菜单
Diagnose a bug, failing test, error, regression, or incorrect runtime behavior by reproducing it and tracing the first causal divergence. Use when the true source is unknown; do not use for speculative cleanup or when a proven fix is already specified.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Draft or revise external communication in the user's concise, direct, informal voice. Use whenever Codex is asked to write, draft, reply, post, or comment on the user's behalf, including posting comments for them without an explicit "as the user" instruction. This skill controls wording only and never grants permission to send or post.
Use when working with a Backlog.md project or its `backlog` CLI/MCP, including creating, reading, refining, updating, completing, or archiving tasks and managing dependencies or milestones.
Refine selected backlog items or whole backlog sources into implementation-ready work. Use when asked to clarify, decompose, or prepare work across Markdown, Backlog.md, GitHub Issues, Linear, or another supported provider.
Resolve and operate on backlog sources consistently across loose Markdown, Backlog.md, GitHub Issues, Linear, and other supported providers. Use whenever a backlog command or skill reads or changes provider-backed work.
Resolve human decisions, missing inputs, stale blockers, and decision-only backlog items. Use when asked to unblock work or persist answers across Markdown, Backlog.md, GitHub Issues, Linear, or another supported provider.
Coordinate multiple local agents working from the same backlog or work queue with bounded Worklease claims. Use when selecting, claiming, heartbeating, checkpointing, handing off, or releasing shared work.
| name | root-cause-investigation |
| description | Diagnose a bug, failing test, error, regression, or incorrect runtime behavior by reproducing it and tracing the first causal divergence. Use when the true source is unknown; do not use for speculative cleanup or when a proven fix is already specified. |
This skill supplies a read-only diagnostic method. It never authorizes edits, commits, pushes, remote updates, issue comments, or delegation; the caller owns those permissions and budgets.
Follow the investigation method. Maintain one hypothesis ledger and choose experiments by information gain. Stop only with ROOT CAUSE PROVEN, ROOT CAUSE NOT PROVEN plus exact missing evidence, or CANNOT REPRODUCE plus the required environment/input/state. Use the diagnosis template or the caller's stricter format.
Apply a caller-authorized fix only when the proven source change is small, unambiguous, and contains no material product/architecture choice. Require the original repro and appropriate behavioral regression coverage. Otherwise hand off the diagnosis and fix direction. Never suppress the symptom, weaken a test, catch-and-ignore the error, or add a fallback that preserves the cause.
