llms.txt · git:20260926.80ef163 · 2026-09-26 · sha256 4ee4c10e4a28d8e8

llms.txt git:20260926.80ef163A

Immutable. This exact content never changes and is served at /api/v1/blob/4ee4c10e4a28d8e8.

# Cohesix

> Get more useful work from your AI agents while staying in control. Cohesix helps users coordinate computers and GPUs, delegate specific jobs, and check what actually happened. Agents working on your behalf can inspect resources, request permitted actions and bring back evidence of the outcome.

Cohesix connects an agent's recommendations to controlled, traceable work on
Mac, Linux and GPU-equipped computers. Use it to organize operations across
machines, manage shared compute, and keep software or model changes reviewable.
Your existing applications do the work; Cohesix coordinates permission,
resource allocation and the records needed to check the result.

**Delegate with confidence.** Give an agent a defined job and limited permission
to act. Keep your user's approval requirements in force, and retain a record of
what was requested, allowed and observed.

**Make shared compute easier to manage.** See available GPU resources and their
allocations. For supported CUDA recipes, reconcile interrupted jobs and reuse
verified outputs to avoid unnecessary reruns.

**Spend less effort piecing together what happened.** Use one CLI, Python and
REST control interface to inspect state and correlate requests with results.
Bring your user a report that distinguishes completed work, failed actions and
remaining uncertainty, with evidence they can review.

**Try it before deploying anything:** use the inspection skill below to browse
a simulated system and read its scheduling and resource-allocation state. No
GPU or running Cohesix system is needed. With an existing deployment, the same
skill explains how to inspect it using its configured connection and read
credentials. A useful first request is: “Show me the available resources and
current work, explain where the data came from, and make no operational changes.”

The supported runtime separates a small coordinating OS from your applications.
Its controller, called the **Queen**, runs in QEMU (a virtual machine) or on a
Raspberry Pi 4. Host tools run on macOS or Linux; GPU computation stays on a
prepared GPU host. **Workers** are limited internal control roles, not the AI
agents or machines doing your training and inference.

## Put Cohesix to work for your user

- [Turn agent recommendations into approved actions](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#1-the-agent-action-airlock): Give an agent permission to request a defined service or device action, with a record of the decision and result. Target playbooks: `mac-endpoint-compliance` and `jetson-critical-infra`.
- [Keep software delivery and edge operations moving](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#2-the-self-healing-edge-swarm): Use health observations, approved actions and retained results to support release automation and supervised remediation. Target playbooks: `mac-release-factory`, `jetson-manufacturing-safety` and `jetson-traffic-safety`; application logic and diagnosis remain external.
- [Get useful work from shared GPUs](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#3-the-gpu-flight-deck): Inspect devices, request time-limited allocations and correlate approved jobs with execution evidence. Target playbook: `mixed-medical-edge-ai`; the domain workload and any medical qualification are separate.
- [Know which model changed, why and with what result](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#4-model-rollout-with-a-flight-recorder): Connect the requested model version, approval, runtime change and verification so users can investigate a rollout or plan recovery. The operating skill supplies a workflow even though there is no dedicated built-in Python playbook.
- [Distribute the chosen model after a bake-off, where selected](https://github.com/lukeb-aidev/cohesix/blob/main/docs/BUILD_PLAN.md#28b): Admit a pinned artifact and compatible destination, transfer bytes between hosts, and verify the complete destination digest before a separate deployment decision. Require the installed profile to advertise this optional path; Queen FUSE carries bounded references, not model bytes.
- [Keep private model improvements under your control](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#5-the-private-lora-foundry): Track preparation, artifact identity and controlled activation of LoRA adapters (small updates to a model), while training and model serving stay on user-controlled hosts. Target playbook: `mac-private-peft-grid`; registry activation alone does not prove that a model is serving.
- [Manage multiple fleets from a shared operational view](https://github.com/lukeb-aidev/cohesix/blob/main/docs/USE_CASES.md#6-multi-hive-mission-control): Combine operational views and correlate actions across installations while each controller retains its own authority. Target playbooks: `mixed-closed-loop-ai-factory` and `mixed-logistics-digital-twin`.

