Expand description
Pull validation and explicit review execution.
Modules§
- agent
- Scoped, sequential Agent reviews with budgets and one tools-disabled verdict repair, and follow-ups that continue a saved review’s conversation.
- auth
- Protected local credential storage, Codex sign-in, and the remote review store’s configuration and token.
- broker
- Run lifecycle, scheduling, cancellation, and budget counters.
- cache
- Fingerprints, reuse keys, execution ownership, reuse, and end-of-review rechecks.
- changes
- Best-effort, state-scoped invalidation hints. SQLite and session files remain authoritative. The first subscriber temporarily hosts a hub; there is no daemon and writers never elect one.
- config
- Inert discovery and strict declarations.
- diagnostics
- Backend readiness and tool checks.
- graph
- Dependency edges, strongly connected components, gates, and obligations.
- graph_
structure - Name-based project structure shared by the graph fingerprint and both graph tools. Paths locate declarations, never identify graph entities. Extraction is flat, so Artifact cycles need no recursive traversal and the graph Artifact includes itself without recursion.
- human
- Human waiting, claim locks, tools, and submission.
- inspect
- Shared non-interactive JSON projections. Interfaces add rendering and paging only.
- limits
- Machine-wide limits in
$STATE/limits.json:backendscaps how many Agent reviews of a backend are in flight at once across everyverifythat shares the state directory, andagentSessionssets whether Agent conversations are saved and how large their store grows. - llm
- Exact-model rig backends with explicit API-key or Codex credentials, and the loopback-only test endpoint override for offline fake providers.
- monitor
- Repository/worktree -> Run -> artifacts/evals -> Detail: the model every interface shows.
- platform
- Operating-system specifics behind one interface: owner-only files and directories, scoped opens and listings that never follow a link, process trees held until admitted, stop signals, hidden terminal input, and host details.
- process
- Literal process launches, bounded capture, and process-group cleanup.
- project
- Project selection, verification orchestration, and status.
- query
- Read-only state queries and result projections.
- remote
- Shared remote review store records: what leaves this machine and how it is mirrored back.
- repository
- Git display identity, separate from the artifactize workspace used for execution.
- review
- Human review: the facts and lifecycle jobs every interface shares. The terminal screens
live in
artifactize-tui. - runs
- Artifactize-owned verification: durable submission, observation and explicit cancellation.
- runtime
- Runtime verdicts come only from ordinary command exit codes. Commands are trusted, not sandboxed; detached descendants can escape group cleanup.
- scope
- Artifact references and canonical scoped paths.
- server
artifactize server: the shared remote review store over its own SQLite file.- store
- The state database and storage maintenance.
- task
- Joining spawned tasks. A task that panicked has a bug, so its panic continues in the task that joins it instead of becoming an ordinary failure; a cancelled task stays an error.
- tools
- Scoped audience-specific tools and language-neutral command protocols.
- types
- Validated identities and persisted states. Their wire representations stay strings.
- workspace
- Workspace paths and external state/output safety checks.