Skip to content

Qualification

AgentMuru release claims come from repeatable gates. Offline contracts and credential-backed live checks remain separate so fixture evidence cannot be mistaken for provider or device evidence.

0.3.0 release evidence

The release candidate was qualified on 18 August 2026 with these results:

  • 182 Python tests across runtime, persistence, providers, server, CLI, packaging, and docs;
  • Ruff and MyPy with no reported issues;
  • 9 frontend unit tests, 2 bundle tests, and 3 Chromium end-to-end flows;
  • zero serious or critical accessibility violations in the Workspace and zero homepage violations;
  • strict MkDocs build in light, dark, desktop, and 390-pixel mobile layouts;
  • wheel and source distribution accepted by Twine 7 with no internal planning files;
  • clean-wheel installation with OpenAI, Anthropic, Google Gen AI, and Databricks extras;
  • all Go tests and vet, a native build, and 40 of 40 correctly routed fixture cases with zero simulated unsafe effects.

Credential-backed model calls, a live Databricks workspace, and reference-device native testing were not executed. Those remain deployment-specific gates and are not implied by the offline results above.

Python and package gate

CI runs Python 3.10, 3.11, 3.12, and 3.13. The gate includes:

  • the full Pytest suite;
  • Ruff and MyPy;
  • provider request, streaming, tool-call, usage, failure, and cancellation contracts without network access;
  • CLI, scaffold, examples, packaging, branding, and documentation contracts;
  • source distribution and wheel construction plus metadata inspection.

Workspace gate

The frontend gate installs from package-lock.json, then runs component tests, lint, TypeScript checks, production build, bundle tests, and bundle budgets. Browser flows cover the real server, session interaction, reconnect, approval states, and restart behavior.

Documentation gate

MkDocs builds with --strict. The contract verifies current installation commands, task-based navigation, provider pages, operational limits, stable references, Labs separation, asset links, and absence of legacy public identity.

Clean-wheel gate

qualification/run_clean_install.py creates an isolated virtual environment, disables user-site and inherited Python paths, installs the built wheel, checks dependency consistency, runs the CLI, imports a generated scaffold, starts the packaged server, and exercises installed runtime paths.

Run it after building:

python -m build
python qualification/run_clean_install.py `
  --wheel dist/agentmuru-0.3.0-py3-none-any.whl `
  --report .tmp/qualification.json

The installed smoke path covers runtime execution, approvals, handoff, workflow, durable SQLite reopen, interrupted-run recovery, and optional integration imports.

Credential-backed checks

Official provider contract tests do not prove that a deployment account can use a model. Before release into an environment, run one restricted live request per selected provider and record:

  • account and region without secrets;
  • model ID and access outcome;
  • text streaming and one harmless tool call;
  • token usage and stable failure handling;
  • quota and billing controls;
  • provider data retention and privacy configuration.

Databricks live checks follow the same opt-in rule.

Native Labs evidence

The action-router fixture gate builds the Go CLI, runs native tests, measures all 40 included cases, requires at least 95% correct routing, validates every result, and requires zero executed effects. It does not qualify a public model, low-end device, Android device, container boundary, or live internet retrieval.