[VasiliyRad]

Summary

Bug-by-omission. Upstream sonichi/sutando shipped PR sonichi#820 (fix(task-bridge): dedup work() submissions with identical text within 2-minute window) on 2026-05-18; the merge that pulled the surrounding code into AG2Platform/stando didn't include this branch, leaving voice over-delegation uncontained here.

Symptom: Gemini retrying work() 3-5 times with identical task text in sub-second succession spawns N parallel task files, each runs independently, each ships its own result DM — the "cascade of duplicate result DMs" the upstream commit calls out.

Fix

Mirrors upstream structure:

• PendingTask grows a taskText field so dedup can match against in-memory state without re-reading task files.
• DEDUP_WINDOW_MS = 2 * 60 * 1000 module constant.
• normalizeTask(t) = lowercase + collapse whitespace + trim + cap at 150 chars. Catches the rewording-noise that triggers the over-delegation pattern (Gemini retrying with extra spaces, casing changes, trailing punctuation past the cap).
• workTool.execute scans _pendingTasks for a normalized-match entry submitted within DEDUP_WINDOW_MS. On hit, returns { status: "duplicate", taskId, message } and skips the task-file write. The message explicitly tells the agent NOT to retry.

Tests

tests/task-bridge-work-dedup.test.ts — 3 integration cases. Invokes the real workTool.execute(...) against the real SUTANDO_WORKSPACE/tasks/ dir; uses the same task-file leak-check harness as tests/task-bridge-format.test.ts so any future regression that writes orphan files is caught.

Case	Pins
second identical call within window returns duplicate with first taskId	Core regression guard — over-delegation containment.
dedup ignores whitespace and case differences	normalizeTask pin — Gemini's retry rewording variants must still collide.
different task texts each produce a fresh taskId	No-false-positive — distinct work doesn't collapse.

3/3 pass locally; existing 12 task-bridge tests (*-format, *-isvoice) still pass.

Bug-finding calibration

1 real bug (port-gap) fixed across 3 tests. Total this round: 7 bugs / ~90 tests ≈ 1:13.

Test plan

• npx tsx --test tests/task-bridge-work-dedup.test.ts — 3/3 pass
• npx tsx --test tests/task-bridge-format.test.ts tests/task-bridge-isvoice.test.ts — 12/12 pass (no regressions)
• npm run typecheck — clean
• Manual: simulate voice over-delegation by calling workTool.execute rapidly with identical text; verify only ONE task file lands in tasks/

🤖 Generated with Claude Code

[bassilkhilo-ag2]

Clean port. A few observations:

Logic: The dedup scan is O(n) over _pendingTasks. In practice this map stays small (tasks expire within their timeout window, typically ≤10 min), so no concern. One micro-nit: normalizeTask(pending.taskText) is recomputed on every loop iteration. Could cache as pending.normalizedText at insertion time, but the savings are negligible here.

Return value: The message field explicitly tells the model not to retry — good anti-footgun. The caller at workTool.execute returns { status: 'duplicate', taskId, message } while a fresh task returns just { taskId, ... } (no status field). Make sure all callers that pattern-match on status === 'duplicate' are covered — the test checks this correctly.

Test isolation gap (non-blocking): afterEach cleans up task files but doesn't reset _pendingTasks (module-level). Each test's first submission stays in the map for subsequent tests. The three cases here use distinct texts so no collision, but future test authors who reuse a prior case's text will see unexpected dedup. Worth adding _pendingTasks.clear() in afterEach — requires exporting the map or a _resetForTest() helper from task-bridge.ts.

Tests: Core regression case, normalization variant, and non-collision case all covered. Leak harness borrowed from the existing format-test pattern — good consistency.

[rickchen007]

Code review: LGTM. Clean port of upstream PR sonichi#820 — normalizeTask collapses the rewording-noise that drives voice over-delegation, the dedup scan + status: 'duplicate' short-circuit is correct, and the 3 integration tests use the real workTool.execute with the leak-check harness.

Not merged — PR is CONFLICTING with main after today's 9 merges. Needs a rebase onto main; once it's mergeable I'll merge it (no code changes needed).

[VasiliyRad]

Closing — current 's src/task-bridge.ts already has the dedup logic (DEDUP_WINDOW_MS, normalizeTask, and the dedup check in workTool.execute). The two implementations are functionally equivalent; main's version landed first via a parallel route. This PR is now redundant.