AgentSkillsCN

investigation-mode

当连续出现2次及以上失败、反复发生错误，或工作进展停滞不前时——暂停实施，转而深入探究根本原因，收集证据，并在经过验证的方案与备选方案确定之后再继续推进。

SKILL.md
--- frontmatter
name: investigation-mode
description: Use when there are 2+ consecutive failures, repeated errors, or the work feels stuck - pauses implementation, switches to root-cause investigation, gathers evidence, and resumes only after a verified plan and options

Investigation Mode

Overview

This skill provides a hard stop and a repeatable workflow when progress stalls or errors repeat. It prevents “random walk” fixes and forces evidence-first debugging.

Use when...

  • •There are 2+ consecutive failures/errors in the same feature or approach
  • •The same error comes back after “fixes”
  • •There is temptation to “just workaround” or change approach silently
  • •The problem statement is unclear and implementation would be guesswork

Symptoms / keywords

  • •“stuck”, “still failing”, “same error”, “again”, “flaky”, “intermittent”, “can’t reproduce”, “works on my machine”
  • •“quick workaround”, “let’s just do it manually”, “skip validation”, “good enough”
  • •CI-only failures, nondeterministic tests, repeating TypeScript/build/lint errors

Rules (verbatim triggers)

Failure response rules

  • •2+ consecutive failures: Switch to investigation mode
  • •Ask before: Using workarounds or alternatives
  • •Explain: Why original approach failed
  • •Options: Use task-direction-approval (2–3 options + trade-offs; ask user when changing direction).

Core: Respect user's original intent. When stuck, find proper solutions rather than taking shortcuts.

After 2 consecutive errors in same feature

  1. •🛑 PAUSE - Stop implementation immediately
  2. •🔍 INVESTIGATION MODE - Switch focus to root cause analysis
  3. •📖 Deep Research - Dispatch a research subagent to execute web fetch for official docs, RFCs, known issues and return a cited Context Package
  4. •🧠 Sequential-thinking - Analyze fundamental misunderstanding
  5. •🧪 Test First - Write comprehensive tests before continuing

Declare mode switch: "🔍 INVESTIGATION MODE: Pausing to research root cause"

Workflow

  1. •Freeze changes: stop making further edits that are not evidence-driven.
  2. •Capture evidence: record the exact error text, stack traces, logs, and minimal repro steps.
  3. •Constrain scope: isolate the smallest failing unit (single test, single endpoint, single build step).
  4. •Run root cause analysis:
    • •REQUIRED SUB-SKILL: Use root-cause-tracing for systematic isolation techniques.
    • •Use uncertainty-verification when the fix depends on exact tool/library behavior.
  5. •Propose options: Use task-direction-approval (2–3 options + trade-offs).
  6. •Ask approval if direction changes:
    • •REQUIRED SUB-SKILL: Use task-direction-approval when switching library/tool/architecture or replacing automation with manual workaround.
  7. •Resume only after selecting a plan and (when applicable) verifying it with a small test.

Common mistakes

  • •Continuing to code while the failure mode is not understood
  • •Changing direction silently instead of asking for approval
  • •“Fixing” by adding retries/timeouts without evidence