[jgluck96]

Summary

Phase 5.1 of the OSS sync roadmap — a selective merge of task-bridge.ts security hardening that OSS sutando has but stando was missing. This is not a wholesale port: stando's deliberate divergences (state-paths.js layout, watch-tasks-stream.mjs watcher, cloud-client telemetry) are left untouched.

Branches off main (independent of the UI-extraction PRs #68/#69).

What changed (3 hardening items)

1. Field ordering (PR fix(agent-api): close task-file field injection via /task body sonichi/sutando#982) — all task writers (writeChatTask, the voice workTool writer, context-drop) now emit task: last and add a priority: field (normal for chat/context-drop, urgent for voice). A multi-line user/Gemini body can no longer forge header fields below it.
   • 🔑 This aligns task-bridge.ts with stando's own agent-api.py:834, which already orders task: last. task-bridge.ts was the lone inconsistent writer, and task_priority.py already exists to consume the priority: field.
2. _isVoiceTask stops at the first task: line — the residual half of fix(agent-api): close task-file field injection via /task body sonichi/sutando#982. Without it, a body line channel_id: local-voice would forge a voice-task classification (affects the offline DM-forwarding gate). Added a FORGED_VOICE_BODY regression test; updated existing fixtures to the new task:-last format.
3. _shouldFallthrough allowlist guard (issue task-bridge: belt-suspenders guard for non-task .txt files (narrow-scope-followup from #1033) sonichi/sutando#1035) — the result-watcher's unconditional onResult() fallthrough is now gated to filenames task-bridge legitimately delivers (task-/voice-/proactive-), so a future per-channel pull namespace won't be injected into voice.

Deliberately NOT included

• Feat/beta monetization wave 1 #4 recent-transcript context block (ASR self-correction) — adds behavior, not hardening. Can follow separately.
• onboarding and screenshare fix #5 context-drop auto-archive (context-drop tasks have no result-delivery path → results pile in results/ sonichi/sutando#969) — requires a companion main.swift change (stando's Swift doesn't write source: context-drop).

Test plan

• tests/task-bridge-isvoice.test.ts — 8/8 incl. new forge-guard regression
• tests/task-bridge-format.test.ts — 5/5 (field matches are line-anchored, order-independent)
• Full TS suite: 241/241 pass
• tsc --noEmit clean

Related

• OSS source: sonichi/sutando src/task-bridge.ts (PR fix(agent-api): close task-file field injection via /task body sonichi/sutando#982, issue task-bridge: belt-suspenders guard for non-task .txt files (narrow-scope-followup from #1033) sonichi/sutando#1035)
• Roadmap: Phase 5.1 (task-bridge) of the OSS→stando backend reconciliation

Made with Cursor

[bassilkhilo-ag2]

Hi Josh — one voice-delivery regression to flag before merge.

PR #70's _shouldFallthrough allowlist is:

return file.startsWith('task-') || file.startsWith('voice-') || file.startsWith('proactive-');

This misses four result prefixes that are explicitly written to be spoken aloud when the voice client is connected:

• question- — written by src/check-pending-questions.py:115
• insight- — written by src/daily-insight.py
• friction- — written by src/friction-detector.py
• briefing- — written by skills/morning-briefing/SKILL.md:49

The discord bridge already codifies these in FALLBACK_PREFIXES at discord-bridge.py:3145:

FALLBACK_PREFIXES = ("task-", "question-", "briefing-", "insight-", "friction-")

As written, this PR would silently stop voice from speaking pending-question pings, daily insights, friction reports, and morning briefings. Suggested one-liner:

export function _shouldFallthrough(file: string): boolean {
  return ['task-','voice-','proactive-','question-','briefing-','insight-','friction-']
    .some(p => file.startsWith(p));
}

PR #94 (unit tests for this guard) also only covers the existing three prefixes — would be good to extend those tests to cover the four new ones too.

Happy to prepare a follow-up commit if helpful.

[bassilkhilo-ag2]

Supporting Bassil's review — here's the exact change needed for _shouldFallthrough:

export function _shouldFallthrough(file: string): boolean {
    return (
        file.startsWith('task-') ||
        file.startsWith('voice-') ||
        file.startsWith('proactive-') ||
        file.startsWith('question-') ||   // check-pending-questions.py:132
        file.startsWith('insight-') ||    // daily-insight.py
        file.startsWith('friction-') ||   // friction-detector.py
        file.startsWith('briefing-')      // morning-briefing skill
    );
}

Source verification:

• question-: src/check-pending-questions.py:132 → path = RESULTS_DIR / f"question-{ts}.txt"
• insight-: src/daily-insight.py docstring → "Output: results/insight-{date}.txt (voice agent can speak it)"
• friction-: src/friction-detector.py → "Output: results/friction-{date}.txt"
• briefing-: skills/morning-briefing/SKILL.md:49 → "Write to results/briefing-{date}.txt so the voice agent can speak it"

Also note: tests/task-bridge-fallthrough-guard.test.ts line 59 currently asserts question- → false (in the "rejects unknown" it() block) — that will need to move to an "allows voice-delivery result types" block alongside the other three new prefixes.