Diagnose and fix a hard bug. Build a reliable reproduction, name the cause, add a regression test, and apply the minimum fix. Use when the user reports a bug, a failure, a flaky test, a performance regression, an error, or a wrong result whose cause is unknown. Use when the user pastes a symptom, a

/pasteurize Use this process for hard bugs. Discipline Iron Law: Build a reliable feedback loop before you form one hypothesis. Stop at each red flag: You name a cause before the loop reproduces the failure. You accept an unrelated failure as the reproduction. You add a test at a mocked seam because the real seam is difficult. You make a fourth fix attempt on the same hypothesis list. You report a clean worktree without the session tag sweep. Rationalization Answer "The cause is obvious." Build the loop. An obvious cause takes one run to confirm. "The loop is too slow to write." A slow loop still beats a wrong fix. "Any failure proves the bug." Match the expected exit code or output. "The mocked seam is close enough." Record the missing seam and route to Mold. "One more attempt will work." Write a new hypothesis list first. Follow [ code intelligence routing.md ](../cheese/references/code intelligence routing.md) when you explore code. Resolve the specification store with artifact path specs . Read the notes about the failed seam from the resolved directory. Read [ harness portability.md ](../cheese/references/harness portability.md) for portable tool use. Use bundled or repository helpers before ${CLAUDE SKILL DIR} . Treat ${CLAUDE SKILL DIR} as an optional host fallback. Use the handoff blocks as the portable contract. Remember that slash commands are host renderings, not the control model. Inputs <input is the reported symptom. Accept a bug report, a stack trace, failing test output, or an artifact path. Accept an investigation request from Affinage, Cheese, or Cook. An investigation request uses these fields: Field Type Required Meaning source string yes The requesting skill, such as affinage . source ref string no The pull request comment or finding identifier. symptom string yes The reported failure in one sentence. expect exit integer no The exit code that shows the failure. expect output string no A regular expression for the failure output. mode string no investigate or fix . The default is fix . Accept these flags: auto runs the phases without the two questions. open pr reaches Plate through Cook. hard reaches the final Hard cheese gate through Cook. Forward open pr and hard to every /cook and /mold command. Do not drop a flag that the caller supplied. Cheese can supply handoff context.wiki hits . Each hit has a page , a line , and a why field. Read each hit before Phase 1. Name any hit that changes the hypothesis ranking. A request with mode: investigate stops after Phase 2. Return reproduced , not reproduced , or inconclusive to the source skill. Do not start Phase 4 for an investigation request. Phase 1: Build the feedback loop Build a fast and reliable pass or fail feedback loop before you diagnose the bug. This feedback loop controls every later phase. Read [ references/feedback loops.md ](references/feedback loops.md) for the ordered loop options. Select the first option that reaches the failed seam. Improve the loop before you continue. Reduce setup and unrelated initialization. Check the exact symptom. Control time, random values, files, and network access. Non deterministic bugs Increase the reproduction rate above 50 percent. Repeat the trigger. Add stress. Narrow the timing window. Add a controlled delay. No usable loop Stop when you cannot build a usable loop. List each method that you tried. Request access to the reproduction environment. Alternatively, request a captured artifact or permission for temporary production instrumentation. Write a status: needs context: <the access you need handoff slug. The orchestrator retries this run after it supplies the access. Do not form hypotheses without a loop. Confirm these conditions before Phase 2: [ ] The loop gives the same result each time. [ ] A flaky loop reproduces the bug more than 50 percent of the time. [ ] One command runs the loop without human action. [ ] The loop checks the exact reported symptom. [ ] The loop completes in less than 30 seconds. Aim for a loop that completes in less than five seconds. Phase 2: Reproduce the bug Run the loop five times. Always name the expected failure mode. Use expect output for a failure with known output. Use expect exit for a failure with a known exit code. Confirm that the result contains reproduced: true . Confirm that matches equals runs for a deterministic bug. Read the results list and confirm each matched run. Increase runs when five runs do not reproduce a flaky bug. Do not continue until the loop reproduces the expected failure. Do not accept an unrelated failure as a reproduction. Symptom gate Classify the symptom before you form hypotheses. Continue at the current model tier for a clean stack trace and a reliable loop. Upgrade the model tier for races, cross module failures, performance regressions, or Heisenbugs. Use the harness command for the model upgrade. For Claude, use /model opus . Then use /effort . Use the named equivalent for Codex or OMP. Use a generic model upgrade on other hosts. Phase 3: Form hypotheses Write three to five ranked hypotheses before you test one. State a testable prediction for each hypothesis. Discard a hypothesis when you cannot state its prediction. Use this format: If <X causes the bug, then <changing Y removes it or <changing Z makes it worse. Show the ranked list through [ handoff gate.md ](../cheese/references/handoff gate.md). Domain knowledge can change the ranking. Continue with your ranking when the user is unavailable or uses auto . Phase 4: Add instrumentation Map each probe to one Phase 3 prediction. Change one variable at a time. Use tools in this order: 1. Use a debugger or REPL when the environment supports it. 2. Add targeted logs at boundaries that distinguish hypotheses. 3. Do not log all data and search it later. Select one session tag for this investigation, such as a4f2 . Prefix each temporary log with the exact token [DEBUG a4f2] . Record the session tag in the handoff slug. Use the session tag to find every temporary log during cleanup. For performance regressions, measure a baseline before you change code. Use a timing harness, profiler, or query plan. Then bisect the failed range. Phase 5: Fix the bug Add the regression test before the fix. Only add the test when a correct test seam exists. A correct seam exercises the real bug at its actual call site. It uses the real data path and failure mode. A shallow or mocked seam gives false confidence. Treat a missing seam as an architectural finding. Record it in the handoff slug. Write the applicable early stop status. Set next: mold . Do not add a test at an incorrect seam. When a correct seam exists, complete these steps: 1. Convert the minimum reproduction into a failing test at that seam. 2. Run the test and observe the failure. 3. Apply the smallest production change that can fix the failure. 4. Run the test and observe success. 5. Run the original Phase 1 loop and confirm that the symptom is absent. Revert an unsuccessful fix before the next attempt. Stop after three unsuccessful fix attempts. Return to Phase 3. Write a new ranked hypothesis list. Do not make a fourth attempt without a new hypothesis. Restart at Phase 4 when you find a new hypothesis. Otherwise, write the fix attempts exhausted status. Then set next: mold . Leave broader changes for /cook . Record those changes in the handoff slug. Phase 6: Clean the worktree Complete this checklist before you write the handoff slug: [ ] Run the original loop and confirm that the bug is absent. [ ] Run the regression test and confirm success. [ ] Record the missing test seam when no correct seam exists. [ ] Remove all [DEBUG ...] instrumentation. [ ] Delete temporary harnesses and prototypes. [ ] Record any retained debug file in the slug. [ ] Record the confirmed hypothesis in the slug. Run the instrumentation sweep with your session tag: session tag matches the exact token [DEBUG a4f2] . changed only scans the files that this worktree changed. The sweep excludes tool output such as .cheese/ , caches, and run logs. Exit status 0 means that the sweep found no tags. Exit status 1 means that the sweep found listed tags. Remove each listed tag before you continue. Do not use the broad tags scan to certify a clean worktree. Identify what could prevent this bug. Record necessary architectural work in the slug. Make this recommendation after you apply the fix. Hand the completed slug to /cook <slug auto . Cook checks the existing diff and runs the press age cure chain. Pasteurize does not commit changes or open pull requests. Fan out sizing /pasteurize currently starts no agents. size pasteurize fanout() defines a future policy in src/easy cheese/shared/fanout/pasteurize route.py . The policy reads review surface descending across the suspect range from the last known good revision to HEAD . A lower evidence score produces more agents because it represents a larger search space. Bug shape Range Repro Agents regression tight, score < 250 deterministic 1 regression tight, score < 250 non deterministic 2 regression wide, score 250 deterministic 2 regression wide, score 250 non deterministic 3 regression no score deterministic 3 regression no score non deterministic 5 race, heisenbug, or performance regression any any 3 cold bug no score deterministic 3 cold bug no score non deterministic 5 A score of 250 defines a tight range. A reviewer reasoned these constants. Runs have not measured them. Reviewer thresholds use 30 repository commits. These constants have no run history. Keep the named constants adjustable. Review them when real runs exist. Bundle only hosts can call the policy with this command: The command reads JSON and writes JSON. It follows the age route bundle convention. Preferred tools Need Preferred tool Fallback Code search and impact semantic caller and dependency search bounded text search with a precision warning Code read fresh bounded read from the write backend bounded read with stable line anchors Instrumentation edit stale safe anchored edit LSP or snapshot edit with stale write detection GitHub context gh local Git history or user links External check /briesearch a clearly marked assumption Continue diagnosis when an optional tool is unavailable. Output Return a short report with these items: The named cause and its confidence. The loop command and the observed result. The considered hypotheses and the confirmed hypothesis. The regression test path. The changed production lines. The instrumentation and temporary file cleanup status. The next command, /cook <slug auto . Use <certain , <speculating , or <don't know for confidence. Handoff slug Write the handoff slug to .cheese/pasteurize/<slug .md . Use this minimum form: Keep the orientation on line four. The parser reads status , next , and artifact before the orientation. It reads no other keyed line in the preamble. Put every diagnostic field in the body after one blank line. Use artifact for the richer report, not for the diagnosis. Follow the [handback contract](../cheese/references/handback contract.md). Use these statuses: Status Disposition Use ok proceed The test, reproduction, and cleanup checks succeed. ok with concerns: <reason proceed The diagnosis is complete, but the work needs Mold. needs context: <reason retry The run needs reproduction