Observability

DynosAI records enough information to explain how a managed coding workflow progressed and why control decisions were made.

Audit records

The local database contains workflow and audit events for significant state changes, scope decisions, validations, state snapshots/restores, and routing/control decisions.

Token and context telemetry

Provider telemetry is normalized into process and phase views. Where exact provider counters are available, they are preferred; estimators are explicitly marked and calibrated separately.

Tracked signals include:

• input, cached input, generated/reasoning tokens;
• phase attribution and live phase deltas;
• MCP calls and MCP time;
• response/checkpoint compaction;
• execution waves and expected round trips saved;
• model-control decisions and transitions;
• acceptance retries and actual spend including retries;
• cost-per-successful-governed-change raw counts (tokens, tools, retries, human gates). Monetary figures appear only when the public catalog matches a provider/model.

Acceptance bundles

dynosai debug acceptance can package normalized logs, provider protocol evidence, workspaces, summaries, Oracle results, runtime audits, and consolidated matrix metrics into a portable ZIP.

The acceptance harness deliberately preserves failed infrastructure attempts when a safe retry is used. Certification totals and actual paid totals are kept separately so retries cannot hide cost or evidence.

Diagnostics

Useful commands include:

dynosai doctor --deep
dynosai env
dynosai diagnose
dynosai usage
dynosai debug acceptance-status
dynosai debug acceptance-logs