A sanitized, public view of 3B (Brandon's Binary Brain) — the agent harness behind this site. It is generated from the architecture model and shows the layers, subsystems, decision history, and blog series. It is not a mirror of the source: private content, paths, and session data are excluded by design.
overview
architecture map
Loading interactive graph…
- 1. Source of Truth (.agents/) 16 nodes
The single canonical store for all shared agent configuration and content stores — rules, skills, command-permission policy, buffer, friction log. Everything is authored here exactly once.
- 2. Generators + Cross-Agent Sync 17 nodes
Node generators that read frontmatter-gated SoT inputs and project them into per-agent profiles, permission files, and ignore lists; pre-commit/CI drift gates keep targets consistent.
- 3. Symlink Mounts + Agent Runtimes 9 nodes
Compatibility back-symlinks/hardlinks that resolve the SoT at each CLI's hardcoded mount, plus the three agent runtimes (Claude Code, Codex CLI, AGY) and the generated docs they load.
- 4. Content + Knowledge Lifecycle 15 nodes
The Zettelkasten knowledge base and the six-layer progressive-refinement pipeline that promotes ephemeral session work upward to durable, publishable entries.
- 5. Governance, HITL + Self-Improvement 8 nodes
The three-gate HITL/audit ladder, the immutable decision (ADR) registry, and the friction lifecycle that turns one-off mistakes into governed, measured fixes.
- 6. Tooling + MCP 10 nodes
The external code-intelligence and doc-conversion stack (graphify, code-review-graph, Serena, context-mode, markitdown) plus model-tier routing and the asdf/brew runtime contract.
layers
1. Source of Truth (.agents/)
16 nodesThe single canonical store for all shared agent configuration and content stores — rules, skills, command-permission policy, buffer, friction log. Everything is authored here exactly once.
2. Generators + Cross-Agent Sync
17 nodesNode generators that read frontmatter-gated SoT inputs and project them into per-agent profiles, permission files, and ignore lists; pre-commit/CI drift gates keep targets consistent.
3. Symlink Mounts + Agent Runtimes
9 nodesCompatibility back-symlinks/hardlinks that resolve the SoT at each CLI's hardcoded mount, plus the three agent runtimes (Claude Code, Codex CLI, AGY) and the generated docs they load.
4. Content + Knowledge Lifecycle
15 nodesThe Zettelkasten knowledge base and the six-layer progressive-refinement pipeline that promotes ephemeral session work upward to durable, publishable entries.
5. Governance, HITL + Self-Improvement
8 nodesThe three-gate HITL/audit ladder, the immutable decision (ADR) registry, and the friction lifecycle that turns one-off mistakes into governed, measured fixes.
6. Tooling + MCP
10 nodesThe external code-intelligence and doc-conversion stack (graphify, code-review-graph, Serena, context-mode, markitdown) plus model-tier routing and the asdf/brew runtime contract.
subsystems
Rules Engine + Routing
Frontmatter-gated rule files decide what loads — always-on vs lazy — per agent.
Skills System
Markdown-defined skills, signal-routed for Claude and mirrored to the other agents.
Cross-Agent SoT + Sync Engine
One canonical config store; a generator projects every edit into each agent profile.
Zettelkasten Knowledge Base
reads private storesAn atomic, forward-linked Zettelkasten knowledge base — backlinks computed, never stored.
Information Layer + Progressive Refinement
A folder-as-destiny pipeline that refines content through progressive layers.
Journal Rollup + Wrap Workflow
reads private storesA session engine: buffer to wrap to daily journal to knowledge, with clean-slate rollups.
Projects + Task Lifecycle + Worktrees
reads private storesTask setup gates work into isolated worktrees with explicit ownership and locks.
Governance, HITL Gates + Audit
A three-gate human-in-the-loop ladder governing every architecture-changing action.
Decision System (ADRs + Rule-6 READMEs)
An immutable ADR registry plus a per-folder decision-doc convention.
Friction Lifecycle
reads private storesCaptured friction accumulates into scored patterns and graduates into governed fixes.
