Diagnose

Skill: /diagnose — Complex Bug Diagnostic Pipeline

When to invoke (and when NOT to)

Use /diagnose when:

• Bug reproduces but root cause is unclear after one read-pass of the failing code
• Previous fix attempts have been reverted ≥ 2 times (symptoms return)
• Failure is intermittent (flaky test, race condition, timing-dependent)
• Failure occurs in unfamiliar code (agent has no prior context)
• User has explicitly requested /diagnose or "deep investigation"
• Circuit Breaker (Rule 14) tripped on the specialist agent that normally handles this domain

Do NOT use /diagnose when:

• Cause is obvious (null ref, typo, missing import, incorrect import path)
• Fix is Investigation -> Verification -> Solution -> Lead Programmer

(signal) (hypothesis) (devil's adv.) (tradeoffs) (assign + exec)

repro/check command investigation.json verification.json solution.json implementation (fast deterministic (rootcause, (status: confirmed | (3 options: (delegates to pass/fail signal) evidence[], refuted | inconclusive, Quick/Strategic/ backend-developer, confidence) reproductionsteps) Future-Proof) qa-engineer, etc.)

Each stage produces a **required artifact** saved to `.investigations//` and a **handoff contract** (per Rule 16) to the next agent.

## Stage 0 — Feedback Loop

**Goal:** Build the fastest reliable pass/fail signal for the exact symptom
before explaining the cause.

The feedback loop is the highest-leverage part of diagnosis. Do not proceed to
root-cause analysis until there is a loop that can reproduce the user's symptom
or a documented reason why no loop is possible.

Try these in roughly this order:

1. Failing test at the seam that reaches the bug.
2. CLI or script invocation with fixture input and expected output.
3. Curl/HTTP request against a running service.
4. Headless browser script with DOM, console, or network assertions.
5. Replay of a captured payload, event, HAR, log, or trace.
6. Throwaway harness that calls the affected code path in isolation.
7. Property/fuzz loop for broad wrong-output symptoms.
8. Bisection or differential loop between known-good and known-bad states.

Improve the loop before investigating:

- Faster: remove unrelated setup and narrow the command.
- Sharper: assert the specific symptom, not merely "does not crash".
- More deterministic: pin time, seed randomness, isolate filesystem/network, or
  raise intermittent reproduction frequency with stress runs.

Do not treat Stage 0 as warm-up. It is the main leverage point. A bad loop
produces fake certainty, weak hypotheses, and symptom-only fixes.

If no credible loop can be built, stop and report what was tried. Ask for access
to the reproducing environment, a captured artifact, or permission to add
temporary instrumentation. Do not proceed on a vibe.

## Stage 1 — Investigation

**Agent:** `diagnostics` (Investigation role)
**Goal:** Produce ranked falsifiable root-cause hypotheses backed by empirical evidence.

### Inputs
- Symptom description (from user or TODO.md bug ID)
- Reproduction steps (or "cannot reproduce" + environment)
- Relevant log lines, stack traces, error IDs
- Feedback loop command/check from Stage 0, or a documented reason no loop can
  currently be built

### Required output — `investigation.json`
```json
{
  "task_id": "BUG-417",
  "symptom": "POST /api/orders returns 500 when cart has ≥10 items",
  "reproduction": {
    "steps": ["...", "..."],
    "frequency": "100% | intermittent (~30%) | once",
    "environment": "staging-eu-west-1"
  },
  "feedback_loop": {
    "command": "npm test -- checkout.e2e.test.ts",
    "signal": "Fails with timeout before hydration marker appears",
    "reliable": true
  },
  "ranked_hypotheses": [
    {
      "rank": 1,
      "cause": "Test clicks #submit before React hydration completes on slow CI runners",
      "prediction": "Waiting for the hydration marker will make the failure disappear without adding a fixed sleep"
    },
    {
      "rank": 2,
      "cause": "Submit button selector matches a hidden stale node",
      "prediction": "Asserting the visible button count will expose multiple matching nodes"
    }
  ],
  "hypothesis": {
    "root_cause": "OrderService.calculateTotal() N+1 query exhausts pool when cart.items.length > 9",
    "confidence": "high | medium | low",
    "falsifiable_by": "Run with pool_size=50; if error disappears, cause confirmed"
  },
  "evidence": [
    {"type": "log", "ref": ".investigations/BUG-417/pg-pool-exhausted.log", "summary": "..."},
    {"type": "code", "ref": "src/services/order.service.ts:142", "summary": "Unbounded .map+await"}
  ],
  "unknowns": ["Why only eu-west-1?", "When did this start?"],
  "next_agent": "diagnostics",
  "next_stage": "verification"
}

Quality gate (Lead Programmer rejects if):

• feedback_loop is missing and no blocked-loop explanation exists
• feedback_loop.signal is vague, broad, or does not isolate the user's symptom
• ranked_hypotheses has fewer than 3 items unless the evidence makes a single

cause unavoidable

• Any hypothesis lacks a falsifiable prediction
• hypothesis.falsifiable_by is vague ("check if it works")
• evidence has fewer than 2 items (unverifiable)
• unknowns is empty but confidence: low (contradictory)

Stage 2 — Verification

Agent: diagnostics (Verification role) Goal: Attempt to refute the hypothesis. Only confirmed if refutation fails.

Inputs

• investigation.json (from Stage 1)
• Access to staging/test environment
• The Stage 0 feedback loop, rerun before and after each meaningful probe

Verification is invalid if it does not go back through the Stage 0 loop. A fix-looking local observation that bypasses the loop is not confirmation.

Required output — verification.json

{
  "task_id": "BUG-417",
  "status": "confirmed | refuted | inconclusive",
  "triangulation": [
    {"method": "reproduce_with_fix_applied", "result": "Error gone with pool_size=50"},
    {"method": "reproduce_without_fix", "result": "Error returns at 10 items"},
    {"method": "adjacent_test_case", "result": "9 items = OK, 10 items = fail → threshold confirmed"}
  ],
  "counter_hypotheses_ruled_out": [
    "DB slowness (ruled out: p99  100`)
2. Select option → write selection to `.investigations//decision.md`
3. Create A2A handoff contract (Rule 16) via `/handoff`:
   - `lead-programmer → backend-developer` (or `frontend-developer`, `data-engineer`)
   - Acceptance criteria derived from `investigation.hypothesis.falsifiable_by`
4. Append ledger entry (Rule 15) to `production/traces/decision_ledger.jsonl`:
```jsonl
{"ts":"2026-04-17T14:22:00Z","session":"main","agent_id":"lead-programmer","task_id":"BUG-417","request":"/diagnose BUG-417","reasoning":"Verified N+1 as necessary+sufficient; selected Strategic per Solution-stage recommendation","choice":"Strategic refactor of calculateTotal()","outcome":"pass","risk_tier":"Medium","duration_s":1840}

