SDD fix-loop redesign: resume-based fix rounds, five-round breaker, controller adjudication, lifecycle restructure - #1998

Who is submitting this PR? (required)

What problem are you trying to solve?

Four problems with subagent-driven-development, reported by the maintainer from real sessions on frontier models:

1. Pathological review loops with no circuit breaker — implement, review, fix, review, review, fix — because the loop was literally "Repeat until approved" and every re-review was a fresh full review of the whole diff, so a nondeterministic reviewer surfaced new findings each round.
2. A three-way contradiction about who fixes findings: the process diagram and dispatch guidance said "dispatch fix subagents"; the Red Flags section said "Implementer (same subagent) fixes them"; implementer-prompt.md assumed re-engagement. Live baseline evidence below shows this produced coin-flip behavior.
3. Accreted structure: 13 top-level sections with guidance for one activity scattered across four of them ("Constructing Reviewer Prompts" held reviewer guidance, fix policy, final-review policy, and plan-conflict adjudication).
4. Red Flags bullet list where the other seven skills use the | Excuse | Reality | rationalization table.

What does this PR change?

Restructures SKILL.md in lifecycle order with a specified fix loop: rounds 1–3 resume the original implementer with the findings, rounds 4–5 dispatch a fresh implementer on a more capable model, and a five-round breaker triggers controller adjudication (park with a written ledger ruling, or STOP as BLOCKED when the finding is load-bearing). Re-reviews are scoped to the findings via a new re-review-prompt.md; the implementer/task-reviewer templates and the Codex reference are aligned with resume semantics; exact ledger line formats are specified; Red Flags becomes a rationalization table. Design spec and implementation plan (with a 33-row verbatim move map) are included.

Is this change appropriate for the core library?

Yes — it modifies an existing core skill only. No new dependencies, no third-party integrations, no domain-specific content.

What alternatives did you consider?

• Fresh fix subagents with full task context (portable, fresh eyes) and keeping dedicated narrow fixers (cheapest) — rejected for resume-the-implementer because the implementer already holds task context and ownership; narrow fixers rebuilding context per finding is plausibly part of why loops didn't converge. Harnesses without agent resume get a specified fallback (fresh dispatch carrying brief + report file + findings).
• Full re-reviews with only a hard cap — rejected: rounds stay nondeterministic and you hit the cap with open findings more often. Scoped re-reviews make the loop structurally convergent; the final whole-branch review remains the broad safety net.
• Human checkpoint at the breaker — rejected by the maintainer: SDD's point is autonomous execution. The breaker routes churn to the ledger and genuine plan defects to the BLOCKED stop that has always existed.
• Targeted surgery instead of lifecycle restructure — rejected: the mixed-concerns complaint was the root issue; the restructure moves eval-tuned sentences verbatim (enforced by the move map) rather than rewording them.

Does this PR contain multiple unrelated changes?

No. Every file serves the fix-loop redesign: the skill, its three templates, the one-sentence Codex reference alignment (implementer close timing must permit resume), and the spec/plan documents for this change.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: fix(sdd): plan-scoped workspace — per-plan artifact dirs, self-identifying ledger, end-of-plan cleanup #1943 (SDD ledger session-scoping — adjacent, deliberately not absorbed; this campaign's mount-contamination incident independently corroborated its failure class and is documented in the evals experiment log). The SDD-collision cluster (for superpowers:writing-plans, wheen create a long plan, try to write more times instead of write a large file once #1947–Harden the BRAINSTORM_OPEN launcher: use execFile instead of parsing the env var into argv #1957 era) is owned by fix(sdd): plan-scoped workspace — per-plan artifact dirs, self-identifying ledger, end-of-plan cleanup #1943. No prior PR attempts a fix-loop/breaker redesign.

Environment tested

New harness support (required if this PR adds a new harness)

Not applicable — no new harness.

Evaluation

• Initial prompt: the maintainer opened the session describing the four problems above ("really pathological 'implement, review, fix, review, review, fix' loops with no circuit breaker", the fix-subagent switch, the missing rationalization table, the structural mess).
• Eval sessions after the change: 7 live quorum runs against this branch (3 new scenarios GREEN + 4 regression), plus 5 baseline (RED) runs against dev, in the superpowers-evals lab. Scenarios, fixtures, and the full campaign log landed on evals main (docs/experiments/2026-07-sdd-fix-loop-redesign.md, evals commit 8192fe2).
• Before/after:
  • Structural finding at the breaker: dev silently downgraded an Important plan-contradiction to Minor and built the dependent task on top of it; this branch stopped and surfaced it as BLOCKED.
  • Cap adjudication: dev left no durable record (no parked/ruling lines); this branch parked the finding with a written ruling, dispatched no round 6, completed the remaining task, and informed the final review — 8/8 deterministic checks.
  • Fix mechanism: dev took opposite mechanisms in two runs of the same scenario (dedicated fix dispatch — including one told to "ignore the trailing-newline requirement", i.e. controller pre-judging — versus same-subagent resume), demonstrating the contradiction; this branch resumed the original implementer with a scoped re-review, including organically (SendMessage ×2) in the planted-defect regression scenario.
  • Regressions: planted-defect catching, YAGNI rejection, broken-plan escalation, and spec-constraint preservation all pass on this branch.
  • Negative results are documented at equal billing in the experiment log (three defeated scenario iterations — the redesigned pre-flight legitimately defuses plan-visible seeds — plus judge stalls and a legacy-check brittleness fix).

Rigor

• If this is a skills change: skill-behavior evaluation was run via the superpowers-evals quorum lab (RED/GREEN/regression campaign above) rather than ad-hoc pressure prompts; adversarial verification came from independent per-task reviews, two whole-branch reviews (one on the most capable model, which mechanically verified all 111 old-content chunks against the 33-row move map), and a final fix wave with scoped re-review
• This change was tested adversarially, not just on the happy path (baseline RED runs, seeded mid-loop fixtures, a never-satisfied-reviewer state, structural-contradiction fixture)
• I did not modify carefully-tuned content without evidence: eval-tuned sentences moved verbatim per the move map; the only licensed rewordings are fix-policy sentences, each enumerated in the plan. One known exception for maintainer waiver: the tail of old Never-item 11 ("the plan's example code is a starting point, not evidence that its weaknesses were chosen") survives in content (§3 pre-judging bullet + task-reviewer template) but not as a verbatim sentence.

Human review

• A human has reviewed the COMPLETE proposed diff before submission

@arittr — could you run evals on this? The campaign log and scenarios live in superpowers-evals main (docs/experiments/2026-07-sdd-fix-loop-redesign.md); the three new SDD scenarios plus the four regression scenarios against this branch are the relevant set. SUPERPOWERS_ROOT should point at this branch's checkout for GREEN and a dev checkout for baseline comparison.