Stopping a Run & Reading the Timeline

How to control an agent run while it's running, and how to read what's happening on screen.

Stopping a run

Click Stop in the AgentTimeline footer while the run is in progress. The runtime will:

1. Notice the stop flag on the next loop iteration
2. Skip whatever step it would have taken next
3. Attempt one final synthesis only if its reservation fits the remaining dollar budget and wallet funds
4. Emit a final answer if synthesis succeeds, or retain the timeline and report why it stopped
5. Persist the terminal run status; a failed synthesis can leave the run marked failed

Worst case: one extra tool runs after you click stop, because the iteration that's already in flight finishes before the flag is checked. That's intentional — the alternative (hard cancel mid-tool) leaves the system in a less clean state.

The synthesized "stopped" answer is sometimes lower quality than a fully-completed one, particularly if the agent had a lot of observations it hadn't yet woven together. If you stop and the answer is bad, just rephrase the goal and start a fresh run.

Navigating away during a run

You can also navigate away from Agent Mode (click Consultations or any other nav button) — the agent will keep running in the background and the result will land in your chat history when it completes. The auto-flip back to chat mode happens via the same mechanism whether you watched it finish or not.

Reading a step card

Each step card has a collapsed and expanded state.

Collapsed

1. 📈 quote_lookup                        $0.00  >

• Step number (1, 2, 3...)
• Tool emoji (or 🛠️ for tools without a known emoji)
• Tool name
• Settled running cost in dollars (with sub-cent precision) (cumulative through this step)
• Expand chevron

Expanded

1. 📈 quote_lookup                        $0.00  v
   Plan: I'll look up SPX directly with quote_lookup since
         it's a market index.
   Args: symbols=["SPX"]
   Result: quote_lookup: 1 results
   tokens: 1240

• Plan — optional. The planner sometimes emits a plan event before the tool call explaining its reasoning. Not all runs have these.
• Args — one-line summary of what the tool was called with (e.g. symbols=["SPX"] for quote_lookup, query="TSMC earnings" for web_search)
• Result — one-line human summary of what the tool returned (e.g. quote_lookup: 1 results, web_search: 5 results, artifact_list: 3 artifacts)
• Tokens — tool observation metadata may be zero; use the run detail API for measured planner input/output tokens

The actual full tool input/output JSON is persisted to the database (agent_steps.input and agent_steps.output) so admins can inspect them via the admin console — but they're not shown in the UI panel because the JSON can be large.

Status icons

What lands in your chat history

When an agent run completes (finishes naturally OR is stopped + synthesizes), two messages are appended to your chat history:

1. A user message with your goal verbatim
2. An assistant message with the agent's final answer, labeled Agent Mode · N steps

Both messages are persisted via the regular messages system, so they're part of the conversation. You can ask a follow-up question that references the agent's answer ("can you elaborate on point 3?") and the regular chat completion will see the agent's answer as context.

What does NOT land in chat history:

• The individual step cards
• The planner's intermediate reasoning
• The raw tool inputs/outputs

Those are in the agent run's full record (the run's persisted agent_steps rows in the database) but the chat history only shows the final answer. This keeps the conversation readable.

The agent's answer is rendered through the same MessageContent component as regular chat replies, so markdown tables, code blocks, ordered lists, headings all work identically.

Note: Currently only the Ridgeline chassis shows the "Agent Mode · N steps" label on the assistant message. Other chassis show their regular chassis-themed label (e.g. "[ STDOUT ]" in Matrix, "Reply" in Paper). The message is still distinguishable because you literally just clicked Agent Mode and watched the run. Per-chassis badges are deferred polish.