Zero-State-LLC · scrimshawlife-ctrl · Sep 4, 2026 · Sep 4, 2026
@@ -0,0 +1,77 @@
+# Agent contract — Hyperlex
+
+You are working in **Hyperlex** (`Zero-State-LLC/Hyperlex`), a standalone
+Hermes skill. This file is the canonical contract for Cursor, Claude,
+Codex, Grok, and OpenClaw. Harness adapters (`CLAUDE.md`) only point here.
+
+`SKILL.md` is the Hermes runtime contract. Do not rewrite it unless an
+accepted intent is explicitly the skill-contract change. Do not copy
+Noema deploy-worker or pin workflows into this repo.
+
+## Intent
+
+Read the active `intent.md` under [`intent/`](intent/README.md) before
+editing. Next stage is `spec.md` / plan — do not skip to code.
+
+If the change is non-trivial and no intent exists, draft one from
+[`intent/_TEMPLATE.md`](intent/_TEMPLATE.md) and wait for a human to
+accept it. Trivial docs and chore fixes do not need an intent file.
+
+## Commands
+
+| Task | Command |
+|---|---|
+| Offline demo | `python3 scripts/hyperlex.py demo` |
+| Health | `python3 scripts/hyperlex.py check` then `doctor` |
+| Smoke | `python3 scripts/hyperlex.py smoke` |
+| Tests | `PYTHONPATH=src pytest -q` |
+| Install dry-run | `bash install.sh --dry-run` |
+
+Do not invent commands. Prefer `${HERMES_SKILL_DIR}` in Hermes `terminal`
+calls. Baseline work uses `--route offline` / `mock`.
+
+## Invariants
+
+- Brier only after operator settlement. Open analysis keeps `brier: null`.
+- Phase 5 / simulate / phylogeny stay **SPECULATIVE**.
+- Fail closed on missing outcomes. Label `OBSERVED` / `INFERRED` /
+  `SPECULATIVE` / `NOT_COMPUTABLE`.
+- Do not invent numeric Brier, auto-settle, auto-register Hermes cron, or
+  rewrite historical receipt integrity.
+- Hyperlex never imports Abraxas. Cron (`risk-schedule`) is advisory only.
+- Do not invent CLI behavior `SKILL.md` does not name.
+- Do not put secrets, tokens, or live credentials in the tree.
+- Do not dispatch leftover Noema-shaped deploy or pin workflows.
+
+## Escalation
+
+If CI, tests, or validators look wrong — missing coverage, silent skips,
+green-but-inert checks, or a suite that contradicts the skill contract:
+
+1. Open a bounded defect issue labeled `bug`.
+2. Do **not** patch tests, fixtures, or CI to force green.
+3. Do **not** weaken an assertion or rewrite `SKILL.md` to match a
+   broken path.
+
+A red honest check is better than a green lie.
+
+Owner-gated (human yes): operator settlement, live ingest, cron
+registration, official Hermes hub submit, license/legal, and spend.
+Agents do not buy things or publish the skill as official.
+
+## Team context
+
+Org and partner status is not this repo. When a task needs it, load
+[Zero-State-LLC/agent-context](https://github.com/Zero-State-LLC/agent-context):
+
+1. `HANDOFF.md`
+2. `STATUS.md`
+3. `DECISIONS.md` only if a prior choice affects this task
+
+Clone: `gh repo clone Zero-State-LLC/agent-context ~/agent-context`.
+Pull `--ff-only` before trusting a local copy.
+
+## Skills
+
+Name only skills that already exist. Do not invent skills or bots.
+The Hermes skill in this repo is `hyperlex` (`SKILL.md`).
@@ -0,0 +1,2 @@
+Follow [AGENTS.md](AGENTS.md). That file is the only contributor and agent instruction source in this repository.
+For a non-trivial change, read the active intent.md before editing. If none exists, draft intent.md and wait for human accept.
@@ -5,6 +5,9 @@ This repo is the single source of truth for **design and specification**.
 
 Implementation details belong in the engine package. Changes here should drive implementation.
 
+Coding agents follow [`AGENTS.md`](AGENTS.md) (`CLAUDE.md` is a thin pointer
+there). Non-trivial work starts with [`intent/`](intent/README.md).
+
 ## How to Contribute
 1. Open an issue or discussion first for significant changes.
 2. Update relevant spec/roadmap file.

