fix(stream): claude tool calls never emit a closing event #182

What happened

Streaming a Claude Code run writes a tool_start record for every tool call and never writes the matching tool_end. execution.ndjson — and the cockpit and --progress views built from it — therefore show each Bash/Edit call as still running, and nothing in the log distinguishes a call that finished from one that was killed.

Two independent causes:

1. src/core/agent-stream.js:227 (mapClaudePayload) treats every type: "user" line as output echo and returns []. In Claude Code's stream-json, that is exactly where tool_result blocks arrive, so the closing event is thrown away.
2. Neither adapter records the CLI's own tool-call id — Codex sends item.id, Claude sends tool_use.id and tool_use_id — so even with the first cause fixed, nothing can pair a result to its call, or tell two concurrent runs apart when the CLI reuses an id.

A run killed mid-call (the default 300s agent timeout, Ctrl-C, a crashed CLI) never emits a result at all, so its calls stay open forever with no terminal status.

What you expected

Every tool_start eventually gets exactly one tool_end for the same call, carrying the CLI's own id, and a call the CLI never finished is closed with a status that says so when the sink shuts down.

Reproduction

Smallest reproduction, no real CLI needed (evalid 0.5.33, Node 24):

import { normalizeAgentEvents } from './src/core/agent-stream.js';

const lines = [
  '{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_1","name":"Bash","input":{"command":"ls"}}]}}',
  '{"type":"user","message":{"content":[{"type":"tool_result","tool_use_id":"toolu_1","content":"ok"}]}}',
];

for (const line of lines) {
  console.log(JSON.stringify(normalizeAgentEvents(line, { adapter: 'claude', seq: 0 })));
}

Output today:

[{"seq":0,"event":"tool_start","event_type":"tool_use","tool":"Bash","args_preview":"{\"command\":\"ls\"}","args_digest":"4cf29611a6693486"}]
[]

The tool_start carries no id and the result maps to nothing.

xflow doctor --project-root .

reports a healthy project — this is a streaming-shape gap, not a configuration problem.

Environment

• evalid version: 0.5.33 (405ffa5)
• Node: 24.21.0
• Python: 3.15.2
• OS: Windows 11
• Agent CLI involved: Claude Code (stream-json); Codex has the same missing-id gap but does emit its closing events

I have a fix with tests ready and will open the PR next — additive record fields only, no consumer changes.