TraceForge: OTel gen_ai spans, tool taxonomy, business success vs HTTP 200, replay.

Active (early) Follow one request across tools, models, and agent steps. TraceForge is the span layer — RCA and required-step correctness. Active (early): schema in agent-trace-collector plus packages under infra-ai-streaming/traceforge (tool-call-analyzer, replay, benchmarks). Full LensAI compose overlay still landing. See DESIGN.md.

When to use

Multi-step agents where tool order and skips matter
You need business success, not just HTTP 200
N+1 tool loops, retry storms, or planner infinite loops

Span model

inferix.call                    # root — keep always (T17)
├── routeiq.decide              # route_decision, model_id
├── agent.tool.<taxonomy>       # tool_name, schema version
├── gen_ai.chat                 # model invoke (OTel gen_ai)
└── hitl.wait                   # optional — tag exclusive waits (T20)

Use OpenTelemetry gen_ai.* attributes. Propagate trace_id / baggage across multi-agent handoffs (T22). Versioned tool taxonomy keeps step names stable (T24).

Business success

HTTP 200 is not success

A refund tool can return 200 while the agent skipped policy.check_refund. TraceForge records business_success from a domain predicate. Dashboards use traceforge.tasks.succeeded and traceforge.steps.skipped_required.

Key metrics

traceforge.tasks.started/succeeded/failed
traceforge.tools.error_rate, tools.p99, tools.retries
traceforge.graph.n_plus_one, graph.exclusive_time_ms
traceforge.steps.skipped_required, replay.available

Operator failure modes

ID	Problem	Next action
T01	Required step missing	Hard gate + alert
T03	Tool 4xx schema error	Bump schema; FineForge adapter
T05	Retry amplification	Cap retries; backoff
T06	N+1 tool loops	Batch tool; cache
T07	Infinite planner loop	Max steps; kill switch
T11	Hallucinated tool args	Validate before execute
T12	Partial success unmarked	Business-success predicate
T14	PII in tool args	Hash args; vault payload
T16	Orphan spans	Fix context propagation in SDK
T23	Cost missing on tool spans	Dual-write LensAI cost

Config

Design target / not shipped

Design target / not shipped Suite observe.traces in inferix.yaml and GET /v1/traces on :4000 are targets. Today: schema package + planned OTLP → ClickHouse path in the TraceForge repo.

API honesty: API reference. Clients: SDKs.

Source

https://github.com/AkshantVats/inferix-traceforge-agent-collector

TraceForge