| name | diagnose |
| description | Opt-in evidence-first causal diagnosis for bugs, browser or app/device failures, flaky behavior, and performance regressions. Activate only when the user explicitly invokes `$diagnose` or explicitly asks to use the named diagnose skill. Do not activate merely because the user mentions a bug, asks why something failed, requests debugging or a fix, or describes unexpected behavior. Once explicitly invoked, locate likely causes, instrument relevant boundaries with extensive structured logging, reproduce the issue, analyze the collected logs, rank causes with confidence scores, and pursue 100% operational confidence while probes can increase confidence. |
Diagnose
Use this workflow only after the explicit opt-in described in the frontmatter. A normal bug report or debugging request does not opt in to this skill.
Find and explain the cause of a symptom with 100% operational confidence.
Start from the user's concrete anchor: file, route, error, log, screen, branch, artifact, or reproduction step. Consult relevant architecture notes, ADRs, glossaries, and test docs. Inspect code, run safe tools or tests, and analyze existing artifacts.
Explicit invocation of Diagnose constitutes approval for behavior-neutral temporary instrumentation within the requested scope. Ask separately only for risky actions, inaccessible-state reproduction, or mutations that affect product behavior or external systems.
References
Load only what applies:
Evidence Loop
-
Locate likely causes. Inspect the code path and existing evidence around the user's anchor. Form the smallest useful set of likely areas and falsifiable candidate causes, including independent or contributing causes when applicable. Give each cause a distinguishing prediction and identify where those predictions diverge.
-
Instrument before reproducing. Load instrumentation.md and add structured logging across the relevant boundaries, extensive enough to reconstruct the causal sequence rather than only record the visible symptom. Instrument competing causes in the same pass when practical so one reproduction can distinguish them.
Before reproducing, define the probe contract:
- the exact trigger and visible symptom
- the questions this run will answer
- each candidate cause's distinguishing prediction
- the events, fields, and correlation IDs that test those predictions
- how the evidence will be tied to the exact visible failure
- the intended process, build, window, document, and runtime identity when applicable
Do not run the reproduction if the probes cannot distinguish the leading candidates or cannot confirm that the exact symptom occurred.
-
Reproduce the instrumented issue. Use the fastest deterministic signal that represents the reported symptom: focused test, script, browser test, trace replay, harness, repeat loop, profiler, app/device automation, or structured human reproduction. Run it when safe and observable. For intermittent failures, run enough repetitions to compare causes.
When reproduction requires inaccessible state, credentials, devices, subjective interaction, or risky actions, first prepare the capture, exact steps, expected observation, and artifact to inspect. Ask the user to reproduce and reply done; confirm the returned evidence matches the reported failure.
If the user corrects the trigger, affected component, lifecycle, or visible symptom, invalidate evidence and probes that target the previous interpretation. Return to cause location and instrumentation before reproducing again; do not continue with a nearby but mismatched reproduction.
-
Collect and analyze. Collect the complete output, correlate events across boundaries, compare it with each prediction, and record evidence for and against every cause. If the logs are incomplete or ambiguous, identify how the likely areas or instrumentation must change before another reproduction.
-
Determine confidence. Rank each cause by role (root, contributing, or alternative), 0-100% confidence, supporting and conflicting evidence, and the next evidence that could change its score or rank. Treat scores as calibrated judgments, not statistical probabilities. Assign 100% operational confidence only when:
- the exact symptom is reproduced
- the causal chain from trigger to failure is observed
- controlling the suspected boundary reliably controls the symptom
- the cause or combination of causes explains every relevant observation
- plausible alternatives are falsified or included as contributing causes
- intermittent behavior is repeated enough to distinguish causality from coincidence
If proof is incomplete, choose the safest practical probe likely to add discriminating evidence and state which results would raise, lower, or redistribute confidence. Revise the likely areas or instrumentation and repeat the loop only while such a probe exists; do not repeat an equivalent pass without changing its inputs, boundary, or instrumentation.
Stop as incomplete when no available probe should increase confidence, a pass adds no evidence and no distinct boundary remains, or the needed evidence requires unavailable access, user action, unacceptable risk, or disproportionate effort. Never label an incomplete diagnosis as proven.
Output
Lead with one status:
- Proven — 100%: every applicable cause meets the proof standard with no unresolved plausible alternative.
- Incomplete: the proof standard is not met.
For Proven, report each cause, role, causal chain, decisive evidence, exact code or runtime boundary, and interaction between causes. For Incomplete, rank candidates with confidence, supporting and conflicting evidence, remaining uncertainty, why the loop stopped, and the smallest evidence needed to continue. Always name the validation signal that preserves the causal evidence.
Temporary Diagnostics
Keep diagnostic logs until the user requests cleanup or they become durable diagnostics. If committing while temporary diagnostics remain, stage only intended durable changes and report what remains unstaged.
Delegation
Use subagents only when the user explicitly requests delegation, subagents, parallel work, or token optimization. Delegate bounded mechanical collection to faster or cheaper agents: known commands or repro loops, logs, screenshots, traces, profiles, extraction, or artifact comparison. Give each task a success condition, output format, runtime boundary, and no-edit instruction unless mutation is explicit.
Keep feedback-loop design, hypothesis ranking, confidence scoring, ambiguous interpretation, and the final diagnosis in the main thread. Verify delegated evidence and inspect the working tree and background processes before continuing.
