spectre-fix-core

name	spectre-fix-core
description	Internal diagnose-and-repair engine for Spectre workflows. Use only when invoked by spectre-fix or spectre-delegate with an explicit phase and authorization contract. Do NOT invoke directly for user requests.
user-invocable	false

fix-core

Purpose

Own one reusable bug flow: reproduce the failure, ground the root cause and behavioral blast radius, then implement a verified RED-before-GREEN repair. Keep user approval in the standalone spectre-fix parent while allowing orchestrated parents to supply scope-bound authorization.

Inputs

• Bug report: error, stack trace, reproduction, and referenced context.
• PHASE=diagnose | repair | full.
• repair requires the exact diagnosis and experience contract plus USER_APPROVED_FIX_CONTRACT=true.
• full requires PARENT=spectre-delegate, PARENT_AUTHORIZATION={scope.md}, AUTHORIZED_SCOPE_SHA256, and ALIGNMENT_MODE=inferred.
• --orchestrated — withhold user-facing routing, never content.

Working Set

• Read affected paths and recent changes just-in-time.
• Invoking this workflow IS the user's request for subagent investigation: dispatch @spectre_analyst for causal and impact traces. Keep returns compact and in-thread, with no scratch reports.
• In parent-authorized full, read the scope artifact before any code write.

Outputs + DONE

Experience-contract rows map technical evidence to product behavior: journey/surface | current experience | expected experience | technical path/consumer | disposition=intended-change|preserved-invariant|collateral-change|unresolved | evidence | verification.

• diagnose → DIAGNOSIS_READY: reproduction, root cause, affected files, candidate repair, evidence-backed experience contract, and regression/invariant opportunities; no code writes.
• repair | full → FIX_COMPLETE: diagnosis, row-level contract evidence, confirmed RED, implementation, GREEN checks, [🪳 TEMP {TOPIC}] diagnostic logging, changed files, and limitations.

DONE when: root cause is grounded and its hypotheses were traced by dispatched analysts; the product and technical direct blast radius was independently explored after the candidate repair boundary was known; authorization is valid with no unresolved row; the intended regression is RED→GREEN; affected checks have no attributable failure; unrelated findings are routed; and the parent receives the result.

Method / guardrails

1. Diagnose. Generate 5–7 hypotheses; reduce to 1–2 via evidence/data-flow/changes/paths; dispatch parallel @spectre_analyst traces; reproduce; synthesize one cause-level repair.
2. Explore product + technical impact. Once root cause and candidate repair boundary are grounded, dispatch ≥1 independent read-only @spectre_analyst; parallelize separable product journeys or technical boundaries. Trace shared callers, state, and data paths through to user/operator-observable outcomes. Return compact experience-contract rows that explicitly identify experiences that change, remain invariant, or are unresolved; synthesize and deduplicate the results.
3. Honor phase. diagnose returns DIAGNOSIS_READY before any code write.
4. Verify authorization before repair.
   • repair proceeds only when the parent supplies the exact diagnosis and experience contract shown to the user plus USER_APPROVED_FIX_CONTRACT=true.
   • full proceeds only when Delegate supplies a readable scope artifact whose recomputed SHA-256 equals AUTHORIZED_SCOPE_SHA256, whose alignment mode is inferred, and the diagnosis, repair, and every experience-contract row remain inside it. Parent authorization replaces only the post-diagnosis approval pause.
   • An unresolved row or unapproved/out-of-scope collateral change blocks repair.
5. Repair test-first. Confirm the intended-change regression RED for the diagnosed reason; capture risk-proportional preserved-invariant checks; implement the smallest root-cause fix plus [🪳 TEMP {TOPIC}] logging; then prove the contract and affected deterministic checks GREEN.
6. Attribute and continue. Classify extras branch-caused|unrelated|indeterminate; repair, route, or focus-check. If failed repair leaves third-party cause unclear, dispatch @spectre_web_research on pinned docs/code/issues, then analogs; revise hypotheses + RED before mutation. Broad baseline red never blocks.
7. Contain scope. Do not broaden behavior, hide failures, weaken assertions, bypass checks, or treat an unrelated defect as authorized. A new or changed experience-contract row or repair boundary returns to authorization.

Handoff

Return the phase status, the hypothesis set, diagnosis, experience contract with row-level evidence, authorization path/hash/mode, RED/GREEN evidence, changed files, checks, limitations, and exact validation steps. The parent owns all user-facing Next Steps.

Escalate-If

• The report is too thin to form testable hypotheses, reproduction is unavailable, no root cause can be grounded, impact evidence conflicts, or desired behavior remains unresolved.
• Authorization is absent, stale, unreadable, hash/mode-mismatched, does not match the diagnosis/experience contract, or the repair exceeds the authorized scope.
• The regression cannot establish the cause, evidence contradicts it, or further safe action needs authority. Never escalate for unrelated red checks, unavailable broad suites, attempt count, or authorized related-file growth.