Nine commits across three repos — dead code removed, retry logic added, grades go private, and two-agent architecture locked.
Three commits closing out technical debt since the worker was first built. Model version bumped to current. The exa_note field — a ghost schema key that caused Claude to hallucinate Exa API content that was never called — removed from both the worker schema and the client UI. Exponential backoff retry added to callAnthropic() so a single 529 no longer kills a review session. max_tokens raised from 4000/6000 to 8000 on both endpoints. A second pass on ic-intake.html removed a 38-line dead SYSTEM constant left over from a prototype era before the worker existed. Both worker and tools deployed and verified live — curl against the live endpoint confirmed 4 keys, no exa_note, full prompts with no truncation.
Two decisions documented in the playbook. Grades go private: AICV is not a public grading authority. Public Snapshots carry analytical substance only; letter grades are reserved for the private Intelligence Review as the upsell hook. Six-surface migration required — Snapshots, worker CHAIR_SYSTEM, JSON schema, IC.md template, pitch language, and methodology copy. Two-agent IC architecture decided: Pre-Council and Post-Council Managed Agents bookending the existing worker, city-scoped memory stores, Rancho Mirage as pilot city. Build deferred to next session.
HANDOFF.md created — first time. SESSION-LOG.md created as a new cross-thread visibility file, append-only, per-thread scope. Session Close Protocol added to CLAUDE.md immediately after the existing Session Start section — symmetric placement, with an explicit proactive trigger so Claude Code offers the protocol without being asked. Grade A downstream sweep run across seven repos — clean, no stale claims.