/pasteurize

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

When exploring the codebase, call the selected source-code backend directly according to

code-intelligence-routing.md

, and check

.cheese/specs/

for any spec or design notes that touch the failing seam.

Portability reference:

../cheese/references/harness-portability.md

. It covers helper resolution, sub-agent dispatch, GitHub operations, and handoff transitions; prefer the bundled or repo-local helper first, and treat

${CLAUDE_SKILL_DIR}

as optional host-provided fallback. The handoff blocks below are the portable contract; slash commands are host renderings, not the control model.

Phase 1 — 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.

Ways to construct one

To pick a loop shape, see

references/feedback-loops.md

for the ten-option ordered menu.

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.)

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. Write a

status: halt

handoff slug (see below) and stop.

Do not proceed to Phase 2 until the loop passes all four checks:

Deterministic — runs the same way every time (or, for flaky bugs, reproduction rate >50% and rising).
Agent-runnable — a single command with no human in the loop.
Asserts the user’s exact symptom — the failure message / wrong output / timing the user reported, not a nearby failure.
Fast — under 30 seconds end-to-end (aim for under 5).

Phase 2 — Reproduce

Run the repro loop N times and verify the failure is consistent:

python3 skills/pasteurize/scripts/pasteurize.pyz repro-rerun --cmd "<repro-command>" --runs 5

Confirm the returned

reproduced: true

and check

failures

matches the expected failure mode. If

reproduced: false

at N=5, the bug is flaky — increase

--runs

before proceeding.

Do not proceed until you reproduce the bug.

Symptom-shape gate

Before forming any hypothesis, classify the symptom shape:

Clean stack trace + deterministic repro — stay at current tier, proceed to Phase 3 normally.
Heisenbug, race condition, cross-module failure, or perf regression — warn to upgrade (harness-detected phrasing: claude
```
/model opus
```
+
```
/effort
```
; codex/OMP named equivalent; generic fallback) before forming any hypothesis. The extra tier buys the wider context window and reasoning depth these shapes need; do not start Phase 3 at the current tier once this branch fires.

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
<X>
is the cause, then
<changing Y>
will make the bug disappear /
<changing Z>
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 through the host routing guide in
../cheese/references/handoff-gate.md
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 or running

--auto

Phase 4 — Instrument

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

Tool preference:

Debugger / REPL inspection if the env supports it. One breakpoint beats ten logs.
Targeted logs at the boundaries that distinguish hypotheses.
Never "log everything and search".

Tag every debug log with a unique prefix, e.g.

[DEBUG-a4f2]

. Cleanup at the end becomes a single content query through the selected search backend. 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.

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 in the handoff slug as an architectural follow-up. The codebase is preventing the bug from being locked down. Skip the test write; do not paper over it. Phase 6's "what would have prevented this bug?" retrospective still applies.

Before writing the test, confirm the seam is correct: verify that the test you're about to write targets the boundary where the bug actually occurs — the real call site, the real data path, the real failure mode. A test at the wrong seam (too shallow, wrong abstraction level, mocked-away side that hides the failure) will pass after the fix but won't catch a regression. If you discover the seam is wrong at this point, treat it as "no correct seam": write the no-correct-seam halt string from Early-stop conditions and route to

/mold

, per the halt path above.

If a correct seam exists:

Turn the minimised repro into a failing test at that seam.
Watch it fail.
Apply the smallest production change that makes the test pass. No scope creep, no "while I'm here" cleanup. If the test still fails, revert and retry — but cap the retries (see After 3 failed fix attempts below).
Watch the test pass.
Re-run the Phase 1 feedback loop against the original (un-minimised) scenario to confirm the symptom is gone, not just the test seam.

After 3 failed fix attempts (3 cycles of "apply change → watch test → revert because test still fails"), stop attempting fixes and re-question the approach: is the hypothesis from Phase 3 actually correct? Is the seam exposing the right failure? Is the bug at a different layer than assumed? Step back to Phase 3 and generate a fresh ranked hypothesis list — do NOT attempt a 4th blind fix. If the re-questioning produces a new hypothesis, restart from Phase 4. If all hypotheses are exhausted, write the fix-attempts-exhausted halt string from Early-stop conditions and route to

/mold

Broader implementation (related cleanup, follow-on changes, anything beyond the minimal fix) is not pasteurize's job. Note it in the slug and let

/cook --auto

pick it up in Phase 6's handoff.

Phase 6 — Cleanup

Before writing the handoff slug, confirm:

Original repro no longer reproduces (re-run the Phase 1 loop).
Regression test passes (or absence of seam is documented in the slug).
All
```
[DEBUG-...]
```
instrumentation removed:
```
python3 skills/pasteurize/scripts/pasteurize.pyz debug-tag-sweep --root .
```
Exit 0 = clean. Exit 1 = tags found (listed in output). Resolve before continuing.
Throwaway harnesses / prototypes deleted (or moved to a clearly-marked debug location and called out in the slug).
The confirmed hypothesis is captured in the slug so the commit message downstream can reference it.

Then ask: what would have prevented this bug? If the answer involves architectural change (no good test seam, tangled callers, hidden coupling), note it in the slug under an architectural-follow-up line. The chain still runs; the user can pick up the architectural work via

/mold

after the fix lands. Make the recommendation after the fix is in, not before.

Once the checklist is green and the slug is on disk, hand off to

/cook <slug> --auto

(default). Cook --auto picks up the post-fix state, runs its taste-test against the applied diff for spec drift / readability / scope creep, produces its package-ready report, and triggers the autonomous

/press → /age → /cure

chain. Pasteurize itself does not commit, open PRs, or drive the chain — cook owns that.

Fan-out sizing

/pasteurize

fans zero agents today.

size_pasteurize_fanout(bug_shape, score, deterministic_repro)

src/fanout/pasteurize_route.py

is the sizing policy for when it does.

The signal is inverted relative to review: a reviewer (

age_route.route

) reads a diff that exists — more diff, more agents.

size_pasteurize_fanout

instead reads the

review_surface

score descending, over the suspect range (last-known-good..HEAD), not over a diff under review — less evidence means more agents, because the search space is what gets fanned over.

Bug shape	Range	Repro	Agents
regression	tight (score <= 250)	deterministic	1 (linear, no fan)
regression	tight (score <= 250)	non-deterministic	2
regression	wide (score > 250)	deterministic	2
regression	wide (score > 250)	non-deterministic	3
regression	score is `None` (no diff to anchor to)	deterministic	3
regression	score is `None` (no diff to anchor to)	non-deterministic	5
heisenbug / race / perf regression	any	any	3
cold bug (no diff to anchor to, score is `None` )	--	deterministic	3
cold bug (no diff to anchor to, score is `None` )	--	non-deterministic	5

Boundary: the code checks

score > 250

, so exactly

counts as tight (

score <= 250

), not

score < 250

as a naive reading of "tight" might suggest.

Every constant above (

WIDE_RANGE_THRESHOLD

_REGRESSION_TIGHT_DETERMINISTIC_N

_REGRESSION_TIGHT_NONDETERMINISTIC_N

_REGRESSION_WIDE_DETERMINISTIC_N

_REGRESSION_WIDE_NONDETERMINISTIC_N

_UNSTABLE_REPRO_N

_COLD_BUG_DETERMINISTIC_N

_COLD_BUG_NONDETERMINISTIC_N

) is reasoned, not measured. Unlike every reviewer threshold in the router -- each validated against 30 commits of real history -- these have no historical validation, because

/pasteurize

fans zero agents today. They are named tunable constants and should be revisited once real runs exist.

On a bundle-only host,

size_pasteurize_fanout

is also reachable as

python3 skills/pasteurize/scripts/pasteurize.pyz pasteurize-route <request.json>

(JSON in, JSON out -- mirrors

age-route

's bundle convention).

Preferred tools and fallbacks

Need	Prefer	Fallback
Code search / blast radius	semantic caller and dependency search	bounded text search with explicit precision loss
Reading code	fresh bounded read from the intended write backend family	native bounded read with snapshot/line anchors
Editing instrumentation	stale-safe anchored edit	LSP or native snapshot edit with stale-write detection
Diff visualization	`delta`	plain `git diff`
GitHub context	`gh`	local git history or user-provided links
External sanity check	`/briesearch`	clearly mark as an assumption

Missing optional tools should not interrupt diagnosis.

Output

Return a short report covering:

The named cause (one sentence, with
```
<certain>
```
/
```
<speculating>
```
/
```
<don't know>
```
calibration).
The feedback loop (command, observed vs expected).
Hypotheses considered and which one held.
The regression test path and the fix's file:line footprint.
Cleanup status (
```
[DEBUG-...]
```
removed, harnesses deleted or relocated).
Suggested next skill —
```
/cook <slug> --auto
```
for the autonomous chain forward.

Handoff slug

Write a minimum-shape handoff slug to

.cheese/pasteurize/<slug>.md

/cook

(and any orchestrator) can resume without re-reading the full report. Schema:

markdown

status: ok | halt: <one-line reason>
next: cook | mold | done
artifact: <path-to-richer-report-if-any>
cause: <one-sentence named cause>
loop: <command or repro path>
seam: <regression-test path:line, or "none — architectural follow-up">
fix: <production diff footprint, e.g. "src/foo.ts:42">
follow_up: <architectural follow-up note, or "none">
<one-line orientation: what pasteurize converged on>

status: ok

when the regression test is green, the original repro no longer reproduces, and cleanup is done.

status: halt: <reason>

when any early-stop condition fires — see Early-stop conditions below.

next:

cook

for the standard chain,

mold

if the diagnosis itself recommends an architectural spec instead of a per-bug fix, or

done

if the bug was caused outside the repo and no follow-up is needed.

Handoff

Pipeline: cheese (debug) → [pasteurize] → cook --auto → press → age → cure → plate

After the report is printed and the handoff slug is on disk, ask through the host routing guide in

../cheese/references/handoff-gate.md

which downstream to run. Lead each option with the verb (what the user wants to do next):

Validate and chain forward (recommended when
status: ok
) —
```
/cook <slug> --auto
```
.
Validate without auto chain —
```
/cook <slug>
```
(cook runs taste-test, then the user picks each subsequent step).

Spec the architectural follow-up first —

/mold <slug>

(when

seam: none — architectural follow-up

Stop — fix is in tree; defer the chain.

Pre-select Validate and chain forward when

status: ok

. The chain default is

--auto

because pasteurize already wrote and verified the fix; the work left for cook → press → age → cure is mechanical validation, not new authoring. Never auto-invoke; the user must still select.

When invoked with

--auto

, skip this host-routed question entirely and chain forward per Auto mode.

Auto mode

--auto

skips Phase 3's user-ranking gate, skips the Phase 6 handoff gate, and invokes

/cook <slug> --auto

directly. Phase 4–5 still run in full.

Early-stop conditions

Phase 1 fails (
```
status: halt
```
written, no loop achievable).
Phase 3 disproves all hypotheses across two rounds (cap at two Phase 3 rounds, then halt).
Phase 5's seam check finds no correct seam — write
```
status: halt: no correct regression-test seam
```
and route to
```
/mold
```
instead of
```
/cook
```
.
The fix breaks an unrelated test that pasteurize cannot reconcile within scope.
Phase 5's fix loop exhausts all hypotheses after 3 failed fix attempts — write
```
status: halt: fix attempts exhausted — architectural re-examination needed
```
and route to
```
/mold
```
instead of
```
/cook
```
.

In every early-stop case, write the halt slug and surface the report. Do not silently downgrade to "best guess".

Rules

Do not skip Phase 1, and do not hypothesise without a reproducing loop.
Phase 5 writes only the regression test and the minimal production change; broader work belongs in
```
/cook
```
.
Do not leave
```
[DEBUG-...]
```
tags in the tree — clean them before the handoff slug is written.
Do not claim "shipped". Pasteurize claims "cause named, regression green, fix in tree, ready for chain". The chain (cook → press → age → cure) claims shipped.
If the bug exposes an architectural gap (no correct regression-test seam), say so in the slug. Do not silently paper over it.

References

skills/pasteurize/scripts/pasteurize.pyz repro-rerun

— run the repro command N times and emit

{exit_code, reproduced, runs, failures}

(Phase 2).

```
skills/pasteurize/scripts/pasteurize.pyz debug-tag-sweep
```
— scan the tree for instrumentation tag prefixes and exit 1 if any survive (Phase 6 cleanup gate).

pasteurize

NPX Install

Tags

SKILL.md Content