Best possible solution:

Land the revised consistency fix once exact-head checks are green or the unrelated gateway-test failures are resolved outside this PR.

Do we have a high-confidence way to reproduce the issue?

Yes. As a source-level reproduction, configure models.providers.openai-codex.models[].contextTokens for gpt-5.5, then run openai/gpt-5.5 through agentRuntime.id: "codex"; current main resolves the runtime context provider as openai.

Is this the best way to solve the issue?

Yes. The PR keeps the public openai/gpt-* model route intact and only redirects context-config lookup to openai-codex across run, status, and directive persistence paths, which is the narrowest maintainable fix I found.

What I checked:

• Current-main bug path: On current main, resolveEffectiveRuntimeModel() resolves resolveContextWindowInfo() against params.provider, and runEmbeddedPiAgent() passes the selected provider without a Codex context-config override, so native openai/gpt-* Codex runs look under openai rather than openai-codex. (src/agents/pi-embedded-runner/run/setup.ts:126, c5fcfa1b56e1)
• Documented route contract: The provider docs define the native Codex route as model ref openai/gpt-5.5 with agentRuntime.id: "codex", while auth/profile commands keep using provider id openai-codex. Public docs: docs/providers/openai.md. (docs/providers/openai.md:216, c5fcfa1b56e1)
• PR run-path fix: The PR diff adds resolveHarnessContextConfigProvider() and passes openai-codex as contextConfigProvider for provider=openai on the codex harness. (src/agents/pi-embedded-runner/run.ts:181, 850369b608e6)
• Earlier review blockers addressed: The revised diff adds the active changelog entry and changes directive persistence to call shared resolveContextTokens() with a Codex-aware context provider, addressing the prior ClawSweeper findings. (src/auto-reply/reply/directive-handling.persist.ts:67, 850369b608e6)
• Real behavior proof: The PR body supplies copied live output from macOS OpenClaw 2026.5.4 after applying the patch, showing /status with Context: 0/1.0m and agentMeta.contextTokens: 1000000 for native Codex openai/gpt-5.5. (850369b608e6)
• Exact-head CI state: GitHub check annotations show check-lint and check-test-types failing in src/gateway/server.cron.test.ts, which is not part of this PR diff; other touched-surface checks and the real behavior proof check are green. (src/gateway/server.cron.test.ts:194, 850369b608e6)

Likely related people:

• steipete: Recent history includes runtime context cap work, status runtime metadata, and model-selection maintenance on the same context/status surfaces. (role: recent maintainer / context-status owner; confidence: high; commits: 58d2b9dd46cc, e5dc0e6d15ea, 59fb9e5ca7fe; files: src/agents/context-window-guard.ts, src/status/status-message.ts, src/auto-reply/reply/model-selection.ts)
• vincentkoc: Recent commits touch status rendering, provider/runtime type boundaries, and embedded runner adjacent files that the PR changes. (role: recent status/harness maintainer; confidence: medium; commits: 46c99cff0b2f, a17d4371d101, 61da711b1a7e; files: src/status/status-message.ts, src/status/status-text.ts, src/agents/pi-embedded-runner/run/setup.ts)
• pashpashpash: Recent docs and runtime work clarified the Codex subscription/native route that defines the user-facing contract this PR repairs. (role: adjacent Codex route/docs owner; confidence: medium; commits: 9fb90f3d2921, 8adbee3a686a, 401ae38f13a3; files: docs/providers/openai.md)

Remaining risk / open question:

• Exact-head CI currently has failing check-lint and check-test-types annotations in untouched src/gateway/server.cron.test.ts; this looks unrelated to the PR diff but still needs CI resolution before merge.

Codex review notes: model gpt-5.5, reasoning high; reviewed against c5fcfa1b56e1.