audit-codemod-safety — Compiles and lints is not behaves the same

Degree of freedom: MIXED — Phases 0–2 [HIGH freedom]; Phase 3 high-risk hand-trace [LOW freedom — run exactly]. Do not re-run the transform.

Read-only. A codemod's danger is scale plus plausibility: it touches hundreds of files and each diff looks trivially correct, so review fatigue waves it through. A mechanical transform proves syntax, not semantics.

Present findings. Do not re-run the transform or edit the diffs. Hand fixes to a follow-up session.

This skill vs neighbors

How to reason

1. Observe — quote the rule, the before/after AST or text, and a concrete diff hunk
2. Interpret — does the swap change runtime meaning, or only spelling?
3. Classify — silent behavior change / incomplete migration / consistent-as-is / needs-a-probe
4. Severity — silent runtime change on a hot path = critical

Worked example

Observe: codemod rewrote value || fallback → value ?? fallback in pricing.ts. Tests still pass. Interpret: 0 and '' now keep the left side; price 0 no longer falls back. Syntax-clean; semantics changed. Classify: silent behavior change (truthiness vs nullish). Severity: critical if money path; else major. Finding: pricing.ts:41 | ||→?? | critical | 0 no longer defaults

Phase 0 — Characterize the transform [HIGH freedom]

Establish what was done and how:

• Tool: jscodeshift / ts-morph / ast-grep / comby / regex sed / AI bulk edit
• Rule in plain terms ("replace X API with Y", "rename Z", "new import path")
• Blast radius (files touched, LOC)
• AST-based (structure-aware) vs text/regex (no scope awareness — higher risk)

Get the diff and the codemod script if it exists. Auditing the rule often finds the gap faster than reading 300 diffs.

Phase 1 — Behavior-preservation [HIGH freedom]

Semantic-equivalence traps — syntactic swap, different runtime meaning:

• Operator / precedence shifts (replacement needed parens and did not get them)
• Truthiness vs nullish (|| → ?? or reverse changes 0 / '' / false)
• Async/await vs .then that drop error propagation or change ordering
• Reference vs value, mutate-in-place vs copy, shallow vs deep
• Import swaps with different defaults, side effects, or tree-shaking
• Type-only changes that alter runtime (enum → const, interface → type where declaration merging mattered)

Comment / string / dynamic false hits (regex sweeps especially) — pattern matched inside strings, comments, template literals, or dynamic property access. Text mods cannot tell code from a string that looks like code.

Formatting-masked logic — a reformat (Prettier bump, quote style) bundled into the same commit as the codemod, so a real behavior change hides inside thousands of whitespace diffs. Flag if logic and formatting were not separated.

Phase 2 — Coverage and consistency [HIGH freedom]

Missed cases — matched the common shape, not the variants: alias, re-export, dynamic import, HOC-wrapped, JSX vs call expression, tests, .vue / .svelte / .mdx the parser never reached. Grep the old pattern's semantic footprint — not just the exact string. Survivors mean a half-done migration (two conventions live at once), which is worse than none.

Inconsistent application — same construct transformed one way here, another way there (passes drifted, or the AI drifted). Cross-check audit-code-quality.

Generated / vendored / ignored — touched generated code that will be overwritten, or skipped files the runtime still loads because the ignore list the codemod respected is not the runtime's.

Orphans — old helpers / imports / types the transform was supposed to remove but left dangling; or new ones it referenced but never created.

Phase 3 — Verify, don't trust [LOW freedom — hand-trace high-risk files]

• Test behavior, not compilation. Did tests exercise the transformed paths, or do they pass because coverage there is thin? Cross-check plan-test-coverage. A codemod over untested code is unverified — say so.
• Pair with visual / behavioral diffing where UI is involved (test-visual-regression). Mechanical JSX / style mods are prime silent-break territory.
• Gate interaction. Did this large mechanical change slip a real bug past CI because its size defeated review and its syntax defeated the linters? If so, that is also a finding for audit-gate-logic.
• Spot-probe. Trace the highest-risk transformed files (auth, money, data mutation) by hand end-to-end. Scale review effort by blast radius, not evenly.

Definition of Done

• Transform characterized: tool, rule, blast radius, AST vs text
• Codemod script / rule reviewed where available
• Semantic-equivalence traps checked (precedence, truthiness, async, refs, import semantics, type-affecting-runtime)
• Regex / text mods checked for matches inside strings / comments / dynamic access
• Formatting confirmed separated from logic (or the coupling flagged)
• Old-pattern footprint grepped for survivors → migration completeness assessed
• Inconsistent application across the repo surfaced
• Generated / vendored / ignored file handling verified
• Orphaned / dangling symbols found
• Transformed paths confirmed actually tested; high-risk files hand-traced
• Gate-bypass interaction noted for audit-gate-logic if the mod hid a bug past CI
• Read-only — findings presented, code not edited

Self-critique before reporting [LOW freedom — do not skip]

1. Evidenced — quoted hunk or surviving old-pattern grep, not "looks fine"
2. Semantics not syntax — compiles/lints is not a pass
3. Severity justified — critical = silent runtime change
4. Right owner — CI waved it through → also audit-gate-logic
5. No re-run — findings only; do not re-apply the transform

Output format

1. Transform summary — tool | rule | files touched | AST or text | tests covering the paths?
2. Behavior-change findings — file:line | trap type | severity (critical = silent runtime change) | why the diff looks fine but isn't
3. Coverage / consistency findings — surviving old-pattern instances | inconsistent conversions | orphans
4. Verification log — high-risk files traced | result; test-coverage verdict on transformed paths
5. Fix plan — silent behavior changes first, then migration completeness, then consistency; handoffs to test-visual-regression, plan-test-coverage, audit-gate-logic, a fix session. Present and stop.