BR self-describes: a machine-queryable, trust-scored Capability Graph (query, don't grep)

2026-08-12

apicapabilitiesmnicapability-graph

LOCKSTEP TRACEABILITY MATRIX --- api_endpoints: ["GET /v1/capabilities"] sdk_methods_updated: ["system.capabilities (generated)"] mcp_tools_updated: ["none (Phase-0 follow-on)"] ---

What We Built

Phase 0 of the ACI revitalization: BR now describes its own capabilities to a machine instead of making an agent grep the codebase to relearn the system every session. GET /v1/capabilities (filters ?query= / ?id= / ?tier= / ?status=) returns a structured Capability Graph — each node a typed contract (what / where / how-to / recovery / next-affordances) plus a trust+action vector (quality, reliability, audit-quality, anomaly-resistance, identity-stability, cost, reversibility-class + blast-radius, dry-run). Every score carries source: measured | declared | unknown so an agent can tell a live signal from a placeholder — nothing is fabricated.

Crucially, the seed promotes the rediscovered "forgotten" substrate to first-class — the exact capabilities the human master-capability-record.mdx had omitted: CAF agent identity (SPIFFE / CA / registry / cryptographic kill switch / anomaly engine), the A2A mesh (authorized proxy + hop-limits + mesh-graph), the Trust Envelope (signed scoped grants with a parent_chain delegation lineage — the mandate substrate), and credential custody (KMS-envelope BYOK vault + local keystore).

Why It Matters

The 2026 frontier calls this the Navigation Paradox: bigger context windows don't fix an agent missing architecturally-critical-but-distant knowledge — only _structural navigation_ does. A knowledge graph served over a machine interface cut agent token use ~10× and tool-calls ~2.1× in independent studies. This is BR turning its own Machine-Native Interface on _itself_: the map is an endpoint you query, not a doc you read. It also seeds the emerging capability trust layer (ERC-8004 / ACHIVX) with signals BR already produces (Thompson quality, governance audit-quality, CAF anomaly/identity).

How It Works

src/capability-graph/{types,seed}.ts hold the data model + hand-seeded graph (the scripts/build-capability-graph.ts generator that regenerates from code + ship-logs is the Phase-0 follow-on). src/api/capabilities/system/capability-graph.ts serves it as a sibling of system.discovery. AGENTS.md/CLAUDE.md and the auto-memory now point agents at the graph first.

Competitive Edge

No competitor exposes a self-describing, trust-scored graph of its own internals for an agent to query. It's the foundation for the ACI reframe — BR as the interface a frontier model _chooses_ to run on (self-describing → trust-scored → reversible/transactional → memory-backed → safe-to-act).

Lockstep Checklist

[x] API Routes: GET /v1/capabilities added as system.capabilities capability.
[x] TS SDK / Python SDK: generated from the capability registry (gen:contract), tagged system.
[x] MCP Schemas: br_capabilities_* tools are the Phase-0 follow-on (would shift the tool count; batched separately).
[x] Master Record: docs/architecture/master-capability-record.mdx gains a Tier-0 "Agent substrate" section + a "query, don't read" pointer to the graph.