23 · Glossary & Sources
Glossary
A2A (Agent-to-Agent). A protocol for agents to call each other. A generated native agent advertises an agent card at /.well-known/agent.json and accepts tasks at /a2a/tasks.
A2UI. The Prometheus combined agent-and-UI protocol, served at /a2ui/session, that streams both agent interaction and UI updates.
AG-UI. A CopilotKit-compatible protocol for streaming agent runs to a UI over Server-Sent Events (agui.* events). Drives the generated React frontend.
AgentSkills.io. The portable skill specification every native skill in the pack conforms to. Defines the SKILL.md frontmatter schema, the name pattern, and the standard directory layout.
Cedar. The policy language used by the Skill-Mutation policy enforcement point. Default-deny; governs skill.mutate, skill.generate, skill.promote, and trace.capture per environment.
Constitution. A per-language set of standards and denied patterns that forge-rs checks code against and injects into enrichment context (constitution_summary).
Constraint manifest. The output of the ZeeSpec interrogator — a scored GO / CAUTION / NO-GO assessment of whether work is specified well enough to proceed.
evolver-bridge.json. The write-back contract between the KBD inner loop (L1) and the evolver (L2). Maps completed changes to evolution items so the strategic loop knows what landed.
forge-rs. The Layer 4 code-enrichment engine. Injects language knowledge before an agent writes code; processes reflections back into the Karpathy loop.
Karpathy pattern. A flat-file Markdown knowledge base, maintained by the AI and queried directly by a long-context model, used instead of vector RAG. Every claim traces to a readable file.
KBD (Knowledge-Based Development). The Prometheus methodology that keeps domain knowledge and code aligned across sessions, via KB priming, hard phase discipline, and waypoint continuity.
KBD control token. A project-scoped, mode-0600, URL-safe bearer token
used only for explicit loopback-TCP Sovereign Sync authentication. The default
local API uses same-user Unix-socket credentials. This token is not a P2P group
secret or an Ed25519 device key.
KbdStateV2. The canonical replayed state containing lifecycle, plan, active phase/stage/change/task path, completion dimensions, decisions, blockers, devices, and commands.
liter-llm. The multi-provider LLM gateway (140+ providers) that does per-phase model routing. Frontier models for reasoning phases, cheap models for status.
Loop levels (L0–L3). L0 harness micro-loop, L1 tactical KBD loop, L2 strategic evolver loop, L3 outer standing loop. State is harness-agnostic; the driver is harness-specific.
loop.json. The definition of an L3 standing loop: goal, feedback sources, termination, cadence, and the backing evolution name.
loop-tick.sh. The runner that advances one outer-loop tick. Exit codes: 0 continue, 1 escalate, 2 terminate, 3 error.
Metaprompting. Designing a system of prompts — routing, critique, evaluation — rather than a single prompt. The basis for critic-context isolation.
PMPO (Prometheus Meta-Prompting Orchestration). The two-loop methodology (inner task loop + outer evolution loop) whose load-bearing constraint is phase discipline.
prometheus-knowledge / pk. The Rust implementation of the Karpathy KB.
Hooks read bounded committed project/shared/global snapshots; the worker ingests
queued learning and reconciles receipts. MCP bridge pk-cherry serves the API.
Progress signals. Mandatory start/completion lines emitted every phase and task, with counts read from progress.json, so multi-session work stays resumable.
Project manifest. .prometheus/project.json, the immutable UUID and
repository fingerprint used to name a canonical KBD runtime.
Compatibility projection. Revision-stamped files such as
progress.json, current-waypoint.json, and position.json rendered from
canonical replay for older readers. They are not mutation authority.
Sovereign Sync group secret. A random 256-bit value persisted in the private P2P identity and transferred only through pairing tickets. It derives the gossip topic and is separate from endpoint, KBD device, and TCP-token identity.
Sycophancy patterns (S-01–S-08). The eight classified patterns the correction server detects, from unprompted affirmation (S-01) to reflect-phase inversion (S-08).
surreal-memory. The semantic knowledge graph (SurrealDB + HNSW) on port 23001. Stores relationships, scoped memory, task streams, and mindmaps.
Waypoint. The resume contract in .kbd-orchestrator/current-waypoint.json / position-reminder.txt that restores an agent's exact position at session start.
ZeeSpec. The Zachman-Framework 5W1H interrogator that gates under-specified work at Layer 1.
Sources
External claims in this guide are validated against the following sources.
Loop engineering and Boris Cherny
- The Anthropic leader who built Claude Code says he ditched prompting — now he just writes loops — The New Stack
- Loop Engineering — Addy Osmani
- Loop Engineering — Cobus Greyling
- Claude Code Loop Engineering: Stop Prompting, Start Designing Autonomous Agent Workflows — TechTimes
- Boris Cherny, Claude Code creator, on shifting from prompting to autonomous loops — Digg
Karpathy knowledge-base pattern and minBPE
- Karpathy shares 'LLM Knowledge Base' architecture that bypasses RAG — VentureBeat
- What Is the Karpathy LLM Wiki Pattern? — MindStudio
- LLM Knowledge Bases — DAIR.AI Academy
- karpathy/minbpe — minimal Byte Pair Encoding for LLM tokenization (GitHub)
Kimi K2.6 / K2.7 Code
- Moonshot AI Releases Kimi K2.7-Code: +21.8% on Kimi Code Bench v2 over K2.6 — MarkTechPost
- Moonshot AI's Kimi K2.7-Code Targets Token Efficiency in Agentic Coding — DevOps.com
- Kimi K2.7 Code: Open-Source Agentic Coding Model — Moonshot AI
Firecrawl vs. Tavily
- Firecrawl vs. Tavily: 2026 guide for RAG and agent pipelines — Apify
- Firecrawl vs Tavily: Complete Comparison for AI Agents & RAG (2026) — Firecrawl
- firecrawl/firecrawl — AGPL-3.0, self-hostable (GitHub)
Metaprompting and critic models
- Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding — arXiv 2401.12954
- The Meta-Prompting Protocol: Orchestrating LLMs via Adversarial Feedback Loops — arXiv 2512.15053
- Meta-Prompting: LLMs Crafting & Enhancing Their Own Prompts — IntuitionLabs
Internal references
The authoritative internal sources are the repository itself: README.md, SKILLS.md, CLAUDE.md, docs/loops-architecture-spec.md, docs/CONTRIBUTING.md, docs/SUBMODULES.md, docs/IMPORTING_SKILLS.md, the article docs/articles/autonomous-loops-prometheus-skill-pack.md, each skill's SKILL.md, each tool's Cargo.toml and source, hooks/hooks.json, .claude-plugin/plugin.json, marketplace/marketplace.json, scripts/mcp-port-table.json, and the sycophancy-correction server's own skill_info output.
Previous: ← 22 · Advantages & Impact · Back to Index