@@ -347,14 +347,16 @@ from hyperlex.compat.abraxas import to_brier_ledger_entry, list_hlx_runes
 | [docs/phase5.md](./docs/phase5.md) | Research simulation |
 | [docs/slang-lineages.md](./docs/slang-lineages.md) | Lineage methodology |
 | [docs/hermes-skill.md](./docs/hermes-skill.md) | Skill model |
+| [AGENTS.md](./AGENTS.md) | Multi-agent SDLC contract (`CLAUDE.md` is a thin pointer here) |
+| [intent/](./intent/README.md) | Intent template for non-trivial work; then spec.md / plan |
 
 Examples: `examples/cron/` · `examples/receipts/golden/` · `examples/calibration/` · `examples/slang-families/` · `examples/case-studies/`
 
 ---
 
 ## Contributing
 
-See [CONTRIBUTING.md](./CONTRIBUTING.md). Keep the settlement rule and offline-first defaults intact. Prefer small, test-backed PRs.
+See [CONTRIBUTING.md](./CONTRIBUTING.md). Agent contract: [`AGENTS.md`](AGENTS.md) (`CLAUDE.md` is a thin pointer there). Non-trivial work starts with [`intent/`](intent/README.md). Keep the settlement rule and offline-first defaults intact. Prefer small, test-backed PRs.
 
 ```bash
 PYTHONPATH=src pytest -q

@@ -25,6 +25,8 @@ Hyperlex is a **Hermes skill** (Python package repo). Relevant Abraxas wire shap
 - [ARCHITECTURE.md](architecture.md)
 - [DESIGN.md](design.md)
 - [SPEC.md](spec.md)
+- [AGENTS.md](../AGENTS.md) — multi-agent SDLC contract (`CLAUDE.md` is a thin pointer there)
+- [intent/](../intent/README.md) — intent template for non-trivial work; then spec.md / plan
 
 ## Examples
 - `examples/slang-families/` — Mermaid family trees

@@ -0,0 +1,23 @@
+# Intent files
+
+Non-trivial work starts here. Copy [`_TEMPLATE.md`](_TEMPLATE.md) to a dated,
+scoped file (for example `intent/2026-09-04-short-slug.md`) and fill every
+section before writing a spec, plan, or code.
+
+Do not invent a parallel process. After the intent is written, continue with
+the flow this repo already uses:
+
+1. `intent.md` — problem, proposed outcome, constraints, verified vs assumed
+2. `spec.md` / plan — existing `docs/` specs, `references/` contracts, and
+   `CONTRIBUTING.md` gates. Revise `SKILL.md` only when the accepted intent
+   is the Hermes contract itself.
+3. Implementation in the allowed trees (skill payload, package, CLI, docs)
+4. PR with honest `OBSERVED` / `INFERRED` / `SPECULATIVE` / `NOT_COMPUTABLE`
+   evidence
+
+Trivial docs, chore, and single-line fixes do not need an intent file.
+`_TEMPLATE.md` is the blank form; do not fill it in place.
+
+An intent is not a spec, not a plan, and not implementation authority. It
+does not authorize a `SKILL.md` rewrite, official Hermes hub publish,
+operator settlement, extra bots, or leftover Noema deploy/pin jobs.
@@ -0,0 +1,39 @@
+# Intent
+
+## Problem
+
+What is wrong, missing, or blocked?
+
+## Proposed outcome
+
+What does done look like? What is explicitly out of scope?
+
+## Affected users / systems
+
+Who or what is affected (Hermes operators, CLI, receipts, score log,
+lineage, Phase 5 research, Pages archive, CI, …)?
+
+## Constraints / non-goals
+
+Hard limits: `SKILL.md` contract, settled-Brier-only, Phase 5 stays
+SPECULATIVE, no Abraxas import, advisory cron only, no invented bots,
+owner-gated settlement / live ingest / hub publish / legal / spend.
+What this work will not decide.
+
+## Open questions
+
+What must be answered before or during the work? Who decides?
+
+## Verified / assumed claims
+
+- **Verified:** facts checked in this repo or on a green run.
+- **Assumed:** beliefs not yet checked. Do not upgrade Assumed to Verified.
+
+| Claim | Label | Basis |
+|---|---|---|
+| \<fact or assumption\> | `OBSERVED` / `INFERRED` / `NOT_COMPUTABLE` | \<source or why it is not computable\> |
+
+## Author / date
+
+- Author:
+- Date:
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		Follow [AGENTS.md](AGENTS.md). That file is the only contributor and agent instruction source in this repository.
		For a non-trivial change, read the active intent.md before editing. If none exists, draft intent.md and wait for human accept.