[akemmanuel]

Summary

Runs that use a Codex model (ChatGPT subscription) now get two extra tools, both backed by the same ChatGPT endpoints the Codex CLI uses:

• web_search calls backend-api/codex/alpha/search and supports search_query, open and find. Search state is scoped per Session, so later open/find calls can reuse ref ids from earlier searches. Citation markers are normalized to [turn0search0]. Results include up to 10 deduplicated http(s) sources, shown as links in the transcript.
• generate_image calls images/generations, or images/edits when reference images are given (gpt-image-2). It supports transparent backgrounds and up to 5 reference images from the Project. Reference paths go through the Host path authorizer. Images are saved to <project>/.opengui-images/<toolCallId>.png; the folder gets its own .gitignore (*) and existing files are never overwritten. The model also gets the image as an attachment, and the transcript shows it pinned below the tool row.

Host policy (Settings → Integrations)

• Web search: live, cached or off. Default: live.
• Image generation: on or off. Default: on.
• Shared offerings may use the Host's ChatGPT subscription. Default: off. Hidden on local single-user Hosts.

The policy is stored in Host settings and exposed at GET/PUT /api/host/codex-tools; PUT requires a human admin.

Harness changes

• combineAgentToolSources merges several AgentToolSources (MCP and ChatGPT tools) and rejects duplicate tool names.
• AgentToolScope now includes the run's model, and invoke receives toolCallId.
• ModelToolDefinition.maxResultBytes lets a tool raise its result cap, clamped to 64KB. Search results use 32KB.

Errors

• After a 401, the credential is refreshed once and the call retried.
• 401, 403, 429 (usage limit), 5xx, timeout and network failures produce clear, redacted messages.

Testing

• vp check is clean.
• New unit tests for the ChatGPT tool source (availability, policy, shared offerings, request shape, 401 refresh, 429, citation normalization, image save and no overwrite, edits endpoint, denied reference path).
• New Host integration tests (a Codex Session gets the tools, the search call uses the Codex credential, the policy survives a restart).
• Updated route auth and validation contracts, harness agent-tool tests, and transcript and tool-call model tests.
• Checked manually against the live ChatGPT backend: a search plus a follow-up open, and a transparent image generated and saved.
• pnpm run slop-check passes.
• The full vp test suite was not run here, only the affected areas (backend codex-tools, host, routes, harness agent tools, transcript, message-list).

🤖 Generated with Claude Code