feat(web): ToolSearch verification UI (panel row + parsed-result card variant)

[helebest]

What

v0.13.0 slice 2 (W4, final slice of the approved extensibility roadmap): the playground's tool search verification UI, pairing with #43's runtime knob.

• Runtime panel row — tool search from /config.toolSearch; null renders as off (e.g. the graceful-degrade state on the MiniMax backend). Tooltip states the honest boundary: this is the configured value; whether a turn actually used tool search is observable only as a ToolSearch call in that turn's stream.
• ToolSearch card variant — search icon, the query as the summary line, and a parsed Matched tools · N list (name + description) from the CLI's <functions> result block (<function>{…json…}</function> per tool, format verified against the CLI binary).
• Never render-fail — parseToolSearchResult is best-effort: unparseable entries are skipped; when nothing parses the card falls back to the raw result text. The block format is a CLI-internal contract, so an unexpected shape must degrade, not break the stream.

Why

Completes the plan's verification story for tool search: the panel shows what is configured, the card shows what actually happened. Pure client-side; the SSE contract and runtime are untouched, and the new DTO field is optional so the playground degrades gracefully against older runtimes.

Verification

• TDD: 6 red-first vitest cases for parseToolSearchResult (well-formed block, surrounding prose, partial-parse, missing/invalid name, raw-text fallback, null/empty).
• biome ci clean on changed files; pnpm -r typecheck / web tests (57) / build all green.
• Browser walkthrough against a deterministic fake runtime implementing the v0.13.0 contract: panel shows tool search auto; a turn renders two ToolSearch cards — one parsing to Matched tools · 2 (mcp__demo__ping/mcp__demo__pong with descriptions), one falling back to raw text; v0.12.0 regressions checked (Task-card nesting, orphan-parent visibility, subagent-delta drop, MCP badges, error card); zero console errors across two passes including a fresh page load.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aafd17c2a4

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Render unset tool search as unset, not off

When RUNTIME_TOOL_SEARCH is unset but a bare ENABLE_TOOL_SEARCH exists in the runtime environment, the runtime intentionally leaves the child env untouched; the test at apps/runtime/test/runtime.test.ts:709-716 asserts that this passthrough still enables values like auto:3, and the README documents /config.toolSearch null as unset/default. Rendering that null as off makes the Runtime panel report tool search as disabled even though the next turn may still use it via ENABLE_TOOL_SEARCH or the CLI default, which undermines this verification UI; please show an unset/default state instead of off.

Useful? React with 👍 / 👎.

[helebest]

Independent pre-merge review (3 lenses: correctness / UI-consistency / scope-tests-docs, each material finding adversarially verified): 0 BLOCKER, 2 SHOULD-FIX confirmed — both fixed in d694d52.

• Fixed — parseToolSearchResult only understood a <functions> text block, but the verifier proved empirically (182 real ToolSearch tool_result payloads across CLI 2.1.183–2.1.199 in local transcripts, bracketing the pinned 2.1.197) that the real wire shape is a tool_reference content-block array — the <functions> encoding exists only in the API's server-side expansion for the model. The parsed card variant would have been silently inert against the real CLI. Now the stringified tool_reference array is the primary parsed shape (+3 red-first tests, 9 lib tests total), <functions> kept as a defensive fallback, and the fake runtime + CHANGELOG corrected to the real shape. Re-verified in the browser: Matched tools · 2 parses from the tool_reference array; the CLI's real no-match string falls back to raw text.
• Fixed — the panel row rendered unset (null) as off, but unset means "the CLI decides" (can be ON on official backends; a bare ENABLE_TOOL_SEARCH passthrough is also invisible to /config) and the row's own tooltip claimed to show the configured value. Now renders default (matching the skills row's null convention), tooltip spells out the semantics; browser-verified against toolSearch: null.

NITs recorded (not fixed): duplicate tool_name entries would collide as React keys (not observed in any real payload); .ts-desc single-line ellipsis truncates long descriptions with no expansion affordance.