Hollow

v1.3.1 — Streaming Task Outputs

submit(wait=True) blocks the calling agent until a task finishes. For tasks taking 30–120 seconds, the agent is frozen. A real OS gives processes non-blocking I/O with completion notifications. This release makes task execution truly async.

Changes to agents/scheduler.py

• Task gains stream_enabled, partial_output, cancelled fields
• submit() accepts stream=True — returns immediately with stream_url and partial_url, task runs in background
• _run_task_streaming() — calls Ollama /api/chat with stream=True, accumulates chunks in task.partial_output, emits:
  • task.token_chunk every 10 tokens: {task_id, chunk, tokens_so_far, model}
  • task.partial_available every 500ms while running
  • task.completed / task.failed / task.cancelled as final event
• cancel(task_id) — aborts queued tasks immediately; sets cancelled flag on running streaming tasks

New API endpoints

• GET /tasks/{id}/stream — SSE endpoint, cursor-based streaming of partial_output
• GET /tasks/{id}/partial — instant snapshot of current partial output, non-blocking
• DELETE /tasks/{id} — cancel a queued or running task

Changes to api/agent_routes.py

• SubmitTaskRequest gains stream: bool = False and wait: bool = True
• Response includes stream_url and partial_url when stream=True

New MCP tools (68 → 69)

task_stream — reads GET /tasks/{id}/partial for current partial output

Integration tests

10 tests, all passing. 5 require Ollama (run automatically when available), 5 run without.

Backward compatibility

All existing wait=True blocking behavior unchanged. wait=False (non-blocking without streaming) also works as before.