openclaw - 💡(How to fix) Fix [Bug]: /context detail shows ~62k untracked provider/runtime overhead on fresh OpenAI Codex gpt-5.5 session [5 pull requests]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Summary

Fresh reset sessions using OpenAI Codex OAuth / openai/gpt-5.5 show ~66k tokens of context usage even though /context detail can only account for ~4k tracked prompt tokens.

Environment

• OpenClaw: 2026.5.22
• Runtime: OpenAI Codex OAuth
• Model: openai/gpt-5.5
• OS: Ubuntu 24.04
• Agent tested: Claw and Reviewer

Observed

After session reset and a trivial prompt like Ping?, /context detail reports:

• Tracked prompt estimate: ~4,240 tok
• Actual context usage: ~66,793 tok
• Untracked provider/runtime overhead: ~62,553 tok

Changing the following did not materially reduce the number:

• tools.profile: full → coding → minimal
• disabled memory-core and memory-wiki
• reduced Claw skills to only summarize/taskflow
• rewrote AGENTS.md to avoid automatic MEMORY.md/daily note loading
• reset session repeatedly
• tested Reviewer as well

Expected

A fresh reset session should report context usage close to the tracked prompt estimate plus known runtime/tool overhead. If the extra ~62k is cacheRead/cacheWrite, provider-side native thread state, or another runtime value, /context detail should label it clearly and it should not trigger misleading context/compaction behavior.

Related issues

• #67419 bootstrap files re-injected / 20–30% startup context
• #72321 cumulative cacheRead/cacheWrite context display bug
• #13853 cache token inflation causes premature compaction
• #13782 /status context display inflated to 200k/200k
• #70972 status bar shows 0% when context is mostly cached
• #75707 compaction/session/tokenization performance loops

Steps to reproduce

1. Start OpenClaw 2026.5.22 with OpenAI Codex OAuth configured and openai/gpt-5.5 as the active model.
2. Use a fresh/reset agent session. In my case, I reproduced this with both Claw and Reviewer.
3. Minimize local context sources:
   • set tools.profile to minimal
   • disable memory-core and memory-wiki
   • reduce Claw skills to only minimal skills, e.g. summarize and taskflow
   • rewrite AGENTS.md to avoid automatic MEMORY.md / daily-note loading
4. Send a trivial message such as Ping?.
5. Run /context detail.
6. Observe that the tracked prompt estimate is small, but actual context usage is much larger.

Observed example:

• Tracked prompt estimate: 16,957 chars (~4,240 tok)
• Actual context usage: 66,793 tok
• Untracked provider/runtime overhead: ~62,553 tok
• Session tokens: 66,793 total / ctx=200,000

This also persists after changing tools.profile from full → coding → minimal, disabling memory plugins, reducing skills, and resetting the session.

Expected behavior

For a fresh/reset session with a trivial prompt, actual context usage should be close to the tracked prompt estimate plus clearly identified runtime/tool overhead.

If the extra ~60k tokens represent provider-side cache tokens, native Codex app-server state, hidden runtime scaffolding, or another expected overhead, /context detail should label the source clearly.

This untracked overhead should not be presented as ordinary context occupancy or used to trigger misleading context warnings/compaction behavior unless it truly consumes the model’s available context window.

Actual behavior

After resetting a fresh agent session and sending a trivial prompt such as Ping?, OpenClaw reports high context usage even though /context detail accounts for only a small prompt.

Observed /context detail output:

• Tracked prompt estimate: 16,957 chars (~4,240 tok)
• Actual context usage: 66,793 tok
• Untracked provider/runtime overhead: ~62,553 tok
• Session tokens: 66,793 total / ctx=200,000

The issue persisted after the following reductions/tests:

• tools.profile changed from full → coding → minimal
• memory-core and memory-wiki disabled
• Claw skills reduced to summarize and taskflow
• AGENTS.md rewritten to avoid automatic MEMORY.md / daily-note loading
• session reset repeatedly
• tested with both Claw and Reviewer

Reviewer showed a similar absolute baseline, around 63.9k / 200k, suggesting the overhead is not specific to one agent.

OpenClaw version

2026.5.22 (a374c3a)

Operating system

Ubuntu 24.04.4

Install method

npm global

Model

openai/gpt-5.5

Provider / routing chain

OpenClaw -> OpenAI Codex OAuth -> openai/gpt-5.5

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information