v1.26.4.0 fix: GSTACK REVIEW REPORT delete-then-append (no more mid-file leftovers)

[garrytan]

Summary

Fixes the bug where ## GSTACK REVIEW REPORT would land mid-file in plan-mode plans whenever a stale copy from a prior /autoplan run was already there. The contradictory write rule in scripts/resolvers/review.ts ("replace it entirely (in place)" vs "always last section / move if mid-file") is collapsed into a single delete-then-append flow with explicit Read-tool verification.

Performance / behavior

• Single rule eliminates the reconciliation step the agent had to perform; the new prompt has no internally inconsistent bullets to weigh.
• Read-tool verification step runs after the append and self-corrects with one retry if the report didn't end up last.

Coverage

• 5 new static assertions in test/gen-skill-docs.test.ts lock the prompt change against drift across all 4 plan-review SKILL.md files + the source resolver. Synthetic regression check (revert → 5 fails, restore → 5 passes) confirms the tests are bound to the prompt change.

Infrastructure

• 6 generated SKILL.md files refreshed (plan-ceo-review, plan-design-review, plan-devex-review, plan-eng-review, codex, devex-review).
• The test/skill-e2e-autoplan-auto-mode.test.ts reverts to its original AUQ-blocked-gate-surface shape after a paid run revealed --disallowedTools AskUserQuestion makes autoplan bail at the Phase 1 premise gate via the plan-file fallback before any report-write path runs. The PTY harness can't drive autoplan through its review phases without auto-progression of AskUserQuestions, so static prompt-text verification carries the regression coverage.

Test Coverage

[+] scripts/resolvers/review.ts (prompt text, no new code paths)
  └── delete-then-append flow              [TESTED] 5 static template assertions

[+] test/gen-skill-docs.test.ts (new tests)
  ├── 4 SKILL.md target tests              [★★★ TESTED] verify new + reject old
  └── 1 source resolver test               [★★★ TESTED] same checks vs review.ts source

[+] test/skill-e2e-autoplan-auto-mode.test.ts (revert to original)
  └── existing AUQ-blocked smoke test      [★★★ TESTED] passed in paid E2E run

COVERAGE: 100% — no new code paths introduced; only prompt text + tests.
SYNTHETIC REGRESSION: stash prompt fix → 5 tests fail; restore → 5 pass.

Tests: 374 → 379 (+5 new static assertions).

Pre-Landing Review

No issues — diff is prompt text + tests only, no logic paths, no SQL, no LLM trust boundary changes, no error handling.

Eval Results

bun run eval:select against the diff shows no prompt-builder pattern matches in the eval gating set (the patterns are Rails-style app/services/*_prompt_builder.rb, etc. — gstack's resolver-based prompts use a different gating system). Free bun test suite exits 0 (379 pass, 0 fail in test/gen-skill-docs.test.ts). Targeted EVALS=1 EVALS_TIER=gate bun test test/skill-e2e-autoplan-auto-mode.test.ts ran during development: 1 pass, 0 fail in 106s.

Plan Completion

Plan at ~/.claude/plans/system-instruction-you-are-working-pure-swing.md. All items DONE or addressed:

• [DONE] Prompt fix in scripts/resolvers/review.ts — delete-then-append flow + verify step
• [DONE] Generated SKILL.md regenerations — 6 files refreshed via gen:skill-docs --host all
• [CHANGED] E2E regression test approach — pivoted to 5 static template assertions after paid E2E run revealed the harness can't reach the bug code path with --disallowedTools AskUserQuestion
• [DONE] Synthetic regression check — verified by stash/restore cycle
• [DONE] Free test suite — exit 0
• [DROPPED] Optional SDK-harness ExitPlanMode-input check — marked optional in approved plan

TODOS

No matching TODOs in TODOS.md to mark complete (1655 lines, none referenced this bug).

Test plan

• Free bun test suite passes (exit 0)
• bun test test/gen-skill-docs.test.ts — 379/379 pass after fix
• Synthetic regression: 5 new tests fail when prompt fix reverted
• Targeted gate-tier autoplan E2E: 1/1 pass in 106s
• In real plan-mode usage with /autoplan, the report should land at the bottom even when an older copy was mid-file

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews

[github-actions]

E2E Evals: ✅ PASS

14/14 tests passed | $2.68 total cost | 12 parallel runners

Suite	Result	Status	Cost
e2e-design	2/2	✅	$0.41
e2e-plan	6/6	✅	$1.45
e2e-review	1/1	✅	$0.35
llm-judge	3/3	✅	$0.06
e2e-design	2/2	✅	$0.41

---

12x ubicloud-standard-2 (Docker: pre-baked toolchain + deps) | wall clock ≈ slowest suite