Summary
Improve the in-thread rendering for reasoning traces and tool calls so users can follow what the agent is doing while it is working, closer to the native Claude Code CLI experience.
Right now the collapsed work-log / reasoning trace rows are too opaque. I want to be able to quickly see what was called or executed, get a useful preview of the output/result, and expand when I need the full details.
Desired behavior
- Reasoning / work-log entries should be easier to scan while a turn is still running.
- Tool calls should show the important metadata in the collapsed row, for example:
- tool name / command type
- the command or tool input that was executed, truncated sensibly
- status: running, succeeded, failed, interrupted
- exit code and duration when available
- Tool results should include a compact output preview in the timeline.
- There should be an explicit expand affordance to show the whole output/result.
- Long stdout/stderr/results should not dominate the conversation by default, but should be recoverable without leaving the thread.
- Rendering should work for command execution, MCP/tool calls, file changes/diffs, subagent activity, and provider-specific tool payloads where the data exists.
- The UI should preserve enough context to understand what the model did without needing to inspect raw logs.
Motivation
The native Claude Code CLI output shown while the model is working is much clearer: it is easy to see the sequence of actions, what is currently running, and enough result context to know whether the model is making progress or going down the wrong path.
In T3 Code, the current rendering makes it harder to answer basic questions like:
- What command/tool did the model just call?
- Did it succeed or fail?
- What was the relevant output?
- Is this worth expanding, or can I ignore it?
- Which hidden trace entry contains the thing I need?
This matters especially for long-running implementation sessions, debugging, failed commands, and sessions with subagents/tool-heavy workflows.
Related upstream work / prior art checked
I searched upstream before filing. Related but not fully covering this request:
This issue is meant to capture the broader UX target: a first-class, live, readable trace/tool-call timeline with compact previews and full expansion, using the Claude Code CLI output quality as the benchmark.
Acceptance criteria
- A running turn shows a readable sequence of reasoning/work-log/tool-call rows.
- Collapsed rows include enough information to identify the call and its outcome.
- Command/tool outputs show a useful preview without flooding the chat.
- Users can expand a row to inspect the complete output/result.
- Failed commands/tools are visually distinguishable from successful ones.
- Large output remains performant and does not cause layout jank.
- The behavior is covered by tests for trace/tool row derivation and rendering, including long output truncation/expansion.
Summary
Improve the in-thread rendering for reasoning traces and tool calls so users can follow what the agent is doing while it is working, closer to the native Claude Code CLI experience.
Right now the collapsed work-log / reasoning trace rows are too opaque. I want to be able to quickly see what was called or executed, get a useful preview of the output/result, and expand when I need the full details.
Desired behavior
Motivation
The native Claude Code CLI output shown while the model is working is much clearer: it is easy to see the sequence of actions, what is currently running, and enough result context to know whether the model is making progress or going down the wrong path.
In T3 Code, the current rendering makes it harder to answer basic questions like:
This matters especially for long-running implementation sessions, debugging, failed commands, and sessions with subagents/tool-heavy workflows.
Related upstream work / prior art checked
I searched upstream before filing. Related but not fully covering this request:
This issue is meant to capture the broader UX target: a first-class, live, readable trace/tool-call timeline with compact previews and full expansion, using the Claude Code CLI output quality as the benchmark.
Acceptance criteria