Skip to content

architecture.md

Latest commit

 

History

History
103 lines (74 loc) · 3.49 KB

architecture.md

File metadata and controls

103 lines (74 loc) · 3.49 KB
Raw

AutoCode Architecture

AutoCode is organized around a local agent runtime that turns a user task into bounded repository context, model calls, tool execution, verification, and a persisted session record.

Main Flow

User task
  -> CLI runtime
  -> repository summary and retrieval
  -> task-context assembly
  -> model client
  -> protocol parser
  -> tool permission gate
  -> tool execution
  -> verification/report/session persistence

Runtime Layer

The runtime layer owns the command-line interface, terminal UI, permissions, task execution, session persistence, reports, and verification state. The main entry point is src/index.ts.

Important runtime areas:

  • runtime/cliTaskExecutor.ts: headless task execution wiring.
  • runtime/tui/: interactive terminal UI.
  • runtime/permissions.ts: approval-mode enforcement.
  • runtime/sessionStore.ts: persisted run snapshots.
  • runtime/taskRunReport.ts: text and Markdown report rendering.
  • runtime/taskRunTypes.ts: task-run mode and result contracts.

Agent Core

The core layer owns prompt construction, the model-response protocol, parsing, recovery, and tool-call orchestration. It is intentionally explicit so that tool use and verification can be audited from session records.

Important core areas:

  • core/agent.ts: main agent loop and tool protocol handling.
  • core/protocol.ts: structured response parsing and turn schema handling.
  • core/structuredRetryTarget.ts: recovery targeting for malformed or failed responses.
  • core/evidenceShapeMatrix.ts: evidence and recovery behavior checks.

Repository Intelligence

The repository layer builds a lightweight model of the target workspace: files, projects, dependency edges, symbols, symbol references, and ignore rules. This repository model feeds retrieval, context assembly, and impact tools.

Important repository areas:

  • repo/repoSummaryBuilder.ts: workspace summary construction.
  • repo/repositoryProjectDetection.ts: project-boundary detection.
  • repo/repositorySymbolGraph.ts: symbol graph extraction.
  • repo/repositorySymbolReferences.ts: reference discovery.
  • repo/workspaceIgnore.ts: ignored-path handling.

Retrieval And Context

The retrieval layer ranks files and neighbors for a user query. The context layer turns that retrieval output into compact task context, optionally with focused compaction settings for large repositories.

Important areas:

  • retrieval/: query terms, scoring, ranking, and graph-neighbor retrieval.
  • context/: task context, memory loading, compaction, and prompt assembly.

Model Adapters

Model access is isolated behind a small ModelClient interface. The default runtime supports:

  • mock: deterministic local test provider.
  • openai-compatible: configurable HTTP provider for compatible endpoints.

Configuration is resolved from shell environment variables and ignored local env files. Public examples should never contain real credentials.

Tools

Built-in tools cover repository inspection, file reads, search, patching, shell verification, dependency paths, symbol graphs, symbol references, and impact-oriented queries. Permission checks happen before tools that can write files or run commands.

Experimental Agent Modes

AutoCode supports multiple task modes:

  • single: one agent loop handles the task.
  • multi: sequential role-style orchestration.
  • graph: fixed-stage graph runtime scaffold.

The multi-stage modes are useful for structured experiments, but the stable baseline for ordinary local use is still single.