Docs navigation
Docs / Plan collaboration

Plan collaboration

Plan mode helps clarify a request and prepare a versioned proposal. The same Agent investigates the workspace, discusses choices and submits a formal plan for work requests. A substantive work request in Plan mode asks the Agent to plan that work, even without the word “plan”. The Agent investigates first, uses structured questions for missing material decisions when available, and submits the proposal when ready. Investigation and other authorized tool work may happen before submission, but it does not replace the proposal or turn a proposal into an approval. Greetings, explanations and requests to keep discussing can finish without producing a new revision. Revising a proposal preserves its earlier revisions in history.

Planning uses the ordinary Agent tool surface and its normal permission, approval and sandbox policies. Commands, tests, builds and other authorized tool calls may happen while the proposal is being prepared; they do not replace the required proposal. For a substantive work request, the planning turn must call submit_plan before it ends. A questionnaire is optional and is only needed for a material user decision that cannot be discovered or reasonably defaulted. Subagents inherit the planning intent and their parent’s permission ceiling. The main task owns formal proposals and Goal controls.

Implement and adapt

Implementing approves a particular revision and starts an ordinary Default Agent task. The task can reorder work, add verification, repair a failed check and publish again after fixing an artifact. Changes in approach within the user’s authorization do not require another proposal approval. Actions outside that authorization still require confirmation through the normal controls. Changing methods or step order does not remove the approved deliverables or acceptance requirements in the proposal’s body and step details.

Before claiming completion, the Agent should inspect the final artifact or resulting state and compare it with those requirements. For example, counting document headings and tables checks structure; it does not establish that the document’s content meets every requirement. Unmet requirements need repair and another check. For documents, reopen the final saved file with a normal reader (for example, python-docx for DOCX) and read its body and tables. Raw ZIP/XML extraction alone does not check the complete document package. Preserve unrelated content and document parts while repairing omissions, and repeat affected checks after any further edits on the version actually delivered. If completion or verification is not possible, the final answer must identify the remaining work and limit its claims to the available evidence.

The proposal and current progress are separate. update_plan replaces an optional task progress list of up to 20 steps. Steps can be added, removed, reordered or reopened. For substantive multi-step implementation, the Agent should report its actual execution list before substantive implementation, showing the first work in progress, and update at meaningful milestones and before the final response when needed. Related changes can be batched; reporting should not all be deferred until the end. Progress does not restrict tools, schedule future turns, or determine whether a task may finish. A successful implementation turn completes its associated run without inventing verification for unreported steps. This runtime status records the turn’s normal completion; it is not independent proof that the deliverable meets its semantic acceptance requirements.

Artifact publication performs the supported format checks and publishes the artifact. It does not establish semantic acceptance, mark plan steps complete or prevent subsequent inspection and repair. When a downloadable file is part of the requested delivery, publish the checked final version and include its returned artifact reference; a workspace path alone is not that download. The user’s request already authorizes that file delivery in the conversation, so it does not need another permission question.

Hide, cancel and continue

  • A normally ended Plan run shows the neutral Execution ended summary with its last reported step statuses and count, then hides after about two seconds. Unreported steps stay unreported, including 0/4; ending the turn does not fill the checklist or add an overall success check. Cancelled, failed and paused runs retain their distinct meanings.
  • Hide plan hides the selected revision’s default card. Its body and history remain available, and the preference survives refresh. Restore it from history to show the card again. A new revision has its own presentation preference.
  • Hiding a proposal does not change collaboration mode or cancel a task. Active progress and Stop remain available separately.
  • Cancel execution and composer Stop use the same ordinary task cancel path. Queued work can be cancelled before model execution; running work stops at the normal cancellation boundaries. The UI confirms cancellation after task termination is acknowledged.
  • Reconnection first recovers the existing task. Retrying an implementation whose acceptance response was lost retains its request identity and destination rather than silently starting another task.
  • Continuing terminated work preserves the conversation, proposal and artifacts. A new ordinary task checks the actual state before acting. An unfinished checkbox is not evidence that an external action must be replayed. Commands are not guaranteed to execute exactly once across process crashes.

A questionnaire keeps its task and tool-call identity while waiting. Other sessions can use the released compute capacity; another task in the same session still waits for its execution lane. Answering reacquires capacity and continues that tool call. A Gateway restart ends the old execution; it does not restore a partially executing coroutine.

Goal mode · Web UI · Sandbox

Edit this page on GitHub OpenSquilla docs · synced from main