Explore mode
Explore mode is the leap from "do this objective" to "find an objective worth doing, then do it". A read‑only scout surveys your repo and proposes small, verifiable objectives; you (Pro) or the scheduler (Max) pick what runs. Nothing ships without the Verification Gate.
Two modes
| Plan | Behaviour |
|---|---|
| Pro | Propose‑only — candidates appear in a queue; you click Run this |
| Max / Enterprise | Autonomous — the scheduler auto‑selects the top candidate, runs it, and re‑scouts when the queue empties |
Enable it in Repo settings → Features → Explore.
Working the queue
Proposals appear in two places:
- Repo settings drawer — per‑repo candidates after clicking Scout now
- Explore tab (sidebar) — all pending proposals across your repos in one view, with a badge count on the sidebar. Accept or dismiss from either place.
How scouting works
- Click Scout now in the repo settings drawer (or let the scheduler trigger it on Max).
- A scout agent does a read‑only survey of the repo (it never opens a PR).
- It writes a short list of candidate objectives, each with a category, a procedural focus, and a suggested verify command.
- Candidates are validated against hard guardrails, de‑duplicated, scored, and added to the queue.
Guardrails (enforced in code, not prompt)
- Allowed categories only: docs, lint, tests, deps, todo, refactor.
- No risky changes: no public API changes, no major dependency bumps, no file deletions, no schema/auth/payment/secret/config edits.
- Size caps: small by estimate (files & lines); the gate re‑checks real size.
- Must be verifiable: every candidate carries a verify command, or it's dropped.
- Per‑day cap: autonomous runs per repo are bounded so a Max repo never burns the whole budget.
Each candidate shows
- Category (docs, lint, tests, …)
- Title and rationale — why it's worth doing
- Suggested verify command — the gate it will run against
- Estimated size — files and lines touched
- Accept & run — applies the candidate's focus + verify and triggers a verified run
- Dismiss — removes it and records the rejection for taste learning
It learns your taste
Every accept (run) and reject is remembered. Future scouts receive that history as hints — they prefer the kinds of work you accept and stop re‑proposing things you've dismissed. Track per‑repo stats in the repo settings drawer: proposed · run · dismissed.
Standing objectives win
While a repo has an active objective, explore mode stands down — the human goal takes priority.