Imported from cortex-loop/cortex-loop (AGENTS.md). Install upstream with npx skills add cortex-loop/cortex-loop. Copyright stays with the author.

Cortex v2 Repo Agent Contract

This file applies to agents editing this repository; it does not define runtime policy for downstream Cortex users. This root AGENTS.md is the only agent contract in the repo.

Agent Briefing

Read this first, every session.

For repo/product judgments in this repository, do not default to affirming the user's ideas and do not default to criticizing them. Do not let prior conversation style, model personality, or training-time preferences decide Cortex positions. Use only the repo's recorded goals and current proof.

Form positions from observable repo truth: docs/CORTEX.md for Cortex identity and narrative fit; the V2 packet docs (docs/CORTEX_V2_*.md) for packet law; internal/truth/cortex_status.json for current operational truth; and cortex/** plus tests/** for implemented behavior and proof.

If you lack doctrine-and-code grounding for a repo position, you do not have that position yet. Read the specific missing surface, or say "I don't know yet; I need to check X." Do not manufacture an answer from the user's latest framing or generic priors.

If context was compacted or your Cortex model feels thin, run python3 internal/workflow/repo_workflow.py orient; use its generated capsule as orientation only, then ground repo positions in the cited docs, code, and tests.

Agreement and disagreement are both acceptable when earned by evidence. Unearned agreement and ungrounded criticism are both failures.

Bootstrap

Bootstrap reads, in order, every session:

1. AGENTS.md
2. docs/CORTEX.md
3. docs/CORTEX_STATUS.md
4. git branch --show-current
5. git status --short --untracked-files=all

Bootstrap reads are preparation, not response content. Do them silently unless the user asks what was read or the result matters to the answer. Begin responses with the substantive answer to the user's request, not with a summary of what you read or what kind of turn you think this is. The operator-split note belongs in turns that involve user empirical work, not in every turn.

Answer First

The substantive answer to the user's request is the primary deliverable. Mission reflection is administrative closure. Do not let the reflection grid determine the answer's structure, ordering, or emphasis.

Mission

Cortex is the shipped rich multi-host executive layer in this repository: an installable executive layer around a model or CLI that adds continuity, focused persistence, context adoption, uncertainty-aware brake, truthful closure, and capability-aware routing after post-training. Lab, eval, archive, and workflow surfaces exist to falsify or prove product seams; they must not become the product identity. docs/CORTEX.md is the canonical narrative for Cortex identity and product fit.

Authority

Authority is scoped, not ordinal. Resolve conflicts by the dimension the question lives in:

• docs/CORTEX.md owns identity, narrative, mission fit, and product-fit rejection.
• docs/CORTEX_V2_CORE_2.md, docs/CORTEX_V2_SRE_2.md, and docs/CORTEX_V2_AUX_2.md own packet law and math.
• internal/truth/cortex_status.json owns current operational truth.
• docs/CORTEX_STATUS.md is the generated human view of that registry.
• docs/internal/REPO_WORKFLOW.md owns branch, session, command, closeout, publication, hook-health, generated-doc, dogfood, and live-CLI mechanics.
• docs/internal/MISSION_REFLECTION_CONTRACT.md owns chat-boundary mission-reflection modes and validation.
• docs/internal/ANTI_DRIFT_RULES.md owns the earned anti-drift rules and their V1 to V2 lesson context.

Non-Negotiables

• Do not turn Cortex into a narrow single-model shell.
• Do not flatten host differences into fake runtime uniformity.
• Do not let lab, eval, archive, governance, or workflow surfaces become Cortex product identity.
• Do not let task identity become product policy. Fixtures, task domains, hidden verifiers, and benchmarks are examples; product code implements abstract capability and state law.
• Do not move active executive policy into the core.
• Do not let shipping truth collapse conformance truth.
• Do not run paid service-lane commands unless the user explicitly approves spend in the current chat.
• Do not set CORTEX_LIVE_SERVICE_SPEND_APPROVED or equivalent spend opt-ins on your own initiative.
• Do not carry forward v1 mechanisms or host hacks without re-earning them under the packet.
• Do not claim product progress unless shipped runtime behavior changed or a direct product blocker was removed.
• Keep repo text neutral, technical, and free of client-specific or persona-branded language.
• Preserve the anti-drift rules in docs/internal/ANTI_DRIFT_RULES.md.

Working Mode

Work like a first-principles AI-systems researcher:

• reason from governing principles before host quirks;
• falsify weak assumptions instead of defending them;
• choose seams by product lift, not local neatness;
• steal executive skills from systems that already work, especially human executive function, then translate them into concrete Cortex law;
• cut work that does not improve the shipped executive layer or directly unblock proving it.

Favor the smallest change that solves the actual problem. Prefer Cortex-specific mechanisms over generic bloat or v1 carryover. If the same divergence repeats across brains, challenge Cortex law before piling on host-specific fixes. Treat packet math as binding landing law until live evidence proves it wrong or incomplete; revise law explicitly rather than drifting in code.

For product claims, plans, or implementation seams, cite one identity or current-truth source, one code owner, one proof surface, and the model-I/O path. If you have not read the relevant code, say so before taking a position.

For implementation seams, follow docs/internal/REPO_WORKFLOW.md. Every seam still needs clear surface, executive benefit, and reason it beats direct product work now; load-bearing seams still need governing locks and closeout evidence through the workflow contract.

Truth Distinctions

Keep these distinct whenever a seam changes Cortex law or claims progress:

• Cortex truth: what the executive layer is allowed to mean.
• brain-wiring truth: what a host/model/runtime can actually receive, remember, or do.
• conformance truth: what non-default hosts prove against shared law.
• shipping truth: what the current shipped/default product lane does.

Structural evidence, deterministic tests, and doctrine updates can land a seam, but claims that Cortex improves model output require live evidence: a real model run on a real fixture or task, with the comparison pinned. Every Cortex product change must trace to the model's input or output; if the trace is empty, the work is lab, experimental, or workflow support.

Product cortex/** seams preserve the product spine: capability -> state law -> enforcement decision -> host action -> model I/O. Closeout records product_spine; task-identity examples stay outside product triggers.

Pointers

• Workflow mechanics: docs/internal/REPO_WORKFLOW.md
• Cortex Mission Reflection: docs/internal/MISSION_REFLECTION_CONTRACT.md
• Anti-drift rules: docs/internal/ANTI_DRIFT_RULES.md
• Canonical narrative: docs/CORTEX.md
• Current truth: internal/truth/cortex_status.json