Skip to content

Compatibility

Runtime requirements

  • Python 3.10 through 3.14.
  • A writable AdaMAST home directory, normally ~/.adamast.
  • Windows and Linux are exercised by CI. Other Python-supported platforms may work but are not release-gated yet.

Codex

  • The conversation host must support Codex hooks and allow the installed hooks.
  • Project and user hook files are supported.
  • Native taxonomy learning uses a subagent in the active Codex task. It does not require a standalone codex executable or separate login.
  • Run adamast-doctor --codex after installation to verify the hooks and native learning configuration.

Claude Code

  • The installed Claude Code build must expose the hook event and blocking contracts checked by adamast-doctor --claude-code.
  • Native taxonomy learning uses one Agent subtask in the active Claude Code session. It does not require a runnable claude -p surface or second login.
  • The browser selector requires localhost access; use --selector-surface inline when that surface is unavailable.

Credentials and usage

Native interactive workers do not require OPENAI_API_KEY, ANTHROPIC_API_KEY, or a second external provider account. Codex and Claude Code use the active task session; each host can still consume its normal included or billed usage.

Provider-backed project installs and direct runtime integrations continue to support OpenAI-compatible, Anthropic, Gemini, and AWS Bedrock credentials. Credential values are read from the environment and are never written to the AdaMAST config.

Current limitations

  • Hooks cannot inject a completion notice into an idle conversation. The notice is delivered exactly once at the next host lifecycle event.
  • Codex uses a compact single-pass Stop checkpoint because a continued desktop turn is not guaranteed to invoke Stop again.
  • Taxonomy rollback and task-group selection are configuration/runtime controls; there is no graphical management surface yet.
  • Automatic redaction is enabled by default, but traces can still contain sensitive task content. Do not place secrets in prompts or tool output.