Replace codex-agent-driver with the official openai-codex SDK

[nabinpkl]

What

Migrates the codex agent runtime off the hand-written codex-agent-driver path-dependency (sibling second-brain repo) and onto OpenAI's official openai-codex Python SDK. The SDK is the harness the driver emulated — it drives codex app-server, exposes native threads, streams typed notifications, supports read-only sandbox + server-enforced output_schema, and bundles the codex binary. Recorded in ADR 17 (supersedes ADR 15's mechanism).

Agent plane (agent-service/)

• One shared AsyncCodex app-server at lifespan; each chat thread is one native codex thread (thread_start / thread_resume). Drops the per-thread codex_home machinery (CODEX_HOME_ROOT), the state-sqlite model-read hack, and the worker-thread/queue bridge in favour of async for over the SDK turn stream.
• codex_config.py — per-thread config overlay mounting the data-plane MCP server (four-tool allow-list) + the built-in-tool lockdown, plus the writable-CODEX_HOME builder.
• codex_events.py — parses the SDK's typed notifications back to their camelCase wire dict, insulating the driver from beta SDK attribute churn.
• Helper calls (constitution gate / eval judge / repeat detector) run on a separate codex helper app-server (own CODEX_HOME, ephemeral threads, no MCP) — preserves the no-bleed separation.
• codex_profile.py removed; tests, smoke script, env, Docker, compose, and CI (agent-service job re-enabled) updated. SDK is Beta, pinned exactly with the accepted risk in docs/dependency-exceptions.md.

Data plane (backend/src/mcp.rs)

The bundled codex (0.137) forwards MCP tool input schemas to the OpenAI API verbatim, which rejects a root $ref schema in strict mode. The three transparent-newtype args (wallet_profile, community_summary, get_token_info) now implement JsonSchema manually to inline the object at the root. Regenerated the tools/list snapshot consumed by the hermetic-eval mock.

Evals

model_assertions_codex.yaml asserted the constitution gate uses the pydantic-ai policy model, contradicting the all-or-nothing helper routing (under codex runtime the gate runs via codex and emits no gen_ai model span). Replaced that probe with a functional "gate fired" check; re-minted the baseline.

Verification

• 345 agent-service unit tests pass (<5s, no live LLM); 360 collect clean; ruff clean on edited files.
• 22 backend mcp:: tests pass (incl. schema snapshot drift test).
• Full stack docker compose up -d --build; live codex turns confirmed: token streaming, wallet_profile MCP dispatch, lockdown, model stamping, constitution gate fires + approves.
• Codex live-eval suite 5/5 pass (turn-root, codex-primary-model-used, no-error, no-builtin-tool-call, constitution-gate-fired).

🤖 Generated with Claude Code