Repository navigation
Confidence-gated dispatch: agent scores task tractability before starting #257
Description
Inspired by Devin 2.1's confidence scores (auto-scores incoming Linear/Jira issues, gates plan approval on confidence).
Before auto-dispatching (#165) or on demand, have a cheap model score a task's agent-suitability from the issue text: clear spec, testable, bounded blast radius. Low-confidence tasks get flagged for human refinement first (pairs with the triage issue); high-confidence ones can skip plan approval. Show the score and its reasoning on the Planboard card.
Activity
currentbits commentedon Aug 17, 2026
Deep-dive detail (Sentry Seer): fixability gate — Seer does not PR every issue; it scores fixability first and declines. The cheap classifier gate ('is this worth an agent run?') before burning tokens is the proven pattern.
currentbits commentedon Aug 18, 2026
Research note (GitHub, Jul 23): a shipped reference design for confidence-gated automation — self-rated confidence with a repo-level threshold, has:suggestions search, bulk accept/decline, rationale preserved as audit trail, and the candid disclaimer that approvals are UX, not a security boundary. Copy the triad (confidence + threshold + rationale) into the dispatch design.
currentbits commentedon Oct 9, 2026
Closing as not planned in the 2026-10-09 board cleanup (#1532). Speculative idea with no work and no tie to the current roadmap. Reopen if it becomes a priority.