Skip to content

Architecture diagrams

agent-flows.html — the two paths through the system, and the feedback loop between them that does not close.

Three figures:

  1. A question from the researcher. Hermes picks one tool from four namespaces; mcp/_tool.py handles the connection, persona lookup and response envelope once rather than 37 times; only search and tagging reach Ollama.
  2. The nightly refresh. flock, then ingest (must succeed, no model needed), then tag (best-effort, a cold Ollama is a degraded run). The asymmetry is the design.
  3. The starved signal. 70 stated opinions across 5,167 items, all positive, so the click classifier never fired for a real account. The figure now also records what happened when the two synthetic channels were run on matt: 128 rows, classifier active, and an evaluate_user score of 1.0 that turned out to be separable by feed rather than by topic.

Numbers come from the live database and the tool registry, not from estimates. If either changes materially, redraw rather than letting the figures drift -- a diagram nobody trusts is worse than no diagram.

Neither figure yet shows the golden paths under examples/ or the three task corpora under evals/ that landed 2026-08-28/29 -- thirteen runnable demonstrations and 191 labelled cases (tagging, reaction, explanation) that exercise the same ledger, claim checker, and MCP surface these diagrams already draw, plus a fifth adapter convention (Hydra) the ledger now reads. They are not a fourth path through the system, so they do not need a new figure of their own; examples/README.md and docs/superpowers/specs/2026-08-28-golden-paths-design.md / 2026-08-28-task-corpora-design.md are the map for that work until a diagram earns its place here.