Skip to content

Deterministic event-sourced LLM pipeline: the model only normalizes intent, while policy/planner/executor workers enforce allowlisted, auditable execution via JSONL events.

Notifications You must be signed in to change notification settings

amcbstudio/jsonllm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

JSONLLM Kernel

Deterministic event-sourced kernel for modular LLM automation.

LLM ingress -> strict events -> deterministic module policy/planner/executor -> append-only JSONL

What this is

  • A kernel that enforces deterministic invariants (schema, lineage, provenance, idempotency).
  • A module system where third parties implement domain logic.
  • A self-serve workflow to scaffold and validate modules without maintainer support.

Core guarantees

  • LLM only emits IntentNormalized (module_id=kernel.ingress).
  • Side effects run only in deterministic workers.
  • Mandatory lineage: IntentNormalized -> IntentAccepted/Rejected -> ActionProposed -> ActionOutcome.
  • ActionProposed.action_id must be allowlisted.
  • actor authority and module_id provenance are validated.

Repository layout

  • src/event_pipeline.py: CLI orchestrator
  • src/jsonllm_kernel/contracts.py: event/config contracts
  • src/jsonllm_kernel/module_api.py: module SDK (BaseModule, types)
  • src/jsonllm_kernel/module_loader.py: module.toml loader/validator
  • src/jsonllm_kernel/scaffold.py: module scaffolding templates
  • src/jsonllm_kernel/conformance.py: module conformance suite
  • src/jsonllm_kernel/project_init.py: workspace bootstrap from packaged templates
  • src/jsonllm_kernel/templates/: packaged default catalog and modules
  • modules/: example modules
  • catalog/: allowlist + security configs

Install

python3 -m pip install -r requirements.txt
export OPENAI_API_KEY="your_key"

Or install as package/CLI:

python3 -m pip install -e .
jsonllm --help

Start in a brand-new directory

mkdir /tmp/my-jsonllm-workspace
jsonllm init-project /tmp/my-jsonllm-workspace
jsonllm --workspace /tmp/my-jsonllm-workspace list-modules

--workspace (or env JSONLLM_WORKSPACE) controls where catalog/, modules/, and data/ are read/written.

CLI

jsonllm --workspace /path/to/workspace list-modules
jsonllm --workspace /path/to/workspace init-module my_domain
jsonllm --workspace /path/to/workspace test-module --module-id my_domain
jsonllm --workspace /path/to/workspace new-intent --request-text "..." --aggregate-id req-1
jsonllm --workspace /path/to/workspace run-policy
jsonllm --workspace /path/to/workspace run-planner
jsonllm --workspace /path/to/workspace run-executor

Equivalent while developing in this repo:

python3 src/event_pipeline.py --workspace . list-modules

Module self-serve workflow

  1. Scaffold module:
jsonllm --workspace /path/to/workspace init-module my_domain
  1. Implement policy/plan/execute in modules/my_domain/module.py.
  2. Update modules/my_domain/module.toml (intents/actions/permissions).
  3. Ensure actions are in catalog/allowed-actions.json.
  4. Run conformance:
jsonllm --workspace /path/to/workspace test-module --module-id my_domain
  1. Run end-to-end pipeline.

Compatibility and standards

  • Module API version is declared by module_api_version in module.toml.
  • Loader rejects incompatible module API versions.
  • See:
    • docs/MODULE_COMPATIBILITY.md
    • docs/MODULE_COOKBOOK.md

Example modules

  • modules/core_builtin: generic routes for search/extraction/classification.
  • modules/math_sum: tangible deterministic computation (math.sum.v1) writing outputs to data/outputs/.

About

Deterministic event-sourced LLM pipeline: the model only normalizes intent, while policy/planner/executor workers enforce allowlisted, auditable execution via JSONL events.

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages