Epic: Automated Task Up-Speccing Pipeline

Goal

Replace manual task up-speccing with an automated multi-pass pipeline that iteratively refines task specifications until an executor model (Sonnet/Haiku) can implement them without clarifying questions.

Background

Current workflow: Ben up-specs tasks manually with Opus, files them, coder agent runs. Many coder failures trace back to underspecified tasks. This epic adds an automated refinement loop.

Architecture

After initial task creation: 1. Send spec to executor model with prompt: "Can you implement this without clarifying questions? If YES: GO. If NO: list every question." 2. If NO-GO: feed questions back to the spec-writing model (Opus/GPT-5.3) to amend the spec 3. Repeat until GO or max 5 passes (circuit breaker) 4. Key addition: executor must "restate the task in its own words" to catch confident misunderstanding

Milestones (see subtasks)

• MVP 0: Baseline metrics on current pipeline
• MVP 1: Executor gate (GO/NO-GO check before coder starts)
• MVP 2: Single-model auto-spec (questions feed back to Opus automatically)
• MVP 3: Diverse review (second large model as devil's advocate)

Key Metrics

• Task success rate (first-attempt pass rate)
• Revision count per task
• Time-to-done
• Passes-to-convergence (starting MVP 2)

Design Principles

• Pass count is emergent, not fixed. Convergence = executor says GO (zero questions).
• Model-as-judge > deterministic checklist for detecting ambiguity
• Large model triages executor questions (skip pedantic ones, amend for correctness)
• If using two large models, sequential roles (drafter + devil's advocate), not adversarial