AGENTS.md@onboarding_assets/codex git:20260924.aacf66bA

Immutable. This exact content never changes and is served at /api/v1/blob/10bf6da6d27da50a.

## Shellbrain

Use Shellbrain for targeted recall during agent work.
Ask about remembered project purpose, decisions, constraints, or past attempts.
Use code search to check current implementation, commands, and paths.

Keep this tuple in mind:

`goal | surface | obstacle | hypothesis`

Emit one `SB:` line when the tuple changes or you hit a boundary state:

- goal, surface, obstacle, or hypothesis changed
- same approach failed twice
- an error repeats
- you are about to make an evidence-bearing decision
- you switch files, subsystems, or strategy
- you are closing out

If prior context may help:

`SB: recall | <goal> | <surface> | <obstacle> | <hypothesis-or-trigger>`

Then run:

```bash
shellbrain recall "<targeted natural-language question>"
```

Recall receives only this query, without your conversation or previous recall questions.
Name the subject and include the task, symptom, or decision you need context for.
Answers can be partial or outdated. Missing information does not prove a feature is absent.

If recall would not help:

`SB: skip | same signature | <one-line reason>`

Then continue. Do not call recall reflexively.

Use `shellbrain recall` for normal task context.

If you changed any files since your last user-facing response, run `shellbrain snapshot` exactly once after validation and immediately before your next user-facing response. Do this on every response cycle where files changed; skip only when no files changed.

Do not call `shellbrain read`, `shellbrain events`, `shellbrain memory`, `shellbrain concept`, or `shellbrain scenario`.

Use the installed `shellbrain` skill for the detailed recall workflow.