Add 20 patterns closing the agentic-design.ai catalog gap

[arst]

Adds the 20 genuinely-missing architectural patterns identified against agentic-design.ai, deliberately excluding its training methods (RLHF/DPO/ORPO/SimPO), model-internals entries (KV-cache, Infini-Attention, latent recurrent thinking) and the 32-entry UI/UX catalogue. Those are legitimate topics that would drift this repo from "an agentic pattern plus a small .NET implementation" toward "every technique associated with AI systems".

All Agent-Framework-only, matching every recent addition.

Reasoning & generation

Pattern	What it demonstrates
ChainOfVerification	Verification questions answered with the draft out of sight
ProactiveClarification	Screened questions, one round, then stated assumptions
GraphOfThoughts	Host-owned DAG, so two lines merge instead of one being pruned
LeastToMost	Ordered subproblems, earlier answers carried forward as facts
StepBack	Principle first, with the question's numbers withheld

Orchestration

Pattern	What it demonstrates
MixtureOfAgents	Layered proposers; layer 2 reads all of layer 1
StateMachineAgent	Host-owned transition table; the model decides only within a state
EventDrivenAgents	Topic subscriptions instead of an orchestrator, generation-capped bus
ControlPlaneAsTool	One execute_capability tool over many backends
AgentRegistry	Capability discovery with signed cards verified before dispatch
SpeculativeToolExecution	Read-only, free-to-discard calls started before the model asks

Knowledge & state

Pattern	What it demonstrates
ContextAssembly	Pinned-first, deduplicated, budgeted; drops carry reasons
MultiSourceContextFusion	Trust then recency, contested fields surfaced not hidden
GraphRAG	Entity graph + community summaries for corpus-level questions
MemoryConsolidation	Recency/importance/relevance retrieval; ripe topics collapse

Production controls

Pattern	What it demonstrates
DualLlm	CaMeL: untrusted content supplies values, never control flow
HumanOnTheLoop	Autonomous by default; silence is not consent when irreversible
MemoryPoisoningPrevention	Untrusted sources propose, corroboration publishes
ContrastiveExplanation	Flip condition re-run against the rule before it is shown
AgentCommunicationFaultTolerance	Retry, dedup, dead letters, reconciliation

Verification

• Solution builds clean, 410 tests pass (80 new). Tests cover only the deterministic host logic — no model calls.
• All 20 samples were run against a live Azure OpenAI deployment. That found five defects compiling could not:
  • StateMachineAgent filled the missing cost centre after asking the model, so the claim died at step 2. Fixed — then a run through Pattern Explorer showed it still intermittently rejected the claim, because the stale Insufficient entry in the fact log misled the model at temperature 0. Now each state carries a one-line brief saying what it decides; stable across three runs.
  • AgentCommunicationFaultTolerance exercised neither its dead-letter nor its dedup path on the old seed — and an absorbed duplicate leaves no trace by construction. New seed (found by simulating the RNG draw order), a DuplicatesAbsorbed counter, and an explicit resend demonstrating the ack-lost case.
  • GraphRAG fragmented into four communities on entity-name drift (checkout vs the checkout service), so the cross-document theme never formed. Extractor now pins naming; manual rollback correctly links INC-101 to INC-104.
  • GraphOfThoughts scorer returned ~0.95 for everything, making Best() a coin flip, and the six-sentence brief was ignored. Length folded into the rubric.
  • ControlPlaneAsTool never reached its own denial because the model refused unprompted — a courtesy, not a control. Now calls the ungranted capability directly.
• Pattern Explorer needed no code change (the catalog globs patterns/*.md; all four categories already exist in CATEGORY_ORDER). Verified against a running instance: 74 patterns served, source viewer resolves the new files, interactive picked up for HumanOnTheLoop and ProactiveClarification, and a live run streamed end to end.

Two deliberate ceilings, both marked with ponytail: comments naming the upgrade path: GraphRAG uses connected components rather than Leiden, and AgentRegistry uses HMAC with one shared registry key rather than asymmetric per-agent signing.

🤖 Generated with Claude Code

https://claude.ai/code/session_0161UFxvL3zPhufoYaQh27Ss