Description
Viewers of the dashboards and Agent Observability conversation lists can tell the traffic is synthetic within a minute. In-app reader questions are seven formal templates crossed with the fixture list, and every follow-up is the literal “Give one more concise detail for claude -p prompt from a pool of 57, a third of them about unrelated open-source repositories, all ending in the same boilerplate. Guard probes arrive at a steady rate. Chosen direction: a hybrid - a richer committed corpus now, plus a job that grows it with a small model and commits additions only after review. Subtasks carry the pieces; this parent closes when they are all Done.
Acceptance Criteria
- #1 Every subtask is Done
- #2 A 30-minute sample of conversations in Agent Observability shows no two first messages with identical text outside deliberate repeats
Definition of Done
- #1 just check
Implementation Plan
Lanes in one checkout, disjoint files, frozen seams (reader corpus apps/corpus/, prompt file format with ‘— followup’ turns, agent-host/codebase/seed-repo). Lane A: .01+.04 (loadgen, site-browser, corpus). Lane B1: .03+.05 (dev-session mechanism). Lane B2: .02 + PII prompt content (codebase, prompts). Lane C: .06 (corpus generator). Main thread wires, runs just check, commits per subtask, releases, bumps the lab consumer stack, then checks the live ACs.
Implementation Notes
All subtask code shipped in v0.2.0 and deployed to the lab 2026-09-29. Open: AC2 (30-minute no-duplicate sample) not sampled at Rob’s request; AIO-0001.05 AC3 waits for the first live multi-turn session. Follow-on: renovate.json now tracks the agento11y CLI/plugin (33ce015).