v0.19.0 — CarAdapter outbound inference

CAR becomes a peer integration target alongside Claude Code and Codex; one OpenAI default-config regression fixed; docs realigned to what the code actually does.

Added

• CarAdapter — route Neo's outbound inference through CAR's unified inference layer. With model=None (default), CAR's adaptive route_model picks local backends (Candle + MLX for Qwen3, Gemma 4) or remote providers (OpenAI, Anthropic, Google) per call. Pin via model="gpt-5.3-codex" etc. Activate with neo --config set --config-key provider --config-value car or env NEO_PROVIDER=car. Default intent_hint={"task":"code"} so the router knows it's serving code work; this is also Neo's local workaround for Parslee-ai/car-releases#52, where route_model defaults are cost-biased rather than quality-biased for code prompts.
• car_inference.get_runtime() singleton — process-wide CarRuntime shared between the inbound A2A host (neo serve) and the outbound CarAdapter. Same state, same policies, same eventlog, single auth handshake. car_host.run_server() pulls from the singleton instead of constructing its own.
• AGENTS.md ecosystem — added Neo's own AGENTS.md (mirror of CLAUDE.md) and auto-ingestion of {project}/AGENTS.md / {project}/.github/AGENTS.md in memory/constraints.py. Codex-spec tools now see the same project rules as Claude Code.
• neo memory prune subcommand — compacts a project's facts file by dropping old invalid tombstones (default: >30 days since last access). Supports --all, --dry-run, --limit, --max-invalid-age-days.
• neo --dry-run "query" — print the assembled context without making the LLM call. Useful for debugging context-gatherer behavior.
• OpenAIAdapter /v1/responses passes max_output_tokens and reasoning.effort so gpt-5*/o-series generation knobs from NeoConfig (and engine memory-driven effort selection) actually take effect.
• Construct library bundled into the wheel as neo/construct_library via [tool.hatch.build.targets.wheel.force-include]. neo construct list now works for pip-installed users; was dev-tree only.

Changed (migration)

• openai>=1.0.0 and pyyaml>=6.0 are now core dependencies, not extras. OpenAI is the default provider, so the base install is runnable on OPENAI_API_KEY alone. The [openai] extra still exists but is redundant; pinning workflows that install only with [anthropic] to avoid OpenAI dependencies should switch to a separate venv.
• Provider env-var selection is now provider-matched, not first-found. Previously NeoConfig.load() picked whichever provider env var was set first (so OPENAI_API_KEY was used even when provider=anthropic). Now it picks the env var that matches config.provider (OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, or AZURE_OPENAI_API_KEY), with NEO_API_KEY as the explicit cross-provider override. Setups that relied on the old leak-through behavior must set the correct env var for their active provider.

Fixed

• OpenAIAdapter regression: temperature rejected by gpt-5/o-series on /v1/responses* (introduced in ad8718e, fixed in aa52ca0). The hardening commit added "temperature": temperature unconditionally to the responses payload. gpt-5/codex/o-series reject it with Unsupported parameter: 'temperature' is not supported with this model — so default config (provider=openai, model=gpt-5.5) was broken on first call. Adapter still accepts the kwarg for LMAdapter ABC compat but no longer forwards it for these model families.
• neo --version no longer eager-initializes FactStore. Was paying the full embedding-pipeline cost (Jina model load + index build) just to read fact counts. Now eager_init=False.
• pipx editable installs were misclassified as INSTALL_EXTERNAL so auto-update silently no-op'd. update_checker._detect_install_method() now checks sys.prefix against the pipx venv path first.
• FileStorage.load() backs up corrupt JSON before re-raising. JSONDecodeError is handled separately from PermissionError/IOError: corrupt files trigger _backup_corrupt_file() so the original isn't silently destroyed by a subsequent save.

Docs

• README, QUICKSTART, INSTALL, and CONTRIBUTING restructured so CAR, Claude Code, and Codex are documented symmetrically. CAR is the lead surface (## Run as an Agent (CAR / A2A) before the plugin sections); inbound (neo serve) and outbound (provider=car) directions both covered with Python and CLI examples; intent_hint knobs and task=code default documented in the LM Adapters CAR subsection.
• docs/tree-sitter-setup.md rewrite: migrated to tree-sitter-language-pack from the deprecated tree-sitter-languages; dropped the obsolete Python 3.13 exclusion; added Ruby/PHP/Swift/Kotlin to the supported list; fixed MAX_CHUNK_LENGTH = 2000 chars (previously claimed >100KB); added the code_smells._ERROR_SWALLOW_DETECTORS registration step.
• README Research & References rewritten to match what's actually implemented. Previously cited five papers (Self-Planning, AdaCoder, As-Needed Decomposition, Multi-Agents Survey, Liu PG-TD) with "Implementation" claims that have zero code references. Now lists the 11 papers from the 0.18 memory architecture wave that ARE wired into code, with file-anchored citations. Wrong CodeSim paper (Xu 2023 vs the implemented Hou 2025) corrected; ReasoningBank moved to "Historical influences" since only the deprecated persistent_reasoning.py uses it. StateBench and memgine cited explicitly (was a one-word parenthetical).
• CLAUDE.md memory hygiene rewritten with correct outcome deltas (ACCEPTED +0.2 / MODIFIED −0.2 / UNVERIFIED +0.1, all ±arch_mod), the full rank_score formula including effectiveness_f, the triple-trigger consolidation gate, and two footgun warnings (new OutcomeType requires updating both outcomes.py and store.detect_implicit_feedback; rank_score is shared between FactStore and ContextAssembler). CLAUDE.md and AGENTS.md are now kept in sync as a deliberate invariant.

Internal

• CarAdapter test coverage: 23 unit/integration tests in tests/test_car_adapter.py — 15 mocked (always run), 4 live integration gated on car_inference.is_available() (skipped if car-runtime not installed), 2 negative-case for runtime failure propagation, 2 wiring. Covers chat-message vs string-prompt paths, default vs explicit intent_hint, real-schema metrics (model_used/prompt_tokens/completion_tokens/latency_ms/trace_id), error propagation (ConnectionError, RuntimeError), bad model id, oversized prompt, and the singleton-shared-with-car_host identity invariant.
• CarAdapter schema validated against car-runtime 0.15.1. The first mocked-only release caught zero schema bugs by accident; running against the live daemon exposed three drift categories (intent_hint→intent_json JSON string; temperature/stop rejected; model/input_tokens/output_tokens/cache_read_input_tokens → model_used/prompt_tokens/completion_tokens/no-such-field). All fixed; FakeRuntime now mirrors the real shape.
• Upstream bugs encountered while wiring CAR integration: Parslee-ai/car-releases#50 (classify mis-routes — CLOSED), #51 (stale v0.8 reference in open_session error — OPEN), #52 (router cost-biased for code prompts — OPEN; worked around in CarAdapter via task=code default).