localize — localize → edit → validate

Do NOT start editing from the first plausible file. Empirically, a staged localize → repair → validate loop resolves real repository issues at a fraction of the cost of broad autonomous exploration (Agentless, arXiv 2407.01489), and the benchmark contract for "resolved" is a change whose tests pass (SWE-bench, arXiv 2310.06770). This skill enforces that discipline: name the target with evidence, make the minimal edit, then prove it with tests.

1. LOCALIZE — produce the target before editing

From the request, find the exact edit site(s). Do NOT edit in this phase.

• Extract the concrete signals from the request: error text, stack trace, symptom, endpoint / field / component name, the file it mentions.
• Search from those signals: grep the error string / symbol; glob the module; read the suspected files. For a symbol you will change, run /harness-kit:blast-radius <symbol> to enumerate everything that depends on it, so the edit scope is known up front (not discovered after a half-fix).
• If tests exist, use the failing test (or write one that reproduces the symptom) to point at the fault — a failing test IS localization evidence.

Localization artifact (write it down, BLUF):

• Targets: each file:line + symbol you intend to change, with the evidence that put it there (the grep hit / the failing test / the reference).
• Impact: the blast-radius set for any changed signature (callsites that must move together).
• Uncertainty: what you are NOT sure about + the candidate you'd check next. Non-empty is honest; a confident-but-wrong target is the failure mode here.

2. EDIT — minimal, scoped to the target

Make the smallest change that addresses the request at the localized site. Update every callsite the blast-radius set named — do not leave a half-migrated signature. Match the surrounding idioms.

3. VALIDATE — tests gate the change

• Focused first: run the specific test covering the change (or the reproducer from step 1). It must go from failing → passing.
• Regression next: run the stack's test command (the repo's detected test_cmd). Report pass/fail counts.
• Real-run evidence for a "feature works" claim: a real input → real output, not just green unit tests (per the kit's 0.3 No premature done).
• Record the commands run + outcomes in the handoff / verify artifact.

Constraints

• The localization artifact (step 1) is REQUIRED before any edit — no "edit first, find the rest later".
• A change is not done until its focused test passes AND the regression run is green (or the failures are pre-existing and named). "Compiles" / "200 OK" is not validation.
• No blind retries: if two edits don't pass, re-localize (the target was wrong), don't loop on the same site — broad retry loops compound error, they don't fix localization.
• Bounded autonomy: this is one change → one validation → the human gate, not an open-ended fix loop.