[Feature Request] SingleAgentWorker lacks unified streaming handling, task-scoped execution context, and structured attempt tracking

[fengju0213]

Required prerequisites

• I have searched the Issue Tracker and Discussions that this hasn't already been reported. (+1 or comment there if it has.)
• Consider asking first in a Discussion.

Motivation

Problem

SingleAgentWorker._process_task has three issues:

• Two separate streaming consumption paths — structured-handler and native-structured-output each inline their own async for chunk in response loop. They differ in content accumulation (delta vs cumulative not handled), TaskResult extraction, and stream_callback forwarding. Bug fixes must be applied twice.
• No task-scoped execution context — worker agents have no way to know which task/attempt they're executing. Downstream (eigent) needs task_id, attempt number, and agent IDs on every lifecycle event for debugging and telemetry, but must patch this externally.
• Flat, minimal attempt records — worker_attempts entries contain only response_content[:50], tool_calls[:50], and total_tokens. Full token usage (prompt_tokens, etc.), tool call count, and execution context are lost.
• Agent pool context leakage — when an agent is returned to the pool, _execution_context from the previous task is not cleared, potentially leaking to the next borrower.

Solution

No response

Alternatives

No response

Additional context

No response

[yugborana]

@fengju0213 May I look into this issue if you are not taking it currently?

[bkdtjw]

Hi @fengju0213, thanks for filing this well-structured issue.
I’m interested in contributing here. Since this covers four related but separable problems, would it make sense to tackle them incrementally? I could start with:

1. Agent pool context leakage (problem 4) — adding _execution_context cleanup on agent return to pool
2. Structured attempt tracking (problem 3) — extending worker_attempts to capture full token usage, tool call count, and execution context
   These two seem self-contained enough for individual PRs without blocking the streaming unification work.
   I’d appreciate any guidance on the expected approach before I start coding. Happy to discuss further.

[bkdtjw]

Just want to confirm — is it okay for me to start with problem 4 (agent pool context leakage)or 4 3 as a first PR?

[fengju0213]

@yugborana @bkdtjw thank you for your interest in this issue. I’m already working on it, sorry about that. If you’re still interested, you could help by reviewing it.