Problem

On a Slack DM, the chat agent repeatedly failed to create a task, reporting a "duplicate-check gate ... isn't letting the new task through even on retry." The deal (Soren Transport) already had 3 open tasks.

Root cause: the duplicate-awareness gate (§3.0) in createTaskTool is guarded by if (runId && …). It's designed to fire at most once and let the agent through on retry , but that bypass only works if the retry "breadcrumb" gets logged. Every chat surface (in-app chat + the Slack/iMessage/SMS/WhatsApp bots) runs with runId = 'noLoggingRunId', a sentinel that:

1. is a truthy string, so it triggers the gate, and
2. is the exact value logToolCall short-circuits on (tool-calls.ts:98), so it disables all tool-call logging.

With logging off, the gate's alreadyLooked query (a prior list-tasks row, or its own duplicateAwarenessGate: true log) can never find anything. So the gate re-fires on every attempt , neither retrying nor calling list-tasks first can clear it. Task creation is permanently impossible on chat for any deal that already has an open task. This affects in-app chat too, not just Slack.

The gate's own comment claimed "interactive chat is unaffected" , that assumption was that chat has no runId, but chat actually passes the truthy sentinel.

Fix

Exclude the sentinel from the guard:

if (runId && runId !== 'noLoggingRunId' && status !== 'recommended') {

This is the same runId !== 'noLoggingRunId' idiom already used across sibling tools (updateTaskTool, the calendar tools, writeDocumentTool). Now:

• Chat surfaces skip the gate entirely , if the user asks the agent to create a task, it always creates it (a duplicate is acceptable, per the chat contract).
• Real logged automation/event runs are unchanged , the gate still protects them, and there its list-tasks/breadcrumb retry mechanism actually works (each event execution logs its calls and handles one conversation).

Also rewrote the stale comment to describe the true invariant.

Tests

Added a regression case to createTaskTool.duplicate-gate.test.ts: with runId: 'noLoggingRunId' and an open task present, the task is created and the gate does not fire.

Verification

• pnpm --filter @zero/server run types → clean on both changed files
• pnpm --filter @zero/server exec vitest run …/createTaskTool.duplicate-gate.test.ts → 8/8 pass
• pnpm deps:check → no violations

Follow-up (not in this PR)

The 'noLoggingRunId' literal is now compared in ~8 files; sibling-draft-guard.ts already localizes it as a const. A shared exported constant would prevent a future typo'd literal from re-opening this class of bug.

🤖 Generated with Claude Code

Greptile Summary

The PR exempts the non-logging run-id sentinel from the task duplicate-awareness gate and adds a regression test for that context. However, the alternate Agent SDK in-app chat path supplies a generated UUID, so it remains subject to the gate.

• Updates the duplicate-awareness guard to skip noLoggingRunId.
• Documents why non-logging chat contexts cannot use the breadcrumb-based retry gate.
• Adds coverage showing sentinel-based chat contexts create tasks despite existing open tasks.

Confidence Score: 4/5

The PR should not merge until the Agent SDK in-app chat path is also made consistent with the intended chat exemption.

Sentinel-based chat calls now bypass the duplicate gate, but SDK-backed in-app chat propagates a generated UUID and therefore still returns the duplicate-awareness response when open tasks exist.

Files Needing Attention: apps/server/src/mastra/tools/task/createTaskTool.ts and its duplicate-gate regression tests

Important Files Changed