oduffy-delphi/systematic-debugging

/systematic-debugging — Root-Cause Discipline for a Single Issue

<!-- provenance: authored 2026-05-30 to replace the retired superpowers:systematic-debugging skill (D-032 conscious uncoupling). Philosophy: docs/wiki/eager-agent-calibration.md — redirection not friction. -->

Walk one identified failure to its root cause and fix it there. The skill exists because the fastest fix that actually holds starts by finding the cause: guessing averages slower, because every wrong patch adds a new variable and can mask the real one. This isn't a discipline tax — it's the shortest path to a fix that survives the next run.

This is the single-issue tool. You have a specific symptom in hand. Distinct from siblings:

• /bug-sweep — searches the whole repo for unknown bugs (broad, discovery).
• /bug-blitz — grinds a known backlog of many bugs (broad, batch).
• /dogfood — exercises a newly-built thing until it works (broad, shakedown).
• /systematic-debugging — one known failure, traced to root cause and fixed (narrow, depth).

If you're holding more than one unrelated symptom, you want a sibling. If you're holding exactly one, you're in the right place.

Announce at start: "Running /systematic-debugging on <symptom> — pinning the premise before I touch anything."

When this earns its keep

Reach for it on any single technical issue where the cause isn't yet known (not merely suspected): test failure, runtime bug, build break, flaky behavior, perf cliff, integration mismatch. It earns the most under exactly the conditions that make guessing tempting — time pressure, a fix that "looks obvious," a previous patch that didn't take, or a borrowed root cause from someone else's repo.

Skip it when the cause is already proven (just fix it), or when you have a pile of issues rather than one (sweep/blitz instead). A simple bug still has a root cause, but if you can already point at it with evidence, you've effectively done Phases 0–2 already — go fix it.

Phase 0 — Pin the premise (the cheap step everyone skips)

Before reproducing anything, confirm the bug you're chasing is the bug that exists. This phase is ~2 minutes and saves hours of debugging a symptom that was already fixed, or fixing in your repo on a cause that lives in someone else's.

1. Is it still broken on HEAD? Stale dogfood logs, predecessor handoffs, and backlog entries describe state that may already be fixed. git log --oneline -- <cited-paths> since the report's date, then re-run the failing case on current HEAD. An unverified "broken today" is a hypothesis, not a signal. → coordinator CLAUDE.md § Verifying Handoff Premises ("Broken today" claims).
2. Is the root cause borrowed? If the cause came from a handoff, a cross-repo memo, or another repo's diagnosis, that framing is hypothesis, not ground truth. Confirm the same shape exists here — read the cited code on your own disk before acting on a diagnosis made elsewhere. The discrimination trap is assuming a matching symptom implies a matching cause. → coordinator CLAUDE.md § Verifying Handoff Premises.
3. Is the signal real? If the symptom surfaced through tool output that was empty, garbled, or contradictory, the channel may have failed rather than the code. Re-read once, solo. Two reads disagreeing → read a third way before treating it as a bug. → docs/wiki/tool-output-flakiness-protocol.md.

If Phase 0 dissolves the bug (already fixed / not reproducible here / phantom signal), you're done — say so and stop. That is the systematic outcome.

Phase 1 — Reproduce and read the evidence

1. Reproduce reliably. Exact steps; does it fire every time? If you can't trigger it on demand, you're gathering data, not yet debugging — get to a reliable repro before forming theories.
2. Read the error completely. The build/error stream is the contract — it under-reports drift far less than compat docs or memory do. Stack traces, line numbers, codes, the whole trace. The answer is often already in it. → coordinator CLAUDE.md § Investigation funnel.
3. Instrument the boundaries in multi-component systems. When the failure crosses layers (CI → build → sign, request → service → DB, MCP → daemon → editor), log what enters and exits each boundary and run once to see which layer breaks — before theorizing about any of them. Evidence first localizes the layer, then you investigate that layer.
4. Check what changed. git diff / recent commits / new deps / config drift. The delta is the highest-prior suspect.

Phase 2 — Trace to the source, discriminate the locus

The bug usually appears far from where it originates. Fixing where the error surfaces is symptom-patching.

1. Trace backward. From the immediate cause, ask "what passed this bad value / state?" and keep walking up the call chain until you reach the original trigger. When you can't trace by reading, add a stack-trace log right before the failing operation, run, and read the chain.
2. Discriminate symptom-locus from root-locus. Name both explicitly: where it manifests vs. where it originates. An audit or a reporter is usually right about the symptom and can be wrong about the locus — verify the producing code before accepting any proposed fix-site. → coordinator CLAUDE.md § Audit symptom is correct; locus may be wrong; § Fix-locus discrimination.
3. State one hypothesis, in writing. "I think X is the root cause because Y." Specific, falsifiable, one at a time. If you have three theories, you haven't finished tracing.

Phase 3 — Fix at the source

1. One change, at the root. Address the originating cause, not the symptom. Resist "while I'm here" bundling — it muddies the verification signal and can introduce the next bug.
2. Guards match the condition, not the container. If the fix is a guard, gate on the actual invalid condition — substring-on-path filters and state-proxy liveness checks reject legitimate cases. → coordinator CLAUDE.md § Guards match conditions.
3. Consider defense-in-depth where the bug was structurally possible. If invalid data flowed through several layers to cause this, a single check can be bypassed by a different path or a refactor. Validating at each layer the data crosses can make the bug structurally impossible rather than merely absent — worth it when the blast radius justifies it.

Phase 4 — Verify with fresh evidence

1. Re-run the repro; prove it, don't assert it. "Never mark complete without proving it works." A test file written to catch this must be run, not just read. → coordinator CLAUDE.md § Verification Before Done.
2. Check for collateral. Did the fix break a neighbor? Run the surrounding tests, not just the one.
3. Land a regression test where one's missing — and, ideally, land it before the fix so you watch it go red→green. A bug without a regression test invites its own return.

When the fix doesn't take

A patch that doesn't work is information, not a prompt to pile on another patch.

• < 3 attempts: return to Phase 1 with what the failed attempt taught you, and form a new hypothesis. Don't stack fixes on top of an unconfirmed one.
• ≥ 3 attempts, each surfacing a new problem somewhere else: stop debugging and question the architecture. When every fix requires "massive refactoring" or spawns a new symptom elsewhere, that pattern is the signal that the shape is wrong, not the patch. Surface it to the PM as an architectural call — this is a different conversation, not a fourth attempt. → coordinator CLAUDE.md § Core Principles (refactor over patch).

A note on tempo

If you catch yourself reaching for a quick patch under pressure — that instinct is fine, it's the eagerness to fix. Spend it on Phase 0 and a reliable repro first; that's where it pays back fastest. The discipline here isn't distrust of your judgment, it's the route that gets your judgment to the right answer in one pass instead of three.

Related doctrine

• docs/wiki/eager-agent-calibration.md — why this skill is shaped as an offer, not an Iron Law.
• coordinator CLAUDE.md §§ Investigation funnel, Verifying Handoff Premises, Verification Before Done, P0/P1 Verification Gate, Convergence as Confidence.
• docs/wiki/tool-output-flakiness-protocol.md — when the symptom might be a failed channel, not a bug.
• Siblings: /bug-sweep, /bug-blitz, /dogfood, /code-health.