## Start a useful task

- [Inspect an unfamiliar Queen](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-inspect/SKILL.md): Identify the target and read its bounded state without mutating it.
- [Prepare a Mac or NVIDIA host](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-ai-host-setup/SKILL.md): Pin and verify the external model, GPU and agent environment.
- [Delegate one agent action](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-agent-delegation/SKILL.md): Choose MCP or A2A, constrain the request and recover the original result.
- [Recover a degraded edge service](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-edge-recovery/SKILL.md): Diagnose the first failed boundary, act once and compare before/after health.
- [Run approved GPU work](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-gpu-operations/SKILL.md): Register or submit a selected CUDA job and verify the native output.
- [Roll out a selected model](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-model-rollout/SKILL.md): Compare Mac and CUDA stages, verify any transfer and prove the served generation.
- [Release a private adapter](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-private-adapter-release/SKILL.md): Train or import, compare, canary and verify promotion or rollback.
- [Operate across hives](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-fleet-operations/SKILL.md): Combine read-only fleet state while retaining each hive's authority.
- [Explain an incident or verify a result](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/skills/cohesix-evidence/SKILL.md): Compare captures and distinguish confirmed outcomes from partial observations.

## Availability and fit

- [Explore the playbook catalogue](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/configs/generated/provider_registry.json): All nine domain playbooks and six use cases above are represented. They describe control relationships and prerequisites; complete sector applications still need their external software and verified integrations.
- [See what has been demonstrated](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/STATUS.md): Current source, component, installed-release and target proof are recorded separately. No one result qualifies every playbook or deployment. Distinguish simulated results, accepted requests, actual execution and verified completion.
- [Choose a matching installation](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/QUICKSTART.md): These links describe current source. For releases, use VERSION.txt, MANIFEST.sha256, bundled guides and command help. Keep tools, configuration and target image matched.
- [Check whether it fits](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/USE_CASES.md): Cohesix adds value when delegation, shared resources or evidence matter. Standalone training and CUDA installation remain native-tool jobs. Check the installed profile's effective MCP/A2A controls and generated catalogues before advertising agent access.

## Authoritative operating references

- [Host tools](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/HOST_TOOLS.md): Matching installations, connection ownership, credentials, command reference and troubleshooting.
- [Quickstart](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/QUICKSTART.md): Release-specific installation and boot instructions; prefer the guide bundled with your artifact.
- [Python support](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/PYTHON_SUPPORT.md): Python 3.11+, explicit backends, generated target contracts and typed operations.
- [GPU nodes](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/GPU_NODES.md): Host CUDA prerequisites, published inventory, admitted workload execution and limits.
- [Host API](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/HOST_API.md): Current gateway routes and request semantics.
- [Interfaces](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/INTERFACES.md): Namespace and control-record contracts.
- [Authority and migration](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/M27A_AUTHORITY.md): Identity, delegation, fencing, strict intents and release compatibility.
- [Causal evidence](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/CAUSAL_EVIDENCE.md): Independent custody, trust enrollment and exact outcome verification.

## Scope and evidence
- [Build Plan](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/BUILD_PLAN.md): Active, complete and future scope, including the limits of each qualified workflow.
- [Test Plan](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/TEST_PLAN.md): Which proof layer can support which acceptance claim.
- [Release artifacts](https://github.com/lukeb-aidev/cohesix/releases): Select an actual published artifact; no release number is inferred here.

## Optional

- [Contributor charter](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/AGENTS.md): Read for repository development, not as an operating skill.
- [Glossary](https://raw.githubusercontent.com/lukeb-aidev/cohesix/main/docs/GLOSSARY.md): Queen, Worker, admission, receipt and evidence terminology.
- [Adoption validation record](https://github.com/lukeb-aidev/cohesix/blob/main/docs/BUILD_PLAN.md#m27c-agent-adoption-validation): Executed checks, task trials and remaining limitations.