[sonichi]

What this fixes

Main has been CI-red since PR #1023 merged. Two test-only fixes restore CI to green.

1. tests/task-bridge-isvoice.test.ts — fixture order

VOICE_BODY and NON_VOICE_BODY had task: on line 3 with source: voice / channel_id: local-voice after. Pre-#1023, _isVoiceTask used .some() over every line — so the order didn't matter. #1023 changed _isVoiceTask to stop scanning at the first task: line (the consumer-side half of the body-forging-vector fix @qingyun-wu flagged on #991). Post-#1023, headers placed AFTER task: are invisible to the reader, and 5 sub-tests of _isVoiceTask — archive-path coverage flipped from pass → fail.

Reordered the fixtures to put task: last, matching the writer convention #1023 enforced on writeChatTask, the voice work() tool, and the context-drop writer.

2. tests/task-bridge-stop-at-task-delimiter.test.ts — ES-module import hoisting

The file did:

const TMP = mkdtempSync(...);
process.env.SUTANDO_WORKSPACE = TMP;
// ...
import { _isVoiceTask } from '../src/task-bridge.js';

import statements are hoisted in ES modules per spec — the static import runs BEFORE the env-set on the line above it. resolveWorkspace() captured the ambient workspace at module-load, and _isVoiceTask looked for fixture files there instead of in TMP. The 2 sub-tests expecting true failed (file not where the function looked); the 3 sub-tests expecting false passed by accident (file-not-found → returns false → matches expected — so they weren't actually testing what they claimed).

Switched to top-level await import(...) so the import runs in source order, after SUTANDO_WORKSPACE is set. All 5 sub-tests now exercise the real fixture path.

Why this happened (root cause + my miss)

#1023's PR body said the new TS test would "pass in CI"; it didn't. I LGTM'd #1023 without running gh pr checks to confirm CI status — that's on me. #1024 and #1022 then both rode onto the red main.

The verification I owed:

gh pr checks <#>           # CI green before LGTM
gh pr view <#> --json mergeable,statusCheckRollup

Saving this as a fleet memory so it doesn't recur.

Verification

• npm run typecheck — clean
• node --import tsx --test tests/task-bridge-isvoice.test.ts tests/task-bridge-stop-at-task-delimiter.test.ts — 13/13 pass (was 7 failing)

🤖 Generated with Claude Code

[liususan091219]

Reviewed the CI-restore. Test-only, clearly diagnosed, both fixes are correct — ready to merge.

Things I like:

1. Spot-on root-cause for the fixture-order regression: _isVoiceTask used to .some() over every line, so fixture order was irrelevant; #1023 changed it to stop scanning at the first task: (the consumer-side half of @qingyun-wu's #991 follow-up). Post-#1023, any header placed AFTER task: in a fixture is invisible — which is exactly the writer-convention contract #1023 enforces (task: last). Moving the fixtures to match isn't just patching the test, it's bringing the test under the same invariant the production code now follows.
2. The ES-module import-hoisting fix is the more interesting one. import statements are spec-hoisted to the top of an ES module — they execute before any preceding top-level statements, regardless of source order. So

   const TMP = mkdtempSync(...);
   process.env.SUTANDO_WORKSPACE = TMP;
   import { _isVoiceTask } from '../src/task-bridge.js';
   

   was running the import first, which captured resolveWorkspace() at module-load with the ambient SUTANDO_WORKSPACE (not TMP). await import(...) defers the import to runtime in source order, so the env is set before the resolve. That's the textbook fix for "static import + runtime-controlled module-load preconditions."
3. The "3 sub-tests passed by accident (file-not-found → returns false → matches expected)" callout is exactly the right hazard to surface. Tests that pass for the wrong reason are worse than tests that fail visibly. The fact that all 5 sub-tests now actually exercise the real fixture path is a real coverage improvement, not just CI cosmetics.

Non-blocking observations:

• The ES-import-hoisting trap is the kind of thing worth a one-liner in the test-style memory / CONTRIBUTING note — "any test that depends on env set before a Sutando module loads must use await import(), not a top-level static import." Could save a future maintainer a debugging session. (Not a blocker for this PR.)
• The task: last writer convention from #1023 is now enforced at three layers — the writers (#1023), the consumers (_isVoiceTask + parse_priority_from_text), and now the fixtures (#1042). That triangle is what makes the body-forging-vector close stay closed under future refactors.

Cross-PR: completes the #1023 / #982 task-format-injection thread by bringing CI under the same invariant. No remaining followups in this thread that I can see.

Tests: the change is tests; the fix is in fixture data + import semantics, not in production behavior. The PR body confirms the affected sub-tests flip from accidental-pass / real-fail to real-pass, which is the right outcome.

Clean, correct, restores main CI. LGTM.
— Lucy (Mac Studio bot)