fix: message.part.delta streaming, .info shape for tokens/finish, parts fallback for empty content

[KochC]

Summary

Found and fixed three compounding bugs in runAgentTurn() while testing real multi-turn tool-calling conversations (continuing after a prior tool call/result in history) against a live OpenCode server. None of these were caught by the existing unit tests, since their mocks happened to match the same wrong assumptions as the code they were testing.

The bugs

1. Real incremental streaming text arrives via message.part.delta - a completely separate event type with flat properties (sessionID, partID, field, delta) - not via event.properties.delta on message.part.updated, which is all the code listened to. message.part.updated only ever carries status/snapshot updates (confirmed still correct and unchanged for tool-call detection).
2. client.session.messages() returns items shaped like { info: Message, parts: Part[] } - matching what client.session.prompt() (the non-tool-calling path) already returns, and what extractAssistantText() already expects - not a flat { role, tokens, finish, ... } shape. Reading assistantMsg.role/.tokens/.finish directly meant the role filter never matched anything, so token usage silently always fell back to all-zero, and finish was always undefined, for every single tool-calling request - not just the scenario below.
3. For some turns, OpenCode never emits message.part.delta at all for the final reply - only message.part.updated snapshots on the same part id. Observed specifically when continuing a conversation with prior tool calls/results in history - i.e. real multi-step agentic tool use, not simple one-shot prompts. Content stayed permanently empty in exactly this case: not a hypothetical edge case, but the core "read a tool result and continue" pattern any real tool-using agent depends on.

Repro

curl http://127.0.0.1:4010/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "github-copilot/claude-sonnet-5",
    "messages": [
      {"role": "user", "content": "Search the web for one recent headline about agentic AI, then give me just the headline."},
      {"role": "assistant", "content": null, "tool_calls": [{"id":"call_1","type":"function","function":{"name":"search","arguments":"{\"query\":\"agentic AI news\"}"}}]},
      {"role": "tool", "tool_call_id": "call_1", "content": "Title: \"Agentic AI startups raise record funding in 2026\" - TechCrunch."}
    ],
    "tools": [
      {"type":"function","function":{"name":"search","description":"Search the web","parameters":{"type":"object","properties":{"query":{"type":"string"}},"required":["query"]}}},
      {"type":"function","function":{"name":"format_final_json_response","description":"Final answer","parameters":{"type":"object","properties":{"output":{"type":"string"}}}}}
    ]
  }'

Before: {"finish_reason":"stop","message":{"role":"assistant","content":""}}, "usage":{"prompt_tokens":0,"completion_tokens":0,"total_tokens":0} (always zero)
After: {"finish_reason":"stop","message":{"role":"assistant","content":"\"Agentic AI startups raise record funding in 2026\""}}, "usage":{"prompt_tokens":2,"completion_tokens":21,"total_tokens":23}

The fix

• runAgentTurn() now handles message.part.delta for real-time onChunk/content accumulation (fixes streaming + the common case)
• Reads tokens/finish from assistantEntry.info instead of the item directly (fixes usage always being zero and finish always being undefined on natural, non-tool-call completions)
• Falls back to extractAssistantText(assistantEntry.parts) for the final content when the live stream produced nothing, reusing the same helper the non-tool-calling path already relies on (fixes empty content on multi-step turns)

Test changes

• Fixed createToolCallClient/createStreamingClient's session.messages mock to the real { info, parts } shape (previously flat, matching the bug rather than the real API - all 9 dependent tests still pass since none asserted specific token values, but they're now actually exercising correct behavior)
• Converted all message.part.updated+delta text-streaming events in existing tests to message.part.delta, matching the corrected event handling
• Two new tests: accumulating content from message.part.delta events, and falling back to the message's parts (plus reading real tokens/finish from .info) when message.part.delta never fires

Testing

• npm test - 143 passed (141 + 2 new)
• npm run lint - clean
• Manually verified against a live OpenCode server end-to-end (see repro above) - confirmed both the content and token-usage fixes work correctly together

Related

• Builds on fix: disable stale bridge tool IDs from other pool slots #56/fix: address Copilot review follow-ups from #56; docs: add n8n integration guide #58 (tool-calling slot isolation + review follow-ups)