maestro-diagnose

Find the cause; change nothing. Any tier may use it.

1. Reproduce. Capture the exact failing command or input and its output. No reproduction: say so and stop; do not guess from symptoms.
2. Localize. Bisect the surface: which layer, module, commit, or input feature flips the behavior. Prefer evidence (logs, minimal harnesses, git bisect) over reading and speculating.
3. Reduce. Shrink to the smallest input or state that still fails.
4. Deliver. Root cause plus its evidence chain, the reduced reproduction, and the smallest fix direction. With a work item: maestro work note <id> carries the one-line cause and the reproduction command; in a bundle, also the NOTES.md Current State. Otherwise the conversation.

Discipline

• Every probe follows a written hypothesis, "I believe X causes this because Y", then the minimal test that can falsify it. No hypothesis, no probe.
• Recall first: maestro search "<symptom>" and maestro attention. A prior failed: note on the same mechanism is evidence about which hypotheses are already refuted.
• Three refuted hypotheses in a row: stop treating this as a local bug. Question the architecture; a structural cause or an unsettled design decision is maestro-design work, not a fourth probe.
• Red flags that mean return to step 1: proposing a fix before the cause is known, changing several things at once, "probably X" without a check.

Diagnosis is read-only toward production code: instrumentation and harnesses live outside production paths or are reverted before delivery.

An explicit fix request routes by target and tier (maestro-bundle tier rule): a quickfix or Light fix proceeds directly with the diagnosis as its first step; Full-tier scope is maestro-work inside the bundle, with the diagnosis as its first phase; a prototype fix belongs to maestro-explore. Never downgrade a requested fix to analysis.