stream-json reports every tool as "name": "unknown" with empty input

[konard]

Summary

With --output-format stream-json, every tool_use event is emitted as "name": "unknown" with an empty "input": {} — regardless of which tool ran. The tool itself executes correctly with the right arguments; only the reported event is wrong.

This makes the stream-json output unusable for its main purpose: observing which tools an agent invoked. It affects any programmatic consumer (CI assertions, audit logs, replay harnesses).

Observed on @link-assistant/agent 0.25.0.

Reproduction

No model, network or API key needed — a ~60-line mock OpenAI-compatible server that streams exactly one tool call is enough.

mock-openai-server.mjs:

import { createServer } from 'node:http';

const port = Number(process.argv[2] || 8931);
const MODEL = 'formal-ai';
const sse = (payload) => `data: ${JSON.stringify(payload)}\n\n`;
const USAGE = { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 };

function toolCallStream() {
  const base = { id: 'chatcmpl-probe', object: 'chat.completion.chunk', created: 0, model: MODEL };
  return sse({
    ...base,
    choices: [{
      index: 0,
      delta: {
        role: 'assistant',
        tool_calls: [{
          index: 0,
          id: 'call_probe_0',
          type: 'function',
          function: { name: 'write', arguments: JSON.stringify({ filePath: 'hi.txt', content: 'hi\n' }) },
        }],
      },
      finish_reason: null,
    }],
  })
  + sse({ ...base, choices: [{ index: 0, delta: {}, finish_reason: 'tool_calls' }], usage: USAGE })
  + 'data: [DONE]\n\n';
}

let completions = 0;
createServer((req, res) => {
  req.on('data', () => {});
  req.on('end', () => {
    const url = req.url || '';
    if (url.endsWith('/models')) {
      res.writeHead(200, { 'content-type': 'application/json' });
      res.end(JSON.stringify({ object: 'list', data: [{ id: MODEL, object: 'model', owned_by: 'mock' }] }));
      return;
    }
    if (url.endsWith('/chat/completions')) {
      completions += 1;
      res.writeHead(200, { 'content-type': 'text/event-stream', 'cache-control': 'no-cache' });
      if (completions > 1) {
        const base = { id: 'chatcmpl-done', object: 'chat.completion.chunk', created: 0, model: MODEL };
        res.end(
          sse({ ...base, choices: [{ index: 0, delta: { role: 'assistant', content: 'done' }, finish_reason: null }] })
          + sse({ ...base, choices: [{ index: 0, delta: {}, finish_reason: 'stop' }], usage: USAGE })
          + 'data: [DONE]\n\n',
        );
      } else {
        res.end(toolCallStream());
      }
      return;
    }
    res.writeHead(404, { 'content-type': 'application/json' });
    res.end('{}');
  });
}).listen(port, '127.0.0.1', () => console.error(`[mock] listening on 127.0.0.1:${port}`));

Run it:

node mock-openai-server.mjs 8935 &

WORK=$(mktemp -d); git -C "$WORK" init -q
cfg='{"provider":{"formalai":{"name":"Mock","npm":"@ai-sdk/openai-compatible","options":{"baseURL":"http://127.0.0.1:8935/api/openai/v1","apiKey":"local"},"models":{"formal-ai":{"name":"Mock"}}}},"model":"formalai/formal-ai"}'

cd "$WORK" && LINK_ASSISTANT_AGENT_CONFIG_CONTENT="$cfg" agent \
  --model formalai/formal-ai --permission-mode auto \
  --output-format stream-json --compact-json --disable-stdin \
  --prompt 'Write hi to a file called hi.txt'

Expected

The tool_use event names the tool the server asked for:

{"type":"tool_use","name":"write","input":{"filePath":"hi.txt","content":"hi\n"}, ...}

Actual

{"type":"tool_use","timestamp":"...","session_id":"ses_...","name":"unknown","input":{},"tool_use_id":"prt_..."}

…and the file is written correctly (cat $WORK/hi.txt → hi), which is what localises this: the call is dispatched with the correct name and arguments, and only the stream-json event loses them.

Scope

• Reproduces both with and without --compact-json — it is the stream-json event, not the compaction step.
• Not provider-specific: the wire format above is plain OpenAI streaming, with name and arguments in delta.tool_calls[].function.
• Cross-checked against another harness on the same server and the same wire: opencode (1.17.20 and 1.18.3) reports the same calls as write and bash, so the tool names are present in the stream being consumed.

Context

Found while building a two-harness E2E in link-assistant/formal-ai#715, where the CLI drove a real 7-round session that wrote its report correctly — every tool_use event in the captured stream reported unknown. Captured artifacts: docs/case-studies/issue-715/agent-cli-learning/.

Incidentally: omitting usage from the final chunk makes the CLI retry the round as a provider API error with exponential backoff (provider returned invalid usage data; retrying as provider API error), which is presumably #249 — the USAGE constant above is there to avoid it.