feat(web): add purpose-built renderers for file, diff, test, and terminal evidence

[sowen1023]

Parent

• [Roadmap] OpenPI Web：像 Codex / Kimi Code 一样的本地浏览器 Agent 工作台 #76

Problem

The Web transcript currently renders thinking, tool calls, and tool results through generic expandable text blocks. This preserves basic evidence but makes common coding-agent output difficult to inspect:

• file reads lose line-oriented navigation;
• write/edit results do not provide a reviewed change view;
• diffs are not rendered as additions/deletions with file identity;
• test results are not grouped by command, status, or failure;
• terminal output lacks a dedicated bounded log view.

A generic raw-text fallback is necessary, but it should not be the primary UI for structured evidence.

Value

• Make actual code changes and validation evidence reviewable.
• Reduce the risk of treating a tool label as proof of success.
• Preserve exact raw evidence while providing useful bounded projections.

Approach constraints

• Use Pi/OpenPI tool events and canonical results as the source of truth.
• Keep user-visible presentation, model-visible context, and persisted runtime evidence separate.
• Every renderer must have explicit byte/line/item bounds and a raw fallback.
• Never infer success, file identity, or changed lines from colors or display labels.
• Escape untrusted content and avoid executing rendered HTML, ANSI, links, or terminal control sequences.
• Unknown and future tools must continue to render safely through the generic path.

Acceptance criteria

• File-read evidence shows canonical path, line context, truncation, and recovery information.
• Write/edit evidence shows affected files and a bounded, reviewable diff when exact evidence exists.
• Test evidence distinguishes command, running/completed state, pass/fail/cancel, and failure excerpts.
• Terminal evidence supports bounded incremental output and exact terminal status.
• Each specialized view can reveal the sanitized raw arguments/result.
• Oversized and malformed results remain bounded and cannot break transcript rendering.
• Unknown tool types retain the current generic fallback.
• Snapshot, SSE replay, and reconnect produce stable renderer input.
• Protocol projection, UI rendering, accessibility, and browser tests cover each renderer.
• bun run check and bun run test pass.

Out of scope

• A browser code editor or filesystem explorer.
• Reimplementing tool execution in the UI.
• Treating presentation state as execution evidence.

[SSK988I]

认领