Skip to main content

Release notes

ActionRail is in public beta. The SDK and Console are released together and must use the same complete version until the SDK-to-Console protocol becomes a separately versioned public contract.

For every upgrade, read the repository changelog, verify the compatibility policy, and follow the upgrade and rollback procedure.

Unreleased​

No changes yet.

0.1.0b9 — operator confidence​

Released July 21, 2026.

  • Adds an opt-in live-model poisoning quickstart while retaining the fully deterministic replay for CI and first runs.
  • Separates manual Python and LangGraph onboarding, teaches Actions in the empty state, and makes agent identifiers directly copyable.
  • Clarifies neutral metrics, Enforce and Monitor modes, validation errors, saved state, runtime endpoint configuration, and anonymous telemetry status.
  • Makes Console refresh visibility-aware and resilient to outages with bounded exponential backoff.
  • Adds WCAG A/AA accessibility scans, keyboard interaction coverage, and light/dark visual regression gates for primary Console workflows.
python -m pip install --pre \
  "actionrail[agents]==0.1.0b9" \
  "actionrail-console==0.1.0b9"

0.1.0b8 — executable action contracts​

Released July 20, 2026.

  • Adds copyable Guard and execute and Decision only snippets for Console-authored actions, while keeping automatic LangGraph actions on the enforce() path.
  • Improves first-run credential handling, manual action arguments, empty-Source guidance, code presentation, and unsaved-change navigation.
  • Publishes the value-poisoning benchmark with model-level results, methodology, limitations, and artifact provenance.
  • Adds deterministic Action Tests and the real MySQL integration-test path described in the repository changelog.
  • Rejects malformed runtime-ingestion batches atomically instead of accepting partial audit data.
python -m pip install --pre \
  "actionrail[agents]==0.1.0b8" \
  "actionrail-console==0.1.0b8"

0.1.0b7 — first-run experience​

Released July 16, 2026.

  • Adds a contextual first-run Console guide without interrupting established installations.
  • Makes nested documentation pages safe to open and refresh directly on static hosting.
  • Uses the canonical ToolJet/ActionRail repository casing in package and documentation metadata.
  • Adds the framework-neutral ActionRuntime, explicit action definitions, and Console-authored actions for applications outside LangGraph.
  • Adds exact, field-gated decimal equality for grounded numeric values without weakening string identifiers.
  • Moves LangGraph and LangChain into the optional agents dependency extra.
python -m pip install --pre \
  "actionrail[agents]==0.1.0b7" \
  "actionrail-console==0.1.0b7"

0.1.0b6 — launch candidate​

Released July 16, 2026.

  • Uses actionrail.ai for the canonical documentation site and hub.actionrail.ai for the documented telemetry receiver.
  • Adds explicit rule-schema versioning and Python 3.11–3.14 compatibility CI.
  • Adds release automation, documentation search wiring, compatibility and upgrade guidance, and the packaged MCP quickstart.
  • Removes pre-launch branding compatibility aliases.
python -m pip install --pre \
  "actionrail[agents]==0.1.0b6" \
  "actionrail-console==0.1.0b6"

0.1.0b5 — first public beta​

Released July 16, 2026.

Runtime and framework​

  • LangGraph action-surface discovery, observation, monitor mode, and inline enforcement.
  • Deterministic allow, hold, and block decisions before the original tool runs.
  • Invocation-scoped trusted context for authenticated identity and tenancy.
  • Version 1 rule documents with save-time validation and last-known-good runtime configuration.

Live verification​

  • SQLite, PostgreSQL, MySQL/MariaDB, HTTP, and MCP Source adapters.
  • Read-only database transactions, named parameter binding, bounded I/O, and TLS configuration.
  • MCP Streamable HTTP discovery, structured results, and mandatory readOnlyHint validation.
  • Field, row-count, cross-argument, trusted-context, literal, and time-based match conditions.

Operations and security​

  • A packaged, local, single-user Console with Agents, Sources, Rules, Activity, Reviews, and runtime health.
  • Hashed agent keys, one-time key display, scope-aware storage, and local secret references.
  • Durable bounded audit outbox with batching, retry, idempotency, recovery, and shutdown flushing.
  • Human review that fails closed when the review service is unavailable.
  • Value-free model and audit messages by default, with explicit preview-field export.
  • Minimal anonymous telemetry with ACTIONRAIL_TELEMETRY_DISABLED=1 and DO_NOT_TRACK=1 opt-out support.

Developer experience​

  • Packaged deterministic LangGraph and MCP quickstarts.
  • Python 3.11–3.14 compatibility CI and safety-critical mutation testing.
  • Docusaurus documentation, GitHub contribution workflows, PyPI trusted publishing, release checksums, and GitHub Pages deployment.

Known beta constraints​

  • LangGraph is the only automatic discovery and wrapping adapter. Other Python applications use the explicit ActionRuntime boundary.
  • The open-source Console is local and single-user; managed cloud is not yet generally available.
  • SDK and Console versions must match exactly.
  • Rule schema version 1 is the only accepted persisted schema.
  • Provenance or taint tracking for values originating in untrusted memory is not implemented.
  • Additional vendor-specific Source adapters will be added based on production demand.

Install this release​

python -m pip install --pre \
  "actionrail[agents]==0.1.0b5" \
  "actionrail-console==0.1.0b5"

Release artifacts and SHA256SUMS are attached to the GitHub release.