Skip to content

Argus 0.7.0 — modular QA runtime and live monitoring

Choose a tag to compare

@chriswu727 chriswu727 released this 19 Oct 07:30
· 154 commits to 335db79c1e66921a658070a6ec44a369951428f3 since this release

Argus 0.7.0 — maintainable QA runtime and live monitoring

This release strengthens Argus as a local, evidence-first MCP testing system while preserving public tool names and profiles.

Highlights

  • Modular protocol/application/domain/adapter structure, explicit runtime ownership and extension points for verification, journal storage and report formats.
  • Serialized actions and lifecycle transitions, validated limits, bounded waits, cancellation cleanup and explicit operational errors.
  • Optional read-only loopback dashboard with health/readiness, status, bounded SSE and Prometheus metrics: argus-mcp --monitor-port 8765.
  • Session-local replay input vault, private capsule files and input-safe persisted steps, including bulk form/paste/selection paths.
  • Shared CLI/MCP regression, atomic artifacts and OS-backed file locks. Inconclusive checks exit 3; storage failures exit 2.
  • Architecture, design decisions and contribution documentation.

Verification

  • 292 local tests passed, including real MCP stdio, real Chromium fixtures and HTTP/SSE integration.
  • Ruff lint/format and scoped core mypy checks passed.
  • Scripted benchmark retained 34/34 seeded findings and resisted 3/3 false-positive baits. This is a tool capability benchmark, not autonomous-model recall.
  • Built wheel/sdist passed Twine validation. An installed wheel outside the checkout loaded packaged resources and drove a real browser successfully.

No paid model evaluation was run. Live native-macOS UI behavior was not exercised for this release. Monitoring is not an authenticated remote service; see the architecture and privacy documents for explicit limitations.

Upgrade notes

Upgrade argus-testing and restart the MCP host to load the new version. Monitoring is disabled unless --monitor-port or ARGUS_MONITOR_PORT is configured.

Exported replay steps no longer contain typed/pasted/selected values. Steps marked requires_input need fresh inputs; do not replay redacted placeholders. Invalid capture settings now fail validation instead of silently falling back.

The release workflow publishes the tested package to PyPI and then updates the official MCP Registry.