cortex-loop 0.1.0a2

Runtime-agnostic enforcement layer for AI coding agents

Cortex

Cortex is a runtime-agnostic enforcement layer for AI coding agents. It verifies completion claims through a deterministic stop path instead of trusting summaries, so "done" only lands when the evidence is real. Coding agents can sound finished before reality has caught up; Cortex turns that completion claim into something the system can prove or deny.

With Cortex, the path of least resistance becomes the path of best practice: the easiest way to finish is to do the work, satisfy the evidence boundary, and let the stop path close honestly.

The stop path is the product. Cortex is not a planning framework, generic memory layer, or workflow wrapper. It exists to make completion claims legible, checkable, and hard to bluff.

Runtime Status

Runtime	Install profile	Status	Notes
Claude Code	cortex runtime install --profile claude	Shipped	Strongest current runtime. The truthful boundary is live-proven; remaining caveats are minor.
Gemini CLI	cortex runtime install --profile gemini	Shipped with watchlist	The truthful boundary is live-proven. One operational watchlist remains: blocked malformed stops can leave Gemini CLI resident until operator termination.
OpenAI Codex App Server (native)	cortex runtime install --profile openai	Experimental	Honest native App Server surface: one turn, then stop audit. Several critical paths are proven, but positive strict close is still not dependable enough for broader support.
OpenAI Codex App Server (assisted)	cortex runtime install --profile openai-assisted	Experimental	Explicit bounded assisted mode with a current committed shared-harness pair: Cortex provides session-start context with a short kernel-owned completion preview, one short evidence-expectation line, and one kernel-derived corrective turn. The lane remains supplemental-only, not native-substitutive proof.

Status labels in this table:

• Shipped: live-proven on the shared harness; remaining caveats are non-critical.
• Shipped with watchlist: live-proven, but one named runtime quirk still creates real operator risk.
• Experimental: several important paths are proven, but one boundary-critical guarantee remains unproven or unstable.

Detailed runtime release evidence lives in docs/ADAPTER_VALIDATION.md plus committed provenance under tests/fixtures/adapter_validation/claude/PROVENANCE.json, tests/fixtures/adapter_validation/gemini/PROVENANCE.json, and tests/fixtures/adapter_validation/openai/PROVENANCE.json.

Current product-proof truth: Cortex is strong on truthful completion boundaries, but the repo still does not earn a net-positive product claim over the raw model. Phase 9 is landed at a truthful-withheld endpoint rather than an earned product-proof win.

This repo state is also the final archival v1 reference point for the enforcement-layer line. The archival package target is cortex-loop 0.1.0a2.

Quickstart

Default public install:

pipx install cortex-loop
cortex init
cortex runtime install --profile claude

Supported advanced alternative:

uv tool install cortex-loop

Before cortex check, choose the baseline that matches the repository or the machine you are using:

• trusted local repo without Docker: add pytest to the promoted tool surface (pipx inject cortex-loop pytest or uv tool install --with pytest cortex-loop) and switch cortex.toml to the trusted-host baseline in docs/SECURE_DEFAULTS.md
• untrusted or container baseline: keep the defaults and ensure Docker is on PATH

Then run:

cortex check

Install, update, uninstall, and live proof details live in docs/INSTALL.md. Guided evaluator onboarding lives in docs/ALPHA_TESTER_ONBOARDING.md. Trust boundary and pack review posture live in docs/TRUST_BOUNDARY.md and docs/PACK_POLICY.md.

For a first evaluation, keep claude. Use gemini when you want the shipped watchlist surface. Use openai only when you are explicitly evaluating the experimental native surface. Use openai-assisted when you want the current bounded assisted OpenAI lane: session-start context with a short kernel-owned completion preview, one short evidence-expectation line, one kernel-derived corrective turn aimed at the unresolved completion gap, and a current shared-harness pair that still remains supplemental-only.

What Cortex Enforces

• Stop-claim normalization and strict payload handling (tests/test_stop_contract.py)
• Challenge coverage across active categories (tests/test_challenges.py)
• Invariant execution outside model output (tests/test_invariants.py)
• Deterministic verdict policy (tests/test_stop_policy.py)
• Session-scoped failure memory and retry controls (tests/test_retry.py, tests/test_graveyard.py)

What Cortex Is Not

• not a generic multi-agent framework
• not prompt theater that trusts summaries
• not a workflow layer whose main job is planning or project management

Read Next

1. START_HERE.md
2. ARCHITECTURE.md
3. MISSION.md
4. docs/README.md
5. CONTRIBUTING.md

Repository Boundaries

• Product-critical code lives in cortex/, cortex_ops_cli/, cortex_repomap/, runtime profile templates, and core tests.
• Extension manifests live in packs/; this repo's pack validation suites live in tests/packs/.
• eval/ is internal and historical validation only. It is not a downstream product interface or product design authority.
• docs/archive/ and _governance_archive/ are historical context, not active authority.

v0.1 remains kernel-first: one hardened kernel, one reference pack, and runtime adapters with capture-backed tests.