Primitives
45 one-of-a-kind primitives
Architectural primitives that appeared in exactly one (or two tightly-related) frameworks across the entire 654-framework corpus. Extracted verbatim from the Phase E executive synthesis. Each represents a genuinely novel engineering decision — not a variation on an existing pattern.
Source: _index/MASTER-BY-DIMENSION-v2.md §19
HTTP-proxy LLM-call interception with hallucination-detection certificates
entroly runs on :9377, compresses context via knapsack DP (Rust), attaches WITNESS NLI certificates (AUROC 0.80 on HaluEval-QA). The only memory-layer framework that sits below the agent rather than beside it.
Auto-backpropagation of runtime failures into the spec
PostToolUse hook detects runtime errors and adds R-numbered acceptance criteria + regression tests back into the originating spec. Living spec, not authored spec.
LLM-evaluated hooks ("type":"prompt")
First surfaced in cc-spec-driven, devteam-michael-harris. Phase D revealed 5+ independent implementations — now its own archetype A21.
Sentinel-driven state machine in stdout
AI emits ✅ Ready / ⛔ Blocked markers; PostToolUse hooks parse stdout into review-gate state.
Ralph-loop XML completion contract
Machine-detectable completion signal embedded in natural-language output: <promise>TASK_COMPLETE</promise>. Independently implemented in ralphy-openspec, cursor-rules-pack-v2.
Adversarial duel review
Two agents independently review the same implementation, then cross-score each other. The only formal adversarial-review protocols in the corpus.
Fuzzy Linear Temporal Logic monitor
Formal LTL ordering contracts at 0.139 ms p50.
Parallel LLM-powered execution sentinels
Asynchronous LLM observers tap the event stream, detect stuck states, trigger recovery.
"Think in Code" — agent writes analysis scripts instead of reading data into context
100× context savings vs any other technique.
JIT semantic context loader via MCP resources
1,675 useWhen scenarios + 4,036 keywords + LRU cache deliver 95–98% token reduction.
TESTABILITY pre-implementation Opus veto
A test-designer Opus subagent returns PASS/WARN/FAIL on the Plan agent's class design before any code is generated. Only example in the corpus of structured design-time veto between agents.
Stop-hook drift-log compliance gate (LLM-on-LLM at Stop)
The Stop hook itself ships session context to another LLM that evaluates whether the agent obeyed its drift log; if not, it blocks completion. The most aggressive use of LLM-evaluated hooks.
Compiled Go binary "don't self-approve" code invariant
The "don't mark your own work done" rule is a Go compile-time check rather than a prompt, removing the possibility of the agent talking itself around it.
PreToolUse banner injection as orchestrator→subagent message bus
Hooks (PreToolUse, UserPromptSubmit, SessionStart) inject [EVO DIRECTIVE] banners that act as a downward communication channel to in-flight subagents. The only framework treating hooks as IPC.
Karpathy autoresearch loop (empirical SKILL.md optimization)
The framework writes PRs against itself to optimize its own skill prompts empirically. Crosses Phase C's ag-coding-flywheel adversarial-review into closed-loop self-modification.
Compiled YAML policy → PreToolUse hooks (5-level escalation)
Policy DSL compiles into a hierarchy of PreToolUse hooks with progressive enforcement: warn → block → quarantine → revert → escalate. The only structural escalation pattern in the corpus.
Loop-detection BREAK-LOOP PROTOCOL
When Claude repeats the same (phase, nextAction) 3+ times, the framework injects a numbered escape menu: try alternative / escalate / upgrade strategy. The only meta-cognitive self-recovery primitive in the corpus.
8-dimension partnership measurement + wave-based file ownership
Quantitative agent-human partnership scoring across 8 dimensions, with file ownership rotated across waves so no single contributor (human or AI) dominates a region.
NixOS atomic rollback + AI root access + post-quantum mesh
Give the AI full root, but make every commit a NixOS generation that can be rolled back atomically. Only OS-distro framework in the corpus.
