A discipline for hard bugs in external systems — engine, scripts, Discord bot, dashboard, edition compile, dispatch ingest, MCP server. Skip phases only when explicitly justified.
Distinct from /self-debug./self-debug is for when the agent is the failing thing (looping, drifting, retrying without progress). /diagnose is for when a GodWorld system is the failing thing. If the failure is "I keep proposing the wrong fix," use /self-debug first; if it's "the engine throws on cycle 92," use /diagnose.
Distinct from /health, /pre-mortem, /tech-debt-audit. Those scan for issues proactively. This structures the response when something has actually broken.
When exploring the codebase, use CONTEXT.md (project glossary) to keep the mental model clear, and check docs/adr/ for prior decisions in the area you're touching.
Phase 1 — Build a feedback loop
This is the skill. Everything else is mechanical. If you have a fast, deterministic, agent-runnable pass/fail signal for the bug, you will find the cause — bisection, hypothesis-testing, and instrumentation all just consume that signal. If you don't have one, no amount of staring at code will save you.
Spend disproportionate effort here. Be aggressive. Be creative. Refuse to give up.
Ways to construct one — try them in roughly this order
Failing test at whatever seam reaches the bug — unit (script function), integration (script-against-fixture), e2e (full cycle replay).
Direct script invocation with a fixture input, diffing stdout against a known-good snapshot. Most GodWorld scripts (buildDeskPackets.js, validateEdition.js, ingestEdition.js, ingestEditionWiki.js, buildCitizenCards.js) accept --cycle N or --file <path> — point them at a known-good cycle.
MCP tool roundtrip. If the bug is in canon retrieval, hit the godworld MCP tool directly (lookup_citizen, search_canon, lookup_cultural) and assert on the response shape.
curl against the dashboard API. Port 3001 — , , etc. Useful when the bug surfaces in the bot or dashboard.
/api/articles?q=...
/api/citizens/<POPID>
Captured-trace replay. Save a real Discord event payload, a real edition compile input, or a real engine-review JSON to disk; replay it through the code path in isolation. editions/cycle_pulse_dispatch_92_kono_second_song.txt is the canonical fixture for dispatch-side bugs (S188-S189).
Throwaway harness. A 30-line script that imports lib/<module>.js and exercises the bug code path with a single function call. Lives in /tmp/ for the duration of the diagnosis, gets deleted in Phase 6.
Property / fuzz loop. If the bug is "sometimes wrong output," loop 100+ random inputs and look for the failure mode. Useful for citizen card generation, name pool sampling, ledger-derivation drift.
Bisection harness. If the bug appeared between two known states (commit, edition, cycle), automate "boot at state X, check, repeat" so you can git bisect run it. Engine-sheet pattern: replay the prior cycle's world_summary_c<N-1>.json and compare current cycle's behavior.
Differential loop. Run the same input through old-version vs new-version (or two configs) and diff outputs. Common for editorial drift — same brief, two reporter prompts, diff articles.
HITL bash script. Last resort. If a human must click (Drive upload, Discord invite, smfs mount), structure the loop with a deterministic shell harness — captured output feeds back to the agent.
Build the right feedback loop, and the bug is 90% fixed.
Iterate on the loop itself
Treat the loop as a product. Once you have a loop, ask:
Faster? Cache setup, skip unrelated init, narrow the test scope. A 2-minute loop is a productivity tax; a 5-second loop is a debugging superpower.
Sharper? Assert on the specific symptom, not "didn't crash." For dispatch parser bugs (E1/E2 from S189), assert on parsedEntities.length === expectedCount, not on exit code.
More deterministic? Pin time, seed RNG, isolate filesystem, freeze network. Engine determinism rules apply (Math.random banned).
Non-deterministic bugs
The goal is not a clean repro but a higher reproduction rate. Loop the trigger 100×, parallelise, add stress, narrow timing windows, inject sleeps. A 50%-flake bug is debuggable; 1% is not — keep raising the rate until it's debuggable.
When you genuinely cannot build a loop
Stop and say so explicitly. List what you tried. Ask Mike for: (a) access to whatever environment reproduces it, (b) a captured artifact (engine-review JSON, production log, Drive PDF, Discord transcript), or (c) permission to add temporary instrumentation. Do not proceed to hypothesise without a loop.
Do not proceed to Phase 2 until you have a loop you believe in.
Phase 2 — Reproduce
Run the loop. Watch the bug appear.
Confirm:
The loop produces the failure mode Mike described — not a different failure that happens to be nearby. Wrong bug = wrong fix.
The failure is reproducible across multiple runs (or, for non-deterministic bugs, reproducible at a high enough rate to debug against).
You have captured the exact symptom (error message, wrong output, slow timing) so later phases can verify the fix actually addresses it.
Do not proceed until you reproduce the bug.
Phase 3 — Hypothesise
Generate 3–5 ranked hypotheses before testing any of them. Single-hypothesis generation anchors on the first plausible idea.
Each hypothesis must be falsifiable: state the prediction it makes.
Format: "If is the cause, then will make the bug disappear / will make it worse."
If you cannot state the prediction, the hypothesis is a vibe — discard or sharpen it.
