Clear stamped code-block height when Hide Info collapses a tool call - #5924

[Claude Code 🤖]

Consuming isDisplayingCode also makes the expand path stamp a height for the first time, and this math over-measures — so as written the PR trades an empty block after "Hide Info" for one after "View Info".

getContentHeight() is Monaco's full content height, but .code-block-editor is capped at --code-block-max-height (250px) in ai-assistant/code-block/index.gts, and the outer .code-block section has no max-height of its own, only overflow: hidden. The stamp is therefore header + contentHeight while the box renders header + min(contentHeight, 250). Monaco's line height here is 1.5 × 12px on macOS (1.35 × elsewhere) plus 16px of padding, so past ~13 lines of tool-call JSON — routine for a patchCardInstance patch or a search filter — expanding leaves a contentHeight − 250 gap. The // max-height is constrained by CSS note on the stamp is what makes it read as safe; that cap is on the child, not on element.

It only bites now because expand previously never re-ran the modifier. And the editor really is measurable on that re-run: Glimmer runs scheduled modifier installs before scheduled updates within one commit, so the freshly created Monaco editor already reports a content height when this update lands.

The stamp is also redundant — the Monaco modifier sizes the editor div and CSS clamps it, so header + renderedEditorHeight is the section's natural height. Dropping it removes both the over-stamp and the need to clear anything:

private scrollBottomIntoView = modifier((element: HTMLElement) => {
  // Consume the toggle flag so this re-runs when the code area opens.
  if (!this.isDisplayingCode) {
    return;
  }
  this.scrollIntoView(element.parentElement as HTMLElement);
});

To keep the stamp instead, measure the rendered box: editor?.getContainerDomNode().offsetHeight in place of getContentHeight().

Either shape carries one decision: expand now reaches scrollIntoView, which it never did on a normal mount, so clicking "View Info" on a message up in the scrollback will scroll the panel. Likely the modifier's original intent, but it is new user-visible behavior — worth confirming you want it.

Regression introduced here, on top of a calculation that was previously only reachable through the remount path. Blocking.

[Claude Code 🤖]

Agreed — real regression, and it's now fixed by dropping the outer height stamp entirely (1410aca). The Monaco modifier already sizes .code-block-editor and CSS caps it at 250px, so the section sizes to its content: no over-measure on expand, and nothing to clear on collapse.

Context that argues for removing it rather than patching the math — the stamp predates the self-sizing editor. It was added in Dec 2024 when the code rendered into a plain monaco-container with no height of its own, so back then the stamp was the only thing giving the code area a size. The self-sizing editor (onDidContentSizeChange) arrived with the CodeBlock refactor in #2859 (Jul 2025), which made the stamp redundant; it just survived the refactor still reading Monaco's uncapped content height. So this removes a leftover, not a safeguard.

On scroll-on-expand: confirmed intentional. "View Info" now brings the code into view, which matches the modifier's original purpose — keeping it.

[Claude Code 🤖]

This pins the collapsed state but not the expanded one, which is where the height math is now newly exercised. The payload here renders 13 lines, just under the 250px editor cap, so it cannot distinguish a correct stamp from an over-stamp.

Add a case whose tool-call payload exceeds --code-block-max-height and assert, while expanded, that the section's height is no greater than the header plus the editor's rendered height. Against the current diff that fails — see the scrollBottomIntoView thread.

Smaller, on these lines: find(...) as HTMLElement is dereferenced immediately after a soft assert.ok, so a null match both fails an assertion and throws; and .tool-code-block carries no CSS rule anywhere in the repo, making it an invisible test-only hook nothing protects. An assert.dom(...).exists() before the read covers the first, and a data-test- attribute alongside the class would make the hook explicit.

Non-blocking; the coverage half is what would have caught the expand-path issue.

[Claude Code 🤖]

Done:

• Added expanding a tall tool call does not leave an empty gap below the code — a 30-line (over-cap) payload asserting that, while expanded, the section height doesn't exceed the sum of its rendered children (and carries no inline height). That fails against the old stamp and locks in the content-sizing behavior.
• The collapse assertion now does assert.dom(...).exists() before dereferencing, so a null match fails cleanly instead of throwing.
• Replaced the class-only .tool-code-block hook with an explicit data-test-tool-code-block attribute on the CodeBlock.