Fable-5 Handoff Prompt

Copy-paste the block below to kick off Fable 5 on PRISM V2. It scopes the first executable package only (Tracks A/B/G-schema/C) — E/F/H stay in brainstorm until the executioner POC lands. Update the two bracketed inputs before sending: **DEADLINE** and **BETA_TARGET** (decisions D8/D7).

You are Fable 5, executing the PRISM V2 build. PRISM is being rebuilt as a standalone, domain-agnostic "Prospect Research Operating System" — a sellable product, not an Algolia-internal tool. Algolia is just the first domain module.

Read first, in order: docs/PRISM-V2/_status.md → 06-v2-execution-map.md (your build spine) → 07-design-system.md (the design language you build every screen against) → 05-role-driven-ia.md (role IA) → docs/research/Discovery-OS-v1.md §9 (the finding/call-plan schema). Do not re-derive decisions already locked in 06 §1.

Ground rules (non-negotiable, from the project CLAUDE.md): best-of-breed stack — Postgres+pgvector + Claude Agent SDK, NOT Algolia-as-DB, NOT Agent Studio. Pydantic on every boundary. Tests ship with every module (unit + integration + contract). Never claim done without running ruff check . && ruff format --check . && mypy src/ --strict && pytest -v and showing output. Evidence + source on every data point — zero fabrication; where data is unknown, render an explicit empty slot, never a guess. Nothing deploys to prism.chowmes.com or main without an attended review.

Do NOT resolve these by invention — stop and flag them (they are 06 §7 blocked decisions): the 3 IA working assumptions (D4), the Discovery-OS gating-rule ambiguity (D5, hard-suppress vs mandatory-M9), executioner self-hosted-vs-managed (D1), durable-retry approach (D2). Also honor: HeyGen/LiveAvatar + Telegram are Phase-2, do not build them.

Your work package, in order:

1 — Design system (Track B). From 07-design-system.md, emit prism-tokens.css (the reconciled §1 tokens) and prism-components.css (ONE button system with primary/secondary/ghost/pill + hover/active/focus/disabled, plus cards/tables/tiles), applying the §1 reconciliations that fix the source's inconsistencies. Retrofit the existing Marketer door (~/prism/marketer/door.html) onto these files.

2 — Role doors, breadth-first (Track A). Build the AE door and BDR door to match the shipped Marketer-door pattern, reading already-published audit-data.json per account — no executioner/VPS dependency. Per-role scope, culls, and components are in 06 §3: AE = 3-stage cockpit (PREP→SS1→SS2, SS3/SS4 culled) + the Discovery-OS single-page call plan as the PREP hero; BDR = signal-ranked queue + micro-Exchange outreach (deep financials/MEDDPICC/stages culled); Marketer already built (finish its stubbed ABM Brief + disabled buttons). ui-validator pass + Playwright verify at 1280/375px. Goal: all 3 doors live and demoable.

3 — Discovery OS schema (Track G-schema). From Discovery-OS-v1 §9 (spec condensed in 06 §4), define the Pydantic Finding / Behavior / Feedback / Branch objects (13 finding categories, confidence, risk_if_wrong, evidence-grounding, the two DISTINCT archetype fields, persona_fit). This schema must be settled before the Postgres schema freezes. Also scaffold the /standards-discovery-os skill.

4 — Executioner + data backend (Track C), POC FIRST. Before building the full executioner, run the e2e POC (R1): prove Claude Agent SDK can do multi-step audit control + subagent dispatch + Postgres-backed SessionStore + status tracking + hooks-as-gates running fact-check+QA after EVERY module run. This needs a real ANTHROPIC_API_KEY — request it if absent, do not stub around it. Only after the POC passes, design the Postgres+pgvector schema (audits/research/state/vectors, using the Track G-schema Finding object) and the migration off the current prism-runner + prism_platform. Reuse the already-proven run_audit/live_status tool mechanism from docs/workspace/hermes-prism-integration/phase-b-cass-agent/L2-execution-orchestration-design.md — do not redesign the agent protocol from scratch.

Context for scoping: deadline pressure = **DEADLINE**; intended first beta user = **BETA_TARGET**. Optimize the MVP for that.

Report format: for each of the 4 items, report done/deferred/blocked with the verification command output. Surface every blocked decision back to Arijit rather than guessing. Chat-as-operator (Track D), modular rearchitecture (E), and productization+GTM+pricing (F/H) are the NEXT package — do not start them in this run.

Before you send it: answer D7 (beta target) and D8 (deadline) so the two brackets are filled — otherwise Fable 5 scopes the MVP blind.