Show the ranked list to Mike before testing. He often has domain knowledge that re-ranks instantly ("engine-sheet just shipped that section S188") or knows hypotheses already ruled out. Cheap checkpoint, big time saver. Don't block on it — proceed with your ranking if Mike is AFK or in another terminal.
Search memory before hypothesising. Per identity rules: claude-mem first, then Supermemory mags, then read code. Past sessions often documented the same bug. Don't re-derive what's already in feedback_*.md or a journal entry.
Phase 4 — Instrument
Each probe must map to a specific prediction from Phase 3. Change one variable at a time.
Tool preference:
node --inspect / REPL inspection if the env supports it. One breakpoint beats ten logs.
Targeted logs at the boundaries that distinguish hypotheses.
Never "log everything and grep."
Tag every debug log with a unique prefix, e.g. [DEBUG-a4f2]. Cleanup at the end becomes a single grep. Untagged logs survive; tagged logs die.
Perf branch. For performance regressions (cycle run > expected, edition compile slow, MCP timeout), logs are usually wrong. Instead: establish a baseline measurement (timing harness, Date.now() deltas, Sheets API quota usage), then bisect. Measure first, fix second.
GodWorld-specific surfaces:
Engine determinism: ledger row counts before/after a phase
Edition compile: brief count vs article count vs slug list
Dispatch ingest: NAMES INDEX line count vs ingestEditionWiki.js parsedEntities count
Discord bot: pm2 logs mags-bot --lines 100 after restart-with-debug-flag
Supermemory drift: npx supermemory tags list count delta cycle-over-cycle
Phase 5 — Fix + regression test
Write the regression test before the fix — but only if there is a correct seam for it.
A correct seam is one where the test exercises the real bug pattern as it occurs at the call site. If the only available seam is too shallow (single-caller test when the bug needs multiple callers, unit test that can't replicate the chain that triggered the bug), a regression test there gives false confidence.
If no correct seam exists, that itself is the finding. Note it. The codebase architecture is preventing the bug from being locked down. Flag this for the next phase — likely an item for Phase 42 Writer Consolidation, ENGINE_REPAIR, or a new ADR.
If a correct seam exists:
Turn the minimised repro into a failing test at that seam.
Watch it fail.
Apply the fix.
Watch it pass.
Re-run the Phase 1 feedback loop against the original (un-minimised) scenario.
For non-test-supported areas (no test harness in tests/ for the affected path): a regression scripts/<bugname>-repro.sh checked into scripts/ counts as a regression check. Future sessions can re-run it to verify the fix didn't regress.
Phase 6 — Cleanup + post-mortem
Required before declaring done:
Original repro no longer reproduces (re-run the Phase 1 loop)
Regression test passes (or absence of seam is documented in ROLLOUT_PLAN as a follow-up)
All [DEBUG-...] instrumentation removed (grep the prefix across the touched files)
Throwaway prototypes deleted (or moved to a clearly-marked debug location and noted in commit message)
The hypothesis that turned out correct is stated in the commit message — so the next debugger learns
Then ask: what would have prevented this bug? If the answer involves architectural change (no good test seam, tangled callers, hidden coupling), file a ROLLOUT entry or an ADR with the specifics. Make the recommendation after the fix is in, not before — you have more information now than when you started.
Memory writes. If the diagnosis revealed a recurring pattern, add a one-line entry to MEMORY.md or write a feedback_<topic>.md file. Past sessions documented S122, S128, S168, S188 dispatch-gap-set this way — each saved a future session from re-deriving the same fix.
GodWorld-specific examples
Bug class: Dispatch parser silently returns 0 entities (S188 gap #4/#5).
Phase 1 loop:node scripts/ingestEditionWiki.js --file editions/cycle_pulse_dispatch_92_kono_second_song.txt --cycle 92 → expect 2 entities, got 0.
Phase 3 hypotheses: (a) NAMES INDEX section header detection fails on flat-body dispatches; (b) CUL-ID lines are filtered out by POPID-only regex; (c) section delimiter is --- not the expected ## NAMES INDEX.
Outcome: (a) was correct (S189 E1 fix). 9 commits to fix across 3 scripts.
Bug class: Discord bot quoting C90 after C92 published (S180 gap).
Phase 1 loop:pm2 restart mags-bot && sleep 3 && pm2 logs mags-bot --lines 50 then ask the bot in Discord — assert on cycle reference in reply.
Phase 3 hypotheses: (a) bot reads stale latest_edition_brief.md; (b) bot reads base_context.json but it's not refreshed post-publish; (c) Supermemory boot profile lags real ingest.
Outcome: (a) was correct — file was deleted, plan written [[plans/2026-04-26-discord-bot-edition-currency]].
Changelog
2026-05-01 — S190, research-build. Initial draft. Adapted from MIT-licensed mattpocock/skills/engineering/diagnose/SKILL.md (Matt Pocock 2026). Six-phase structure preserved; localized to GodWorld surfaces (engine scripts, MCP tools, dashboard API, dispatch fixtures, Supermemory tags). Disambiguated from /self-debug (agent is failing thing → that skill) at the top. Source attribution preserved per MIT.
