hermes-operator-stack

Overview

Use this skill to keep Kevin's persistent Hermes setup coherent across machine, model, memory, browser, messaging, cron, and wiki writeback layers. Architecture intent is not deployment state: never call Hermes or an external memory provider active until the target machine has a version, config, doctor, behavior, and recovery receipt.

Required Reads

Read these before making recommendations or edits:

1. wiki/architecture/hermes-mac-mini-agent-os.md
2. wiki/tools/hermes-agent.md
3. wiki/architecture/hermes-harness.md
4. wiki/tools/hindsight.md
5. wiki/tools/honcho.md
6. wiki/tools/agent-browser.md
7. wiki/tools/agent-client-protocol.md
8. wiki/meta/capability-routing-map.md

Audit Workflow

Check the stack in this order:

1. Runtime - Hermes is the persistent operator; Codex, Claude Code, OpenCode, and local models are workers, not the source of truth.
2. Memory - Kevin-Wiki/qmd is active canonical memory. Hindsight is the selected external runtime-memory candidate/default only after the Hindsight admission gate below passes. Honcho is the separate profile/modeling layer via MCP/SDK or a separate profile.
3. Browser - agent-browser is the default browser automation surface. Use logged-in Chrome for auth-heavy flows. Use Playwright only for committed regression tests or existing repo test suites.
4. Interfaces - Telegram topic threads are the primary ops console; Hermes Desktop over SSH is the rich local console; Discord, WhatsApp, terminal, and ACP are secondary surfaces.
5. Schedulers - Use no-agent launchd or Hermes cron for deterministic monitors. Wake Hermes only when summarization, ranking, triage, or drafting requires reasoning.
6. Intake - Route feeds, Raindrop, YouTube, X/Last30Days/BirdClaw, LinkedIn, Economist, and FT through summaries first, deletion second, deep engagement selectively.
7. Writeback - Durable facts, preferences, failed guesses, decisions, and reusable workflows must land in wiki pages, skills, automations, or logs.

Web extraction benchmark gate

Hermes/Scrapling launch claims, including the saved “up to 60x” claim, are candidate evidence rather than a default-route change. Before routing a source class through Hermes scraping, freeze representative public URLs and compare it with direct fetch/reader extraction and agent-browser where rendering is required. Record successful extraction, required fields, citation/locator fidelity, latency distribution, token and provider cost, retries, blocks, fallbacks, and terms/data handling. Promote only the measured source class and configuration. If the benchmark artifact is absent or current releases have an open performance regression, keep the incumbent browser route.

Preferred Commands

Use only when Hermes or agent-browser is installed in the target environment:

hermes config check
hermes plugins
hermes memory setup
hermes acp --check
hermes send --list
agent-browser doctor
agent-browser batch "open https://example.com" "snapshot -i" "close"

For repo changes in kevin-wiki, close the loop with:

npm run skill-registry
npm run routing-doctor
npx tsx scripts/build-index.ts
qmd update
qmd embed

Hindsight Admission Gate

Before recommending an install, read the complete current Hindsight source page and name the runtime job that qmd/wiki and Hermes built-in memory do not satisfy. Before saying Hindsight is active or healthy:

1. record the pinned Hindsight core and harness-integration versions;
2. verify the actual executable/config paths without downloading a package as a side effect merely to run status;
3. declare API/database, extraction/reflect model, embeddings, reranker, and all provider egress independently;
4. define a worktree-aware per-repository or per-user bank map; use a shared bank only with mandatory source/repository tags and strict matching;
5. diff every hook, MCP, skill, or agent configuration mutation and keep an uninstall/rollback path;
6. pass readiness plus retain → recall → reflect smoke tests and record the recall-versus-reflect latency/cost difference;
7. pass invented cross-bank and cross-tag leakage cases with zero violations;
8. compare no-memory, raw-context, recall-only, and reflect arms under the same corpus, answer model, budget, and scorer; and
9. export a representative bank, restore it into an empty target, re-run frozen queries, verify provenance/scope, and prove durable corrections still promote through Kevin-Wiki owners.

Do not install or configure Hindsight merely because a source review finds it useful. Installation mutates the machine and agent harnesses; it belongs to the environment-bootstrap workflow and requires the user's requested target scope.

Decision Rules

• Prefer Mac Mini M4 Pro 48GB for the always-on local operator box; treat older Apple Silicon as a prototype lane.
• Keep model routing pluggable. Do not bake one model into the architecture without an eval or local availability check.
• Keep Hindsight and Honcho separate unless Hermes explicitly supports a combined provider mode. One active memory provider means one active slot.
• Prefer recall when the acting agent should reason over bounded evidence. Use reflect only for deliberate synthesis with citations, model/cost receipt, and an unavailable/degraded state.
• Treat facts, observations, and mental models as derived runtime memory. Preserve immutable source revisions outside Hindsight and require governed writeback before changing canonical behavior.
• Prefer one bank per user, authority, or repository. A shared bank must stamp provenance, use strict tag matching, and pass leakage tests; never infer isolation from “one local instance.”
• Treat local API transport, local inference, local embeddings/reranking, and local storage as separate claims. A loopback endpoint backed by a cloud model is not fully local.
• Treat export support as a capability, not recovery proof. Require a fresh-target restore and frozen-query replay before provider promotion.
• Keep qmd/Obsidian as the durable knowledge repository. Runtime memory is cache and recall until written back.
• Use hermes send or no-agent cron for script output. Use an agentic cron only when the message content requires reasoning.
• Ask before enabling account-mutating browser actions, adding new credentials, or scheduling unattended paid model calls.
• Do not repeat upstream scraping multipliers as local facts without the frozen corpus, comparator, versions, run artifacts, and reproduced distribution.