Agentic QA — Phase 4 & 5 (Claude in Chrome)

Drive/verify the v0.4 intelligence + dataset work by pointing the Claude-in-Chrome extension at the live dashboard.

Scenarios#

cic-v4-rework-lineage — fetch the rework_lineage metric and have the agent contrast it with the rework_cycles heuristic (authoritative vs gameable).
cic-v4-active-idle — read the active-ratio card + a per-session ratio (reliable on the R2 lossless store).
cic-v5-dataset — fetch /plugins/velocity-insights/dataset; confirm scope=single_tenant, lineage + event_count on sessions, and quote the boundary note.
cic-v5-boundary — the honest negative: the agent looks for cross-team/multi-tenant data and confirms there is none (the gated Phase-5 lines were not faked).

Honest scope: rework_lineage and the dataset are API surfaces today (the Velocity page still shows the older heuristic card), so these verify via the authed JSON endpoints, not a UI widget. Additive to the Go integration tests; 🟡 agent-driven, NOT counted toward the automated-coverage number. Canonical prompts + pass criteria: docs/v0.4/qa-claude-in-chrome.md.