Skip to content

agent-team: add opt-in human checkpoint after spec (state:spec-approved) to prevent drift chains #49

@verkyyi

Description

@verkyyi

Summary

Today the pipeline runs spec → plan → impl in one unbroken chain: the spec agent posts, then immediately dispatches planner, which immediately dispatches implementer. For sparse or ambiguous issue bodies, that chain compounds interpretation errors and burns a full round-trip (spec → plan → impl → review kickback) before the human can course-correct. The spec is also where misinterpretation is cheapest to fix — one comment, ~30 seconds of human attention.

Seen in the verkyyi/agentfolio#105 dogfood run: the spec was fine, so the auto-chain worked. But on ambiguous issues, a single-stage checkpoint would let the human adjust before planner/impl tokens are spent.

Proposal

Introduce a state:spec-approved label as the real handoff into planner, controlled by an opt-in:

  • Default (with human checkpoint): spec agent posts spec, adds state:spec-needs-approval, and stops. Human either (a) edits the spec inline in the comment and adds state:spec-approved → a new spec-approval-handler workflow dispatches planner, or (b) comments /agent-team go.
  • Opt-out (today's behavior): if the issue carries an agent-team:auto label when the spec agent finishes, skip the checkpoint and dispatch planner directly.

Acceptance

  • catalog/agent-team/spec-agent.md posts spec + state:spec-needs-approval label when agent-team:auto is absent, and skips direct dispatch.
  • New catalog/agent-team/spec-approval-handler.md (triggered by labeled: state:spec-approved) dispatches planner.
  • catalog/agent-team/README.md documents both flows and when to prefer each.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions