Bug Reproducer

Bug Reproducer

Turn a codebase or bug report into evidence. When no bug is supplied, discover concrete candidates from the code and its contracts, then try to prove the strongest candidates with focused tests. Never present a suspicion as a confirmed bug.

Enforce two approval gates

Treat invocation of $bug-reproducer as consent to inspect only. Even when the user says find and fix, apply, or do it, do not interpret the initial request as approval for unseen changes.

Before either approval, do not create or edit project files, install dependencies, run formatters or migrations, modify configuration, generate reports, create worktrees, or execute commands likely to mutate project state. Read source, configuration, documentation, existing tests, user-supplied logs, and Git history. Run an existing targeted test only when it is clearly safe and does not require project changes.

Gate 1 — test the candidate

After read-only discovery or report inspection, present and stop at:

## Bug candidates

| # | Candidate | Contract evidence | Trigger | Location | Confidence |
|---|---|---|---|---|---|
| 1 | ... | ... | ... | file:line | high/medium |

## Proposed bug test

- Candidate(s) to test:
- Why each could be a real bug:
- Exact files to create or edit:
- Minimal fixture/input:
- Test or harness command:
- Signal that will confirm each bug:
- Main risk or uncertainty:

No project files have been modified. Do you want me to create and run these tests?

List only candidates backed by a reachable code path, a defensible expected behavior, and a specific triggering input or state. Approval covers only the displayed reproduction files and commands. Test one strong candidate by default; batch at most three only when their files and commands are all explicit.

Gate 2 — fix a proven bug

After a candidate is genuinely reproduced and the root cause is isolated, present and stop at:

## Proposed fix

- Reproduction status: REPRODUCED
- Proven bug:
- Root cause:
- Exact production files to change:
- Proposed transformation:
- Behavior that must remain identical:
- Regression and broader test plan:
- Main risk:

The bug is proven by a failing test, but no production fix has been applied. Do you want me to apply this exact fix?

Continue only after explicit approval. If the file list, dependencies, public behavior, or approach changes, request fresh approval. Never bundle unrelated cleanup or another fix into the approved patch.

Choose the workflow

• hunt-and-prove — Default when no specific bug is supplied. Discover, rank, and test likely correctness bugs.
• reproduce-only — Stop after a minimal failing case is proven. Use when the user asks only to find, reproduce, isolate, or write regression tests.
• prove-fix — Use for a reported bug or after a discovered candidate is proven and the user wants it fixed. Apply both approval gates and prove red to green.
• flaky — Reproduce timing, ordering, seed, isolation, or concurrency-dependent failures across repeated controlled runs. Do not label one intermittent failure as reproduced.

Run hunt-and-prove

1. Map the project read-only

• Identify runtime, entry points, public interfaces, data boundaries, test framework, existing commands, and high-change or high-risk modules.
• Derive intended behavior from tests, types, schemas, documentation, callers, UI copy, API contracts, and invariants. Prefer explicit contracts over personal style preferences.
• Inspect recent changes only when repository history is available; do not assume old code is correct.
• Read references/bug-discovery-playbook.md before ranking candidates.

2. Generate concrete candidates

• Trace user-controlled or boundary inputs through real reachable paths.
• Look for falsifiable defects: off-by-one boundaries, inverted conditions, missing empty/null handling, unsafe state transitions, ordering or deduplication errors, stale cache keys, permission gaps, precision/time-zone mistakes, async races, inconsistent validation, and error paths that violate the surrounding contract.
• For every candidate, identify the exact path, triggering input/state, expected result, likely actual result, and evidence for the expectation.
• Reject vague possibilities such as “this could be null” unless the codebase permits null and the path mishandles it.
• Do not call security posture, performance, formatting, maintainability, or missing features a bug unless they violate a concrete product contract.

3. Rank before testing

Rank candidates using:

1. Contract strength — Is expected behavior explicit in tests, types, docs, schemas, callers, or consistent nearby behavior?
2. Reachability — Can a real caller or valid input reach the path?
3. Reproducibility — Can a small deterministic test distinguish correct from incorrect behavior?
4. Impact — Does it affect outputs, data, permissions, crashes, or a user-visible workflow?

Show no more than five candidates and label confidence high, medium, or low. Do not propose testing low-confidence candidates while a stronger one exists. If no candidate survives this filter, return NO_BUG_PROVEN and identify the next useful evidence instead of inventing a bug.

4. Request Gate 1 approval

• Show the ranked shortlist and the exact test plan for the strongest one to three candidates.
• Name every file and command that approval would cover.
• State that no project files have been modified, then stop and wait.

5. Test approved candidates

• Create only the approved test or harness.
• Use deterministic inputs and the smallest fixture that reaches the real production path.
• Assert the intended contract, not the current implementation.
• Capture the narrow command with scripts/capture_command.py when appropriate:

python3 scripts/capture_command.py --label reproduced --output reproduced.json -- npm test -- path/to/regression.test.ts

• Confirm that a failure matches the predicted behavior and cause. Syntax errors, missing dependencies, invalid fixtures, unrelated failures, and assertions built on an unsupported assumption reject or invalidate the candidate.
• Return REPRODUCED, NOT_REPRODUCED, or INCONCLUSIVE for each tested candidate. Delete or clearly separate rejected speculative tests unless the user asks to retain them.
• Stop after evidence when the user requested discovery or reproduction only. Do not slide into production changes.

6. Isolate and offer the fix

• Trace a reproduced failure to the smallest responsible condition and distinguish root cause from the line where it surfaces.
• Show Gate 2 with exact production files, transformation, preserved behavior, checks, and risk.
• Stop and wait for separate approval.

7. Apply only the approved fix

• Preserve the regression test unless the user explicitly requests a temporary harness.
• Change only the approved production files and necessary test expectations.
• Prefer the smallest causal fix over symptom suppression.
• Do not weaken assertions, catch and ignore errors, add arbitrary retries, disable tests, or broaden tolerances merely to turn the test green.

8. Prove red to green

• Run the exact same targeted command used for failing evidence.
• Run the broadest relevant suite, typecheck, lint, or build.
• Capture and classify the fixed run:

python3 scripts/capture_command.py --label fixed --output fixed.json -- npm test -- path/to/regression.test.ts
python3 scripts/compare_evidence.py reproduced.json fixed.json result.json --reproduction confirmed --full-suite passed

• Classify a passing targeted test with a failing relevant suite as FIX_REGRESSION.
• Classify a passing targeted test without broader evidence as FIX_UNVERIFIED, unless the project genuinely has no broader checks and that limitation is explicit.

Handle an existing bug report

When the user supplies a symptom, error, screenshot, stack trace, or failing behavior:

1. Restate expected behavior, actual behavior, trigger, environment, frequency, and last-known-good version.
2. Separate observed facts from assumptions and remove secrets or personal data.
3. Map the report to likely paths and existing test conventions.
4. Skip broad discovery, but use the same Gate 1, reproduction integrity, Gate 2, and red-to-green proof.
5. Read references/reproduction-playbook.md when selecting the smallest credible layer.

Use evidence labels precisely

• REPRODUCED — A focused case fails deterministically for the predicted or reported reason.
• NOT_REPRODUCED — The approved case does not produce the predicted failure.
• NO_BUG_PROVEN — Read-only discovery or approved candidate tests did not prove a correctness defect.
• INCONCLUSIVE — Environment, nondeterminism, missing access, or ambiguous signals prevent a conclusion.
• STILL_FAILING — The same reproducer continues to fail after an attempted fix.
• FIX_UNVERIFIED — The reproducer passes, but broader correctness evidence is missing.
• FIX_REGRESSION — The reproducer passes or changes, but relevant correctness checks fail.
• FIX_PROVEN — The same reproducer goes red to green and relevant broader checks pass.

Never claim a bug from code inspection alone. Never claim a fix without red-to-green evidence.

Create the native report

After a completed approved workflow, create:

• outputs/bug-reproducer-evidence.json — captured test evidence and classification.
• outputs/bug-reproducer-report.md — native report that opens directly in Codex and renders on GitHub.

Prepare context JSON following references/report-schema.md, including discovery scope and ranked candidates when using hunt-and-prove, then run:

python3 scripts/generate_report.py result.json context.json outputs/bug-reproducer-report.md

Include discovery evidence, tested candidates, minimal reproduction, root cause, approvals and scope, changed files, red/green commands, broader checks, limitations, and residual risks. Link the Markdown report in the final response.

Hand off

Lead with the strongest evidence label, then report:

1. Scope inspected and candidates considered
2. Candidate tests and outcomes
3. What reproduced and why the failure is credible
4. Root cause, if proven
5. Approved fix and red-to-green evidence, if requested
6. Broader checks, limitations, and residual risk
7. Link to outputs/bug-reproducer-report.md

If no bug is proven, preserve the project and say so plainly. A clean hunt is evidence about the inspected scope, not proof that the entire codebase has no bugs.