openclaw - 💡(How to fix) Fix [Bug]: Preflight compaction required but failed: Already compacted (infinite loop)

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

After upgrading OpenClaw from 2026.5.20 to 2026.5.22, every agent session message fails with Error: Preflight compaction required but failed: Already compacted. The compaction fires on every message (triggered by as little as 61 tokens over an internal ~45K budget), but then immediately fails with reason=already_compacted_recently. This creates an infinite retry loop. Downgrading to 2026.5.20 resolves the issue immediately — confirmed as a regression in 5.22.

Steps to reproduce

1. Install or upgrade to OpenClaw 2026.5.22 (a374c3a) 2. Start gateway: openclaw gateway run --auth none --port 18789 3. Configure agent with provider deepseek/deepseek-chat 4. Open dashboard and send any message to the agent 5. Wait 40-100 seconds → message fails with compaction error 6. Check gateway logs for already_compacted_recently

Expected behavior

Messages should be processed normally. Compaction should only trigger when the session approaches the actual context window limit (1M tokens for deepseek-chat), not when it exceeds an internal budget of ~45K by 61-172 tokens.

Actual behavior

Every message triggers this cycle: 1. Preflight check detects overflow (61-172 tokens over a ~45K internal budget) 2. Compaction fires but immediately fails: already_compacted_recently 3. Tool result truncation does not help 4. Message fails after 40-100 seconds 5. Agent becomes completely unusable

OpenClaw version

2026.5.22 (a374c3a)

Operating system

WSL 2 (Ubuntu 26.04) on Windows 11

Install method

npm (npm install -g openclaw)

Model

deepseek/deepseek-chat

Provider / routing chain

DeepSeek via openai-completions API (https://api.deepseek.com/v1)

Additional provider/model setup details

Model configured with contextWindow: 1000000 (1M tokens) in models.json and openclaw.json. Also tried setting agents.defaults.contextTokens=200000 and compaction.budgetMargin=50000 — neither helped.

Logs, screenshots, and evidence

[agent/embedded] [context-overflow-precheck] sessionKey=agent:xiaoqi:main
       provider=deepseek/deepseek-chat route=compact_then_truncate
       estimatedPromptTokens=47439 promptBudgetBeforeReserve=45536
       overflowTokens=172 toolResultReducibleChars=244
       reserveTokens=20000 effectiveReserveTokens=20000

     [compaction-diag] end sessionKey=agent:xiaoqi:main
       trigger=overflow provider=deepseek/deepseek-chat attempt=1 maxAttempts=3
       outcome=failed reason=already_compacted_recently durationMs=2345

     [context-overflow-recovery] Tool result truncation did not help:
       no oversized or aggregate tool results
     Auto-compaction failed. Preserving existing session mapping.

     Impact and severity:
     High — Agent is completely unable to process any messages. The only workaround is
     downgrading to 2026.5.20. Impacts all users on 5.22+ with large system prompts (our
     system prompt is ~38K).

Impact and severity

High — Agent is completely unable to process any messages. The only workaround is downgrading to 2026.5.20. Impacts all users on 5.22+ with large system prompts (our system prompt is ~38K).

Additional information

No response