senior-bug-test

Prompt file imported from Nom-nom-hub/senior-agent (.codex/prompts/senior-bug-test.md). Fill in {{arguments}} before use. Copyright stays with the author.

Arguments

{{arguments}}

Resolves slug=<slug> like /senior.bug-fix (single directory auto-resolves).

What this stage is

The senior closes the loop: re-run the ORIGINAL reproduction, through the outfit the user would experience it through, and grade what actually happened. This stage never edits source code. It exists because "the fix commit exists" is not evidence that the bug is fixed. Output: .senior/bugs/<slug>/test.md.

Execution steps

1. Prerequisites: fix.md must exist. Read assessment.md + fix.md for: original reproduction steps, verdict, deviations from assessment, "honest limits" section.

2. Re-run the original reproduction — the exact steps/scenario from assessment.md, not the fix's own new test. Use the reproduction method closest to the user's experience (CLI invocation, real endpoint call, UI flow). Record every command run and the real observed output verbatim.

3. Cross-checks beyond the symptom:

   • the regression test from the fix fails against the pre-fix code if that can be arranged cheaply (revert-in-branch/cherry-pick check) — confirming it locks the bug out permanently, not just today;
   • affected tests + typecheck + lint + full suite (report what was run and what was skipped, separately);
   • deviations from fix.md get honestly re-checked (each deviation is a place where the fix differs from the plan — re-verify those paths specifically).

4. Write .senior/bugs/<slug>/test.md:

   # Bug Verification: <slug>
   
   - **Verified**: <date> | **Verdict**: verified | partial | failed | not-run
   
   ## Original reproduction (re-run)
   
   | Step | Expected | Observed | Run? |
   |---|---|---|---|
   
   <Verbatim commands + output snippets>
   
   ## Suite results
   
   - regression test: <name, result>
   - affected tests / full suite / typecheck / lint: <results>
   
   ## Unverified remainder
   
   <what could not be exercised (environment gaps, flaky infra, prod-only conditions) and
   how much risk that leaves>
   
   ## Verdict rationale
   
   <one paragraph: why graded this way>

5. Grade:

   • verified: original reproduction now behaves correctly by the direct method, and all locally-runnable suites are green.
   • partial: fixed in the paths tested, but some reproduction internals (or deviation paths, or environment-dependent steps) were not directly exercised.
   • failed: reproduction still misbehaves on any run-through — a failed verdict routes BACK to /senior.bug-fix (or re-assessment if the hypothesis was wrong); it never ships with a "passes for now" note.
   • not-run: could not execute at all (environment). Unverified ≠ verified. The user must be told plainly, with the exact gap preventing the run.

Rules

• Read-only on source. Writes only in .senior/bugs/<slug>/.
• Never manufacture observed outputs; every row is either a real run or an explicit Run? no-entry with the reason.
• The test stage is independent: its opinion is formed from what it observed, not from what fix.md claims.

Done When

• test.md written with reproduction table, suite results, unverified remainder
• Verdict justified against the grading criteria
• On failed: user given an explicit routing decision (re-fix or re-assess)
• User told the final verdict plainly — including partial and not-run, in so many words