Hook timeline

reflect does nothing on its own. Everything it does happens inside hooks the harness fires. This timeline replays one scripted session: pick a harness, scrub or press Play, and watch which lane lights up at each hook.

How to read it

Lane	Meaning	Costs model tokens?
Recall injection	The hook sends additionalContext into the model’s context before it acts.	Yes: the injected text is input tokens (see token economics).
Capture	Cheap state work: arm a watcher, write a mini-learning, update slots.	No.
Drain enqueue	Append the transcript to ~/.reflect/pending_reflections.jsonl after the $0 gate.	No.
Drain run	The detached reflect-drain-bg.sh process turns queued transcripts into learnings.	Yes: this is the only LLM spend in the write path.

Two rules fall out of the wiring, and the timeline makes both visible:

• Producers never reflect. Stop, PreCompact, SessionEnd and SubagentStop are plain shell commands. They can only enqueue. The queue holds one entry per session even though several hooks try to add it (the dedupe is by session id).
• The drain is spawned from SessionStart. The next session’s start hook launches it detached, so reflection never delays the session that earned the learning.

Harness differences

Wiring comes straight from each harness’s hook file. The counts include every event reflect registers, not just the ones in the scripted session.

Harness	Hook file	Events	Not wired
Claude Code	.claude-plugin/plugin.json	13	none
Codex	plugin/codex-hooks.json	10	Notification, PostToolUseFailure, SessionEnd
Copilot CLI	plugin/copilot-hooks.json	13 (camelCase names, agentStop for Stop, plus errorOccurred)	postCompact

Codex arms the failure watcher from PostToolUse because posttooluse_minilearning.py detects failed results itself. Without SessionEnd, Codex relies on Stop and PreCompact to enqueue, which is why the dedupe matters.

Where in the code

What	Path
Claude Code wiring	.claude-plugin/plugin.json
Codex wiring	plugin/codex-hooks.json
Copilot wiring	plugin/copilot-hooks.json
Hook scripts and the producer/consumer contract	plugin/hooks/README.md
Recall hooks (SessionStart, UserPromptSubmit)	plugin/skills/recall/hooks/
Enqueue and dedupe	plugin/hooks/hook_common.py (enqueue_reflection)
Drain consumer	plugin/hooks/reflect-drain-bg.sh

Note

The counts under “state” (3 learnings at start, 2 on prompt 1, and so on) are illustrative. The caps behind them are real: SessionStart injects at most 3 learnings; each prompt at most 3 new ones within 1500 characters (the header line is not counted).

For the full hook reference see the hooks reference; for what happens inside the drain see drain.