Codex bug 修復閉環：先重現再修，附可重跑的驗證

Fix this bug using a strict reproduce-first, verify-after loop. Do not skip steps. Do not claim it's fixed without showing a passing test run.

## The bug
{{BUG_DESCRIPTION}}

Expected behavior: {{EXPECTED}}
Actual behavior: {{ACTUAL}}
How to trigger (if known): {{REPRO_STEPS}}

## Loop you MUST follow

**1. Reproduce FIRST (red).** Before changing any production code, write or identify a test that fails *because of this bug*. Run it and paste the real failing output. If you genuinely cannot write an automated test (e.g. it needs a live external service), say so explicitly and instead give the exact manual command + observed wrong output that demonstrates the bug. A bug you can't reproduce is not yet understood — keep investigating until you can.

**2. Find the root cause.** Trace from the failing symptom back to the line(s) responsible. State the root cause in 1-2 sentences: *what* is wrong and *why* it produces the observed behavior. Do not pattern-match a fix onto the symptom — explain the mechanism. If there are multiple plausible causes, name them and say how you ruled the others out.

**3. Apply the MINIMAL fix.** Change only what's needed to fix the root cause. No drive-by refactors, no reformatting unrelated lines, no renaming. If you spot adjacent issues, list them at the end as 'noticed but not changed' — do not fold them into this fix.

**4. Verify (green).** Re-run the SAME test from step 1 and paste the now-passing output. Then run the broader suite (`{{TEST_COMMAND}}`) and paste the result to prove no regression. If the suite was already partially failing before your change, note which failures are pre-existing and unrelated.

**5. Report.** Summarize: root cause, the one-line essence of the fix, files touched, and the verification command anyone can re-run.

## Hard rules
- Steps 1 and 4 MUST include actual command output, not a description of what you expect to happen. 'Should pass now' is not acceptable — run it.
- If after a reasonable effort you cannot reproduce, STOP and report what you tried and what info you need, rather than guessing at a fix.
- Keep the diff small and reviewable. The test added in step 1 stays in the codebase as a regression guard.
- Respect existing project conventions (read AGENTS.md / CLAUDE.md if present for the right test command and style).

Fix this bug using a strict reproduce-first, verify-after loop. Do not skip steps. Do not claim it's fixed without showing a passing test run.

## The bug
{{BUG_DESCRIPTION}}

Expected behavior: {{EXPECTED}}
Actual behavior: {{ACTUAL}}
How to trigger (if known): {{REPRO_STEPS}}

## Loop you MUST follow

**1. Reproduce FIRST (red).** Before changing any production code, write or identify a test that fails *because of this bug*. Run it and paste the real failing output. If you genuinely cannot write an automated test (e.g. it needs a live external service), say so explicitly and instead give the exact manual command + observed wrong output that demonstrates the bug. A bug you can't reproduce is not yet understood — keep investigating until you can.

**2. Find the root cause.** Trace from the failing symptom back to the line(s) responsible. State the root cause in 1-2 sentences: *what* is wrong and *why* it produces the observed behavior. Do not pattern-match a fix onto the symptom — explain the mechanism. If there are multiple plausible causes, name them and say how you ruled the others out.

**3. Apply the MINIMAL fix.** Change only what's needed to fix the root cause. No drive-by refactors, no reformatting unrelated lines, no renaming. If you spot adjacent issues, list them at the end as 'noticed but not changed' — do not fold them into this fix.

**4. Verify (green).** Re-run the SAME test from step 1 and paste the now-passing output. Then run the broader suite (`{{TEST_COMMAND}}`) and paste the result to prove no regression. If the suite was already partially failing before your change, note which failures are pre-existing and unrelated.

**5. Report.** Summarize: root cause, the one-line essence of the fix, files touched, and the verification command anyone can re-run.

## Hard rules
- Steps 1 and 4 MUST include actual command output, not a description of what you expect to happen. 'Should pass now' is not acceptable — run it.
- If after a reasonable effort you cannot reproduce, STOP and report what you tried and what info you need, rather than guessing at a fix.
- Keep the diff small and reviewable. The test added in step 1 stays in the codebase as a regression guard.
- Respect existing project conventions (read AGENTS.md / CLAUDE.md if present for the right test command and style).