as-fix（作为修复）

Fix

Purpose

Resolve concrete bugs safely through structured diagnosis, minimal fixes, and verification.

This skill is for bug-fix execution when scope is clear and risk is manageable.

Scope Gate (Required Before Starting)

Use this skill only when all conditions below are true:

1. Concrete issue signal exists

- Clear failing behavior, error, or regression is identified - Enough evidence exists to begin diagnosis (logs, stack trace, failing test, or reproducible steps)

1. Likely localized fix path

- Issue appears constrained to a small area - No expected cross-system redesign or migration

1. Low-to-moderate architectural risk

- Fix can be made without foundational changes - No phased rollout/feature-flag strategy is expected

1. Straightforward verification path

- Reproduction and post-fix checks can be run in practical scope - Success criteria can be validated directly

If any condition fails, escalate to brainstorm or write-plan.

Hard Stop Escalation Criteria

Immediately stop and switch to planning if any of these occur:

• Root cause remains unclear after focused investigation
• Fix requires cross-cutting architectural or data-model changes
• Scope expands into large refactor or multi-module migration
• Security/compliance-sensitive behavior is involved
• Verification reveals broad regressions requiring phased mitigation
• Multiple fix attempts fail without converging on a root cause

Escalation action:

1. Stop all implementation activities.
2. Output the exact message: "This fix exceeds safe single-pass debugging limits. Ask, Recommend brainstorm or write-plan first to define phased diagnosis, implementation, and risk controls."

Workflow

Step 1: Intake and Contextualize

1. Understand the reported bug and define explicit expected behavior.
2. Collect/confirm minimum bug report fields:

- Title - Expected behavior - Actual behavior - Reproduction steps - Evidence (logs, trace, screenshot, failing test output) - Environment (branch/OS/runtime/version if relevant) - Impact/severity

1. Load project context per the shared Context Loading Protocol.

Step 2: Reproduce and Diagnose

1. Reproduce the issue consistently.
2. Locate the failure point (file/function/line range/subsystem).
3. Trace control/data flow to identify the root cause.
4. Form a fix hypothesis and confirm it explains observed behavior.

Guidelines:

• Read surrounding code, not only the failing line.
• Prefer root-cause correction over symptom patching.
• Add temporary diagnostics only when needed; remove them after use.

Step 3: Decide Small vs Bigger Fix

Classify the work before coding:

Small fix (continue in this skill)

Most are true:

• Root cause is clear and validated
• Change surface is narrow (typically a few files)
• Regression risk is limited and testable quickly

Bigger/risky fix (escalate to brainstorm or write-plan)

Any are true:

• Root cause is uncertain
• Multiple subsystems must change together
• Requires migration, rollout sequencing, or broad refactor
• Risk cannot be reasonably controlled in a single pass

If it is a bigger fix, follow Hard Stop Escalation Criteria.

Step 4: Implement (Small Fix Path)

1. State a brief 1-3 bullet implementation plan.
2. Apply the smallest targeted change that resolves the root cause.
3. Keep scope strict; avoid unrelated refactors.
4. Add/update regression tests when applicable.

Step 5: Verify

Run validation in increasing scope:

1. Focused checks for the changed behavior/module
2. Nearby regression checks
3. Relevant project checks (lint/typecheck/tests/build) as needed

Fix is complete only when:

• Reproduction no longer fails
• Expected behavior is confirmed
• No critical regressions are introduced

If verification indicates broader impact, escalate to brainstorm or write-plan.

Step 6: Complete and Report

Provide a concise completion report with:

• Root cause
• What changed
• Why this fix works
• Verification performed and results
• Residual risks / follow-ups

If behavior or documentation-relevant rules changed, update the minimal relevant docs. If architecture changed, this should have been escalated.

Rules

• Do not guess when key context is missing; ask via Question Tool.
• Always prioritize root-cause fixes.
• Keep blast radius minimal.
• Do not mark done without verification.
• Escalate early when scope/risk exceeds this skill.

Optional Bug Report Template

Use when the report is incomplete:

• Title:
• Expected behavior:
• Actual behavior:
• Reproduction steps:
• Error logs/stack trace:
• Environment:
• Impact/severity:
• Additional context: