Trustabl Agent OpenTelemetry Rules — v0.1

A small Apache-2.0 OpenTelemetry process-metadata profile (trustabl.* span attributes and events) plus a deterministic runtime rule format, published as data so any observability backend can evaluate it against live agent trace data.

What this is not

• Not loaded, parsed, or evaluated by trustabl scan. The agent-reliability-analyzer engine never imports or fetches anything from this repository. Nothing here is part of that engine's ScanResult, RulesSkipped, or SupportedSchemaVersion compatibility contract.
• Not related to agent-reliability-rules, the engine's own static-analysis rule pack repo. Those rules are pulled and evaluated in-process by the trustabl binary against source code it just parsed. The rules in this repository are evaluated by third-party observability backends (Grafana, Langfuse, Datadog, OpenLIT, …) against live OpenTelemetry trace data — the trustabl binary is never in that loop.

Layout

• attribute-profile.yaml — the v0.1 trustabl.* span attribute and event reference.
• rule-ir-schema.yaml — the rule YAML shape reference.
• rules/ — all twelve v0.1 rules, each as a YAML file: tool_loop, retry_storm, policy_absent, policy_drift, contract_drift, irreversible_unguarded, constraint_hit (shadow), false_idempotent, handoff_unbound, instrumentation_degraded, plus two from the DSPM addendum: dspm_unbound, denied_then_answered.
• translations/ — worked translations of the rules into vendor-native formats. Confidence varies by backend — see each subdirectory's own README for specifics:
  • translations/grafana/ — dashboard JSON + alerting-rule provisioning YAML, shaped for self-publish to Grafana's own catalog.
  • translations/langfuse/ — a score-config JSON. Apply against your own Langfuse project via their API.
  • translations/datadog/ — a Monitor API JSON object. Apply against your own Datadog org.
  • translations/dunetrace/ — a detectors.yml adapter mapping onto Dunetrace's existing TOOL_LOOP detector. Speculative — no verified access to Dunetrace's real config format.
  • translations/phoenix/ — a Python span evaluator function. Speculative — shape inferred from Phoenix's evaluator API.
  • translations/langsmith/ — a saved-filter query. Most speculative of the seven — LangSmith's filter grammar cannot fully express the sliding-window condition in one filter.
  • translations/openlit/ — a ClickHouse SQL detection query against OpenLIT's otel_traces table. Uses an exact call-count window via row_number() — closer to the rule's intent than the time-window approximation in the Grafana/Datadog packs.
• scripts/ — create-datadog-monitor.py: translates any rule YAML into a Datadog composite monitor via the Monitors API.

v0.1 rules

All twelve rules are authored as YAML files in rules/, each following the shape defined in rule-ir-schema.yaml:

Rule	Severity	Fix class	Shadow
tool_loop	high	policy.cap_tool	no
retry_storm	high	policy.cap_retries	no
policy_absent	high	policy.require_bundle	no
policy_drift	high	policy.pin_bundle	no
contract_drift	high	policy.pin_contract	no
irreversible_unguarded	critical	policy.require_constraint	no
constraint_hit	medium	policy.review_constraint	yes
false_idempotent	high	policy.audit_tool	no
handoff_unbound	high	policy.require_bundle_on_handoff	no
instrumentation_degraded	medium	instrumentation.add_emitter	no
dspm_unbound	high	policy.require_bundle	no
denied_then_answered	high	policy.require_constraint	no

constraint_hit is shadow-only: constraint fires are expected behavior, not a defect — the rule surfaces them for trend analysis without paging anyone.

The Grafana translation in translations/grafana/ is fully worked for tool_loop. The other eleven rules follow the same mapping; translating them is the next translation-pack slice.

Status

v0.1, first slice. No translator tooling exists yet — every vendor translation is hand-authored to prove (or, for the three speculative packs, sketch) the rule-IR-to-vendor-format mapping, not generated. Licensed Apache-2.0.