporter.
• --reporter claude forces the Claude reporter.
• Agents without a dedicated reporter fall back to the default human-readable output.

• Go 1.26+
• mise for project tasks

mise run fmt
mise run lint:fix
mise run test

MIT