fix: correct Codex context window (266K) and default fallback order

[dorukardahan]

Changes

Codex context window: 200K → 266K

The Codex 5.3 context window was listed as 200,000 tokens across benchmarks.json and SKILL.md. The actual value is 266,000 tokens (per OpenAI docs and confirmed in production configs).

Updated in:

• benchmarks.json — context_window field
• SKILL.md — model tier table

Default fallback order: Codex before Pro

The default fallback chain was Opus → Pro → Codex → Kimi. Reordered to Opus → Codex → Pro → Kimi.

Rationale: Codex has a higher Intelligence score (51.5) than Pro (48.4), making it a better general-purpose fallback. Pro is a specialist (GPQA 0.908, 1M context) — great when you specifically need research or long context, but Codex is the safer all-rounder when Opus is unavailable.

Updated in:

• examples/full-stack/openclaw.json — defaults and main agent
• references/provider-config.md — reference config block
• SKILL.md — Full Stack fallback table
• README.md — Cross-Provider Fallback Chains table

Impact

No behavioral change for users with custom fallback chains. Only affects the default/example configs.

[dorukardahan]

Thanks for the factual corrections — this PR improves reliability docs materially. I reviewed diff scope for README.md, SKILL.md, benchmarks.json, examples/full-stack/openclaw.json, and references/provider-config.md.

What looks good:

• Codex context window is now consistently 266K in both SKILL.md and benchmarks.json.
• Main-agent fallback order (Opus -> Codex -> Gemini Pro -> K2.5) is now aligned across README table, SKILL full-stack routing table, examples/full-stack/openclaw.json, and references/provider-config.md.
• No secrets/private paths/credentials introduced (good for public repo).

Detailed consistency check:

• benchmarks.json: openai-codex context_window updated to 266000
• SKILL.md model catalog: Codex context updated to 266K
• README.md fallback matrix: main row reordered to Codex before Gemini Pro
• examples/full-stack/openclaw.json defaults + main agent fallbacks reordered
• references/provider-config.md recommended config reordered

Potential follow-up (non-blocking):

• There may still be stale fallback-order wording in narrative prose outside the changed blocks (if any). I did not see contradictions in this PRs touched lines, but a repo-wide grep for the old order string (Gemini Pro before Codex in reasoning/main fallbacks) could prevent future drift.

Overall: LGTM from a factual accuracy + cross-file consistency perspective.