Brainerd Reflect

Persist durable repo memory from the current session and route to the correct harness-specific flow.

Detect the harness first

1. If the run exposes brainerd_current_session and brainerd_apply_changes, you are in the Pi branch. Stay inside the Pi tool flow.
2. If CODEX_THREAD_ID is present, you are in the Codex branch. Use the Codex wrapper.
3. If BRAINERD_CLAUDE_SESSION_ID or BRAINERD_CLAUDE_TRANSCRIPT_PATH is present, you are in the Claude branch. Use the Claude wrapper.
4. If none of those signals match, stop and report that harness detection failed. Do not guess.

Pi branch

1. Read the ambient brain context already injected into the run.
2. Open additional brain/ files only when needed.
3. Call brainerd_current_session.
4. Distill the smallest durable change that helps future sessions.
5. Prefer updating an existing principle; otherwise create one focused note.
6. Apply changes only through brainerd_apply_changes.
7. End with a visible section that starts exactly with Brainerd summary:.

Codex and Claude branches

1. Read brain/index.md and brain/principles.md.
2. Run the matching wrapper:

./scripts/brainerd-codex.sh current-session
./scripts/brainerd-claude.sh current-session

3. Distill the smallest durable change that helps future sessions.
4. Prefer updating an existing principle; otherwise create one focused note.
5. Write a small JSON payload to /tmp/brainerd-reflect.json with a changes array.
6. Apply changes only through the matching wrapper:

./scripts/brainerd-codex.sh apply-changes --input /tmp/brainerd-reflect.json
./scripts/brainerd-claude.sh apply-changes --input /tmp/brainerd-reflect.json

Guardrails

• Do not store secrets, one-off task state, or generic skill instructions.
• Do not hand-edit generated entrypoints like brain/index.md or brain/principles.md.
• If nothing durable happened, say so and stop.
• End with a short summary of what changed and why.