Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
55 lines (39 loc) · 2.53 KB

ARCHITECTURE.md

File metadata and controls

55 lines (39 loc) · 2.53 KB

Azycode Architecture

Runtime

Azycode is a dependency-free Node.js CLI. The main process owns:

  • command routing in src/cli.js
  • provider presets in src/providers.js
  • persisted configuration and state in src/config.js
  • OpenAI-compatible model calls in src/llm.js
  • coding harness loop in src/agent.js
  • filesystem/shell/git tools in src/tools.js
  • mission execution in src/missions.js
  • custom subagent registry in src/subagents.js
  • repository snapshots in src/context.js
  • long-lived local notes in src/memory.js
  • local git-diff review heuristics in src/local-review.js
  • terminal UI primitives in src/ui.js
  • interactive TUI in src/tui.js
  • centralized logging in src/logger.js

It is intentionally not a wrapper around any existing local azy-code binary or repository. The CLI entrypoint is bin/azycode.js, and all runtime behavior lives under this repository's src/ directory.

Config

Default config directory is ~/.azycode; set AZYCODE_HOME to isolate a project, test run, or CI environment.

The CLI stores keys with 0600 file permissions. It does not hardcode user keys and can also read provider-specific environment variables.

Config, state, and todos are loaded with an in-memory cache keyed on file mtime to reduce repeated disk reads during a single process. Config values are validated on load: unknown modes, reasoning levels, permission profiles, and tool policies are normalized to safe defaults.

Agent Loop

The loop sends:

  1. mode-specific system prompt
  2. user task
  3. OpenAI tool schemas
  4. tool results until the model returns a final assistant message

Risky tools ask for approval unless alwaysApprove is enabled or policy is changed. Git guard is separate from approval and still blocks protected branches unless disabled.

Plan and review artifacts can be saved with --save; sessions are persisted to state and can be exported for debugging or audit trails. Agent runs emit progress events, persist recent tool execution records, and expose formatted transcripts for review. Optional context packs gather bounded, ignore-aware source excerpts into the system prompt with --context. The chat command provides an interactive loop with slash commands for mode, reasoning, context, progress, local review, and status.

Provider Strategy

The model client supports:

  • OpenAI Chat Completions for OpenAI-compatible providers.
  • Anthropic Messages normalization for OpenCode Go models that use /messages.
  • Model-based routing for OpenCode Go, because its documented endpoints differ by model family.