fix(components): keep serialized tool-input JSON out of Markdown math

[tommy0103]

Related issue

Closes #710

Problem / pressure

Expanding a tool call showed its raw input JSON mangled. Tool calls report their raw input as a text content block holding serialized JSON (e.g. {"command":"…","timeout":60}); routing that text through the Markdown pipeline lets single-$ inline math treat shell fragments like $(git -C …) as TeX and drop characters like the & in 2>&1. The expanded view displayed a command that was not the command that ran — the stored history and the executed command were both intact, so the damage was render-time only.

Summary

• Add detectToolCallJsonText (packages/components/src/lib/tool-call-json-text.ts): returns the original text, trimmed but otherwise untouched, when a text block parses as a JSON object or array; null otherwise (prose, primitives, broken JSON).
• JSON.parse only validates — the payload is never re-serialized, so integer lexemes beyond 2^53 (64-bit IDs) and forms like 1e10 keep their exact digits. (Addresses the P1 review finding: a parse/stringify round trip would corrupt large numeric IDs and recreate the same displayed-differs-from-actual defect.)
• StandardToolContentBlock's text case renders a detected payload in the same monospace <pre> treatment already used for raw tool output; everything else continues to MarkdownBlock unchanged.
• The classifier lives in a leaf lib module so the regression test can assert it without pulling the whole view.tsx import graph into the test module graph (package rule).
• Regression test covers the real-world payload shape (a git commit --amend command with $(...), 2>&1, nested quotes), integer lexemes beyond 2^53, exponent forms, plus prose/primitive/malformed-JSON negatives.
• Agent Note (EN+ZH): .agents/notes/implemented/bug-fix/2026-09-15-tool-call-json-markdown-math.md.

Visual explanation

Simple change: one pure classifier plus one new branch in a render switch; the diff is the structure, and no multi-step flow changes.

Before / after

Before	After
Raw tool-input JSON rendered as Markdown: $(...) spans became italic TeX (user.name="(git−C/…)"), 2>&1 lost its & — displayed command ≠ executed command	Detected JSON payloads render byte-faithfully in a monospace block: shell fragments, big integer IDs, and exponent forms all preserved exactly; prose still gets full Markdown including $...$ math

Test plan

• packages/components: tests/tool-call-json-text.test.ts 8/8 pass; related suites (markdown-math-delimiters, markdown-autolink, assistant-turn-render-blocks, markdown-streamdown-mode, assistant-message-render-items) pass.
• packages/components pnpm typecheck: clean.
• oxlint on touched files: 0 errors (pre-existing warnings only); Prettier: clean.
• Root lint:i18n, check:code-collab-imports, check:platform-boundaries, check:public-boundary, docs check: all exit 0.
• Skipped: full root pnpm check (monorepo test:ci). In a fresh worktree, markdown-idle-rerender and markdown-streaming-reparse fail identically on the unmodified base (jotai storage environment), so they are unrelated to this change; verified by A/B against a stash of this branch.
• Not verified: live UI click-through; the mangling was observed on recorded history predating this branch, so rendering is verified by code inspection plus the classifier contract test.

Context handoff

Instructions for reviewing agents

• Review focus: packages/components/src/lib/tool-call-json-text.ts (classifier contract, validation-only parse) and the text case of StandardToolContentBlock in packages/components/src/components/ai-gui/view.tsx (new render branch).
• Decisions to challenge: any JSON object/array text is rendered as code rather than Markdown (a tool returning legitimate JSON-shaped prose loses Markdown rendering); single-line payloads stay single-line (no pretty-print) to keep numeric lexemes exact; the resource text case was deliberately left on the Markdown path and can still hit the artifact.
• Plausible failures / evidence gaps: no live UI verification on this branch; the two failing suites listed in the test plan fail on the unmodified base too — re-check if CI disagrees.

Authoring context

• User goal / directives: investigate a mangled tool-call JSON display, confirm upstream had no fix, then implement the fix on a clean worktree branch and open a PR; a P1 review finding on numeric literals was addressed by switching from pretty-print to verbatim display.
• Constraints / non-goals: minimal scope — producer (CLI/ACP) data format untouched; resource text blocks untouched; no new user-visible copy (no i18n changes); no new AGENTS.md rules (owning files are near their size gate; rationale lives in the Agent Note).
• Risk-bearing decisions: render-path classification changes what some tool text looks like (JSON-shaped prose now renders as code); renderer-side fix chosen over producer-side because existing recorded history already carries these blocks.
• Destructive or irreversible behavior: none; display-only change, no storage or wire format changes.
• Deliberately not done or tested: full monorepo test:ci (see test plan); live UI click-through; resource text case; lexeme-preserving pretty-printer (rejected as disproportionate tokenizer complexity for a display path).
• Unknowns / confidence: high confidence in the classifier contract (unit-tested against the real payload and large-number forms); moderate confidence there is no other consumer relying on Markdown rendering of JSON text blocks.

Original user prompt

Show original prompt

我的意思就是这个，这个显示肯定不是你点开 toolcall 时所期望的
你看一下上游有没有人修这个问题

你拉一个干净的 worktree 分支来做这个修复

开一个 PR 吧

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f669025f6e

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".