Codewhale
GitHub
Codewhale documentation

Find the guidance you need.

Start with the getting-started guide and installation, or go straight to vocabulary, modes, permissions, tools, providers, Fleet, hooks, MCP, and the Runtime API. These pages are the product documentation; each one cites the source documents in the repository for full detail.

Install CodewhaleBrowse source docs ↗

Hooks

Hooks attach your own commands to Codewhale's lifecycle: inject context before a message is submitted, enforce policy before a tool call, and audit turns or sub-agent activity. This page describes what currently ships; docs/rfcs/1364-hooks-lifecycle.md is the design RFC for this surface, and docs/CONFIGURATION.md carries the full configuration schema.

Hooks are configured under [[hooks.hooks]] entries in config.toml; run /hooks in the TUI to see every configured hook grouped by event — name, command preview, timeout, and condition — plus the global [hooks].enabled state.

message_submit (mutable)

Runs before a submitted message is added to history or sent to the model. The hook receives JSON on stdin; exit 0 with stdout JSON carrying a non-empty text field replaces the submitted text, and exit 2 blocks the submission before the turn starts. Multiple hooks run serially in config order, each receiving the previous hook's output. Hooks marked background = true are observer-only and cannot transform or block.

tool_call_before (decision)

Runs before each tool call executes. Beyond the exit-2 hard deny (which always wins), a foreground hook may print a JSON decision on stdout with exit 0: allow / deny / ask, plus updatedInput to rewrite the tool input and additionalContext appended to the tool result the model sees. When several hooks match, precedence is deny > ask > allow; tool_name conditions support * globs (mcp__* matches every MCP tool). Full Access does not open tool-approval prompts, so ask does not downgrade that posture.

turn_end (observer)

Fires after each model turn ends, once usage, cost, notifications, receipts, and queue-recovery state have settled. The stdin JSON carries fields such as status, duration_ms, usage, totals, and queued_message_count. Stdout is ignored and failures are warn-only — the hook cannot block input, mutate the transcript, or change the next queued follow-up.

subagent_spawn / subagent_complete (observer)

Observe sub-agent start and completion with bounded JSON metadata on stdin (agent_id, status, truncated prompt/result previews). Failures are warn-only and never block scheduling or change prompts or results; use the transcript handle returned by agent when full detail is needed.

Project-local hooks

Repositories can ship policy in <workspace>/.codewhale/hooks.toml. Because project hooks are executable shell configuration, Codewhale loads them only after the workspace is trusted through the trust prompt or a trust_level = "trusted" entry in user-owned config — session /trust on and legacy .deepseek/trusted markers do not enable project hooks by themselves. Once trusted, project hooks are appended after the global hooks from config.toml, so they run last and win updatedInput ties. A malformed trusted project file logs a warning and startup falls back to global hooks only.

Source documents: docs/rfcs/1364-hooks-lifecycle.md (design RFC), docs/CONFIGURATION.md (configuration schema) · Update docs-map.ts when changing.