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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/prime-radiant-inc/toil --skill debug-run命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Use when monitoring an in-progress toil run, checking if a run is stuck, or observing live workflow execution
Use when writing, editing, validating, or debugging toil workflow YAML — node/edge structure, ForEach, decision tags, expression syntax (${workflow_input.X} vs ${input.X}, supported node fields, required ! refs), or analyzing why a run misbehaved
基于 SOC 职业分类
| name | debug-run |
| description | Use when a toil run has failed or produced unexpected results and you need to investigate what happened |
Systematic post-mortem investigation of toil workflow run failures.
Resolve the run to investigate:
/ui/runs/{id} or /runs/{id} pathGET /runs returns []string of IDs (not objects). Fetch GET /runs/{id} for recent IDs, find most recent with status: "failed" by comparing started_at timestamps. Do NOT sample randomly — work backwards from the end of the list.Base URL: TOIL_URL env var, default http://localhost:8080.
Single agent. Do this first — escalate to Tier 2 only if inconclusive.
GET /runs/{id} — check status, scan nodes map for status: "failed", read both message and error fields on failed nodesGET /runs/{id}/events — JSONL (one JSON object per line, NOT a JSON array). Parse line-by-line. Find node_failed and run_failed events. Build a timeline.@resources/failure-patterns.md for known signature match. Do this BEFORE proposing a fix. If the error matches a known pattern, use the documented fix.node_failed event is usually the root cause. Later failures may be consequences of retrigger attempts. Report them separately.parent_run field, this is a child run — always check the parent run's state to understand why this child was spawned and whether the parent has additional context. Do NOT debug a child run in isolation.Escalate to Tier 2 if: error is in a child run, root cause is ambiguous, no pattern match, or the failure involves multiple interacting nodes.
Multi-agent fan-out. Run in addition to Tier 1 findings.
GET /runs/{id}/compound-graph — discover all child runs and their topology. Do NOT manually trace child runs from state/events — use this endpoint.GET /runs/{child_id} + GET /runs/{child_id}/events (JSONL). Build timeline per child. Aggregate results back in the parent agent.GET /workflows/{workflow_id} — returns raw YAML (text/plain, NOT JSON). Cross-reference the failing node's definition: check its decisions, outputs, inputs, and edges. Understanding what a node declares tells you what validation it triggers.GET /health — server status, uptime, active run countsdata for file paths)decisions or outputs.When debugging a child run, always trace upward to the parent. A child run failure may be a symptom of a parent-level issue. Check parent_run field in the child's state.
When a node declares decisions or outputs in workflow YAML, its JSON output MUST include all three at the top level:
{"decision": "done", "message": "completed successfully", "data": {}}
The data field must be a JSON object (not null, not missing). Do NOT nest decision/message inside data. They are siblings, not children.
Validation is skipped entirely for nodes that declare neither decisions nor outputs.
After Tier 2, if root cause is novel:
@resources/failure-patterns.md — does this signature already exist?@resources/toil-api.md — endpoint reference@resources/data-model.md — run/node states, event types, validation contract@resources/failure-patterns.md — known failure signatures