DO NOT MERGE YET: harden Codex structured-output routing

[PhilosophiMoonbeam]

Status

Draft. Do not merge without a fresh end-to-end Codex build and non-Codex smoke.

Summary

• Use Codex native structured output with --output-schema and --output-last-message.
• Use per-call schema/output temp dirs to avoid concurrent same-repo collisions.
• Sanitize schemas for strict Codex structured output.
• Route QA synthesizer through provider-aware router.harness() so bare Codex model names work.
• Use an explicit HITL choice model instead of a free-form dict schema.
• Preserve configured permission_mode through DAG git operations: setup, merge, integration test, cleanup.
• Default Codex to --dangerously-bypass-approvals-and-sandbox when permission mode is omitted, empty, auto, default, danger-full-access, or bypassPermissions; explicit read-only and workspace-write still pass through.

Why

• Codex does not use AgentField's file-write path for .agentfield_output.json.
• Shared output files collide during concurrent same-repo calls.
• Strict structured output rejects defaults and free-form object maps.
• Dropped permission mode made Codex fall back to workspace-write, which blocked git metadata writes in SWE-AF merge/setup paths.

Verification

• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/python -m pytest -q -> 1025 passed, 1 skipped
• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/python -m pytest tests/test_coding_loop_regressions.py tests/test_codex_harness_patch.py tests/test_model_config.py -q -> 61 passed
• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/ruff check swe_af/reasoners/execution_agents.py swe_af/runtime/codex_harness_patch.py swe_af/execution/dag_executor.py tests/test_coding_loop_regressions.py tests/test_codex_harness_patch.py -> passed
• git diff --check origin/main..HEAD -> passed

Still Needed

• One clean end-to-end SWE-AF build with runtime: "codex" on a disposable repo.
• Smoke claude_code and open_code paths.

[CLAassistant]

All committers have signed the CLA.