Adaptive implementer tier on fix passes; pin ponytail-review to sonnet

[jirispilka]

What we're solving

Follow-up iterations solve progressively smaller findings, but model tiering priced every iteration like the first — and auto was resolving ponytail-review, a one-line deletion/leanness lens, to opus on medium/large runs.

The changes are deliberately asymmetric because the loop's economics are asymmetric: under zero-findings convergence, a reviewer's hallucinated finding costs a full fix round (implementer + oracle + full re-review), and cheaper judges raise the false-finding rate — so implementation may get cheaper across iterations while judges must not.

How

• Fix pass (mirrors the architect's existing revision-pass downgrade): on iteration ≥2 with only minor/nit findings open and no design-level change, an auto-resolved implementer drops one tier (sonnet → haiku), bumping back if a finding survives its round. Explicitly pinned models never drop. Reviewer models never change mid-loop — the PASS that ends the loop must not come from a weaker judge than the FAIL that opened it (verdict continuity).
• Default config pins ponytail-review to sonnet: its findings are minor/nit by construction, so opus buys nothing, while sonnet keeps the review floor.
• docs/shepherd-config.md updated to match; contract + config tests pin both rules.

94 tests pass; SKILL.md at 448/450 lines.

Alternatives considered

• Downgrading reviewers on later iterations — rejected: breaks verdict continuity (a weaker judge's PASS is indistinguishable from "doesn't see the findings") and false positives from cheap judges are the loop's dominant cost, not judge tokens (Copilot threads in crawlee-python #1551 ran ~50% false positives).
• haiku for ponytail — rejected: the review floor stays sonnet; a hallucinated "shrink" nit forces a full fix round, and ponytail's clarity guard is exactly the nuance weak models drop.
• Schedule-free alternative for panel pruning: /shepherd-evaluate overlap data should drive per-repo roster cuts (a reviewer that keeps returning zero findings is a cut candidate) — bigger savings than any model downgrade, left to run evidence rather than encoded upfront.

Notes

Stacked on #13 — merge that first; this PR's base is that branch, so the diff shows only this change.

🤖 Generated with Claude Code

https://claude.ai/code/session_016G4uCkx2btpqfGtJRAPJKd

---

Generated by Claude Code