Design persistent conversation goals and controlled automatic continuation

[Lewik]

Status

Preliminary design/research proposal, not a final specification or authorization to implement. Inspect the current runtime and issue history, compare options, and agree on the behavior with the user before implementation.

Motivation

Support requests such as "keep investigating until this regression is reproduced and fixed, or report a concrete blocker" without requiring the user to send "continue" after every intermediate result.

A goal should be a persisted objective with lifecycle controls and a continuation policy, not just a reminder added to every prompt. Ordinary one-off requests should remain ordinary requests.

Direction to investigate

• Store the objective, completion criteria, constraints, lifecycle state, and progress/accounting in conversation-scoped state.
• Let the Server schedule another normal agent turn at an idle boundary while the goal remains active. Reuse existing durable conversation scheduling rather than creating a second execution loop on a Worker.
• Give queued user input, cancellation, pending tools, and required user decisions precedence over automatic continuation. Define which Agent owns a goal in a multi-participant Conversation.
• Provide visible status and explicit create, pause, resume, and clear controls. Decide which transitions belong to the user/runtime and which the model may request.
• Define progress and stopping policies: completed, blocked, paused, optional budget/time limits, repeated failures, and turns that produce no useful action.
• Preserve the objective across compaction and reconnects. Explicitly decide restart/resume behavior; never replay an action with an unknown outcome merely to advance the goal.
• Require a completion report linked to tests, artifacts, or other evidence. Separate model-declared completion from any independently enforced verifier; a status update alone is not proof of success.

Research deliverables

1. A concrete state machine and turn-by-turn example, including an interruption and a blocker.
2. Integration points in current scheduling, queue, persistence, and UI contracts.
3. An evaluation plan covering successful completion, no-progress loops, queued corrections, cancellation, restart recovery, and honest partial results.
4. A small proposed first scope, with costs and alternatives. Do not turn every task into unlimited autonomous work.

References and related work

• Codex Goals: design and continuation model.
• Codex goal extension implementation, pinned source. Inspect scheduler, prompt, and tool handler separately rather than assuming every documented rule is mechanically enforced.
• Plan agent runtime and tool execution improvements #84: runtime and tool-execution planning.
• Preserve queued messages when stopping active work #95: preserving queued messages when stopping active work.
• Plan instruction and temporal-context improvements #86: instruction and temporal-context planning; reminder placement is related but is not the goal scheduler itself.