Skip to content
Codexa

AI agent

Provider abstraction for Anthropic and OpenAI, Build/Ask/Fix modes, tool-based edits, sandbox validation, a two-attempt repair loop and streamed progress.

By Codexa Team1 min read

What ships

  • Provider abstraction with adapters for Anthropic and OpenAI, plus a mock provider for tests. The operator configures the key and default model; without a key the product shows a not configured state and the API returns ai_not_configured.
  • Modes: Build (full tools), Ask (read-only tools), Fix (repair prompt).
  • Tools: list_files, read_file, search_files, write_file, edit_file, delete_file, run_command (allow-listed install/build/test/typecheck), get_logs, propose_plan, finish.
  • Stages streamed over SSE: queued, loading_context, planning, working, validating, repairing, done; events stage, text, tool_call, tool_result, file_change, plan, usage, done, error.
  • Bounds per run: 8,000-character prompt, 40 tool calls, 60 turns, 25 minutes, 16k output tokens per call, 2 repair attempts, 2 provider retries.
  • Snapshot before the first write; review diff, revert and cancel.
  • Conversations per project with the last ten messages as history.
  • Usage recorded against workspace AI request and token quotas, checked before the provider is called.

Documentation: AI agent, Generations API.