debug

/debug - Diagnose one TASK failure

<input_contract> Expected $ARGUMENTS: one TASK-NNN-TN-FT-NNN-WN with an observed failure bound to the current Execution Attempt and actual change surface.

Read the indexed task card, direct task-linked specs, current-attempt protocol and handoff, functional/semantic verdict evidence, actual changes, and only the callers or boundaries needed to trace the failure. Missing reproduction or evidence is a valid diagnostic input and must remain explicit.

After root-cause confirmation, inspect only relevant prior diagnostic, execution, verification, bug, or repository-history evidence needed to compare the mechanism. Keep missing history explicit. </input_contract>

<hard_invariants>

• Diagnosis is read-only over implementation, tests, specs, task records, protocols, scheduler state, and external systems. The only allowed write is this command's report under .tasks/<TASK_ID>/.
• Do not change status, tier, dependencies, scope, Execution Attempt, retry budget, or permissions.
• Do not issue functional/semantic verdicts, accept product/spec/architecture decisions, implement a correction, or create BUG/follow-up work.
• Recurrence and prevention are advisory evidence. Do not infer recurrence from similar symptoms or absence of a prior match, create a durable rule without a confirmed material recurring mechanism, or expand current task authority.
• A probe may mutate only known local isolated or disposable state with safe rerun and cleanup. It grants no production, destructive, privileged, secret, or external permission. When safety or initial state is uncertain, do not run it and keep the diagnosis inconclusive.
• Keep tracing task-scoped. Neighboring callers and boundaries are evidence, not authority to expand the task. </hard_invariants>

<agent_discretion> Choose the smallest evidence-producing path. Useful tactics include reproducing the symptom with exact inputs/environment, reading the complete failure, tracing symptom -> immediate cause -> caller/source -> first violated invariant, and testing one evidence-backed hypothesis at a time with a distinguishing experiment. Condition-based waiting is usually preferable for async or flaky behavior. No fixed phase model, checklist, script, or mandatory failing test is required. </agent_discretion>

<required_outputs> Write exactly:

.tasks/<TASK_ID>/<TASK_ID>-S-DEBUG-final-report-docs-01.md

Include:

• symptom and reproduction, or the exact evidence gap;
• current Execution Attempt and actual change surface;
• confirmed root cause plus first violated invariant, or inconclusive;
• experiments and only materially useful rejected hypotheses;
• minimum recommended correction;
• suitable regression check;
• recurrence: repeated_confirmed|no_prior_evidence|unclassified, with inspected evidence scope. repeated_confirmed requires a prior failure with the same causal mechanism and first violated invariant; no_prior_evidence is scoped to inspected history; insufficient evidence is unclassified;
• for repeated_confirmed, prior evidence locators, the previous correction and why it failed to prevent recurrence, or the exact correction-history gap;
• for a material repeated_confirmed, one cheapest sufficient prevention guardrail — test, invariant, contract, lint/tooling check, runbook condition, or architecture boundary — with canonical artifact/owner, owning workflow, and adoption check; otherwise no guardrail and the reason;
• residual uncertainty and next owner;
• exactly one standalone marker:

DIAGNOSIS: CONFIRMED|INCONCLUSIVE