Artifact storage

All intermediate reports MUST be saved to .investigations//:

.investigations/
└── BUG-417/
    ├── investigation.json     # Stage 1 output
    ├── verification.json      # Stage 2 output
    ├── solution.json          # Stage 3 output
    ├── decision.md            # Stage 4 — human-readable rationale
    ├── evidence/              # logs, screenshots, traces referenced in reports
    └── handoffs/              # A2A contracts (copied from .tasks/handoffs/)

Retention: Keep until bug is closed + 30 days, then archive to .investigations/archive/.

Escalation paths

| Trigger | Escalate to | | ------------------------------------------------- | ---------------------------------------------- | | diagnostics circuit OPEN in Investigation (Rule 14) | Surface raw symptom to user; skip pipeline, request manual triage | | Verification stage returns inconclusive twice | Surface to user; request manual reproduction | | Solution stage cannot produce 3 distinct options | Escalate to technical-director — scope unclear | | User rejects all 3 options | Return to Investigation stage; hypothesis likely wrong | | Bug reoccurs after fix merges | Restart /diagnose with new task_id; link prior investigation in unknowns[] |

Integration with coordination rules

• Rule 6 (Layered Recovery): If any stage fails once, retry with fresh context before escalating
• Rule 14 (Circuit Breaker): Read production/session-state/circuit-state.json before invoking each agent
• Rule 15 (Decision Ledger): Every stage transition logs to decision_ledger.jsonl
• Rule 16 (A2A Handoff): Stage 1→2, 2→3, 3→4 each require a handoff contract in .tasks/handoffs/

Concrete example — "Flaky checkout test"

Symptom: checkout.e2e.test.ts fails ~20% of CI runs; local always passes.

/diagnose flaky-checkout-e2e
  ↓
Stage 1 → Investigation (diagnostics)
  hypothesis: "Test clicks #submit before React hydration completes on slow CI runners"
  evidence: [CI traces showing hydration marker missing, local has DevTools overhead masking timing]
  confidence: medium (cannot reproduce locally)
  ↓
Stage 2 → Verification (diagnostics)
  triangulation:
    - Inject 500ms delay before click → test passes 50/50 runs ✓
    - Remove delay → fails 9/50 ✗
    - Check for hydration marker instead of fixed delay → passes 50/50 ✓
  status: confirmed
  ↓
Stage 3 → Solution (diagnostics)
  Quick: add sleep(500ms)              [masks issue]
  Strategic: waitFor hydration marker  [addresses root cause]
  Future-Proof: custom test util that always waits for RSC boundary [reusable]
  recommendation: Strategic
  ↓
Stage 4 → lead-programmer
  selects Strategic; assigns to qa-engineer
  handoff contract: "qa-engineer updates checkout.e2e.test.ts to use waitFor(hydrationMarker)"
  acceptance_criteria: ["10 CI runs in a row pass", "no sleep() in test"]
  ledger entry written

Common pitfalls

| Pitfall | Fix | | --------------------------------------------- | --------------------------------------------------------------------------- | | Hypothesizing before building a loop | Return to Stage 0. A diagnosis without a signal is speculation. | | Keeping a slow or blurry loop | Rework Stage 0 until the signal is fast, specific, and trustworthy. | | Skipping Verification ("cause is obvious") | Verifier exists specifically to catch "obvious but wrong" hypotheses | | Investigator produces only 1 hypothesis | Reject unless evidence makes alternatives impossible; require ranked hypotheses and predictions | | Solver picks Quick fix without naming tradeoff| Reject — all 3 options required for explicit tradeoff comparison | | No artifact written to .investigations/ | Reject — verbal diagnosis is not auditable | | Running /diagnose in parallel on same bug | Only one active investigation per task_id; concurrent runs create race |

Output to user

After Stage 4 completes, summarize in ≤ 5 lines:

/diagnose BUG-417 complete.
Root cause: Unbounded .map+await in OrderService.calculateTotal() exhausts pg pool.
Selected: Strategic (batch via IN-clause, ~40 LOC, Medium risk).
Assigned: @backend-developer; acceptance = load test with 50 items passes.
Artifacts: .investigations/BUG-417/

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tranhieutt
• Source: tranhieutt/softwaredevelopment_department
• License: MIT

Install and usage instructions live in the source repository linked above.