Skip to content

fix(cx): stabilize Codex continue sessions - #1004

Open
josephvoxone wants to merge 4 commits into
decolua:masterfrom
josephvoxone:codex/pr-865-continue-session
Open

josephvoxone wants to merge 4 commits into
decolua:masterfrom
josephvoxone:codex/pr-865-continue-session

Conversation

@josephvoxone

Copy link
Copy Markdown

Summary

  • Builds on feat(cx): add reasoning effort variants for GPT-5.5 and GPT-5.4 #865 by keeping Codex session_id stable across resumed conversations.
  • Fixes node-machine-id loading in the Next server bundle.
  • Normalizes text-only Claude/OpenAI payloads and raw NDJSON stream parsing uncovered during local verification.
  • Preserves GPT-5.5/GPT-5.4 Codex effort aliases and review metadata from the original PR branch.

Verification

  • npx --yes vitest run tests/unit/translator-request-normalization.test.js tests/unit/codex-image-fetch.test.js tests/unit/provider-models-codex-effort.test.js
  • npm run build
  • Local dev server on http://localhost:20128
    • /api/health returned {"ok":true}
    • / redirected to /dashboard
    • /api/v1/models included cx/gpt-5.5-medium and cx/gpt-5.4-xhigh
  • Local Codex continuation simulation confirmed session_id stays stable while model suffix maps to upstream model and reasoning effort.

mugnimaestra and others added 4 commits May 4, 2026 11:41
Add withCodexEffortVariants() helper and generate -xhigh, -high, -low,
-none model entries for gpt-5.5 and gpt-5.4 in the Codex provider.

Consumer tools using @ai-sdk/openai-compatible (e.g. opencode) cannot
pass 'none' or 'xhigh' reasoning effort because the SDK's zod schema
only validates ['low', 'medium', 'high']. The model name suffix approach
(gpt-5.5-xhigh) bypasses SDK validation entirely since reasoningEffort
never appears in the request body — the effort level is encoded in the
model name, which codex.js already parses and strips generically.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants