Plan mode

For most chat request that needs more than one tool call to satisfy, the in-app assistant writes a short prose plan before doing anything. You see the plan in a card with two buttons — Approve & run and Abort. This page describes what triggers plan mode, what the card shows, and how to keep it out of your way for quick one-shot questions.

Why plan mode exists

Complex workspace operations — setting up a new study, triaging a result batch, importing a CSV and wiring it into a document — are not a single tool call. They require the assistant to first understand its surroundings (which environments and tables exist, which protocol the user means, what the relevant document already contains, what tools are available for the job) and then choose a coherent sequence of actions that gets to the result. Plan mode is the dedicated step where that grounding and sequencing happen.

During planning the assistant searches the wiki, reads relevant workspace documents, checks the tool catalogue, and drafts the steps it intends to run. The plan then surfaces in a card with Approve & run and Abort buttons — so you can intervene on the shape of the answer once instead of correcting tool calls twelve times. The card is a consequence of planning, not its motivation: plan mode is primarily how the model thinks through a complex request, and only secondarily a confirmation surface.

What triggers plan mode

Plan mode is automatic. The assistant decides based on the request:

• "What does our IFN-γ protocol look like?" — direct answer, no plan.
• "Search the wiki for the column-type cheat sheet." — direct answer, no plan.
• "Set up a PK study schema for DLA-7 and import the dose groups CSV." — plan.
• "Triage these 40 result rows: flag any with CV > 15% and write a summary block at the bottom of the doc." — plan.

You can also explicitly ask the assistant to plan first ("draft a plan, don't execute yet") for anything.

What the plan card shows

The card has a header (the word Plan, a status badge like Needs approval / Executing / Done, and a rev N chip when the assistant revised an earlier draft of the plan), followed by the plan content itself:

Rationale

Optional one-line italic summary above the body — the headline form of the approach. Lets you scan the intent without reading the full markdown.

Plan body

Markdown — a few numbered or bulleted steps, each one sentence. References specific tools where relevant ("call data_objects.bulk_create with the parsed CSV") and specific resources by name.

Plan envelope

Optional collapsible sections that scaffold the plan: Context (why this work), Assumptions (load-bearing facts the assistant is taking as given), Critical files (resources to be touched), Reuse (existing functions / tools the plan will lean on rather than reinvent), and Verification (end-to-end steps you can follow afterwards). Empty sections are hidden.

Approve & run / Abort

Buttons that only render while the plan is awaiting approval. After approval they disappear and the live execution checklist takes over.

Live checklist

Once execution starts, todos render below the plan with circle / spinner / check icons as each step progresses. Driven by the assistant's update_todos tool, not by the plan itself — the plan is the upfront commitment, the checklist is what got done.

Not every plan uses every part. The plan card is and Plan envelope are important tools for checking the assumptions and expected behaviours from the models perspective.

Before you see the card

Before emitting the plan card, the assistant grounds the request against the wiki, webresources and your workspace, drafts an approach, double-checks the assumptions, and only then writes the card.

After you approve

The assistant executes step-by-step. Each updating tool call still gets its own approval card (see AI overview) unless you previously chose "Always approve this action type" for the same tool. Read-only steps run straight through.

Revising a plan

There's no inline plan editor — the only two choices on the card are Approve & run and Abort. If the plan needs changes, hit Abort and send a follow-up message describing what to change ("don't touch the protocol doc — write the summary to the daily debrief doc instead"). The assistant replans from the updated conversation context on the next turn.

Aborting a plan

Abort discards the plan but keeps the conversation. The assistant acknowledges and waits for the next message. Nothing has run, nothing has changed.

What's next