Give completed Agent thoughts descriptive labels

[marksftw]

Closes #646

What changed

• Track newly completed Agent reasoning phases without changing timeline order or content.
• Generate one bounded, phase-specific present-tense label with auto:quick.
• Keep Thinking visible while the label is pending, fall back to Thought after five seconds or on failure, and fade between label changes.
• Validate and locally persist only single-line labels, reuse cached labels, and clear them with the related task, turn, or Agent history.
• Coalesce duplicate in-flight requests and invalidate pending work safely when history is replaced or deleted.

Why

Completed reasoning rows previously used the hard-coded Thought label, which gave no indication of what a phase had done. The generated-label prompt now treats the user request as background and prioritizes the current phase's latest concrete action, producing more useful labels without adding conversation history, tool data, or an immediate-action field.

Impact

Label generation is non-blocking and does not delay Agent activity, approvals, final responses, or run completion. Invalid or failed responses remain a silent Thought fallback, and loading old history does not trigger new requests.

Validation

• bun run test — 170 passing
• bun run typecheck
• bun run lint — 0 errors (12 existing warnings)
• bun run format:check

The desktop binary was not built.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 0cac15e9-33a2-466a-83dc-e6b467fd31ef

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/issue-646-thought-labels

---

_{Comment @coderabbitai help to get the list of available commands.}

[AnthonyRonning]

Just tested it, had a few thoughts on the approach.

When it comes to the thinking thoughts, I know there's a couple approaches. One is a continual summarizer that runs on whatever tokens have already streamed in. Updating after a few seconds with the latest amount. This gives an appearance like it is actively thinking.

Another is sort of combining the thoughts + content + tool calls + repeat into the collapsed form and doing something similar with the continuous summarizer. Though I kinda like that they stream in first, and then collapse down.

Separately of the approach, I think whatever prompt you're using for the summarization would need some more examples of what it should output. These make it look like it is making the tool calls, but it is just thinking about the tool calls it might want to make.

We should use llama or gemma for this, not gpt-oss, It has random problems at times.

[marksftw]

Thanks, this makes sense. I agree the current labels can sometimes sound like the agent is already making a tool call when it is really only considering one. I'll adjust that.

As for the concept of collapsing the reasoning and tool calling into UI phases, that is already part of the broader [Agent timeline exploration](#635). I'm trying your suggestion from last week to break UX experiments into smaller chunks instead of trying to redesign the whole timeline at once. I also like the direction of letting activity stream in first and collapsing it afterward.

For this pass, I’m going to try gemma, strengthen the prompt with examples that distinguish intent from actual execution, and test a bounded version of the continual summarizer: one provisional label during longer thoughts, followed by a final label when the thought ends. If gemma isn’t fast enough, I’ll try llama and retune the prompt for it.

[AnthonyRonning]

Makes sense. Make sure to set reasoning off for gemma. I have an example of that with my auto approver classifier for tool calls.