Flaky Test Diagnoser

Runs competing-hypothesis experiments to isolate the true root cause of a flaky test. Produces a diagnosis with reproducible evidence — or an honest inconclusive label. Never promotes a guess to "root cause."

This is a tracing skill, not a fixer. The goal is to explain why the test is intermittent with falsifiable evidence, not to jump into patching code.

Execution contracts

Subagent watchdog: any time this skill spawns a test-running subagent via run_in_background=true (multi-run protocols, ordering bisection, distinguishing experiments), it MUST be armed with a staleness monitor per _shared/subagent-watchdog.md. Use Flavor A (Monitor tail per spawn) with thresholds STALE=15 min, HUNG=45 min — test suites can legitimately take a long time, especially with N≥10 reruns, but a 45-min silent window is pathological. TaskOutput status reports PID liveness only; output-file mtime is the ground truth for progress. Contract inheritance: timed_out_heartbeat joins this skill's termination vocabulary as a peer of blocked_by_environment (watchdog kill is a different failure class — the test-runner subagent hung, not the test environment). A watchdog-killed experiment contributes zero runs to the N≥10 count; the skill must retry or escalate.

Workflow

1. Bootstrap run state — create flaky-diag-{run_id}/ in CWD with state.json, hypotheses/, runs/, experiments/. See STATE.md.
2. Detect test runner — identify the project's test framework and runner command from config files. See RUNNERS.md.
3. Confirm flakiness — run the target test N>=10 times in isolation, record pass/fail per run, compute fail rate. Record each run under runs/. See EXPERIMENTS.md.
4. Generate competing hypotheses — produce >=3 deliberately different hypotheses from the 6 categories (ORDERING, TIMING, SHARED_STATE, EXTERNAL_DEPENDENCY, RESOURCE_LEAK, NON_DETERMINISM). Each hypothesis gets its own file in hypotheses/ following the schema in FORMAT.md. See ANALYSIS.md for category signatures.
5. Gather evidence for AND against each hypothesis — run the structured experiments (isolation, ordering bisection, timing, environment, code reads) and record findings as evidence_for / evidence_against entries per hypothesis. See EXPERIMENTS.md.
6. Design a distinguishing experiment per surviving hypothesis — an experiment whose outcome uniquely discriminates that hypothesis from its rivals. An experiment that "merely supports" all hypotheses does NOT count. See FORMAT.md.
7. Run distinguishing experiments — record outcomes under experiments/. Down-rank hypotheses whose distinctive prediction failed. See GOLDEN-RULES.md for the falsifiability gate.
8. Rebuttal round — let the second-ranked hypothesis present its best rebuttal against the leader. The leader must answer with evidence, not assertion.
9. Attempt red-green-red demonstration — if a leader has been selected, describe the minimal fix and have the user apply it (or apply it in an experimental branch). Rerun the multi-run protocol (N>=10). Only label root_cause_isolated_with_repro if the red-green-red cycle succeeds.
10. Emit structured output and diagnosis report — each hypothesis gets a STRUCTURED_OUTPUT_START / STRUCTURED_OUTPUT_END block. Final report follows REPORT.md.

Honest termination labels

Pick exactly one. Never invent synonyms.

Never use: "diagnosed", "likely root cause", "probably", "should be", "root cause found" as a terminal label without a red-green-red demo. See GOLDEN-RULES.md.

Self-review checklist

Before delivering the report, verify ALL:

• State file exists at flaky-diag-{run_id}/state.json with generation >= 1
• >=3 competing hypotheses were generated, each with its own file in hypotheses/
• Every hypothesis has both evidence_for AND evidence_against entries (never empty)
• Every hypothesis has an explicit uncertainty_level in {high, medium, low}
• Every hypothesis has a distinguishing_experiment field (or is explicitly marked "no distinguisher available within budget")
• Flakiness was confirmed: at least one fail AND at least one pass in N>=10 runs
• Isolation vs in-suite comparison completed (both were run)
• If a hypothesis is promoted to root_cause_isolated_with_repro: a red-green-red demonstration exists with N>=10 post-fix runs (all green)
• Report uses one of the 4 honest termination labels — never "diagnosed" without red-green-red
• Each hypothesis evaluation has a STRUCTURED_OUTPUT_START / STRUCTURED_OUTPUT_END block per FORMAT.md
• Anti-rationalization counter-table was consulted before finalizing the label (GOLDEN-RULES.md)

Golden rules

Hard rules. See GOLDEN-RULES.md for the full text, anti-rationalization counter-table, and the falsifiability gate.

1. Never promote a hypothesis to root cause without a distinguishing experiment that empirically rejects all rivals.
2. "Diagnosed" requires red-green-red. Test fails → fix applied → test passes deterministically across N>=10 runs. Without it, the label is narrowed_to_N_hypotheses.
3. Always run isolation before ordering. If the test fails in isolation, ordering analysis is irrelevant.
4. Bisect, never brute-force when searching for an interfering test.
5. Capture exact commands verbatim. Every experiment logs its shell command so the user can reproduce it.
6. Minimum 10 runs for any statistical claim. Flaky tests can have fail rates under 10%.
7. Never modify test code during diagnosis beyond temporary instrumentation that is reverted before the report is emitted.
8. Evidence for AND against every hypothesis. A hypothesis with only supporting evidence is unfalsifiable and must be flagged.
9. Coordinator does not self-approve. The hypothesis ranking is based on experiment outcomes recorded in files, not coordinator opinion.
10. Structured output is the contract. Per-hypothesis evaluations live between STRUCTURED_OUTPUT_START/STRUCTURED_OUTPUT_END markers; free-text commentary is ignored when computing the verdict.

Cancellation / resume

If interrupted mid-run: on next invocation read flaky-diag-{run_id}/state.json, identify the highest-generation completed stage, and replay from the next stage. Never recompute completed experiments — trust the recorded run artifacts. See STATE.md for the replay protocol.

Reference files

Execution routing (sagaflow-first — Contract 6)

Sagaflow is the default execution path. The in-session workflow above is the FALLBACK, used only when the sagaflow worker is confirmed unavailable. See _shared/execution-model-contracts.md Contract 6.

Routing sequence (mandatory before any in-session work):

1. Run sagaflow doctor
2. If healthy → launch via sagaflow below. Stop. Do not run in-session.
3. If unhealthy → log SAGAFLOW_UNAVAILABLE, proceed with in-session fallback.

Launch command:

Bash(
  run_in_background=true,
  command="sagaflow launch flaky-test-diagnoser --arg test='<TEST_ID>' --arg command='<RUN_CMD>' --arg n_runs=<N> --await"
)

Both test and command are REQUIRED. Substitute <TEST_ID> with the test identifier (e.g. tests/test_foo.py::test_bar), <RUN_CMD> with the exact shell command that runs that test, and <N> with the stability-run count (default 10). The workflow writes ~/.sagaflow/runs/<run_id>/report.md with pass/fail table, isolation+ordering results, timing analysis, top hypothesis with evidence, and proposed fix or mitigation.