Summary
When OpenClaw's history compaction runs on long conversations with Qwen models (Qwen-Long, Qwen-3-Coder-Plus), the outbound messages[] sometimes ends up with no role: user or role: assistant entries — only role: system (or only role: tool) survives.
Qwen upstream rejects with HTTP 400:
Role must be user or assistant and Content length must be greater than 0Role specification invalid
The user-facing symptom (in deployments that route through a proxy that wraps 4xx) is a generic stream-failed message that hides the real error. Without proxy wrapping the operator sees the 400 directly, but the request is always invalid.
Environment
- OpenClaw: 2026.4.2 (planning to upgrade to 2026.5.22)
- Models affected:
Qwen-Long (10M context), Qwen-3-Coder-Plus (1M context)
- Provider type: openai-completions compatible
- Trigger threshold: ~20K tokens of history before LCM compaction fires
Repro pattern
- Start a session with
Qwen-Long or Qwen-3-Coder-Plus
- Run 15-20+ turns with mixed tool use + direct chat to push past the
softThreshold
- Wait for LCM compaction
- The next user turn → outbound request has
messages: [ { role: 'system', ... } ] (and occasionally a trailing role: tool block), with every prior user/assistant turn gone
- Qwen 400 → request fails
Why this looks distinct from already-tracked compaction issues
| Existing issue |
What it covers |
Why this is different |
#32759 (closed) |
"no conversation messages at all" guard, returns 400 vs graceful skip |
This produces a message array that has some entries, but none of them user/assistant. The guard at #32759 doesn't fire. |
#69756 (closed) |
"chat.history omitted: message too large" replacement |
That replacement preserves placeholder text; our case strips entries entirely. |
#74073 (open) |
auto-compaction race for readSubagentOutput |
Race-condition focus, single-call; our case is the steady-state result of compaction summarization. |
Suggested fix direction
After compaction transforms messages[], the serializer should validate that at least one entry has role: 'user' or role: 'assistant' before transmitting. If not, either:
- (a) Re-run compaction with a less-aggressive ratio, preserving the most recent user/assistant pair verbatim, or
- (b) Fall back to a user-visible "history exhausted, please start a new session" error rather than letting upstream return an opaque 400
Option (a) is closer to graceful UX, option (b) is a clearer signal to the operator.
Offered
Happy to provide sanitized session jsonls + outbound request bodies if useful for a regression test. We hit this enough times in production that we can produce captures on demand.
Summary
When OpenClaw's history compaction runs on long conversations with Qwen models (
Qwen-Long,Qwen-3-Coder-Plus), the outboundmessages[]sometimes ends up with norole: userorrole: assistantentries — onlyrole: system(or onlyrole: tool) survives.Qwen upstream rejects with HTTP 400:
Role must be user or assistant and Content length must be greater than 0Role specification invalidThe user-facing symptom (in deployments that route through a proxy that wraps 4xx) is a generic stream-failed message that hides the real error. Without proxy wrapping the operator sees the 400 directly, but the request is always invalid.
Environment
Qwen-Long(10M context),Qwen-3-Coder-Plus(1M context)Repro pattern
Qwen-LongorQwen-3-Coder-PlussoftThresholdmessages: [ { role: 'system', ... } ](and occasionally a trailingrole: toolblock), with every prior user/assistant turn goneWhy this looks distinct from already-tracked compaction issues
#32759(closed)#32759doesn't fire.#69756(closed)#74073(open)readSubagentOutputSuggested fix direction
After compaction transforms
messages[], the serializer should validate that at least one entry hasrole: 'user'orrole: 'assistant'before transmitting. If not, either:Option (a) is closer to graceful UX, option (b) is a clearer signal to the operator.
Offered
Happy to provide sanitized session jsonls + outbound request bodies if useful for a regression test. We hit this enough times in production that we can produce captures on demand.