Stop ingestion from manufacturing spurious or auto-trusted tasks

[marcelsamyn]

Problem

A daily-brief conversation ingest (life updates + loose planning) was turned into a Task node titled "Sunday Morning Check-In: Zouk Social, Masterclass, and Weekly Planning" with the conversation summary stuffed into its description — and it landed in the trusted commitment band, not marked tentative.

Two root causes in src/lib/extract-graph.ts:

1. Nothing stopped task invention. The "CURRENT OPEN TASKS" prompt section told the model to mint a new Task node for "brand-new tasks" with no guidance on what qualifies. So assistant suggestions, recaps, and loose planning got packaged into tasks — including meta "check-in" tasks that just restate that the conversation happened.
2. Nothing forced the tentative flag. "Tentative" is stored as assertedByKind = 'assistant_inferred' (this splits trusted open commitments from candidate commitments). New tasks never had their provenance pinned, so _resolveAssertedByKind fell back to "user" (trusted).

Changes

Prompt hardening (_formatOpenCommitmentsSection):

• A task must be a concrete, actionable item with a clear owner.
• Explicit DO-NOT list: no tasks from assistant suggestions/proposals, recaps/summaries, uncommitted planning/brainstorming, or the existence of the conversation itself (no "check-in" / "weekly planning" / "review" meta-tasks).
• Task labels must be short imperative actions, never a conversation summary/recap.

Deterministic tentative policy (code):

• Every brand-new Task node created during document/conversation ingestion is now recorded as assistant_inferred (tentative) — unconditionally, regardless of source type or what the model asserts. Background ingestion never autonomously creates a firm commitment.
• Status updates against existing tasks are untouched, so the candidate→trusted promotion flow (explicit user confirmation in a later source) and the commitment write APIs still produce firm commitments.

Tests (extract-graph.test.ts): two regression tests — a new task with omitted provenance lands tentative, and a new task the model marks "user" is still forced tentative. All 8 tests in the file pass.

Notes

• Typecheck + lint clean on the changed files.
• 3 failures in commitments.test.ts / open-commitments.test.ts are pre-existing — confirmed identical on the pristine tree; they don't touch extraction.
• This only prevents recurrence; the existing bad task is still in the store.

https://claude.ai/code/session_01ByQzXHcjjLCgUepJn5VER7

---

Generated by Claude Code

[gemini-code-assist]

Code Review

This pull request ensures that any brand-new tasks created during conversation or document ingestion are deterministically recorded as tentative (using the assistant_inferred assertion kind), preventing passive background ingestion from creating firm commitments. It also updates the graph-extraction prompt instructions to explicitly forbid manufacturing spurious tasks (e.g., from assistant suggestions, recaps, or meta-conversations) and requires task labels to be short imperative actions. Corresponding unit tests have been added to verify this behavior. There are no review comments, so I have no feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.