PreToolUse-on-Task hook for sub-agent context injection
When the parent agent calls Task to spawn a sub-agent, the PreToolUse hook injects curated JSONL spec files into the sub-agent's prompt. "Specs injected, not remembered."
Self-evolving plugin meta-linter (auto-PRs new rules)
Detects new failure patterns and auto-PRs new lint rules into the framework itself. The closest the corpus comes to a self-modifying agent runtime.
Hard USD threshold enforcement
Actually stops execution at a configurable USD cap rather than warning. Only example in the corpus.
Compose-lang: framework that only defines a new language
Ships a DSL spec and a parser, no agents, no skills. Meta-level "what if the framework is a syntax?" answer.
Only true peer-to-peer swarm
Agents handoff to each other without a central coordinator; swarm pattern in literal sense. Phase C's claude-flow was queen+workers, not P2P.
Claude-Code-as-containerized-executor inversion
A Go binary embeds V8 and containerizes Claude Code itself as a capability being orchestrated rather than the orchestrator. Inverts the normal relationship.
~75% token reduction via symbolic encoding + 6-step backprop
Caveman notation §G/§C/§I/§V/§T/§B compresses spec text by ~75%; every test failure backprops into a §B entry and §V invariant.
Sonnet-classifier-at-UserPromptSubmit shaping response mode/tier
The UserPromptSubmit hook fires a Sonnet classifier on the user prompt itself and selects which response tier/persona Claude operates under. Mode-aware agent.
88% token efficiency replacing MCP tool-calling with code-exec
Instead of MCP tool calls, the agent generates code that calls a library. 88% fewer tokens than equivalent MCP usage.
LLM-evaluated Stop hook
The Stop hook ships full session context to an LLM that returns block/continue.
Bug Council 5-analyst parallel escalation
"type":"prompt" hooks with 5-analyst parallel escalation and SQLite per-task model history.
ISA-95 industrial-controls methodology applied to AI workflows
Borrows enterprise integration standard from manufacturing.
27-tool MCP self-awareness server
The MCP server's tools are tools to inspect the running MCP server itself. Meta-introspection.
50+ module-scoped CLAUDE.md files
Pushes Claude Code's CLAUDE.md hierarchy further than any other framework — one per module, scoped by directory.
LSP linter with 423 validation rules
The only LSP server in the corpus dedicated to linting agent context files (CLAUDE.md, AGENTS.md).
AGENTS.md ecosystem-win over AGENT.md (270× stars)
Demonstrates that naming convention adoption can win despite technical inferiority. The community converged on AGENTS.md (plural) even though AGENT.md (singular) had slightly better technical design.
state:agent-ready self-application prohibition
The agent cannot set its own ready state; only an external signal can.
Cross-vendor model routing (Claude→GPT-5/Qwen)
7+ frameworks now route across competing vendors; A16 archetype. ccg-workflow, droids-curtis, switchboard, octomind-muvon, golembot.
Auto-merge-on-CI-pass
If CI passes, the PR auto-merges. The only framework with merges_automatically: yes plus CI gating.
Self-hosting / self-developing
The framework is used to develop the framework itself, dogfooding the full SDLC loop.
policy-as-code DSL (safe-outputs separation)
Agentic workflows as policy: every write goes through a separately sandboxed job; safe-outputs separation. The most security-conscious framework in the corpus.
Tap registry pattern (community specialists as one-file TOML manifests)
octomind tap <github-repo> installs a community agent like a Homebrew tap.
Epistemic lease / decay model
Cross-module trajectory knowledge decays over time and must be renewed via 4 escalation levels.
13-tier cost-based model routing
T0 local free → T12 Opus $25/M, with iCPG and Telos IFS = F1×F2×F3 quality scoring. Most elaborate multi-model config in the corpus.
rtk-token-killer (54k★ beating every orchestration framework)
Single-purpose token-efficiency tool. Proves the community values cost optimization over orchestration.
Cross-vendor routing as visual kanban
Routes Gemini Flash for cheap tasks, Claude Opus for expensive ones, via a VS Code kanban webview.