Dashboard/PRISM (Second-Brain)/wiki/decisions/2026-06-28-prism-is-chowmes-prism-hermes-instance.md
PRISM (Second-Brain)

wiki/decisions/2026-06-28-prism-is-chowmes-prism-hermes-instance.md

ADR: PRISM = Chowmes-PRISM (a dedicated Hermes instance), not a custom SaaS

Date: 2026-06-28 Status: Accepted

Context

PRISM was an idea for a custom-built prospect-intelligence platform; only Wave-1 intel modules were ever built and the pipeline never ran end-to-end. Meanwhile the manually-run algolia-* Claude skills already produce every deliverable (proven by the published PetSmart audit). The question: build PRISM as bespoke software, or orchestrate the existing skills?

Decision

PRISM ≡ Chowmes-PRISM: a dedicated Hermes agent instance (Nous Research Hermes, the same runtime as the personal "Chowmes"/Athena agent) standing on the VPS, that (1) runs the algolia skills to generate audits and (2) answers grounded questions over the resulting report. "PRISM"/"prism" now means exclusively this instance (naming canon).

Sub-decisions: - Separate Hermes instance, not a profile inside the personal Chowmes, own container (hermes-prism), /root/.hermes-prism volume, dashboard :9120, own Telegram bot. Shares Scout + Caddy on the same VPS. - Execution runs on the VPS (standalone, works when the laptop is off). - Temporal DROPPED, it belonged to the abandoned deterministic-module plan; Hermes-native kanban + cron + durable subprocess cover queue/schedule/retry. The stale VPS Temporal stack was deleted. - Two model planes: control plane (the Hermes chat/coach brain) = Gemini-direct (paid key, cheap, high-frequency); execution plane (running the 22 skills) = headless Claude / Anthropic (the skills are tuned for Claude Code; proven; lower-frequency, higher cost).

Rationale

The skills already produce every deliverable, so orchestrating them reaches a working product far faster than reimplementing 6 waves. A separate instance gives blast-radius isolation + independent lifecycle for a multi-rep product without polluting the personal agent. Splitting models keeps the cheap chat on Gemini while preserving Claude fidelity for the audit generation that is prospect-facing.

Alternatives Considered

Option Why rejected
Build PRISM as custom SaaS Far from done (Wave 1 only); skills already do the job
A Hermes profile (not a separate instance) Shares the personal container's fate; wrong for a multi-rep product
Keep Temporal for orchestration Second orchestrator alongside Hermes = duplicated state, no gain
Run skills on Gemini too (no Anthropic) Skills tuned for Claude Code; quality unproven on Gemini

Consequences

  • New VPS container hermes-prism + /opt/chowmes-prism/docker-compose.yml; data at /root/.hermes-prism.
  • Generation needs an Anthropic key on the box (real per-audit cost).
  • Multi-rep isolation handled inside the instance later (per-rep profiles or API+frontend).
  • Project plan + workstreams (W-A..W-F) live at repo docs/workspace/hermes-prism-integration/03-plan.md and the prism GitHub repo. Related: 2026-06-28-grounded-report-qa-gate, 2026-06-27-prism-as-ai-sales-toolkit-stage-cockpit.