hoatv2211 Bug Hunt Swarm
Overview
Explore an unknown failure through independent read-only hypotheses, then converge on the smallest discriminating reproduction or instrumentation plan.
When to use
Use for intermittent crashes, unclear subsystem ownership, multi-service failures, race conditions, protocol mismatches, or symptoms with several plausible root causes.
When NOT to use
Do not use for reviewing an already known patch, implementing fixes, or parallel lanes that would need to mutate the same environment.
Required inputs and context discovery
Collect symptom, expected behavior, frequency, environment, timeline, known-good snapshot, logs, subsystem map, active processes, safe read-only commands, and integrator owner.
Safety and risk level
Lanes are read-only. They may propose instrumentation but cannot patch files, start or stop services, change databases, or modify shared fixtures.
Workflow
- Freeze one symptom statement and shared evidence packet. Completion criterion: all lanes investigate the same observable failure.
- Assign disjoint hypotheses such as data, timing, config, protocol, rendering, or environment. Completion criterion: each lane has a falsifiable question and suspect paths.
- Run independent read-only inspection and reproduction attempts. Completion criterion: each lane returns evidence, counterevidence, and confidence.
- Rank hypotheses by explanatory power and cost of the next experiment. Completion criterion: the integrator selects one discriminating action.
- Hand the selected action to
evidence-first-debuggingfor mutation or instrumentation. Completion criterion: the swarm ends before any lane writes files.
Evidence and output contract
Produce bug packets containing hypothesis, evidence, counterevidence, reproduction status, suspect paths, proposed experiment, confidence, and integrator rank.
Handoff contract
Record symptom, shared snapshot, lanes, commands, ranked hypotheses, unresolved evidence, and the next single experiment with its owner.
Pitfalls and anti-rationalization
- Lanes may not patch “obvious” suspects.
- Multiple opinions without falsifiable experiments are not evidence.
- Intermittent reproduction remains Unverified or BLOCKED.
- Do not let every lane inspect the same logs with the same hypothesis.
Verification checklist
- Symptom and snapshot are shared.
- Lanes are disjoint and read-only.
- Hypotheses are ranked with counterevidence.
- A single next experiment is selected.
- No lane mutated the repository or environment.
References and scripts
Pair the output with evidence-first-debugging. Use project logs and process/port inspection only when those commands are read-only.