Agent Registry & Capabilities
Relevant source files
The following files were used as context for generating this wiki page:
Introduction
The Agent Registry & Capabilities system within IDAHO-VAULT defines the multi-agent ecosystem designed to assist in journalism research, repository maintenance, and data orchestration. The system operates under a strict hierarchy where Logan (the human principal) remains the sole authority, and agents serve as specialized software tools with varying tiers of access and responsibility.
The registry facilitates coordination through defined persona dotfolders, capability tiers, and communication protocols. It ensures that every agent—whether it has direct repository access or operates externally via chat—follows non-negotiable core principles, such as maintaining a public record and respecting human-directed execution.
Sources: CONTEXT-PASTE-BUNDLE.md:46-47, READY-STATE-2026-03-23.md:37-38, _353main.md:23-25
Agent Classification & Tiers
Agents are categorized by their Capability Tiers, which dictate the extent of their interaction with the vault repository. These tiers range from read-only advisory roles to high-level repository management.
Capability Tiers Matrix
| Tier | Access Level | Description |
|---|---|---|
| PERMANENT: AUTHORITY: CODE | Direct Write | Responsible for repository operations, automation, and terminal mechanics. |
| PERSISTENT: ADMINISTRATION | Draft Only | Provides constitutional and governance support; cannot commit directly. |
| GitHub Copilot (ADMIN GitHub) | Multi-repo Admin | Manages PRs, GitHub maintenance, and multi-repo orchestration. |
| ADVISORY | Read/Analysis | External agents without repo access; produce drafts for Logan's verification. |
Sources: CONTEXT-PASTE-BUNDLE.md:89-98, READY-STATE-2026-03-23.md:54-61, _353main.md:15-17
Role vs. Persona Decoupling
The vault enforces a strict decoupling of agent identity variables to prevent software calcification into permanent authority.
- NAME: The vendor/model instance (e.g., Claude Code).
- OFFICE: The functional position (e.g., Code Authority).
- TITLE: The symbolic moniker used in handoffs (e.g., The Abhorsen).
- ROLE: Transient operational descriptor (e.g., Executor).
Sources: VAULT-CONVENTIONS.from-pushable-2026-05-08.md:387-402
Multi-Agent Ecosystem Workflow
Coordination among agents is facilitated through GitHub Issues and Pull Requests. Each agent typically works on its own branch, and PRs serve as the primary deliverable for human review.
Interaction Flow Diagram
The following diagram illustrates how a user request is routed through the swarm to the repository.
A brief explanation of the flow: Logan assigns tasks via labeled GitHub Issues; agents work in isolated branches and submit PRs, which must be manually approved and merged by Logan.
Sources: CONTEXT-PASTE-BUNDLE.md:100-101, _353main.md:162-167, VAULT-CONVENTIONS.from-pushable-2026-05-08.md:368-372
Registry Implementation
The system uses specific configuration files and directory structures to track agent state and instructions.
Core Registry Files
agents.json: Tracks every active node including name, role, and last-seen status.AGENTS.md: The root cross-tool pointer and narrative agent roster.swarm.json: A machine-readable registry and promotion ruleset for the broader swarm..claude/CLAUDE.md: Operational instructions specific to the Claude implementation.
Sources: READY-STATE-2026-03-23.md:46-47, _353main.md:183-185, VAULT-CONVENTIONS.from-pushable-2026-05-08.md:656-658
Persona Dotfolders
Each agent is assigned a "body" or persona dotfolder (e.g., .claude/, .gemini/) which serves as its personal continuity surface. These folders contain three types of surfaces:
- OWNER: Writable by the specific agent only.
- SHARED: Protocol surfaces other agents may write to under explicit contract.
- ARCHIVE: Read-only historical memory.
Sources: VAULT-CONVENTIONS.from-pushable-2026-05-08.md:92-96, VAULT-CONVENTIONS.from-pushable-2026-05-08.md:133-145
Operating Rules for Agents
Agents are governed by the Standing Engine axes, which provide epistemological operating rules to ensure truthfulness and restraint.
| Axis | Operating Rule |
|---|---|
| Truthfulness | Report what is actually present; training-data patterns are not emanation sources. |
| Provenance | Always show where a claim originated (e.g., "I read X.md"). |
| Restraint | Stop before touching non-delegated surfaces. |
| Handling | Place output in the correct surface (PR, staging, etc.) without premature promotion. |
| Jurisdiction | Act only within delegated scope; do not counterfeit authority. |
Sources: _353main.md:103-120
External Agent Orientation (LEVELSET)
External agents without repository access must undergo a LEVELSET initialization. This requires the agent to answer a six-point questionnaire:
- WHO YOU ARE: Name and platform capability.
- WHAT YOU KNOW: Explicit list of pasted context absorbed.
- WHAT YOU'VE DONE: Summary of current session output.
- WHAT IS UNRESOLVED: Pending task ambiguities.
- WHAT YOU NEED: Additional required decision/excerpts.
- COLLISION RISKS: Potential for conflicting changes or stale assumptions.
Sources: CONTEXT-PASTE-BUNDLE.md:21-25, READY-STATE-2026-03-23.md:73-77
Conclusion
The Agent Registry & Capabilities system serves as the foundational governance layer for IDAHO-VAULT’s silicon workforce. By enforcing strict capability tiers, identity decoupling, and standardized orientation protocols, the system ensures that multi-agent assistance remains auditable, secure, and always subordinate to human direction.