openclaw - 💡(How to fix) Fix Codex GPT-5.5 context budget can be overestimated as 1M, delaying compaction until context overflow

Summary

On OpenClaw 2026.5.12, long-running Feishu sessions using GPT-5.5 through the Codex/OAuth path were cached and reported with a 1,000,000 token context budget. OpenAI's GPT-5.5 launch docs distinguish the surfaces: GPT-5.5 in Codex has a 400K context window, while the API has a 1M context window.

That mismatch made the active context engine (lossless-claw) compact too late. Eventually the user-facing symptom was the generic channel reply:

⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

After lowering only the GPT-5.5 Codex-backed agents/sessions to contextTokens: 350000, the same Zero/Feishu route returned to normal and status reported the expected smaller budget.

Environment

• OpenClaw: 2026.5.12 (f066dd2)
• OS: macOS 26.5 arm64
• Gateway: local LaunchAgent, 127.0.0.1:18789
• Channel: Feishu
• Agent: main / Zero
• Model config: openai/gpt-5.5 with Codex runtime/OAuth route
• Context engine: lossless-claw
• lossless-claw config at time of workaround:
  • contextThreshold: 0.7
  • proactiveThresholdCompactionMode: inline
  • freshTailMaxTokens: 100000

Observed behavior

Before workaround, active GPT-5.5 Codex sessions were shown/cached as 1000k context:

agent:main:feishu:default:...  gpt-5.5  OpenAI Codex  117k/1000k (12%)
agent:main:feishu:zero:...     gpt-5.5  OpenAI Codex  135k/1000k (14%)
agent:balao-wealth:main        gpt-5.5  OpenAI Codex  239k/1000k (24%)

openclaw models list --agent main --json also included:

{
  "key": "openai-codex/gpt-5.5",
  "name": "GPT-5.5",
  "contextWindow": 1000000,
  "available": true
}

The session then eventually failed in Feishu with the generic Something went wrong message instead of compacting early enough.

Expected behavior

For GPT-5.5 when served through Codex/OAuth runtime, OpenClaw should budget against the Codex surface limit, not the API limit.

Expected behavior would be one or more of:

• openai-codex/gpt-5.5 defaults to the Codex-specific context budget documented by OpenAI.
• Canonical openai/gpt-5.5 routed through Codex/OAuth uses the effective Codex budget for session persistence, status, and context-engine token budgeting.
• Existing sessions with stale contextTokens: 1000000 are recomputed or capped when the effective runtime is Codex.
• Context overflow surfaces a clear context-overflow diagnostic and triggers compaction/retry where possible, instead of only the generic channel error.

Local workaround

Applied only to GPT-5.5 Codex-backed agents, not globally:

{
  "agents": {
    "list": [
      { "id": "main", "contextTokens": 350000 },
      { "id": "neo", "contextTokens": 350000 },
      { "id": "balao-wealth", "contextTokens": 350000 }
    ]
  }
}

Also updated existing stale session entries where model == "gpt-5.5" and contextTokens > 350000 to 350000, then restarted the gateway.

After workaround:

agent:main:feishu:default:...  gpt-5.5  OpenAI Codex  117k/350k (33%)
agent:main:feishu:zero:...     gpt-5.5  OpenAI Codex  135k/350k (39%)
agent:balao-wealth:main        gpt-5.5  OpenAI Codex  239k/350k (68%)

A direct smoke test returned normally:

openclaw agent --agent main --message "请只回复：ok"
# ok

Why this matters

A global lower contextThreshold is not a good workaround because other agents use different models with smaller native windows, e.g. MiniMax around ~200K. The durable fix should be model/runtime-specific context budgeting, so Codex GPT-5.5 can compact at a safe point without affecting MiniMax or other agents.

Related issues

This seems related to the broader Codex context metadata/history, but the observed failure mode is specifically the 1M API budget being applied to Codex/OAuth GPT-5.5 sessions and delaying lossless-claw compaction:

• #71403: 272k vs 400k metadata inconsistency
• #81816: compaction trigger inconsistency with 1M contextTokens
• #77858: contextTokens handling in native Codex harness

References

OpenAI's GPT-5.5 announcement states:

• GPT-5.5 in Codex: 400K context window
• GPT-5.5 in API: 1M context window

https://openai.com/index/introducing-gpt-5-5/