“/research” gets stuck

[wangtsiao]

When conducting in-depth research, the "/research" command may get stuck and waste tokens after spawning the agent, which is the supervisor stage.

[wangtsiao]

Investigation result: /research stuck in supervisor stage

Root cause

The wait_agent tool returns individual streaming token deltas (assistant_delta events) instead of accumulated final output from child worker agents.

In crates/server/src/runtime/agents.rs:885-897, each ItemDeltaKind::AgentMessageDelta from a child worker is pushed as a separate AgentOutputEvent with kind: "assistant_delta". When the supervisor calls wait_agent, it receives a vector of these token-level fragments rather than a coherent worker result.

How it manifests in the research pipeline

Research pipeline: Clarify → Brief → Supervisor (stuck here) → Compress → FinalReport

1. The supervisor spawns research workers via spawn_agent (workers have web_search, webfetch, read, write, apply_patch tools)
2. Workers run searches and fetch pages, producing streaming output
3. Each streaming chunk gets pushed as an individual AgentOutputEvent to the parent's output buffer
4. The supervisor calls wait_agent to collect results
5. It receives a massive list of individual delta events instead of usable worker output
6. The model either can't make sense of the garbled input, keeps re-polling via wait_agent (wasting tokens), or fills its context window with raw delta events

Related code

• crates/server/src/runtime/agents.rs:885-897 — each streaming delta pushed as individual assistant_delta event
• crates/server/src/runtime/agents/coordinator.rs:16-41 — wait_agent_inner returns raw events from buffer
• crates/server/src/subagent.rs:306-334 — SubagentOutputBuffer::wait_after returns all events after a sequence
• crates/server/src/runtime/research.rs:745-788 — run_supervisor_stage consumes wait_agent results
• crates/server/src/runtime/research_stages.rs:6-12 — supervisor only has agent coordination tools
• crates/core/prompts/research/supervisor.md:8 — supervisor prompt says "Always call wait_agent for every spawned worker"

Fix needed

This is directly linked to Issue #138 (wait_agent output as streaming tokens). The fix there will resolve this issue too. Additionally:

1. Accumulate child output: Instead of forwarding raw assistant_delta events, the output buffer should accumulate deltas into coherent text blocks before returning them to the caller
2. Status consolidation: A "worker completed" status event should be pushed after the worker finishes, so the supervisor can tell when a worker is done
3. Iteration guard: Add a max spawn_agent/wait_agent iteration limit in SupervisorQueryCapture to prevent runaway token consumption if the model keeps looping

References

• Issue Agent Runtime: Wait agent output returned as raw streaming tokens — makes agent think sub-agent is unresponsive #138 — wait_agent output as raw streaming tokens (same root cause)
• Rollout file: rollout-2026-06-30T08-43-32Z-019f17b2-a79e-7502-837c-974b4cd08a1d.jsonl lines 17-18

[wangtsiao]

PR's refactoring solved the concurrent deadlock BUG.