1 / 6
Keys: → / space / N next · ← / P back · Home restart · A autoplay
CLI fixture replay
==[ ACT 1 · THE QUESTION ]=====================================
real-user-0 → support_agent: "How many widgets are in stock?"
session_id 7e352c34-4c1c-4395-acd5-fb3c8f215346 eval_id 7e352c34
==[ ACT 2 · THE TRACE ]========================================
USER_MESSAGE_RECEIVED → INVOCATION_STARTING → AGENT_STARTING b7ad6b7169203331 → (silence)
missing: TOOL_STARTING · check_inventory · AGENT_COMPLETED
sibling ab7535a5 answered: "There are 0 widgets in stock."
==[ ACT 3 · NATIVE IMPORT ]====================================
$ bq-agent-sdk evalbench-native-import \
--source-table test-project-0728-467323.bqaa_e2e_real.agent_events \
--job-id mvp-e2e-real-traces ...
{ "status": "imported", "event_row_count": 27, "score_row_count": 7 }
source: production agent_events · destination: BQAA-owned snapshot
EvalBench configs/results/scores are not read.
==[ ACT 4 · FAILED_SESSIONS ]==================================
import_identity evalbench-native-import:mvp-e2e-real-traces:v1:7e352c34
failed_count 1 of session_count 7
taxonomy_categories ["task/planning", "finalization", "tool blockers"]
failing_scores { "goal_completion": 0.0 }
(session-level denominator — span labels localize, they do not replace this)
==[ ACT 5 · SPAN-G1 ]==========================================
label_native_run(run, policy=EvalScorePolicy({"goal_completion": 1.0}))
three frozen G1 names localize on span b7ad6b7169203331
target_kind gap_after_span · no synthetic span ids · no turn_index
library localization on the native snapshot — not persisted BQ rows
judge trap: correctness 1.0 · llm_feedback null · pass_rate 1.0
(nothing to judge — failed_sessions is the denominator)
==[ ACT 6 · PUNCHLINE ]========================================
Span-G1 localizes all three to AGENT_STARTING span b7ad6b7169203331
(gap_after_span). Next debugging action: inspect that span.
We did not need EvalBench tables to see this.
Native agent_events was enough; span-G1 tells you where.
Watch the CLI version
Adapter-path CLI recording (PR 462 freeze). Native span-G1 path is this page.