[allozaur]

Overview

Tool call blocks stay folded while the model writes the call and while it runs; they only open on their own with Always show tool call content, which then leaves them open forever.

Adds Show tool call in progress (Settings > Display, on by default): a tool call is expanded by default while its arguments stream and while it executes, then folds back on its own once its result is in — the tool-call counterpart of Show thought in progress. A block you open or close by hand keeps your choice.

Because in-flight calls are now on screen by default, they have to render while in flight. The rest of the PR closes the cases where they didn't: a running exec_shell_command folded itself the moment it started, its command appeared only once the args JSON was complete, and write_file / edit_file showed an empty block until a path had streamed.

Details

• New showToolCallInProgress checkbox in Settings > Display (default true), listed above Always show tool call content; that one's help text now spells out the split — in-progress covers the streaming/executing phases, always show keeps finished calls expanded and wins on its own.
• Default expansion moves into a pure isSectionExpandedByDefault() in $lib/utils so the rules are testable without DOM. A manual toggle always overrides the default, both ways. Existing installs inherit the default via the config merge, no migration.
• exec_shell_command creates an empty tool result before reading output, so its section already reads TOOL_CALL while live. A section whose toolCallId matches the executing call therefore stays expanded regardless of type, and releases when the loop reports the call done.
• That command now streams into the terminal title from the first character (args parsed partially), with a blinking cursor until the JSON parses; then "Executing command..." (was "Running...") holds until the first output chunk so the block never goes blank. The prompt shows the conversation cwd until the result message carries the real x-tool-cwd (new conversationCwd prop).
• write_file / edit_file no longer bail out when the args carry no path — it's display-only, so content/edits render with filePath / fileName undefined (now optional), falling back to pretty-printed raw args JSON, then "Receiving arguments..." while nothing has arrived, then "No content" / "No edits". Path-alias and basename handling collapses into shared extractToolArgPath() / fileNameFromPath().
• New unit + client tests for the setting, the expansion truth table, the streaming titles, and the path-less bodies; existing parser tests updated.

Notes

tools/ui only — no server, API, or storage changes. Only default expansion is affected; nothing per-block is persisted, and collapsed blocks still skip parsing the args/result blobs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, llama-ui (ggml-org/Qwen3.8-Flash-Next:Q8_0)¹, Delta (zai-org/GLM-5.3-Flash)

Footnotes

1. https://huggingface.co/datasets/allozaur/llama-ui-traces/blob/main/2026-10-16_17-14-02_conv_add1290b_create_new_branch_in.jsonl ↩

[Copilot]

🟡 Changes recommended

Streaming shell calls collapse before execution finishes, hiding live output.

2 open findings

• Keep actively executing tool calls expanded · New
• Test collapse transition when a call receives its result · New

What changed in this PR

Adds a default-enabled Display setting to expand tool calls while running, similar to reasoning blocks.

Changes:

• Registers the setting and updates help text.
• Adjusts expansion defaults while preserving manual choices.
• Adds settings and component tests.

File	Description
tools/​ui/​tests/​unit/​settings-tool-call-in-progress.test.ts	Checks setting registration and default.
tools/​ui/​tests/​client/​tool-call-in-progress.svelte.test.ts	Tests expansion settings and manual toggles.
tools/​ui/​src/​lib/​constants/​settings.constants.ts	Adds checkbox and help text.
tools/​ui/​src/​lib/​constants/​settings-keys.constants.ts	Defines the setting key.
tools/​ui/​src/​lib/​components/​app/​chat/​ChatMessages/​ChatMessageAgenticContent.svelte	Applies progress-based expansion defaults.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ServeurpersoCom]

There's a catch with exec_shell_command: to stream the terminal output live, it creates an empty result message first and fills it as the command writes, so the UI sees a result and folds the block right when the command starts, hiding the live output.

The component already knows which call is actually running (currentlyExecutingToolCallId), so keeping the block open while its id matches would fix it.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

13 open issues

• {:else if writeFileBody?.content} treats an empty string as “no content”, which can mis-render… · New
• {:else if writeFileBody?.content} treats an empty string as “no content”, which can mis-render… · New
• {:else if writeFileBody?.content} treats an empty string as “no content”, which can mis-render… · New
• {:else if writeFileBody?.content} treats an empty string as “no content”, which can mis-render… · New
• {:else if writeFileBody?.content} treats an empty string as “no content”, which can mis-render… · New
• The section.toolArgs fallback will typically be hit whenever edits.length is 0, which makes… · New
• The section.toolArgs fallback will typically be hit whenever edits.length is 0, which makes… · New
• The section.toolArgs fallback will typically be hit whenever edits.length is 0, which makes… · New
• The section.toolArgs fallback will typically be hit whenever edits.length is 0, which makes… · New
• The section.toolArgs fallback will typically be hit whenever edits.length is 0, which makes… · New
• commandStreaming can re-run very frequently during args streaming, and JSON.parse on a growing… · New
• commandStreaming can re-run very frequently during args streaming, and JSON.parse on a growing… · New
• Using :has() can cause cross-browser issues depending on the project’s supported browser matrix… · New

2 resolved since last review

• Keep actively executing tool calls expanded
• Test collapse transition when a call receives its result

_{Review effort: Lite}

[ServeurpersoCom]

Triaging the Copilot pass:

• :has() browser support: fine, it is Baseline since Dec 2023 and the UI already uses it (markdown-content.css, table cells).
• JSON.parse in commandStreaming: fine, it only runs while the model is typing the call, on a few hundred bytes.
• "No edits" unreachable in edit_file: don't care, showing the raw args of a malformed edit is more useful anyway.
• Empty write_file content: the only real one, cosmetic. If you feel like it, keep content undefined in the parser when the key is missing and test !== undefined in the template, so an intentionally empty file renders as an empty block with "0 bytes". Here or in a follow-up, your call.