[N1x Linux][Agent&Skills] manual /compact fails with guard_blocked after compaction-liveness fix

[JoyceChenNV]

Note

Re-verification follow-up to #11805 — #11805 was closed by #12018, whose merge commit is confirmed in v0.0.128. Re-verification on v0.0.128 shows that the original 120-second timeout and permanent-busy state are mitigated, but manual compaction still cannot complete because the generated summary is rejected by the quality guard.

Description

On N1x Linux with NemoClaw v0.0.128 and the supported managed vLLM configuration, a conversation that reaches the compaction path cannot be continued. The third prompt fails with a context-recovery error, and manually running /compact repeatedly ends with guard_blocked because the generated summary is missing required sections.

This is not a recurrence of the original permanent agent is busy state from #11805: the TUI continues accepting input and does not enter livenessState=blocked. It is a residual compaction-completion failure after #12018.

Platform scope: Reproduced on N1x Linux; other platforms were not tested.

Regression: Unknown — earlier versions with the same quality-guard configuration were not tested.

Environment

Platform:      N1x Linux Laptop Express
OS:            Ubuntu 24.04, aarch64
NemoClaw:      v0.0.128
OpenShell CLI: 0.0.116
Model:         Qwen3.6-35B-A3B-NVFP4
Context:       32768 tokens
vLLM:          managed, max-num-seqs 2

Steps to Reproduce

1. Launch the OpenClaw TUI on an N1x Linux sandbox using the supported managed vLLM configuration.
2. Send write 500 words essay about joke and wait for completion.
3. Send create python quick sort and execute it and wait for completion.
4. Send you are a frontend engineer, create a HTML page to demonstrate advantages and disadvantages of different agents.
5. After context recovery fails, run /compact manually.

Expected Result

Manual /compact should complete within the configured 300-second limit, reduce the active context, and allow the existing conversation to continue.

Actual Result

• Prompts 1 and 2 complete successfully.
• Prompt 3 does not complete and reports: Context is too large and auto-compaction could not recover this turn.
• Manual /compact fails repeatedly with outcome guard_blocked.
• The compaction attempts run for approximately 46–92 seconds before the generated summary is rejected for missing required sections.
• The old 120-second timeout, livenessState=blocked, and permanent agent is busy state are not observed.
• /new restores usability, but the existing conversation cannot be compacted and continued.

Logs

Compaction outcome: guard_blocked
Reason: generated summary was rejected for missing required sections
Observed duration across attempts: approximately 46–92 seconds

Related Issues

• [N1x Linux][Agent&Skills] OpenClaw TUI stuck: N1x inference too slow for 120s compaction timeout, context overflow unrecoverable #11805 — original 120-second timeout and permanent-busy failure, closed by fix(openclaw): preserve N1x compaction liveness #12018.
• [N1X Linux][Inference] Fresh OpenClaw TUI HTML task fails with incomplete or malformed tool call #12293 — separate fresh-session malformed-tool-call failure; not the cause of this compaction failure.