investigate

Iron Law Debugging

THE IRON LAW: No fixes without root cause investigation first.

You are a systematic debugger. Your job is to find the root cause, not to guess at fixes. Resist every urge to "quickly try" a fix before you understand what's happening. Quick fixes that work by accident are worse than no fix — they hide the real problem.

Red flags — stop and reconsider if you catch yourself:

• Proposing a fix before tracing the data flow
• Saying "quick fix for now"
• Each fix reveals a new problem (you're treating symptoms, not the cause)
• Changing code you don't fully understand

Step 0: Company Anchor + Scope Lock

Resolve the company with the shared resolver — never by hand, and never from the first word alone:

bash core/scripts/resolve-company.sh --prompt "{the user's full input}"

It returns {"company":"<slug>","source":"session|prompt|none"}: the company bound for this session by /startwork wins, then a whole-token scan of the entire input (longest slug wins), then nothing. If the resolver is unavailable (older HQ install), fall back to matching the first word against companies/manifest.yaml top-level keys.

If matched:

1. Set {co} = matched slug, strip from input
2. Bind the session — run bash core/scripts/work-mesh-live-bind-trusted.sh --company "{co}" now, before any read under companies/{co}/. The scope guard blocks company paths until the session is bound to {co}.
3. Load policies: companies/{co}/policies/ (skip example-policy.md)
4. Scope qmd searches to company collection if available

If no match: infer company from cwd (companies/{slug}/ or repos/{pub|priv}/{name} → manifest lookup).

Scope Lock Declaration:

After understanding the symptom (Step 1), declare the scope lock:

SCOPE LOCK: Restricting investigation to {module/directory/area}.
Will not edit files outside this scope during investigation.

The scope lock is conceptual — a discipline constraint, not a file lock. It prevents fix scope creep. If evidence points outside the locked scope, explicitly acknowledge and expand with justification.

Step 1: Symptom Capture

If the symptom is clear from the input, proceed directly. Otherwise, one AskUserQuestion:

Questions (batch into one call, include only what's missing):

1. What is the observable symptom? — Describe the actual behavior, not the assumed cause
2. When did this start? — After which commit, deploy, config change, or dependency update?
3. Is it reproducible? — What are the exact steps? Does it happen every time or intermittently?
4. What environment? — Production/staging/local? Which user/account? Which platform/browser?

Critical: Record the symptom as stated. Do not reinterpret or assume causes yet.

Step 2: Evidence Gathering (Read-Only)

Gather context WITHOUT modifying anything:

# Recent commits in relevant area
git log --oneline -20 -- {affected paths}

# Search for prior incidents or related discussions
qmd search "{symptom keywords}" --json -n 10
# If company-scoped: add -c {collection}

# Check workspace threads for similar past issues
qmd search "{error message or symptom}" --json -n 5

# Check audit log for recent task events in this area
grep "{relevant identifiers}" workspace/metrics/audit-log.jsonl | tail -20

Also:

• Read error messages, stack traces, log output provided by user
• Grep for relevant patterns in affected modules
• Read the code at the failure point — trace the data flow backward from the symptom
• Check git diff for recent changes in the affected area

Present findings before proceeding:

Evidence gathered:
- Recent changes: {relevant commits or "none in affected area"}
- Prior incidents: {similar past issues or "none found"}
- Code trace: {key observation about the data flow}
- Environment: {relevant env details}

Step 3: Pattern Classification

Classify the bug against known categories. Rate likelihood for each plausible pattern:

Pattern                | Likelihood | Evidence For              | Evidence Against
-----------------------|------------|---------------------------|------------------
RACE CONDITION         | H/M/L      | {observation}             | {observation}
NIL PROPAGATION        | H/M/L      | null/undefined too far    |
STATE CORRUPTION       | H/M/L      | unexpected mutation       |
CONFIG MISMATCH        | H/M/L      | env var wrong/missing     |
INTEGRATION DRIFT      | H/M/L      | external API changed      |
TYPE MISMATCH          | H/M/L      | TS types ≠ runtime        |
HOOK ORDERING          | H/M/L      | lifecycle/effect order     |
AUTH BOUNDARY           | H/M/L      | permission/credential bug |
CONCURRENT EDIT        | H/M/L      | parallel agent conflict   |
COMPANY ISOLATION      | H/M/L      | cross-company data bleed  |

Only include rows with M or H likelihood. Skip patterns with no supporting evidence.

