scriptorium/docs/policy/architecture.md

Architecture

This document is the development architecture policy for Scriptorium.

It is for developers and LLM coding agents. User-facing behavior belongs in README.md and the docs under docs/ that target operators/users.

Project Shape

Scriptorium is a narrow prompt-execution application with three entry paths:

• CLI run
• CLI render
• HTTP POST /v1/runs through serve

Domain behavior is centralized in internal/usecase and internal/domain.

Core Principles

• Keep orchestration narrow: Scriptorium executes one prompt request; it is not a multi-step workflow engine.
• Keep adapter logic thin: adapters map external shapes to domain requests/results and should not hold domain decisions.
• Keep boundaries explicit: repositories/loaders/renderers/validators/LLM client stay behind package interfaces.
• Keep config strict: YAML/JSON decoding for external inputs should reject unknown fields.
• Keep secrets out of payloads: raw API key values must not be accepted or emitted.

Package Boundaries

Current package map:

• cmd/scriptorium: process entrypoint.
• internal/adapter/cli: command parsing, app wiring for CLI commands, output behavior.
• internal/adapter/http: HTTP DTO mapping and error/status mapping.
• internal/config: application settings loading and CLI override precedence.
• internal/defaults: compile-time default constants.
• internal/domain: core request/result and contract types.
• internal/usecase: Runner prepare/run orchestration and repair-hook boundary.
• internal/promptdef: filesystem prompt-definition repository.
• internal/profile: filesystem execution-profile repository.
• internal/artifact: artifact reference readers.
• internal/prompt: template renderer.
• internal/llm: provider-neutral LLM client interface and OpenAI-compatible implementation.
• internal/validate: validator interfaces and standard implementation.
• internal/format: prepared-run output formatting.

Detailed component behavior is documented in:

• docs/internal/runner.md
• docs/internal/adapters.md

Configuration And Precedence

Application settings are resolved as:

1. built-in defaults
2. config file values
3. CLI overrides

config.yml is for application wiring (directories, server address, render default format), not prompt/profile runtime execution settings.

Profile selection and runtime model resolution remain use-case concerns.

State And Persistence Policy

Scriptorium has no durable run-state store.

• No built-in resume/checkpoint/archive behavior.
• Recovery model is rerun after correcting inputs/config/environment.

External Integration Policy

Current external contracts:

• inbound HTTP contract: POST /v1/runs
• outbound model contract: OpenAI-compatible chat completions subset
• subprocess contract for integrators: CLI run/render

Integration docs belong under docs/integrations/.

Error Handling And Logging

• Wrap errors with domain/operation context.
• Map domain errors to adapter-appropriate statuses/codes without leaking sensitive internals.
• Keep stderr summaries concise for CLI success/error paths.
• Never emit raw secret values.

Testing Expectations

• Core runner behavior should be covered with isolated unit tests and fixture-based integration tests.
• Adapter behavior should be tested for parse/mapping/error semantics.
• Config parsing, prompt/profile loading, validator behavior, and LLM client error handling should remain covered by package tests.
• Repository-level docs/examples that claim runnable behavior should be validated by tests or smoke commands.

Documentation Expectations

• Document implemented behavior only outside docs/roadmap/.
• Keep canonical reference locations stable (docs/cli.md, docs/config.md, docs/operations.md, docs/troubleshooting.md, docs/internal/).
• Update docs in the same change when architecture-relevant behavior changes.

Architectural Invariants

• Runner.Run reuses Runner.Prepare flow.
• CLI and HTTP currently instantiate Runner without a repairer.
• Artifact reading supports inline and file references.
• Unknown input fields in config/prompt/profile/http JSON should be rejected by strict decoding.
• Raw API key values must not be accepted through config/HTTP payloads.

Non-Goals

• Do not move orchestration responsibilities from external callers into Scriptorium.
• Do not add adapter-specific business logic in internal/adapter/* packages.
• Do not bypass repository/renderer/validator/LLM boundaries by introducing cross-package coupling.

Work that is not implemented belongs in docs/roadmap/.