Inspect Runs, Tools and Usage
Read persisted results, provider usage and operational warnings for harness runs.
Inspect the stored run after execution, including when the client disconnected:
curl "http://127.0.0.1:7777/agents/support-agent/runs/$RUN_ID?session_id=$SESSION_ID"Check terminal status, content, tool calls/results, metrics and warning metadata. A completed response does not by itself prove that every native transcript entry was mirrored successfully.
Identify the harness and model
Session list/detail responses and run reads under /sessions/{session_id}/runs expose metadata.sdk, such as claude-agent-sdk or codex. This identifies the harness; it is separate from the model. Existing session metadata is preserved, and sessions without a recorded harness identity do not gain an inferred one.
External run records also include model and model_provider when known. Use those fields alongside the provider usage details below when inspecting a run.
Provider metrics
Both harnesses report the usage supplied by their native SDK. Agno adds duration and, for streamed output, time to first token. Metrics are available on the run and terminal completion event; completed runs contribute to stored session metrics.
| Provider | Interpretation |
|---|---|
| Claude | Input/output tokens, cache usage and reported cost. Per-model details can include helper models and subagents. |
| Codex | Input/output tokens and available cached/reasoning usage. Cost is not reported by this integration; unset cost does not mean zero spend. |
Do not use these fields as a billing ledger for every failed attempt or native transport retry. Verify what the pinned SDK reports, including error and cancellation paths, before reconciling usage with provider billing.
Tool history and lineage
Inspect tool IDs, arguments, results and errors separately from the model's final text. Claude subagent tool records can include parent_tool_call_id, preserving the relationship to their parent call. A saved result is evidence of a recorded tool response; use the business service's transaction ledger to verify external effects.
Retries retain recorded tool history. See Retries and Tool Effects for why that is different from exactly-once execution.
Operational warnings
| Signal | Action |
|---|---|
transcript_persistence_failed | Investigate native recovery state; avoid repeating successful mutations to repair the transcript mirror. |
| Retry warning | Separate failed-attempt output from the next attempt in the client. |
| Authentication failure | Repair credentials and perform a real provider check; API health is insufficient. |
| Missing native Codex thread | Expect bounded-history recovery, not exact native resume. |
| Error or cancellation | Inspect tool effects and stored run state before deciding whether to retry. |
Native run/tool records and metrics are available independently of external tracing exporters. The OpenInference external-agent integration and Claude compaction tracing are separate upstream follow-ups; do not depend on them for the initial POC's run inspection.