MVP 2: Single-model auto-spec loop

t-675.3·WorkTask·
·
·
·omni.hs
Parent:t-675·Created1 month ago·Updated1 month ago·pipeline runs →

Dependencies

Description

Edit

When the executor says NO-GO (from MVP 1), instead of bouncing to a human, feed the executor's questions to a large model (Opus) to automatically amend the task spec. Loop until the executor says GO or max 5 passes.

Flow: 1. Task enters the gate (from MVP 1) 2. Executor says NO-GO with questions Q1, Q2, Q3... 3. Send to Opus: 'Here is the task spec. The executor had these questions: [Q1, Q2, Q3]. Amend the spec to answer these questions. Do not answer in a side channel — incorporate the answers into the spec itself.' 4. Opus returns amended spec 5. Send amended spec back to executor gate 6. Repeat until GO or 5 passes

Implementation:

  • Add a spec-revision loop in the orchestrator between task claim and coder spawn
  • Store each revision of the spec (for debugging and metrics)
  • Circuit breaker at 5 passes — if still NO-GO, bounce to human with all accumulated context
  • The amended spec replaces the task description (or is appended as a 'refined spec' section)

Key design decisions:

  • Opus triages executor questions: skip style/preference questions (naming, formatting), only amend for correctness-affecting questions
  • Each pass appends to a revision log on the task (comment or structured metadata)
  • Track: passes-to-convergence, convergence rate, circuit-breaker hit rate

Acceptance criteria:

  • NO-GO tasks enter auto-spec loop without human intervention
  • Loop converges (executor says GO) for the majority of tasks
  • Circuit breaker fires at 5 passes and bounces to human
  • Each revision is logged and traceable
  • Spec amendments are in-place (the final spec is self-contained, not a chain of Q&A)

Timeline (0)

No activity yet.