@openagentaudit/adapters

Source-format adapters that map external trace formats into the OpenAgentAudit canonical event model.

Each adapter is a versioned pure function: (source) → CanonicalEvent[]. Adapters never fabricate data and never call out to the network. See docs/adapter-contract.md for the full contract.

Adapters

| Adapter id | Source format | File | Status | |---|---|---|---| | aep-v0.2 | @wasmagent/aep AEPRecord (aep/v0.2) | src/aep-v0_2.ts | implemented | | bscode-rollout-v1 | bscode RolloutWireRecord | src/bscode.ts | implemented | | compliance-eval-record-v0.1 | @wasmagent/compliance records | src/compliance-eval-record.ts | implemented | | otel-genai-v0.1 | OpenTelemetry GenAI spans | src/otel.ts | implemented | | langfuse-export-v0.1 | Langfuse trace export | src/langfuse.ts | implemented | | langsmith-export-v0.1 | LangSmith trace export | src/langsmith.ts | implemented |

Usage

import { aepV0_2, bscode } from '@openagentaudit/adapters';

// AEP v0.2
const run = aepV0_2.AepV0_2Adapter.beginRun(aepRecord);
const events = aepV0_2.AepV0_2Adapter.toEvents(aepRecord);

// Extract run-provenance metadata and attach to ReportMeta
const prov = aepV0_2.getProvenance(aepRecord);
const meta: ReportMeta = { ..., aep_provenance: prov };

// bscode rollout
const run = bscode.BscodeAdapter.beginRun(rolloutRecord);
const events = bscode.BscodeAdapter.toEvents(rolloutRecord);

Batch conversion with toEventsBatch

When auditing multiple AEP records from the same session or across runs, use toEventsBatch to produce a single continuous event stream with hash-chain continuity across record boundaries:

import { aepV0_2 } from '@openagentaudit/adapters';

const records: AEPRecordInput[] = [record1, record2, record3];
const events = aepV0_2.toEventsBatch(records);
// or via the adapter object:
const events2 = aepV0_2.AepV0_2Adapter.toEventsBatch!(records);

The method signature on the SourceFormatAdapter interface:

toEventsBatch?(records: TSource[], initialPrevHash?: string): CanonicalEvent[];

Warning: Calling toEvents per record and concatenating the results will break the hash chain (prev_hash of the first event in record N+1 will not match the last hash of record N). Always use toEventsBatch when you need aggregate reporting across multiple source records.

AEP v0.2 adapter — what is preserved, what is rejected, where the boundary is

What the adapter preserves

| AEP field | Canonical field | Notes | |---|---|---| | actions[].tool_name | event.tool.name | Direct mapping | | actions[].capability_decision.capability | event.tool.capability | Only when present | | actions[].input_taint_labels + output_taint_labels | event.tool.risk_tags | Merged into single array | | capability_decisions[].decision | event.policy.decision | dry_run → allow (closest semantic) | | capability_decisions[].reason_code | event.policy.reason | Empty string when absent | | verifier_results[] (failed only) | event type observation | Passed verifiers are silent | | signature.sig / signature.key_id | event.evidence.signature / signer_key_id | Carried on every event | | run_id, model_id, run_context.agent_id | run.run_id, run.model_id, run.agent_id | Via beginRun() | | repo_commit, runtime_version, policy_bundle_digest, tool_manifest_digest | AepProvenance struct | Via getProvenance() |

What is rejected (silently dropped)

• input_refs, output_refs — not yet mapped to a canonical field
• budget_ledger — inventory data, not an event
• run_context.environment_digest, dependency_lock_digest — no canonical equivalent yet
• Passing verifier results — only failures emit an observation event

Schema mismatch boundary

The adapter accepts both aep/v0.1 and aep/v0.2 records. Missing required fields (run_id, schema_version, created_at_ms, signature.*) cause an actionable error rather than a silent partial parse:

AEP adapter: missing required fields [run_id]. Ensure the AEPRecord was
produced by a compliant emitter (aep/v0.2).

Mapping

AEPRecord.actions[]               → CanonicalEvent type:"tool_call"
AEPRecord.capability_decisions[]  → CanonicalEvent type:"policy_decision"
AEPRecord.verifier_results[] (failed only) → CanonicalEvent type:"observation"

RolloutWireRecord.tool_call_sequence[] → CanonicalEvent type:"tool_call" / "observation"
RolloutWireRecord.final_answer    → CanonicalEvent type:"final_answer"
RolloutWireRecord.build_result    → CanonicalEvent type:"observation" source:"build_verifier"

Test fixtures

Real-world AEPRecord fixtures from both upstream emitters are committed under examples/traces/:

| File | Source | Key fields | |---|---|---| | aep-wasmagent-fixture.json | [email protected] — signed with Ed25519 | 2 actions, 1 cap-decision, 2 verifier results (1 failed), all 4 traceability fields | | aep-bscode-fixture.json | [email protected] via buildAEPEvidence() — signed with Ed25519 | 2 actions, 1 cap-decision, 1 verifier result (passed), all 4 traceability fields via resolveRunProvenance() |

The adapter test suite (src/aep-v0_2.test.ts) exercises both fixtures end-to-end and asserts canonical-event output shape, taint propagation, provenance extraction, and error messages for invalid inputs.