Author or wire a new hook emission in Nova's LangChain agent loop. Use when adding observability for a new tool, when promoting a logged audit event to a structured receipt, or when extending the agent's self-evaluation storage. Nova's hook surface today is `RunJournal` (`backend/packages/harness/deerflow/runtime/journal.py:38`) which implements `BaseCallbackHandler` (on_tool_start / on_tool_end / on_chain_start / on_chain_end); there is no separate `PreToolUse`/`PostToolUse` enum, so any new hook emission must go through `RunJournal.record_middleware` or a new `BaseCallbackHandler` subclass. Triggers on phrases like "add observability for X", "log when tool Y runs", "instrument Z", or before shipping any new middleware.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Author or wire a new hook emission in Nova's LangChain agent loop. Use when adding observability for a new tool, when promoting a logged audit event to a structured receipt, or when extending the agent's self-evaluation storage. Nova's hook surface today is `RunJournal` (`backend/packages/harness/deerflow/runtime/journal.py:38`) which implements `BaseCallbackHandler` (on_tool_start / on_tool_end / on_chain_start / on_chain_end); there is no separate `PreToolUse`/`PostToolUse` enum, so any new hook emission must go through `RunJournal.record_middleware` or a new `BaseCallbackHandler` subclass. Triggers on phrases like "add observability for X", "log when tool Y runs", "instrument Z", or before shipping any new middleware.
Identify the event class. Is it a tool.start / tool.end / chain.start / chain.end (use BaseCallbackHandler)? Or a domain event that needs its own handler (define a new BaseCallbackHandler subclass and wire it in worker.py:235)?
If a LangChain hook: extend RunJournal in backend/packages/harness/deerflow/runtime/journal.py. Add a method like _record_tool_event(tool_name, tool_call_id, payload) that writes a RunEvent to the store.
If a domain event: add a new BaseCallbackHandler subclass (see subagents/token_collector.py:16 for the shape). Register it in worker.py:235 alongside journal.
Wire SSE emission so the frontend can consume the event in real time. Add the event type to core/threads/hooks.ts:onCustomEvent (currently recognises task_running, llm_retry, task_progress, verify_result, llm_error).
Update the frontend consumer. Add the new event type to the switch in core/threads/hooks.ts:859-960; route the payload into a new tab in the Agent's Computer panel; if a MetricCounter is appropriate, add a pill in runtime-capabilities-bar.tsx.
Test contract
Unit:backend/tests/test_journal.py should grow a test that calls the new hook emission with a sample payload and asserts the right RunEvent is written.
Replay-golden: the SSE event shape contract (backend/tests/_replay_fixture.py:113) must include the new event in write_read_file.ultra.events.json if the agent graph exercises the new event path during the recorded scenario.
Frontend:frontend/tests/unit/core/threads/hooks-*.test.ts should add a parser test for the new event.
Anti-patterns (read before authoring)
❌ Don't add a new record_middleware call — the function exists but is dead. If you must log, log via RunJournal._record_* methods.
❌ Don't patch langgraph middleware event signatures — that's an upstream contract; instead, expose a new BaseCallbackHandler that observes the existing events.
❌ Don't add an asyncio.Queue listener per event type — use the existing RunEventStore so the event flow is replayable.
❌ Don't write to stdout / stderr in a hook — use logger.debug or push to RunEventStore. stdout breaks the test capture path.