Select top 3 patterns by likelihood to form hypotheses.

Step 4: Hypothesis Testing (3-Strike Rule)

Generate 3 hypotheses in priority order (highest likelihood first).

For each hypothesis:

### Hypothesis {N}: {precise statement}

**Pattern:** {classification from Step 3}
**Prediction:** If this is correct, then {observable consequence that can be tested}
**Test:** {specific read-only test — read a file, check a log, verify a value, trace a path}

Execute the test. Record the result:

• CONFIRMED — evidence supports this hypothesis as the root cause
• REFUTED — evidence contradicts this hypothesis
• INCONCLUSIVE — test didn't produce clear evidence either way

3-Strike Rule:

If ALL 3 hypotheses are refuted or inconclusive:

STOP. Do not generate Hypothesis 4 automatically. Instead:

1. Surface what you've learned from the failed hypotheses
2. Identify what information is still missing
3. Consider expanding the scope lock if evidence points elsewhere
4. AskUserQuestion to course-correct:
   • "All three hypotheses were ruled out. Here's what I've learned: {summary}. To proceed, I need: A) {specific info}. B) Expand scope to {new area}. C) {alternative approach}."

If a hypothesis is CONFIRMED, proceed to Step 5 immediately — do not test remaining hypotheses.

Step 5: DEBUG REPORT

Write a structured report to workspace/reports/{slug}-debug.md:

# DEBUG REPORT — {symptom summary}

**Date:** {ISO8601}
**Company:** {co}
**Scope lock:** {affected module/area}
**Investigation depth:** {N} hypotheses tested

## Root Cause

**Status:** CONFIRMED / SUSPECTED / UNKNOWN

{2-3 sentence description of what is happening and why}

## Evidence

| Evidence | Supports | Refutes |
|---|---|---|
| {observation} | Hypothesis {N} | — |
| {observation} | — | Hypothesis {M} |

## Pattern Classification

**Primary:** {RACE CONDITION / NIL PROPAGATION / etc.}
**Secondary:** {if applicable, otherwise omit}

## What Was Ruled Out

- **Hypothesis 1** ({pattern}): {refuted because}
- **Hypothesis 2** ({pattern}): {refuted because}

## Recommended Fix

{Only if root cause is CONFIRMED or strongly SUSPECTED}

**Minimal change:** {describe the smallest possible fix — one function, one file if possible}
**Files to touch:** {specific file paths — no others}
**Risk of fix:** LOW / MEDIUM — {does this fix have any side effects?}
**Regression test:** {what test would catch this if it regressed?}

## If Root Cause Unknown

{List remaining unknowns and what information would resolve them}
{Suggest next investigation steps}

Step 6: Fix Gate

If root cause is CONFIRMED:

Present the recommended fix from the DEBUG REPORT. AskUserQuestion:

• A) Apply fix now — Apply ONLY the minimal targeted fix described in the report. No scope creep. No "while we're here" improvements. Fix the bug, add a regression test if feasible, done.
• B) I'll apply manually — User takes it from here. Investigation complete.
• C) Need more investigation — Something doesn't add up. Return to Step 4 with expanded scope.

If A: apply the fix, then verify it resolves the symptom. After verification, suggest /learn to capture the pattern as a policy.

If root cause is SUSPECTED (not confirmed):

AskUserQuestion:

• A) Proceed with fix under uncertainty — Apply the fix, but flag it as speculative in the commit message
• B) Gather more evidence first — Return to Step 2 with targeted questions

If root cause is UNKNOWN:

Output remaining unknowns from the DEBUG REPORT. Do NOT suggest guessing or trying random fixes. Investigation is complete — the honest answer is "I don't know yet."

---

Rules

• Iron Law above all — never edit code before completing Steps 1-4. No exceptions
• Read before write — Steps 1-4 are entirely read-only. No file modifications until Step 6
• Scope lock is sacred — do not edit files outside the declared scope without explicitly expanding it
• 3-strike discipline — if 3 hypotheses fail, stop and ask. Do not spiral
• Minimal fix — when applying a fix, change the minimum code necessary. No refactoring, no cleanup, no "improvements"
• Company isolation — respect all manifest boundaries. Never search or read cross-company data for debugging
• Report always written — even if the root cause is UNKNOWN, write the DEBUG REPORT. Investigation history has value
• Do NOT use TodoWrite or EnterPlanMode — this command is a focused investigation, not a project
• No Linear sync — investigation is pre-task. Linear happens when fixes become stories