Hai Debug
For Chinese readers, see SKILL.zh_CN.md. English is the execution source of truth.
Status: trial. The method has not yet demonstrated improvement across representative real tasks.
Principle
An explanation must account for the observed failure and predict a check that could disprove it.
A plausible patch is not causal evidence. Begin with the cheapest check that separates likely causes.
Workflow
- Establish expected versus actual behavior, scope, timing, environment/version, and requested
action (diagnose or fix). Read available logs, errors, code, recent changes, and existing tests
before asking for information that is already available.
- Reproduce on an isolated local fixture when possible. Record inputs, preconditions, command,
and observed result. For intermittent failures, capture frequency/order/concurrency and a
bounded repeat strategy; do not loop indefinitely or hammer live services.
- Trace the symptom backward through callers, state/config, persistence, and external boundaries.
Compare working and failing cases. Distinguish observations, assumptions, and historical claims.
- Form a small set of plausible competing hypotheses. For each, record supporting/conflicting
evidence and one cheap discriminating check. A trivial obvious cause needs no ceremonial table.
- Run the highest-information safe check, update confidence, and eliminate contradicted
hypotheses. Instrument narrowly if authorized by the fix request; diagnose-only work uses
read-only checks or separate disposable reproductions, not edits to user code.
- Explain the causal chain: precondition → faulty decision/state → observable symptom.
Cite actual file:line/log/test evidence. A failed reproduction leaves the cause unconfirmed;
report the best next check and missing evidence rather than guessing.
- If only diagnosis was requested, stop with the cause and a proposed repair/regression target.
If fixing was requested, continue through the smallest complete repair and verification:
use
hai-tdd when a meaningful behavioral RED exists; otherwise use honest targeted
integration/runtime checks. Reproduce the original case again and check nearby contracts.
- Deliver using
references/output-template.md. Separate symptom mitigation from root-cause
repair and executed checks from proposed checks.
Constraints and stop rules
Respect the user's environment and authority. Reproduction does not authorize destructive
production operations, external messages, paid calls, or traffic load. Prefer isolated fixtures.
After repeated checks yield no new evidence, change the hypothesis or evidence source; identify
the exact blocker if progress needs unavailable data. Elapsed time is not proof.
Do not broadly refactor, suppress errors, add blind retries, or change expected behavior merely
to make the symptom disappear. A mitigation may be useful but must be labeled and bounded.
Preserve unrelated work and remove only temporary instrumentation created for this task when
its diagnostic purpose is complete.
Handoffs
- Cause known, implement a behavior change →
hai-tdd when suitable, otherwise normal execution.
- System hard to change without a specific failure →
hai-architecture.
- Review a completed change →
code-review-and-quality.
- Prove the full change against requirements →
write-technical-acceptance-report.
- Multi-phase repair with unresolved execution dependencies →
hai-goal.
1---2name: hai-debug3description: Diagnoses an unexplained software failure through reproduction, competing hypotheses, discriminating checks, and a causal explanation. Use for 排障、根因分析、偶发失败、环境差异、重复执行, or a bug whose cause is unclear. Return observed evidence, ruled-out hypotheses, cause/confidence, and a regression target; fix and verify only when the request includes fixing. Use hai-tdd when the cause and desired behavior are already known, and hai-architecture for change complexity without a malfunction.4---56# Hai Debug78For Chinese readers, see `SKILL.zh_CN.md`. English is the execution source of truth.9Status: trial. The method has not yet demonstrated improvement across representative real tasks.1011## Principle1213An explanation must account for the observed failure and predict a check that could disprove it.14A plausible patch is not causal evidence. Begin with the cheapest check that separates likely causes.1516## Workflow17181. Establish expected versus actual behavior, scope, timing, environment/version, and requested19 action (diagnose or fix). Read available logs, errors, code, recent changes, and existing tests20 before asking for information that is already available.212. Reproduce on an isolated local fixture when possible. Record inputs, preconditions, command,22 and observed result. For intermittent failures, capture frequency/order/concurrency and a23 bounded repeat strategy; do not loop indefinitely or hammer live services.243. Trace the symptom backward through callers, state/config, persistence, and external boundaries.25 Compare working and failing cases. Distinguish observations, assumptions, and historical claims.264. Form a small set of plausible competing hypotheses. For each, record supporting/conflicting27 evidence and one cheap discriminating check. A trivial obvious cause needs no ceremonial table.285. Run the highest-information safe check, update confidence, and eliminate contradicted29 hypotheses. Instrument narrowly if authorized by the fix request; diagnose-only work uses30 read-only checks or separate disposable reproductions, not edits to user code.316. Explain the causal chain: precondition → faulty decision/state → observable symptom.32 Cite actual file:line/log/test evidence. A failed reproduction leaves the cause unconfirmed;33 report the best next check and missing evidence rather than guessing.347. If only diagnosis was requested, stop with the cause and a proposed repair/regression target.35 If fixing was requested, continue through the smallest complete repair and verification:36 use `hai-tdd` when a meaningful behavioral RED exists; otherwise use honest targeted37 integration/runtime checks. Reproduce the original case again and check nearby contracts.388. Deliver using `references/output-template.md`. Separate symptom mitigation from root-cause39 repair and executed checks from proposed checks.4041## Constraints and stop rules4243Respect the user's environment and authority. Reproduction does not authorize destructive44production operations, external messages, paid calls, or traffic load. Prefer isolated fixtures.45After repeated checks yield no new evidence, change the hypothesis or evidence source; identify46the exact blocker if progress needs unavailable data. Elapsed time is not proof.4748Do not broadly refactor, suppress errors, add blind retries, or change expected behavior merely49to make the symptom disappear. A mitigation may be useful but must be labeled and bounded.50Preserve unrelated work and remove only temporary instrumentation created for this task when51its diagnostic purpose is complete.5253## Handoffs5455- Cause known, implement a behavior change → `hai-tdd` when suitable, otherwise normal execution.56- System hard to change without a specific failure → `hai-architecture`.57- Review a completed change → `code-review-and-quality`.58- Prove the full change against requirements → `write-technical-acceptance-report`.59- Multi-phase repair with unresolved execution dependencies → `hai-goal`.
hylarucoder