Issue Station
Sign In
Tools hereToolbox

openclaw - 💡(How to fix) Fix GPT-5.4 context window resolves inconsistently after reboot/fresh session (200k / 272k / 1050k disagreement)

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

On macOS with OpenClaw 2026.4.15, openai-codex/gpt-5.4 can resolve to different context limits depending on which layer you look at, especially after reboot or fresh session creation.

Observed split:

  • fresh Telegram DM session seeded at 200000
  • Mac app model catalog showed 272000 for openai-codex -> gpt-5.4
  • provider/runtime layer knew GPT-5.4 should be 1050000

This makes the system feel broken even when the configured model is correct.

Root Cause

OpenClaw 2026.4.15 can correctly route openai-codex/gpt-5.4 while still incorrectly seeding or displaying its context budget as 200k or 272k after reboot, because session seeding, status display, and app catalog resolve context from inconsistent paths.

Fix Action

Fix / Workaround

Observed value before local patch:

Local workaround that fixed the live session

Code Example

resolveContextTokensForModel(... allowAsyncLoad: false, fallbackContextTokens: 2e5) ?? 2e5

---

resolveContextTokensForModel(... allowAsyncLoad: false) ?? 2e5

---

lookupContextTokens(resolved.model, { allowAsyncLoad: false }) ?? 2e5

---

OPENAI_CODEX_GPT_54_DEFAULT_CONTEXT_TOKENS = 105e4
RAW_BUFFERClick to expand / collapse

Summary

On macOS with OpenClaw 2026.4.15, openai-codex/gpt-5.4 can resolve to different context limits depending on which layer you look at, especially after reboot or fresh session creation.

Observed split:

  • fresh Telegram DM session seeded at 200000
  • Mac app model catalog showed 272000 for openai-codex -> gpt-5.4
  • provider/runtime layer knew GPT-5.4 should be 1050000

This makes the system feel broken even when the configured model is correct.

Environment

  • OpenClaw: 2026.4.15
  • Install: pnpm
  • OS: macOS 26.4.1 arm64
  • Model path: openai-codex/gpt-5.4
  • Config: no agents.defaults.contextTokens, no GPT-5.4 context override in ~/.openclaw/openclaw.json

Symptoms

  1. OpenClaw app / status showed Telegram DM session at 200k
  2. Mac app model-catalog cache showed 272k
  3. Gateway/provider layer had GPT-5.4 default/native context at 1050000
  4. Reboot + fresh session brought the bad values back even after a previous live-session fix looked correct

Repro

  1. Set default model to openai-codex/gpt-5.4
  2. Restart gateway
  3. Start a fresh Telegram DM session or use /new
  4. Check:
    • openclaw status
    • ~/.openclaw/agents/main/sessions/sessions.json
    • ~/Library/Application Support/OpenClaw/model-catalog/models.generated.js

Observed result

  • live session may be created with contextTokens: 200000
  • app catalog may still report 272000

Expected result

GPT-5.4 should resolve consistently to the full 1M-class context budget across:

  • fresh session creation
  • app display
  • openclaw status
  • cached model catalog

Findings

1. Fresh session seeding path falls back to 200k

In:

  • /opt/homebrew/lib/node_modules/openclaw/dist/session-store.runtime-D0Wf4U9r.js

Observed logic:

resolveContextTokensForModel(... allowAsyncLoad: false, fallbackContextTokens: 2e5) ?? 2e5

If model context lookup does not resolve synchronously, the session gets seeded with 200000.

2. Inline/directive persistence path also falls back to 200k

In:

  • /opt/homebrew/lib/node_modules/openclaw/dist/directive-handling.persist.runtime-XUaMaVxD.js

Observed logic:

resolveContextTokensForModel(... allowAsyncLoad: false) ?? 2e5

3. Session defaults display path also prefers sync lookup

In:

  • /opt/homebrew/lib/node_modules/openclaw/dist/session-utils-hDe0hdis.js

Observed logic:

lookupContextTokens(resolved.model, { allowAsyncLoad: false }) ?? 2e5

4. App-side cached model catalog can drift from provider/runtime truth

In:

  • ~/Library/Application Support/OpenClaw/model-catalog/models.generated.js

Observed value before local patch:

  • openai-codex -> gpt-5.4 -> contextWindow: 272000

5. Provider/runtime layer had the correct larger value

In:

  • /opt/homebrew/lib/node_modules/openclaw/dist/openai-codex-provider-QEuYACvP.js

Observed value:

OPENAI_CODEX_GPT_54_DEFAULT_CONTEXT_TOKENS = 105e4

Why this is confusing

Three layers disagree:

  • provider/runtime says one thing
  • app catalog says another
  • live session store says a third

So users can be correctly configured and still see wildly wrong context numbers.

Local workaround that fixed the live session

As a local test, changing the session/runtime lookup paths from allowAsyncLoad: false to allowAsyncLoad: true, aligning the cached model catalog, and repairing the live session entry made the current Telegram DM session show correctly as 130k/1050k in openclaw status.

That suggests the upstream issue is not model support itself, but inconsistent context resolution paths.

Suggested fix

  1. Stop seeding fresh sessions with 2e5 when model lookup is unresolved but provider/model metadata is known.
  2. Make session creation, status display, and app model catalog derive GPT-5.4 context from one consistent source.
  3. Invalidate or refresh model-catalog cache when provider/runtime metadata changes.
  4. Add regression coverage for openai-codex/gpt-5.4 across:
    • fresh session creation
    • reboot/relaunch
    • app model catalog display
    • openclaw status

One-line summary

OpenClaw 2026.4.15 can correctly route openai-codex/gpt-5.4 while still incorrectly seeding or displaying its context budget as 200k or 272k after reboot, because session seeding, status display, and app catalog resolve context from inconsistent paths.

extent analysis

TL;DR

The most likely fix is to update the session seeding and status display logic to consistently derive the GPT-5.4 context from the provider/runtime metadata.

Guidance

  • Review the resolveContextTokensForModel function to ensure it prioritizes asynchronous model lookup and uses the correct fallback value.
  • Update the session creation and status display logic to use the same consistent source for GPT-5.4 context resolution.
  • Consider implementing a mechanism to invalidate or refresh the model-catalog cache when provider/runtime metadata changes.
  • Add regression tests to cover openai-codex/gpt-5.4 context resolution across different scenarios, including fresh session creation, reboot, and app model catalog display.

Example

// Example of updated resolveContextTokensForModel function
function resolveContextTokensForModel(model, allowAsyncLoad = true) {
  // Prioritize asynchronous model lookup
  if (allowAsyncLoad) {
    // Use async lookup to retrieve the correct context tokens
    return asyncLookupContextTokensForModel(model);
  } else {
    // Use a fallback value that is consistent with the provider/runtime metadata
    return getDefaultContextTokensForModel(model);
  }
}

Notes

The issue seems to be related to inconsistent context resolution paths, and updating the session seeding and status display logic to use a consistent source should resolve the problem. However, further testing and verification are necessary to ensure the fix works as expected.

Recommendation

Apply the suggested fix by updating the session creation and status display logic to consistently derive the GPT-5.4 context from the provider/runtime metadata, as this should resolve the inconsistent context resolution issue.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#integration issue#index setup#retrieval issue#search optimization#API routing

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

openclaw - 💡(How to fix) Fix GPT-5.4 context window resolves inconsistently after reboot/fresh session (200k / 272k / 1050k disagreement)