What happened?

Running /review <pr> --comment (seen with /review 5096 --comment) on an OpenAI-compatible / DashScope (Aliyun Bailian) backend fails before it produces any review:

1. Step 4 of the bundled /review skill launches its 9 review agents in a single response (Agent 1 Correctness … Agent 7 Build & Test, with 6a/6b/6c personas).
2. The model passes a name argument to every spawn (agent1-correctness, agent2-security, …).
3. Every spawn fails immediately:

   Cannot spawn teammate "agent1-correctness": no active team.
   Use team_create to start a team first, or omit "name" to launch a one-shot subagent.
   

4. The model correctly notes it should drop name ("I need to use one-shot subagents instead"), but instead of recovering it hallucinates that the never-started agents are "still running" / "standing by" and repeats the same tool call across consecutive rounds.
5. The backend's repetitive-call guard then aborts the turn:

   [API Error: <400> InternalError.Algo.InvalidParameter: Repetitive tool calls detected
   in the conversation history. The same tool call with identical name and arguments has
   been repeated across multiple consecutive rounds. ...]
   

Net effect: /review is unusable in this configuration — it never produces a review and burns the whole turn into a hard API error.

What did you expect to happen?

/review should launch its review agents as ordinary one-shot subagents and complete the review. At minimum, passing name when no team is active should degrade gracefully to a one-shot subagent (treating name as a label, with a warning) instead of hard-failing every spawn and cascading into a provider-side abort.

Client information

$ qwen /about
# paste output here

Login information

OpenAI-compatible / DashScope (Aliyun Bailian) backend.

Anything else we need to know?

Root cause — an inconsistency introduced by #4844 (Agent Team)

• The Agent tool's name parameter is advertised in the schema unconditionally (packages/core/src/tools/agent/agent.ts, initialSchema, ~L543), described as "spawn as a named teammate … requires an active team context."
• The team guidance that explains how to use it (call team_create first) is gated behind isAgentTeamEnabled() (~L622). So the visibility of the param and the guidance for using it correctly are out of sync.
• Passing name with no active team is a hard error (~L1639-1654), not a fallback.
• The bundled /review skill (packages/core/src/skills/bundled/review/SKILL.md, Step 4) gives every review agent a name/number and tells the model to launch them all "in a single response". The model maps those names onto the name param. The skill has zero awareness of teams (no team_create step, no mention of name), and "all in one response" is structurally incompatible with "team_create must come first in a prior turn".

So the model is invited to fill name — by the always-present param, and (when agentTeam is enabled) actively pushed by the team guidance, which explicitly says "if the user asks for agents to review each other's work, or produce a consolidated result — create a team." That is exactly what /review is. But nothing makes the model call team_create first, and the failure is fatal instead of recoverable.

Reproduction conditions

• /review <pr> --comment (or any /review invocation that reaches the multi-agent Step 4).
• Most reliable with experimental.agentTeam: true / QWEN_CODE_ENABLE_AGENT_TEAM=1 (the team guidance steers the model toward named teammates). Can also occur with the flag off, because the name param is always in the schema and the skill names its agents.
• More likely on weaker models that do not self-recover from the per-call error.

Suggested directions (not prescriptive)

1. Gate the name param's schema visibility behind isAgentTeamEnabled(), matching the team-guidance gate — so the param disappears when teams are off.
2. When name is passed without an active team, fall back to a one-shot subagent + warning, instead of a hard error.
3. Make /review team-aware (insert a team_create step when teams are enabled) or explicitly instruct it to use one-shot agents without name.
4. Optional: a runtime circuit-breaker on N consecutive identical tool calls, so a degenerate loop fails fast locally instead of relying on the provider's guard.

Related

• Introduced by feat: add Agent Team experimental feature for parallel sub-agent coordination #4844 (Agent Team experimental feature).
• Background: Feature Request: Agent Team - Multi-Agent Collaboration #1815 (Agent Team feature request).
• 长程任务下，出现大量工具重复调用情况，导致会话被终止 #5019 (mass repetitive tool calls terminating long-running sessions) — the loop in step 4-5 here is a specific, reproducible instance of that class, but the trigger is the agent-team / /review collision rather than long context.
• Feature Request: Port Dynamic Workflows / Ultracode from Claude Code 2.1.160 #4721 (Dynamic Workflows) is a separate code path (vm-sandboxed agent() orchestration) and is not involved.