Autoresearch

Invocation:

/skill:autoresearch <issue-or-target> goal="<outcome>" [ui=<registered-spec>] [max=<1..5>]

Read ../../../docs/procedures/autoresearch.md, then the regular ../plan/SKILL.md and ../exec/SKILL.md. Autoresearch controls their sequence; it does not replace their contracts.

1. Require a tracked issue, measurable goal, bounded target, exact functional proof, optional exact registered UI spec, and iteration cap (default 3, maximum 5). Resolve missing intent through the plan procedure and stop while any owner decision remains open.
2. Apply the plan procedure once. In a dedicated worktree, capture iteration 0 from the unchanged target and bind its evidence to the commit and tree.
3. Apply the exec procedure once as the sole writer. Internally repeat the canonical combined-review loop: select at most three ordered findings, fix, prove, and independently re-review the resulting revision. A selected fix spends one iteration even when proof remains red.
4. Stop early at success, stalled, or blocked-owner; otherwise stop after the declared cap as cap-exhausted. Produce one terminal owner handoff and never merge.

Complete only when every iteration is recorded, deterministic proof and reviewer dispositions target the same final revision, the terminal state is explicit, and the owner has a runnable artifact or exact validation path.