PRISM — Overview

Updated 2026-06-29. PRISM was redefined on 2026-06-28. Earlier framings in this wiki (a custom v2 SaaS with Temporal/Postgres; a built web UI) are superseded — see the superseded-banner ADRs.

What PRISM is now (canon)

PRISM ≡ Chowmes-PRISM: a dedicated Hermes agent instance on the VPS that is the Algolia sales team's prospect-intelligence engine + sales coach. It (1) runs the algolia-* skill suite to generate a search-audit + sales deliverables for a prospect, and (2) acts as the Ultimate Algolia Sales Coach — answers questions grounded only in the audit report and turns findings into Discovery-OS call plans — reachable from Telegram (and, later, an in-SPA chat).

"PRISM" / "prism" from 2026-06-28 means only this instance — not the old custom-SaaS idea, not the personal Chowmes/Athena agent. See 2026-06-28-prism-is-chowmes-prism-hermes-instance.

Architecture (two planes + a gate)

• Control plane = the Hermes instance (chat/coach brain) on Gemini-direct (paid key); own container hermes-prism, /root/.hermes-prism, dashboard :9120, own Telegram bot.
• Execution plane = a headless Claude (Anthropic) worker at /opt/prism-executor that runs the 22 algolia skills (the skills are Claude-Code-tuned). Generation is standalone on the VPS.
• Acquisition plane = Scout where PRISM needs web evidence before interpretation: /scrape, /structure, /harvest, source artifacts, and citation-grade acquisition outputs. Scout acquires and structures evidence; PRISM interprets, coaches, and generates sales deliverables.
• Orchestration = Hermes-native (kanban/cron), no Temporal (dropped + deleted).
• Grounding gate = the prism-report-qa Hermes plugin: pre_llm_call injects the bound report; transform_llm_output runs a Gemini judge that blocks/rewrites unsupported factual claims (coaching is allowed if anchored to cited facts). No Hermes fork. See 2026-06-28-grounded-report-qa-gate.
• Data layer = /opt/data/reports/<slug>/audit-data.json + index.json. Reports come from the executor (new audits) or imported from the algolia-arian-v2 hub (existing audits).

What's true today (2026-06-28)

• ✅ Chowmes-PRISM live on the VPS; Telegram connected; Gemini-direct.
• ✅ Data layer with PetSmart + Home Depot Mexico reports imported.
• ✅ Grounded report-QA verified over PetSmart (real fact 15.98% cited; absent fact refused; coaching allowed + anchored).
• ⏳ Generation (W-A) executor installed (/opt/prism-executor: node/deno/claude-cli/playwright
• 35 skills) — blocked on an Anthropic key on the box + 3 MCP keys.

Personas

Algolia AEs and BDRs (Arijit = operator/builder). Methodology backbone = the FY27 gated motion (PREP→SS1→SS2→SS3→SS4); Constructor.io = the primary competitor. The coach voice uses Discovery-OS ("Exchange": calibrated hypotheses + named moves F1–F6 / M1–M10).

Still-valid prior knowledge

• Search-vendor detection (live network-packet inspection) — the BuiltWith replacement / "Golden Angle" basis. 2026-06-27-search-vendor-packet-detection (unchanged, valid).
• The FY27 AE motion + AI-Sales-Toolkit framing survive as the coach's domain knowledge (the built-UI parts of the old toolkit ADR are superseded).

Where state lives

• Plan + workstreams (W-A..W-F): repo docs/workspace/hermes-prism-integration/03-plan.md.
• Code/docs: GitHub arijitchowdhury80/prism.
• Decisions: wiki/decisions/ (this vault). Build chronicle: wiki/dev-log.md.
• Cross-session recall: Claude memory (Chowmes-PRISM canon + the session's findings).