fsd-diagnose

name	fsd:diagnose
description	Disciplined diagnosis loop for hard bugs and performance regressions in the @flow-state-dev repo. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression. For flow execution failures specifically, prefer `fsd:debug-flow`.

Diagnose

A discipline for hard bugs. Skip phases only when explicitly justified.

Repo context

Before diagnosing, orient yourself:

• docs/architecture/overview.md — system architecture and package roles
• docs/architecture/<area>.md — deep dives (e.g. items.md before touching items/rendering/stream, streaming.md, execution-and-errors.md, state-and-scopes.md)
• docs/contributing/architecture-reference.md — locked contracts quick reference
• docs/contributing/best-practices.md — universal standards + situational index (BP-001–BP-039); situational rule text lives in docs/contributing/best-practices/<category>.md
• AGENTS.md — process protocol and code style rules

Related skills — pick the right one:

Symptom	Use
Flow execution misbehaves (wrong items, failed block, schema mismatch in a .step() chain, generator/tool errors during a run)	fsd:debug-flow — it has the NDJSON trace reader, fsdev block isolation workflow, and the FSD failure-pattern lookup table.
Anything else (typecheck regression, build break, store adapter bug, devtool UI bug, CLI bug, perf regression, flaky test, package boundary violation)	This skill.
Mixed — flow is broken and the root cause is upstream (e.g. a core builder regression breaking many flows)	Start here for the discipline; use fsd:debug-flow inside Phase 1 to build the loop.

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

Loops that work well in this repo, fastest-to-slowest signal:

1. Vitest filter at the package level. A single test file is the sharpest, fastest loop available:

   pnpm --filter @flow-state-dev/<pkg> test path/to/file.spec.ts
   pnpm --filter @flow-state-dev/<pkg> test -t "name fragment"   # run by test name
   

   Watch mode for tight iteration: pnpm --filter @flow-state-dev/<pkg> test:watch. Skips unrelated init across the monorepo.
2. Typecheck loop for type-level bugs (declaration drift, generic regressions, package boundary breaks):

   pnpm --filter @flow-state-dev/<pkg> typecheck
   pnpm typecheck   # whole monorepo + boundary validator (scripts/validate-package-boundaries.mjs)
   

3. fsdev block for single-block isolation. One-shot JSON-in/JSON-out. Returns structured schemaValidation, execution.durationMs, and error.stack. Ideal for handlers, utility generators, and routers when you can name the input shape:

   pnpm --filter @flow-state-dev/cli fsdev block <path-to-block.ts> -i '<json>'
   

4. fsdev run with NDJSON capture when the bug only appears in a real flow (sequencer composition, state-scope interactions, generator + tool loop). Pipe stdout to a file so you can diff between runs:

   pnpm --filter @flow-state-dev/cli fsdev run <flowKind> <action> \
     -i '<json>' --flow-dir <path> > /tmp/run.ndjson 2> /tmp/run.log
   

   For deep flow analysis, hand off to fsd:debug-flow — it has the NDJSON event-type reader and failure-pattern table.
5. Persistent session replay. fsdev run -s <session-id> reuses session state across invocations; --seed-session / --seed-user / --seed-project pre-populate state. Use this to reproduce bugs that only appear after specific state.
6. DevTool / kitchen-sink browser run for UI-layer bugs (React hooks, renderers, devtool itself). fsdev dev for the live devtool; the kitchen-sink app under apps/ for SSR/streaming assertions. Watch the browser console and the server stderr.
7. Replay a captured trace. Save a real NDJSON stream or HTTP request to disk; replay it through the code path in isolation (e.g. by feeding it to a unit test fixture).
8. Property / fuzz loop. For "sometimes wrong output" bugs — drive the block/builder with 100–1000 random inputs and assert invariants. Especially useful for sequencer composition and schema strictness (BP-016).
9. Bisection harness. If the bug appeared between two known good/bad states, automate "boot at state X, check, repeat" so you can git bisect run it against a vitest filter.
10. Differential loop. Run the same input through old-version vs new-version (or two configs) and diff stdout. Particularly effective for generator output regressions and serialization changes.
11. HITL last resort. If a human must click — UI bug, real provider call, OAuth flow — drive them with a structured prompt and capture their report back into the loop. Don't proceed to Phase 2 without some loop.

