Open Model Agent Stack

Status: beta. The guarded workflow passes its offline suite and one bounded live Ollama Cloud canary on Linux aarch64. Clean-environment timing and production-repository behavior are not yet validated.

This public companion asset for The AI Architect lets a builder-operator test one bounded question: can a low-cost orchestrator and executor complete a tiny objective fixture with less subscription dependence?

It does not cancel a subscription, configure a production repository, expose credentials, or prove that one fixture generalizes to professional work.

Implemented beta path

Role	Route	Model
Orchestrator	Ollama Cloud through the user's local Ollama daemon	kimi-k3:cloud
Executor	Ollama Cloud through the user's local Ollama daemon	deepseek-v4-flash:cloud
Authentication	User-owned Ollama CLI flow	ollama signin
Local GPU	Not required	—

The model IDs and Ollama Pro price were checked against official Ollama pages on 2026-08-22. Catalog, limits, and pricing can change; see docs/evidence.md. The redacted live canary receipt is preserved at evidence/live-canary-receipt.json. The separate Neon Velocity receipt and its system-comparison boundary live in docs/neon-velocity.md.

What the workflow does

doctor → plan → approval → apply → verify → receipt → optional rollback

• doctor inspects Python, OS/architecture, the Ollama CLI, and the loopback API without writing anything.
• plan writes one proposed canary configuration and hash-bound diff under ignored .generated/ state.
• apply requires the printed plan SHA-256 and creates only this plan's internal canary file; it never reads or replaces user configuration.
• verify requires the same SHA-256 as a separate two-request cloud-usage approval. It asks Kimi for a bounded plan and DeepSeek for structured outputs, then checks exact expected results without executing model-generated code.
• receipt excludes credentials, account identifiers, and raw model responses and identifies itself as locally mutable evidence.
• rollback removes the exact unchanged internal canary while retaining its evidence files.

Quick start

Prerequisites: Python 3.11 or newer, Ollama, and an Ollama account with enough cloud usage for two small requests.

python3 -m open_model_agent_stack doctor

python3 -m open_model_agent_stack plan

Inspect the printed proposed.diff, plan.json, and approval SHA-256. Apply creates only .generated/PLAN_ID/applied/stack-config.json; it does not touch a harness or user configuration:

python3 -m open_model_agent_stack apply \
  --plan .generated/PLAN_ID/plan.json \
  --approve PLAN_SHA256

Authentication remains manual and user-owned:

ollama signin

verify makes two Ollama Cloud requests. Included limits apply, and extra usage may cost more. Run it only after inspecting the plan and accepting that boundary:

python3 -m open_model_agent_stack verify \
  --plan .generated/PLAN_ID/plan.json \
  --approve-usage PLAN_SHA256

python3 -m open_model_agent_stack receipt \
  --plan .generated/PLAN_ID/plan.json

Rollback remains available while the internal canary is unchanged:

python3 -m open_model_agent_stack rollback \
  --plan .generated/PLAN_ID/plan.json \
  --approve PLAN_SHA256

Every command supports stable --json output. A failed, uncertain, or rolled-back attempt cannot reuse its cloud-usage approval; start over with python3 -m open_model_agent_stack plan --new-attempt and inspect the new plan and SHA-256.

Objective fixture

The supplied normalize-route-v1 fixture is deliberately small. The executor returns exactly three structured input/output rows. The verifier rejects unknown, missing, duplicate, extra, or oversized values and compares every output with deterministic expected data. No model-generated code is executed, and model self-report is ignored.

This fixture answers whether the transport, role split, structured-output contract, and acceptance loop work. It does not prove production-code quality.

Not yet supported

• automatic Ollama installation or sign-in;
• any subscription or billing mutation;
• Windows;
• external harness or user-configuration mutation;
• production-repository mutation;
• local Qwen installation or runtime setup;
• portable performance claims beyond Tyler's separately documented DGX Spark evidence;
• the intended 30-minute clean-environment first win.

Read SETUP.md, AGENTS.md, and stack.json before extending or live-testing the workflow.