LLM workflows

LLM workflow runner for pipeline automation.

Project structure

src/
  index.ts        bin shim
  cli/            argv parsing and one module per command
  workflows/      one folder per capability (vertical slices) plus the registry
    review/       the code review workflow (agents, prompt, tools, report)
    ...
  engine/         workflow-agnostic agent engine (guardrails, config, prompt, tools)
  runtime/        LLM runtime port plus the pi adapter
  integrations/   external systems (ports plus adapters)
  shared/         cross-cutting infrastructure (env, logger)

Architecture

flowchart LR
  CLI[CLI] --> Registry[Workflow registry]
  Registry --> Workflow[Selected workflow]
  Context[External integrations<br/>for context]
  Workflow --> Context
  Workflow --> Engine[Agent engine]
  Engine --> Sessions[Agent sessions]
  Sessions --> Runtime[Runtime interface]
  Runtime --> Provider[Model provider]
  Workflow --> Result[Output, errors,<br/>and telemetry]
  Result --> CLI

  subgraph EngineDetails[Agent engine]
    Sessions
    Concurrency[Session concurrency]
    Tools[Engine tools]
    Guardrails[Guardrails]
    Sessions --- Concurrency
    Sessions --- Tools
    Sessions --- Guardrails
  end

Loading

Environment variables

Application-wide variables are discovered from cleanEnv schemas. Workflow-specific variables are documented with their workflow. Per-agent overrides use the uppercased agent id as the prefix.

Environment variable	Type	Default	Choices	Description
<AGENT>_*	Varies	Varies	Varies	Per-agent overrides for ENABLED, MODEL_NAME, MODEL_SAMPLING_PARAMS, CONTEXT_WINDOW, MAX_OUTPUT_TOKENS, TIMEOUT_MS, INPUT_TOKEN_BUDGET, OUTPUT_TOKEN_BUDGET, TOOL_LOOP_THRESHOLD, TOOL_FAILURE_THRESHOLD, SOFT_LIMIT_RATIO. Defaults come from the corresponding DEFAULT_* variables.
LOG_LEVEL	str	INFO	DEBUG, INFO, ERROR	The log level for the application.
LOG_DIR	str	/tmp/llm-workflows	—	The directory for log files. Holds combined.log and one file per agent. An empty value disables file logging.
MODEL_API	str	openai-completions	openai-completions	The model API to use for the agent.
MODEL_PROVIDER	str	custom	custom, openrouter	The model provider to use for the agent.
MODEL_BASE_URL	str	—	—	The base URL for the model API.
MODEL_API_KEY	str	—	—	The API key for the model API.
REPO_DIR	str	—	—	The path to the repository directory.
DEFAULT_MODEL_NAME	str	—	—	The default model name for agents.
DEFAULT_MODEL_SAMPLING_PARAMS	json	{}	—	The default model sampling parameters for agents.
DEFAULT_CONTEXT_WINDOW	num	131072	—	The default model context window for agents.
DEFAULT_MAX_OUTPUT_TOKENS	num	32768	—	The default maximum model output tokens for agents.
DEFAULT_TIMEOUT_MS	num	300000	—	The default wall clock timeout in milliseconds for an agent session. Zero disables the timeout.
DEFAULT_INPUT_TOKEN_BUDGET	num	32768	—	The default input token budget for an agent session. Zero disables the budget.
DEFAULT_OUTPUT_TOKEN_BUDGET	num	32768	—	The default output token budget for an agent session. Zero disables the budget.
DEFAULT_TOOL_LOOP_THRESHOLD	num	5	—	The default number of consecutive identical tool calls before an agent is terminated. Zero disables the limit.
DEFAULT_TOOL_FAILURE_THRESHOLD	num	5	—	The default number of consecutive failures of the same tool before an agent is terminated. Zero disables the limit.
DEFAULT_SOFT_LIMIT_RATIO	num	0.8	—	The default ratio of a hard guardrail limit at which a soft warning is issued.
AGENT_MAX_CONCURRENCY	num	5	—	The maximum number of agent sessions that may run concurrently. Zero disables the limit.