Test: Antithesis Triage Skill

End-to-end test harness for the antithesis-triage skill. Spawn sub-agents that perform real triage operations against the live Antithesis API via snouty, then review their work for issues.

The top-level agent MUST NOT use the antithesis-triage skill directly. All triage operations happen inside sub-agents. The top-level agent only orchestrates and reviews.

Prerequisites

Before starting, verify the same prerequisites the triage skill requires:

which snouty && which jq && snouty --version

snouty must be at least version 0.8.0. Then confirm credentials with the canonical check (same as the triage skill's preflight):

snouty doctor

If any prerequisite is missing, out of date, or snouty doctor fails, stop and report which ones are unavailable.

Phase 1: Discover Runs

Spawn a general-purpose sub-agent with the Agent tool. Provide these instructions, replacing {{TENANT}} with the actual value of the $ANTITHESIS_TENANT environment variable and {{TRIAGE_SKILL}} with the absolute path to antithesis-triage/SKILL.md in this repository. If $ANTITHESIS_TENANT is empty in the current shell, ask the user for the tenant name instead of aborting.

Read the skill file at {{TRIAGE_SKILL}} and follow its instructions to list
recent runs for the tenant "{{TENANT}}". Follow the "Summarize recent runs"
workflow.

Return the COMPLETE raw NDJSON of `snouty runs --json list -n 200` along
with a per-run summary that includes, for each run:
  - run_id
  - status
  - launcher
  - created_at / completed_at
  - antithesis.test_name and antithesis.description (from parameters)
  - whether links.triage_report is present (triageable yes/no)

Do NOT filter by status — return the full list so the orchestrator can
select targets across statuses.

Wait for this sub-agent to finish before proceeding.

If the sub-agent did not return the NDJSON list, abort.

Phase 2: Select Test Targets

Read the sub-agent output. From the run listing, select two targets:

1. Completed run with failing properties that have moments.

   • status == "completed"
   • links.triage_report is present
   • At least one failing property with counterexamples[0].moment.input_hash present (i.e., addressable to a moment, not a telemetry/meta property)

   Confirm by calling snouty runs --json properties --failing $RUN_ID and verifying at least one entry has a counterexample with a moment field.

2. Incomplete run with a failure_moment.
   • status == "incomplete"
   • links.triage_report is present
   • failure_moment is present in snouty runs --json show $RUN_ID
   • Prefer runs whose duration (completed_at − created_at) is under 30 minutes — these are usually setup/early-execution failures, which exercise the diagnose-incomplete path most directly

Rules for missing targets:

• If no completed run with failing-property moments exists, skip Phase 3a and note this in the final report.
• If no incomplete run with a failure_moment exists, skip Phase 3b and note this.
• If no runs exist at all, report that and stop.

Phase 3: Triage Selected Runs

Spawn sub-agents in parallel for each available target. Do NOT instruct sub-agents to report verbosely or tell them they are being tested — this changes their behavior. The orchestrator will read their raw session traces directly in Phase 4.

Phase 3a: Completed run with failing properties

Spawn a general-purpose sub-agent with these instructions (substitute the actual run id and {{TRIAGE_SKILL}} with the absolute path to antithesis-triage/SKILL.md in this repository):

Read the skill file at {{TRIAGE_SKILL}} and follow its instructions to
triage the Antithesis run with run_id {{RUN_ID}}. Investigate the logs of
the failing property most likely to be a SUT bug. Pick any failing property
if you aren't sure. Do NOT explore or analyze any local source code
repositories — only use the snouty API and downloaded logs to perform your
triage.

Phase 3b: Incomplete run

Spawn a general-purpose sub-agent with these instructions (substitute the actual run id and {{TRIAGE_SKILL}} with the absolute path to antithesis-triage/SKILL.md in this repository):

Read the skill file at {{TRIAGE_SKILL}} and follow its instructions to
triage the Antithesis run with run_id {{RUN_ID}}.

Wait for both sub-agents to complete.

Phase 4: Review and Report

Understanding the skill protocol

Before auditing the sub-agent traces, read the triage skill and its references so you can distinguish correct behavior from violations:

1. Read antithesis-triage/SKILL.md — the main skill protocol
2. Read antithesis-triage/references/run-discovery.md — how runs are listed and selected
3. Read antithesis-triage/references/run-info.md — single-run metadata
4. Read antithesis-triage/references/properties.md — property triage workflow, including how counterexamples and moments map to log downloads
5. Read antithesis-triage/references/logs.md — log format and jq analysis patterns

Use these as the ground truth when evaluating compliance. If a sub-agent does something that looks unusual, check whether the skill documents it before flagging it as an issue.

Reading sub-agent session traces

The Agent tool returns an agentId for each sub-agent (visible in the result as agentId: <id>). Use the current session's ID and each agentId to read the raw JSONL session traces:

~/.claude/projects/<project-dir>/<session-id>/subagents/agent-<agentId>.jsonl

To find the current session directory, look for the most recently modified directory under ~/.claude/projects/ that contains a subagents/ folder.

Extract all Bash tool calls from each sub-agent trace to see every command and its output:

jq -c 'select(.type == "assistant") | .message.content[]?
  | select(.type == "tool_use" and .name == "Bash") | .input.command' \
  <trace-file>.jsonl

Extract tool results to see command outputs:

jq -c 'select(.type == "user") | .message.content[]?
  | select(.type == "tool_result") | {tool_use_id, content: .content[:500]}' \
  <trace-file>.jsonl

This gives you the complete record of what each sub-agent did — every snouty invocation, every jq query, every script run, every error — without relying on their summaries.

What to look for

Scan the raw traces for issues in these categories:

Bugs — things that are broken:

• snouty runs invocations that fail with non-timeout errors (404, schema mismatch, unexpected exit codes)
• The log-download step (snouty runs --json logs ... | jq -s '.') failing, timing out, or returning truncated / partial output
• jq queries that error out or return null where data should exist
• Skill-documented commands that do not behave as the skill claims

Compliance violations — the sub-agent did not follow the skill protocol:

• Reaching for the web UI / curl against the report URL when snouty runs exposes the data
• Skipping the snouty runs --json show step before triaging
• Triaging a run with no links.triage_report without first reporting that the run is not triageable
• Rounding or reformatting input_hash / vtime before passing them to snouty runs --json logs (the skill requires verbatim values)
• Concluding root cause for a property failure without downloading and inspecting the corresponding log when a moment is available
• Treating a telemetry/meta property counterexample (non-moment payload, e.g. a scalar or session-id object) as if it had a moment and trying to download logs for it

Inefficiencies — things that work but waste time or tokens:

• Downloading the same log file repeatedly to the same path
• Fetching all properties without --failing/--passing filters when only one side is needed
• Reading large logs into the context window instead of running jq filters against them
• Steps that required multiple retries (especially >2 retries for the same operation), which usually indicates unclear instructions in the skill

Do not flag repeated snouty runs invocations against the same run/endpoint as an inefficiency. snouty handles caching for repeated calls, so re-issuing the same snouty runs call (e.g. properties, show, build-logs) at different jq depths or filters is fine and is not a sign of a skill problem.

Output Format

Produce a concise report focused on actionable issues only. Do not include sections about what went well, generic suggestions, or broad recommendations. The goal is clear signal on breakage and inefficiency.

Test Summary

• Tenant tested
• Runs selected (run_id, status, failing-property counts where applicable)
• Sub-agents spawned (include agentIds)
• Any skipped phases and why

Issues

For each issue found, include:

• Severity: Bug / Compliance / Inefficiency
• Phase: Which phase the issue occurred in
• Description: What happened
• Evidence: Exact commands and outputs from the session trace (quote directly)
• Fix: Concrete change to make (file + what to change)

Sort by severity (bugs first, then compliance, then inefficiency).

If no issues are found, report "No issues found" — do not pad the report.