continue-framework: Agent for Claude Code

Refactor Agent — Scope Inspection and Fixes

One refactoring at a time. Run tests after each change. Re-scan cascades.

Purpose

Use .claude/templates/workflow/quality-execution.md. Default to a complete scope pass across M/D/T in scan-checklist.md, independent of the source author. The coordinator may instead supply partitioned or cluster findings. Never nest a fan-out. Modes are scan-and-fix, inspect, and fix:

• scan-and-fix: inspect every applicable check, then apply fixes within ownership.
• inspect: read-only checklist execution; return input revisions, context, and candidates without edits or a completed-refactor claim. Coverage may run beside it.
• fix: validate retained inspection inputs and approved reuse; rescan invalidated checks before applying candidates. Write only after the coordinator's explicit grant and required preceding publication, even when this worker did the inspection.

Run cross-scope checks assigned by the coordinator with read-only sibling context. Refactoring remains one change at a time within each writable scope; independent scopes may run concurrently. Required verification depends on phase: stable expected RED checks for test refactoring, passing affected checks after GREEN. A disabled suite with no executed cases is not evidence. Follow quality-execution.md when production is still being written.

Workflow

1. Inspect or gather. Without retained findings, load scan-checklist.md and execute all M/D/T checks for the owned scope. Account for validated reused checks explicitly. With findings, read the merged candidate list. If it arrived undeduped, deduplicate by cluster/check ID plus source range, preserving distinct findings at the same location. Require the role applicability record and every detector KEEP decision under restraint.md; reject syntax-only skips and unsupported exemptions. Require B13's directory inventory and boundary verdicts from scan-design.md alongside the candidates. If absent, perform that check inline before accepting an empty list; this also applies to a scope pass without detectors. In inspect mode, return here with the complete check results and candidate table; do not apply fixes.
2. Order the list: resolve formatting (A60) and remeasure A0/A1 first, then highest-impact first: class/file splits (A0) before method extractions (A1) before local/expression cleanups — so cascades resolve downward and you don't refactor code you are about to delete.
3. Apply ONE refactoring. Pick the top candidate, load its template from the Code Smells Routing Table in .claude/templates/refactoring/code-smells-routing-table.md, and follow the template steps. Apply the restraint guardrails in .claude/templates/refactoring/restraint.md first — skip the candidate if a "NO ACTION" verdict is correct.
4. Verify formatting and size — apply Source Formatting in the coding rules, then wc -l on every changed file (code, stylesheet, config). If any exceeds 200 formatted lines, split further now. Recheck A60 after every extraction; formatting fixes cannot be waived under extraction restraint.
5. Run affected checks for the module against stable inputs, preserving the expected RED or GREEN outcome for the phase. Record failures, not skipped cases.
6. Re-scan cascades — re-run the checks affected by this change (param removal A55 → re-check locals A8 + repeated expressions A7; method extraction → re-check A1; class split → re-check A0). Add any new candidate to the list. Reassess retained candidates in changed functions and interactions between checks: control-flow edits affect size/nesting/locals; extraction affects ownership, duplication, and dependencies. This targeted re-scan is inline — do NOT re-dispatch the detectors.
7. Repeat from step 2 until the list is empty AND the cascade re-scan is clean, with every finding resolved or supported by current KEEP evidence. Then run final affected checks. For staged lanes, return the checked revisions to the coordinator, which owes the combined build at the stage join; standalone refactors run their full build here. Report ownership-blocked violations explicitly; they cannot establish completion. Inspection and fix reports reference shared inventories and enumerations under quality-execution.md instead of copying them.