Skip to content

ObneyAI/orc / orc-skills/orc-design-tree

orc-design-tree

Verified

Design a hand-assembled ORC behavior tree from a knowledge-work goal, including known specialist selection and revisiting, typed context, success conditions, control nodes and bounded recovery.

references/specialist-discovery.md

View on GitHub
# Discover registered specialists before dispatch

Use this when a capability catalog is too large for direct finite selection. For a small known roster, described options and semantic selection may be simpler; retrieval adds a recall failure boundary.

## Separate eligibility, retrieval and selection

1. Keep an authoritative registry of callable child sheets, described capability contracts, input compatibility, effect permissions and versions. A retrieved pattern or document does not register an executable specialist.
2. Apply host permissions and contract compatibility before a bounded shortlist can discard useful eligible candidates. If the search API cannot filter, rank a permitted catalog or use an explicitly bounded overfetch policy with a visible exhaustion outcome. Removing forbidden results only after a tiny global top-k can leave no valid choice even when useful capabilities exist.
3. Use retrieval as candidate evidence. ColBERT late interaction can rank described capabilities; graph traversal can expand known relationships. Record the actual query, index/model version, scores, seeds, edge meanings and expansion bound. Graph edges require a maintained meaning, not the desired answer disguised as a route.
4. Give the selector the original goal, relevant evidence and self-contained descriptions of offered eligible capabilities. Include an honest no-match outcome. Revalidate the selected registry identity and permissions before dispatch, and map typed inputs into the actual registered delegate.
5. Check the resulting artifact against the parent goal. Selecting a plausible specialist and completing a tick do not prove completion. Route unresolved work through the controller's declared recovery policy rather than falling through to an unrelated tool.

ColBERT retrieval, finite model selection, delegated work and parent acceptance are separate quality boundaries. Measure them separately. A capable alternative may complete a task despite missing the obvious specialist; that does not establish retrieval recall. Empty or poor shortlists are not proof that a decision model cannot route.

## Focused evidence from an authored consumer

A five-capability corpus ranked a forbidden publishing capability first for two read tasks. Filtering after top-one selection erased all eligible choices. Applying eligibility first fixed this starvation mechanism without changing task wording or descriptions. A fixed top-two shortlist still omitted the calculation capability for a numeric goal; ranking quality remained a separate concern.

Indexed and direct scoring matched exactly. Shorter query padding did not repair ranking; identical-description controls ranked correctly, with valid markers and dimensions. These observations did not establish an encoder or ORC implementation defect. Preserve the queries, descriptions, checkpoint identity and failed rankings before changing retrieval policy. Independent scorer comparison against the same checkpoint bytes is a useful next diagnostic when implementation correctness remains uncertain.

This small consumer does not establish scalable retrieval advantage, automatic ontology harvesting, dynamic capability registration or better task performance than direct selection. It uses explicit registered children and an application-owned relationship graph. Prior-pattern guidance for adaptive researchers is a different path: see [pattern reuse](../../orc-researcher-pattern-reuse/SKILL.md).
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall