Skip to main content
QM support is GA with explicit partial coverage. The adapter and qm-snapshot/v1 format are platform-independent. The full transport acceptance matrix is certified against xerg_export/v1 on Fly.io and QM commit 5eb3393315b45b338b860572ab516db9f6eae6da. Xerg does not claim complete QM model economics or provider-invoice reconciliation. QM currently omits some auxiliary model paths from its durable request ledger. An audit describes recorded observations, not every possible provider call. Other boundaries:
  • run_activity is pruned after approximately one hour. One-shot snapshots cannot reconstruct older tool history.
  • Pi and OpenCode model created_at values are flush/capture times. Claude values are closer to result-step recording. Codex rows are turn aggregates.
  • Retry attempts exist as run lifecycle evidence, but model spend is not attributable per attempt.
  • Tool calls/results are not defensibly associated with model observations, so monetary tool-loop is disabled.
  • Iteration counts and cache carryover are unavailable.
  • openrouter/auto is unpriced without the actual routed response model.
  • Continuous follow capture, polling watermarks, and interrupted-follow recovery are not currently supported.
  • Local, Docker, AWS, external PostgreSQL, and other deployments can use snapshot mode or strict direct mode when their database supports a genuinely restricted role, but those platforms have not completed the Fly acceptance matrix.
  • Fly Managed Postgres uses a disclosed operator-contained process boundary. It does not use Fly’s broad managed reader and does not claim database-level least privilege.
  • QM’s current Fly Sprites backend resumes persistent Sprites without applying the configured sandbox OCI image. Offline private-Slack audit therefore requires an explicitly approved, version-verified Xerg CLI bootstrap inside that private Sprite; this limitation does not affect the core exporter or platform-independent adapter.
  • A cold QM request to set up https://xerg.ai/skill.md can explain prerequisites, but it cannot grant database privileges, persist an admin skill pack, or provision secrets by itself.
  • A QM Slack agent can audit only an authorized pre-created snapshot. It cannot start live database or Fly collection, and no live Slack-triggered bridge is currently supported.
Upstream proposals for routed response model, cost authority, and broader content-free telemetry are useful but do not block the current one-shot adapter.