DO NOT MERGE YET: harden Codex structured output handling

[PhilosophiMoonbeam]

Status

Draft only. Do not merge without further live Codex/SWE-AF runtime testing.

Why

The upstream Codex fixes address auth/default model behavior, but AgentField SDK 0.1.96 still does not appear to drive Codex through native --output-schema / --output-last-message. In SWE-AF runs this can still leave AgentField relying on the generic .agentfield_output.json write-tool contract, which is fragile for Codex and related to empty/null structured outputs.

Changes

• Use Codex native structured-output flags and write the final answer to AgentField's expected .agentfield_output.json path.
• Keep non-Codex providers on AgentField's original output-file prompt suffix.
• Preserve model selection and project directory handling for Codex.
• Make HITL form choice options an explicit Pydantic model so strict structured-output schemas remain valid.

Permission-mode note

This draft intentionally preserves SWE-AF's existing Codex permission mapping. permission_mode: "auto" still uses --dangerously-bypass-approvals-and-sandbox; explicit read-only, workspace-write, and danger-full-access values are still passed through; and unspecified/unknown values still fall back to --sandbox workspace-write.

We briefly tried mapping auto to workspace-write, but live SWE-AF git initialization can fail because Codex cannot create .git/refs/...lock files inside that sandbox. This PR does not attempt to change permissions; safer sandboxing should be treated as separate future work requiring a git-management design that does not need Codex to mutate .git metadata.

Local verification

• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/python -m pytest tests/test_codex_harness_patch.py tests/test_ask_user.py tests/test_environment_scout.py -q -> 40 passed
• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/python -m ruff check swe_af/runtime/codex_harness_patch.py swe_af/hitl/ask_user.py tests/test_codex_harness_patch.py tests/test_ask_user.py tests/test_environment_scout.py -> passed
• AGENTFIELD_SERVER=http://localhost:9999 .venv/bin/python -m pytest tests/test_model_config.py tests/test_fatal_error.py tests/test_empty_build_guard.py tests/fast/test_app.py -q -> 123 passed
• Full suite: 1017 passed, 1 skipped, with one expected guard failure because this branch intentionally modifies non-swe_af/fast files.

Still needed before merge

• Live end-to-end SWE-AF run using runtime: "codex".
• Confirm HITL pause/resume behavior still works during active child executions.
• Confirm no regressions for claude_code or open_code providers.

[CLAassistant]

All committers have signed the CLA.