MCP + Graph + Runtime Tooling
A layered code-intelligence and doc-ingestion stack, each layer owning one waste vector.
decision history
- 2026-05-30 ADR-039 Universal-Rule Native-Load Suppression — never-match sentinel on the 3…
- 2026-05-30 ADR-040 F2 Agent-Managed Memory Layer — split-tier candidate → HITL-promotion → tracked…
- 2026-05-29 ADR-038 Two-Layer Commit Discipline — branch WIP checkpoints (Layer 1) + atomic…
- 2026-05-27 ADR-036 Untrack — worktree project-name collision under the shared daemon (partially…
- 2026-05-27 ADR-037 Cross-Agent SSoT — Codex network-on baseline (the axis approval never touched)…
- 2026-05-23 ADR-035 Flat Worktree Paths — drop the level; canonical (supersedes ADR-026 on path…
- 2026-05-22 ADR-033 AGY Migration — Replace Gemini CLI with Antigravity CLI; D1-D5 + 6-phase…
- 2026-05-22 ADR-034 Close emit vs lint reformat drift loop (issue #134) — surgical wrappers around…
- 2026-05-18 ADR-031 Gates A+B Concurrency + HITL Protocols — 2 canonical rule bodies…
- 2026-05-16 ADR-030 3B Layered Self-Improvement Architecture (2-plane + L0-L6 + 9-step loop +…
- 2026-05-11 ADR-029 Information-Layer Schema Additive ( + + ) — Phase E1
- 2026-05-06 ADR-027 Per-Agent Content Filtering Convention (v0.1.0 markers ignored, v0.2.0 filtered)
- 2026-05-06 ADR-028 Initiative #51 (3B-redesign) Closure — phases shipped (D4 deferred outside…
- 2026-05-03 ADR-024 D2 concurrency: cross-journal carry-forward for + mkdir-lock
- 2026-05-03 ADR-025 Custom Cache Keepalive (Claude-only after scope narrowing)
- 2026-05-03 ADR-026 Cross-Agent Worktree Strategy (agent-namespaced path)
- 2026-05-02 ADR-022 Gemini Phase 3 completion — generator + markers + MCP + parity
- 2026-05-02 ADR-023 wrap cross-agent: single skill + agent detect + adapter pattern; HIGH bias #1…
- 2026-05-01 ADR-021 Shared Serena MCP daemon at (HTTP) for Codex + Gemini reuse
- 2026-04-30 ADR-020 3B partial-tracking: memories + project.yml in, cache out
- 2026-04-29 ADR-019 Graph-tool integration: graphify hooks + scoped rule + cross-agent propagation
- 2026-04-28 ADR-016 Cross-agent command permission policy generator
- 2026-04-28 ADR-017 claude-mem definitive removal (attempt #3) after silent billing leak
- 2026-04-28 ADR-018 Codex skill discovery boundary for source vs adapters
- 2026-04-27 ADR-014 Cross-Agent SoT Folder; Phase 1 + Phase 2 sub-phases) executed 2026-04-27
- 2026-04-27 ADR-015 Codex broader cross-agent sharing via routing + sync generator
- 2026-04-21 ADR-013 claude-hud v0.0.12 pin + three-defense pattern
- 2026-04-18 ADR-011 Retire Storm Brainstorm Engine (superseded by graphify)
- 2026-04-18 ADR-012 Per-project Claude rules in 3B subdirectory, symlinked
- 2026-04-17 ADR-009 Retire Agent Teams + Restore Symlink
- 2026-04-17 ADR-010 Retire Pattern A (Forge Theme Saturation) for Apply-over-Shelve
- 2026-04-09 ADR-008 RTK Hook Re-enable with Defense
- 2026-04-08 ADR-006 Storm Brainstorm Engine
- 2026-04-08 ADR-007 Claude-Forge Merge into 3B Workspace
- 2026-03-29 ADR-005 QMD MCP Server Removal
- 2026-03-21 ADR-004 RTK Hook Removal
- 2026-03-12 ADR-003 Empty Deny, All in Ask
- 2026-02-19 ADR-002 Regex Permission Hook
- 2026-02-12 ADR-001 Friction Feedback Loop
blog series
1 of 10 published
- 2. Rules That Route Themselves: Frontmatter as the Loaderplanned
- 3. Same Skill, Three Transport Physicsplanned
- 4. Forward Links Only: A Zettelkasten Where Backlinks Are Computedplanned
- 5. Folder Is Destiny: A Six-Layer Information Lifecycleplanned
- 6. The Session Engine: /wrap, Clean-Slate Rollups, and a Cross-Session Bufferplanned
- 7. Parallel Agents Without Collisions: Tasks, Worktrees, and Locksplanned
- 8. Three Gates and an Audit Log: HITL for an Agent Harnessplanned
- 9. A Harness That Fixes Itself — And Prunes Its Own Fixesplanned
- 10. The Token Stack: Four Layers of Code Intelligence Without Re-Scanningplanned
snapshot
Generated from the 3B architecture model. Sanitized — it contains no private files, paths, or session data.