fix(desktop): render tool calls in the transcript instead of blank "Tool" rows

[agentforce314]

Tool activity reached the desktop transcript as contentless rows labelled "Tool", and raw content-block JSON leaked into assistant bubbles. The transcript looked nothing like the TUI's tool output.

Three mismatches against what the renderer actually reads (ui-desktop/src/lib/chat-messages.ts):

sent	read by renderer	effect
(no name)	payload?.name || "tool"	every row labelled "Tool"
output	payload.result	finished tools rendered empty
is_error	payload.error	failures looked like successes

Changes

• Tool names carry across frames. A tool_result names only its tool_use_id, so DesktopSession now keeps an id → name map populated from the assistant frame's tool_use blocks and consumed when the result arrives.
• Renderer field names. tool.complete sends result/error. The agent's trimmed display envelope (tool_use_result: structuredPatch, filePath, duration_s, inline_diff, summary, preview) is forwarded as well — that is what lets an Edit row show a diff instead of raw text, and what puts the elapsed time on a long-running row.
• No JSON in prose. Payloads where the UI expects text ran through coerceGatewayText, which JSON-stringifies an object it cannot read — that is how a bare {"type":"tool_use",…} block ended up printed inside an assistant bubble. New as_text() flattens content blocks and keeps only the text ones, so a tool block can never reach the transcript as prose.

Testing

• New tests/server/test_desktop_translate.py pins the payload contract: name carried across the tool_use → tool_result pair, result/error field names, display envelope forwarded, and as_text flattening a tool block to the empty string.
• Full Python suite: 9892 passed, 15 skipped.
• Verified live against a real agent turn — labelled Read/Glob/Write rows with their output and duration, a rendered approval bar, and no tool_use JSON anywhere in the transcript.

🤖 Generated with Claude Code