agentsclimarketplace

Generate triage plan

Skill qa-veritas/skillpack/skills/generate_triage_plan

Agent capability that loads only when the task needs it. The Skills layer of QA Veritas.

Install
npx -y skills add qa-veritas/skillpack --skill generate_triage_plan

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Produce an ordered investigation plan from a failure classification and the available evidence sources, with explicit stop conditions so the investigation ends when the next step would not change the conclusion. Use at the start of any non-trivial failure.

SKILL.md

3.0 KB, as published. Nobody here has run it

generate_triage_plan

Purpose

Unplanned triage wanders. It re-reads the same log, chases the loudest error instead of the first one, and stops when the engineer is tired rather than when the question is answered. This skill produces a cheap, ordered plan with stop conditions before any deep investigation starts.

Inputs

  • failure_classification — a first-pass class (PASS / HANG / TEST-BUG / INFRA / FRAMEWORK / PRODUCT) plus the error signature.
  • available_evidence — what can actually be inspected: result metadata, logs (and their sizes), support bundles, live cluster access (yes/no), prior-incident search, knowledge base.

Outputs

ordered_steps:
  - step: "Parse result metadata for exact phase + counts"
    cost: cheap
    yields: "Confirm class; narrow to a phase"
  - step: "Slice logs around the first error timestamp"
    cost: cheap
    yields: "The emitting code path"
  - step: "Search prior incidents for the error signature"
    cost: cheap
    yields: "Known-issue match or novelty"
  - step: "Inspect live cluster state (if access available)"
    cost: expensive
    yields: "Current vs expected state"
stop_conditions:
  - "A prior incident matches the signature exactly -> attach and stop."
  - "Root cause confidence is high and the next step is expensive."
  - "Evidence is exhausted -> report best hypothesis + what's missing."

Prompt template

You are planning a triage. Inputs: classification {failure_classification}, evidence {available_evidence}.

  1. Order steps cheapest-first and highest-information-first. Parse hard facts (counts, phases, exact signatures) before any reasoning that depends on them.
  2. For each step, state what it costs and what it would yield. Skip steps whose evidence source is unavailable; degrade gracefully (logs-only if there is no live access).
  3. Define stop conditions up front: a prior-incident match, a high-confidence cause before an expensive step, or exhausted evidence. The goal is to end the investigation deliberately. Output the ordered plan and the stop conditions.

Examples

  • HANG with logs only, no live access. → plan centers on the last progress line before the hang and the timeout owner; skip live-state steps; stop when the blocking call is identified.
  • INFRA-class with live access. → cheap metadata + log slice first, then targeted live checks (disk pressure, node readiness) only if the logs point there.

Composition

This is the entry-point skill. It schedules summarize_logs, explain_failure, analyze_cluster_health, check_disk_pressure, and finally identify_root_cause.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.