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:

• Can I make it faster? (Cache setup, skip unrelated init, narrow the test scope.)
• Can I make the signal sharper? (Assert on the specific symptom, not "didn't crash".)
• Can I make it more deterministic? (Pin time, seed RNG, isolate filesystem, freeze network.)

A 30-second flaky loop is barely better than no loop. A 2-second deterministic loop is a debugging superpower.

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 the user for: (a) access to whatever environment reproduces it, (b) a captured artifact (HAR file, log dump, core dump, screen recording with timestamps), or (c) permission to add temporary production 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 the user 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 the user before testing. They often have domain knowledge that re-ranks instantly ("we just deployed a change to #3"), or know hypotheses they've already ruled out. Cheap checkpoint, big time saver. Don't block on it — proceed with your ranking if the user is AFK.

Phase 4 — Instrument

Each probe must map to a specific prediction from Phase 3. Change one variable at a time.

Tool preference:

1. Debugger / REPL inspection if the env supports it. One breakpoint beats ten logs.
2. Targeted logs at the boundaries that distinguish hypotheses.
3. 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, logs are usually wrong. Instead: establish a baseline measurement (timing harness, performance.now(), profiler, query plan), then bisect. Measure first, fix second. For flow-level perf, the NDJSON stream already carries per-block durationMs in block_output items — diff between a known-good and known-bad run before reaching for a profiler.

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.

If a correct seam exists:

1. Turn the minimised repro into a failing test at that seam.
2. Watch it fail.
3. Apply the fix.
4. Watch it pass.
5. Re-run the Phase 1 feedback loop against the original (un-minimised) scenario.

FSD test placement — match the existing convention:

• Block / pattern unit tests live next to the source: packages/<pkg>/src/.../foo.ts → packages/<pkg>/src/.../foo.spec.ts. See the fsd:write-block-tests skill for the mock-context idiom.
• Cross-package or flow-level regressions belong in packages/integration-tests/ (Tier 1 suite). Add a flow there when the bug needs more than one package to reproduce.
• Generator output-schema bugs: add a makeSchemaStrict walker assertion (see labs/trading-desk/test/output-schemas-strict.spec.ts). BP-016 requires this guard for any schema that's reachable from a generator output.
• Verification per CLAUDE.md: if the change is flow-logic, the default verification is fsdev run, not pnpm test. Reach for pnpm test for unit-level changes; reach for the browser for UI-layer changes.

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)
• All [DEBUG-...] instrumentation removed (grep the prefix)
• Throwaway prototypes deleted (or moved to a clearly-marked debug location)
• The hypothesis that turned out correct is stated in the commit / PR message — so the next debugger learns
• Per-package check passes: pnpm --filter <pkg> typecheck && pnpm --filter <pkg> test for any package you touched. Run pnpm typecheck at the root if your fix crossed package boundaries — it also runs the boundary validator.
• User-facing doc updates (CLAUDE.md rule): any new or changed end-user functionality requires updates to the relevant packages/*/README.md and any affected apps/docs pages in the same change set.
• BP-007: if you added a new file, give it a header comment and doc-comment every export.

Then ask: what would have prevented this bug? If the answer involves architectural change (no good test seam, tangled callers, hidden coupling between packages, missing capability), capture the specifics — the right place depends on scope:

• Implementation-level concern (BP violation, missing test seam, recurring pattern bug): consider whether it warrants a new BP (in its home per the update policy — universal in docs/contributing/best-practices.md, situational in docs/contributing/best-practices/<category>.md), or a follow-up Linear issue.
• Architecture-level concern (documented contract is ambiguous or wrong, cross-package drift, missing seam): update the relevant docs/architecture/<area>.md in the same change set, and consider whether fsd:improve-codebase-architecture should run on the area afterwards.

Make the recommendation after the fix is in, not before — you have more information now than when you started.