[witbrock]

Merge decision: ready — Von can choose a bounded background inquiry, keep foreground conversation moving, and later use its sourced task product without replacing newer conversation state.

User outcome

Implements native task #V#task_agent_42b6d9f8b35a707c5a43bf57428c4c8e. An unfamiliar idea or useful connection can merit quiet investigation; novelty alone need not trigger work. The prior path shared the source conversation's execution fence and supplied no conversational inquiry/product-consumption capability.

Material changes

• Reuse canonical task creation, launch/outbox reconciliation and the existing dispatcher with a server-derived independent conversation carrier. Preserve and revalidate source actor, organisation, creator and conversation authority.
• Add an optional model-selected inquiry capability with exact source mention/turn, bounded context, stable retry identity and a compact pending receipt. It cannot recursively expand into more inquiries.
• Project up to three actor-scoped current products into later foreground context, retaining source references, full-content hashes and truncation status. Foreground judgement reconciles corrections, declines and whether to raise a finding; background completion does not write its situation.
• Document mechanism, actual-model evidence and limitations in docs/engineering/conversational_rumination.md. No scheduler, worker process, parallel product store or live represented-policy migration.

Evidence

144 targeted execution, queue, task-product, conversation-projection and inquiry tests passed. The eight inquiry tests passed again after compacting the submission receipt. The ordinary-turn suite had 280 passes and two database-authentication failures independently reproduced with unchanged source. PDM lock, minimum Python 3.11 grammar and diff checks passed.

Actual authenticated isolated-fixture conversations used provider-observed gpt-6-astra. Two ordinary title requests quietly launched inquiries; foreground answers returned in 26–29 seconds while background work continued for about four minutes. An unrelated foreground request completed in 10 seconds during research. Canonical products retained source context, attempted/retrieved URLs, uncertainty and retrieval evidence. The positive case retrieved arXiv 2304.03442's actual abstract; a later turn cited it, distinguished supported claims from extrapolation and offered a useful study question. Correction-before-completion retained “Cyc, not Scone”. Two turns after a decline did not raise the report again. Another ordinary conversation chose no inquiry.

Ship boundary

• Minimum criteria met: independent execution, durable sourced product, later grounded use, varied initiative/no-question cases, and correction/stale-context handling. Exact retry and authority boundaries have targeted tests.
• Stop-ship conditions checked: foreground waiting for research, invented retrieval, source-situation loss and duplicate launches on transport retry.
• Limits: successful retrieval evidence is abstract-level, not full-paper research or broad role competence. Sparse fixture vocabulary increased product-persistence recovery: background runs used 24–32 calls and 0.77–1.14 million reported tokens, predominantly cached input. Exact pricing was unavailable. Tavily retrieval lacked a fixture key; those failures were retained honestly. No cost-efficiency claim.
• Publication only. No public deployment, credentials, production data/membership changes, infrastructure or database migration is requested.