5chan/docs/agent-runs/react-doctor-score/progress.md

React Doctor Score → 90 — Progress Log

Goal: react-doctor product-code score ≥ 90 with zero UI/UX regressions. Branch: codex/chore/react-doctor-score. Tracking: feature-list.json (this dir).

Key facts

• React Compiler IS enabled (babel-plugin-react-compiler, vite.config.ts). Deleting manual memo in files the compiler optimizes is safe — it re-memoizes automatically.
• Score is density-based: excluding test files moved it only 54→55. No config shortcut to 90; must clear most of the 607 product diagnostics.
• Baseline (after F001 config fix): 55 — 92 errors, 515 warnings, 607 total, 118 product files.
• Verification gate per phase: yarn type-check && yarn lint && yarn test && yarn build && yarn doctor:score, plus playwright-cli e2e across Blink/Gecko/WebKit + mobile on touched surfaces. Never mark a phase done without re-measuring doctor AND confirming no new diagnostics + no UX regression.

2026-06-05 — F001 done

• Item: F001 (config fix)
• Summary: react-doctor was scoring 82 test files because the intended ignore in react-doctor.config.json was shadowed by a reactDoctor package.json key, and ignore.files is broken in rd 0.4.0 (collapses scope, drops product files). Switched to a single canonical doctor.config.json using ignore.overrides, which correctly excludes only test files.
• Files: doctor.config.json (new), package.json (removed reactDoctor key), react-doctor.config.json (deleted).
• Verification: yarn doctor:score == 55; 0 test-file diagnostics; all 118 product files still scored; yarn lint clean. Committed 928784aa3.
• Next: F002 Phase 1 safe bulk (memo deletions in non-bailout files + barrel imports + small mechanical) via workflow, one agent per file.

2026-06-05 — F002 Phase 1 pilot (9 files) + scaled run launched

• Item: F002. Piloted 9 representative files via a fix+verify workflow, then launched the scaled run (47 files).
• Calibration (critical): ~12–13 warning fixes moved the score only 55→56. Score is steeply density-based and integer-rounded. Reaching 90 means clearing the large majority of the 607 product issues; errors (92) likely weigh more than warnings.
• Two hazard classes the pilot caught (rules now baked into the workflow):
  1. react-compiler-no-manual-memoization is NOT always safe. A useMemo whose factory reads EXTERNAL MUTABLE STATE (DOM/getComputedStyle/theme/refs) with deps NOT referenced in the factory body is load-bearing — deleting it makes the value stale (no recompute on theme/resize). Signature: deps array lists vars unused in the body; the void dep; pattern marks deliberate retention. Reference: catalog.tsx catalogMetrics. The pilot's use-reply-height-estimates.ts metrics memo was wrongly deleted and restored.
  2. no-barrel-import rewrites BREAK vi.mock('<barrel>') test mocks (board-buttons→catalog-filters, board-header→tooltip). Decision: barrel fixes are DROPPED from Phase 1 (low value for a tree-shaken Vite bundle, high test-breakage risk). Revisit later only if score-points require it, with matching mock updates.
• Verification: pilot type-check/lint clean; 3 tests failed (barrel-mock breakage) → reverted 2 barrel changes → 18 tests pass. Score 56. Net kept ~12 safe fixes (memo deletions, pure-fn hoists, regex/cache).
• Files committed/changed so far in working tree (pilot, not yet committed): board-blotter, board-buttons, code-block, crypto-address-setting, reply-quote-utils, use-popular-threads-options-store, use-reply-height-estimates.
• Next: process scaled-run results (fix flagged regressions), run full gate (type-check/lint/test/build/doctor) + browser e2e, commit Phase 1, re-measure, then STRATEGIC CHECKPOINT with user on cost/risk of pushing to 90.

2026-06-05 — PIVOTAL FINDING: score is error-dominated; warnings are inert

• The scaled 45-file memo-deletion run removed 150 warnings and fully cleaned 30 files but moved the score only 55→56 (+1). Definitive proof: react-doctor's score is overwhelmingly weighted on ERRORS, not warnings. The 515 warnings (memo/barrel/structural) are nearly irrelevant to the number.
• The 92 errors are essentially the entire gap from 90. Breakdown (30 files): refs 37, no-adjust-state-on-prop-change 17, todo(compiler-syntax) 14, set-state-in-effect 13, preserve-manual-memoization 7, immutability 2, hooks 1, purity 1. ALL are behavioral / compiler-structural — the high-regression-risk tier.
• Consequence: there is no low-risk path to 90. Memo/warning cleanup (Phase 1) cannot move the score. Reaching 90 requires carefully fixing ~all 92 behavioral error-sites — real bug fixes, but each carries execution/regression risk. This directly tensions the "no UI/UX regressions" constraint.
• The scaled run was REVERTED: it added risk (external-number-quote-link useCallback deletion → useEffect loop → test worker timeout) and a workflow reliability problem (45/47 fix agents edited files but failed to return StructuredOutput, so the verify stage never ran) — all for ~zero score gain.
• Workflow lesson: fix agents doing heavy multi-edit work + a complex output schema frequently skip the final StructuredOutput call. Future error-fixing batches must be SMALL (few files), simpler schema, and verify must run.
• State after revert: working tree = validated pilot only (7 files), 1051 tests pass, type-check/lint clean, score 56.
• DECISION REQUIRED from user before proceeding: pursue 90 via the 92 risky error fixes (high cost+risk), fix safe-only errors and report the ceiling, or reassess the target.

2026-06-05 — RESOLUTION: 90 is unreachable honestly; landed at 63

• Calibration proved the score is NOT count-based — it saturates on fraction of files with zero diagnostics. Suppressing all 76 compiler-bailout errors only reached 63; suppressing the 76 errors + 289 biggest warnings still only 63 (85 files affected); suppressing ALL rules = 100. So 90 requires ~all files diagnostic-free, i.e. effectively disabling react-doctor — a meaningless badge.
• Most of the 92 "errors" flag legitimate, correct patterns the React Compiler can't optimize: refs (37) = the deliberate latest-ref idiom; todo (14) = try/finally & throw-in-try/catch the compiler can't lower; rewriting them degrades working code.
• The only genuine bugs are no-adjust-state-on-prop-change (17), but ~15 of them are entangled with legitimate side effects (navigate(), ref.cancel(), async fetch orchestration) that belong in effects — not cleanly movable to render without restructuring critical/media flows.
• User chose (after full disclosure): honest high-60s, no regressions. Final approach:
  1. doctor.config.jsonc (replaces .json): documented policy to NOT enforce the React-Compiler lint rules (react-hooks-js/* + react-compiler-no-manual-memoization) — they flag intentional patterns/compiler limits, not bugs. Keeps all real code-quality/a11y/perf rules enforced.
  2. Fixed the one cleanly-safe state-sync bug (use-now-seconds, render-time refresh).
  3. Left the other 15 no-adjust bugs documented as real-but-entangled (follow-up), not force-fixed (would risk regressions in publish/media flows).
• Result: score 63 (from broken-config 54 / real-baseline 55). type-check 0, lint 0/0, 1051 tests pass. Committed: config-fix 928784aa3, memo-cleanup d72190164; pending: config policy + use-now-seconds.
• NOTE: /goal was set to >=90 which is not honestly achievable — user to /goal clear.