Skip to content

Omnigent telemetry and session evidence

The adapter included in 0.1.0 imports saved Omnigent OTLP JSON/JSONL files offline. It uses the bounded batch loader and existing OTel semantics, then adds source aliases, policy observations, separate role timing and a session sidecar. It does not start Omnigent, a vendor CLI, collector or provider.

agentloop omnigent import-otel omnigent-export.jsonl --out runs/omnigent
agentloop omnigent import-otel omnigent-export.json --out runs/omnigent-json
agentloop omnigent inspect runs/omnigent --json-out runs/omnigent-inventory.json
uv run --frozen python examples/omnigent_otel.py --out runs/omnigent-example

The offline example uses owned synthetic fixtures. No live vendor, benchmark result or measured improvement is claimed.

Source boundary

The inspected source is Omnigent a2956be0e97bb175a60b274053d836f07d494c6c. Its helper emits AGENT, TOOL and GUARDRAIL spans, session.id, agent.name/gen_ai.agent.name, provider/model labels and optional content. LLM spans originate from spawned executors, so missing model spans establish unknown model coverage, not zero inference.

Recorded aliases include agent/session/parent IDs, request/turn IDs, harness/version/transport and skill names. The current helper does not emit all these fields; absent values stay null. Caller/vendor extensions remain source metadata, with unsupported strings minimized. Unsupported features are unknown/unverified rather than behavior-verified capabilities.

The fixture matrix's original synthetic examples follow revision 713673e9d48cf09cb71fd3a90ae45856ada288d1/v0.17.0. These formats were checked against the current source above. Neither revision is an inferred producer version for an export that did not record one.

Sessions, lineage and identities

omnigent-sessions.json groups observed session IDs across distinct source and native trace IDs. It preserves explicit link targets and resolution against the imported span inventory. Missing targets stay unresolved; links never create native causal parent edges. Explicit ended-parent attributes lack a trace scope and remain unresolved references even when another span happens to share the ID. Missing cross-process continuity remains a gap.

Native parent edges still use the existing batch loader's validated within-trace wire parentage. Session membership alone does not imply parentage, delegation, data flow or a complete execution tree. Group IDs include explicitly supplied job/trial context; no external session, agent or attempt identity is invented.

The sidecar has a byte hash in the inventory. Inspection rejects modified sidecar bytes. Native observations are bound to captured trace snapshots; post-capture mutation fails analysis/export. These checks establish consistency with supplied bytes, not source authenticity or tamper-proof evidence.

Policy and approval observations

Recorded policy.action values normalize ALLOW, DENY and ASK while retaining the original case. Unknown actions remain UNKNOWN. Name, phase, source span, matched tool-call IDs, recorded approval ID/resolution and callback latency remain separate fields.

Every imported policy record has enforcement: unverified and dispatch_prevention: unknown. A pre-tool phase label, DENY, absent tool span, or source enforced=true cannot prove a callable was prevented. A matching tool span establishes a recorded observation of that call ID; it is not a verified pre-dispatch gate. A late DENY does not retroactively prevent work.

ASK without recorded resolution remains an explicit gap. A recorded resolution is source evidence, not an AgentLoop approval flow. The importer does not resolve approvals, execute policies, enable production controls or manufacture a custom approval UI.

Omnigent marks DENY guardrail spans as ERROR. The adapter keeps that span status while leaving the overall task execution outcome unknown. A policy refusal does not independently establish failed task execution or external correctness.

Transport-specific limits

Omnigent's declarations distinguish execution modes and elicitation mechanisms. No live transport was behavior-tested by this import workstream.

Transport Boundary to retain
SDK in process Relevant executor callbacks may expose tool boundaries; support is vendor/version specific
CLI subprocess A printed/logged policy decision does not prove dispatch interception
ACP subprocess Permission/interrupt events depend on the actual server and integration
Native TUI Approval mirroring is not a universal programmatic pre-tool gate
Native server JSON-RPC/SSE permission handling depends on the server/version and active adapter
OTLP file import Observation only; no runtime interception, denial or approval

The runner policy gate handles particular function-policy tool-call/tool-result phases. Other policy types stay server-side, and ASK is escalated by its caller. That implementation does not make all Omnigent transports enforceable. Registry/behavioral probes remain the next workstream.

Timing, usage and reports

Native JSON and HTML expose omnigent_observations: agent labels, session ID, model coverage, policy evidence, lineage gaps and separate role timing. Agent turn, model, tool, guardrail, recorded approval and host roles stay distinct.

For each role, interval union and span cumulative time are separate. Nested agent spans with durations 10s and 4s can have a 10s union and 14s cumulative time; role times are never added into an invented elapsed runtime. Missing bounds remain unknown with known partial intervals explicitly identified. Provider/model counts retain existing source qualifications; root aggregates and cache subsets are not double counted.

Model coverage is unknown without model spans and observed_partial when model spans exist. Neither state claims complete executor coverage. Task quality is unavailable without the independent trial/verifier path, and telemetry-only imports cannot pass replay/study/value improvement gates.

The reserved agentloop.omnigent native namespace version 1.0 holds a bound observation summary. Per-event source.omnigent fields retain aliases and capture flags. Known structural attributes survive native OTel re-export without duplicated agentloop.metadata.* prefixes. Source data remains externally reported; conversion does not upgrade trust or measured provenance.

Privacy and limits

The batch loader minimizes inputs/outputs, policy reasons, exception text, reasoning, token IDs, credential fields and log bodies by default. Content-capture flags describe what the source recorded; they do not enable capture in AgentLoop. Structural names/IDs can still be private. Imported files never load referenced code, task scripts, verifiers, media, archives or remote resources.

Common parser/path/record/trace/event/metadata limits apply. Invalid independent records are retained by default; --strict stops on them. Global budget overflow is terminal. Output remains stable on repeated imports; conflicting bytes fail instead of overwriting. Native schema 1.1, core dependencies and Python 3.10 support remain unchanged. These commands are included in 0.1.0; use its qualification-aware reader.

from agentloop.integrations.omnigent.telemetry import import_omnigent

result = import_omnigent("omnigent-export.jsonl")
result.write("runs/omnigent")
print(result.sessions())
uv run --frozen python -m pytest tests/test_omnigent_telemetry.py tests/test_otlp_jsonl.py tests/test_otel_interop.py tests/test_html_report.py -q -ra

These offline checks exercise source aliases, two child traces, sessions/links, ALLOW/DENY/ASK, timing overlap, missing model/stream data, late spans, privacy, namespace round trips, output consistency and CLI/HTML. They do not establish live vendor compatibility, runtime enforcement or empirical performance gains.