| name | scramble |
| description | Tactically assemble a short-term plan of many parallel options, then pursue as much local action as possible in the moment — rebases, protogen, PR pushes, tracker state updates, sqfan fleet dispatch. Not a survey (orient/sitrep) and not a single-threaded task: a multi-option sprint where the agent fans out everything it can do locally right now. Triggers on "scramble", "max delta-v", "30 minute sprint", "parallel push", "many options at once", "as much local action as possible", "unblock and ship", "velocity push". |
scramble
Time-boxed sprint (default 30 minutes): build an option board of everything that could move, execute the largest independent set in parallel, locally, then score what changed state. Not sitrep (one goal → A/B/C call), not orient (sweep → gestalt), not one task run end-to-end. Bias to doing: no reports, tracker comments, or planning docs unless asked. If an architectural decision must precede code, plan first — not a scramble.
Hard rules
- Options first — scan everything that can move before committing to any single thread.
- Local action wins — protogen, rebases, tests, pushes, tracker states, batch writes run on host; never wait on squire for mechanical work.
- Execute yourself — never end with "you should run …".
- Linear: state transitions only, no comments (unless overridden).
- Link PRs at every mention.
- Fan-out limit is independence — options sharing a dirty tree or branch serialize.
Workflow
1. Pin the window (30s). Time budget (default 30m), scope, metric = count of state transitions (merged, pushed, opened, dispatched, tracker state changed).
2. Option board (2–3 min). Tracker items in scope + real-world signals (PRs, branches, CI, blockers). One line per option: what, type, Local?, depends-on, estimate. Mark Local? honestly — remote-blocked options stay on the board but never enter the execution set.
