feat(integrations): extract Codex SDK harness adapter from evals

[miguelg719]

Stacked on #2748. Part 3 of the harness consolidation stack. Reshaped after review feedback: thin session layer only.

What

New thin package @browserbasehq/stagehand-integrations-codex-sdk: loadCodexSdk + runCodexSession — the startThread/runStreamed event loop (thread config, tool-step budget with abort + listener detach, usage/stop-reason capture) with explicit options and no EVAL_* env reads. Evals' codex runner and the codex facade example both call it.

Deliberately NOT extracted (stays in evals): the loopback code bridge and mount machinery — they expose evals' in-process tool surfaces to codex and have no integrations consumer. codexCodeBridge.ts is back in evals unchanged (still using the shared redaction from #2746).

The example keeps its posture (read-only sandbox, on-failure approvals, harness-tuned default model) and now exits non-zero on failed sessions instead of printing nothing and exiting 0.

Net diff +674/−420 — scaffolding plus the session module; logic is a move.

Verification

• Full gates ✅; session tests moved with the code
• Connected smoke re-run on the reshaped code — results in PR comment; step-budget semantics previously verified identical to main (50 counted command executions)

[changeset-bot]

⚠️ No Changeset found

Latest commit: b8e63d9

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[cubic-dev-ai]

All reported issues were addressed across 22 files

Architecture diagram

sequenceDiagram
    participant E as Evals Runner
    participant C as Codex SDK Harness
    participant B as Bridge Server
    participant S as Codex Session
    participant M as MCP Server
    participant A as Codex Example

    Note over E,C: NEW: Evals delegates to shared harness

    E->>C: runCodexSession(prompt, model, thread config)
    C->>C: Create AbortController, forward signal

    alt Handle-based mount (code bridge)
        E->>B: startCodeBridge(handles, context)
        B-->>E: bridge port
        E->>M: Prepare mount session, write browser_run.mjs
        M->>B: POST /run with snippet code
        B->>B: Execute snippet with handles in scope
        B-->>M: Result or error (redacted)
        B->>E: onRunExecuted notification
    else MCP-based mount
        E->>M: Prepare mount via mcp_servers config
        M->>E: codexConfig forwarded
    end

    E->>S: Start Codex thread (sandbox, approval, webSearch disabled)
    S-->>C: Stream events
    C->>C: Count tool steps (command_execution + mcp_tool_call)

    alt Step budget exceeded (default 50)
        C->>S: Abort stream with budget exhaustion reason
    end

    alt Agent produces completion
        C-->>E: finalMessage, token usage, status=completed
        E->>E: Build transcript, parse result
    else SDK error or turn failed
        C-->>E: iterationError, status=sdk_error
        E->>E: Mark eval as failed
    end

    Note over A,C: NEW: Codex example uses shared harness
    
    A->>C: prepareCodexMountSession(mcp mount)
    C-->>A: Prepared session (cwd, codexConfig)
    A->>C: runCodexSession(read-only sandbox, on-failure approval)
    C-->>A: Result
    
    alt Status not completed
        A->>A: Throw error (non-zero exit)
    else Completed
        A->>A: Print final message
    end

Loading

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[miguelg719]

Connected smoke on the reshaped code (restored evals bridge/mount + thin session package): b:webvoyager --harness codex --tool stagehand_code -e browserbase — 3/3 trials passed (100%), all complete in 23–46 steps with real answers. (Baseline before the stack was 1/3 on this task; the bridge round-trips, step budget, and grading all flow through the new session layer.)