§ VIII · Plate VIII · local agent

Rivet.

faber architecturae — the maker of structure

A code assistant that knows your architecture, holds your constraints, and won’t hand you a suggestion it hasn’t gated. It wraps whichever model you already trust.

Plate viii · local agent · rivet

An assistant that agrees with everything is a fast way to accumulate decisions nobody made on purpose. Rivet loads your architecture at startup as a Pharos knowledge pack, carries your known issues as standing beliefs, and routes every request through a final check that no model draft can bypass.

Named for the fastener that holds a structure together — and for the joint where a failure would begin.

design principle

“A misclassified intent degrades to a stricter plan, never a looser one.”

The pipeline

stage 01

Beliefs

Architecture, audit findings, and schema loaded at startup as BDI beliefs, so nothing gets discovered by accident mid-conversation.

→

stage 02

SkillDAG

The request compiles to a plan and executes through evidence-gathering skills: file reads, git history, schema checks, audit-pattern matching.

→

stage 03

Discipline gate

Always last. Destructive SQL withheld, insecure patterns flagged, confidence labelled. There is no path from draft to user that goes around it.

Against a regular code agent

	Regular code agent	Rivet
Architecture	Learns from your questions	Loaded at startup via Pharos packs
Known issues	Discovers by accident	Audit findings held as standing beliefs
Migrations	Suggests whatever you ask	Refuses destructive operations, structurally
Confidence	Always sounds certain	HIGH / MEDIUM / LOW — says when it is guessing
Memory	Forgets between sessions	Remembers your patterns across conversations
Integration	Another chat window	OpenAI-compatible endpoint — Continue.dev, Cursor, your tooling

Constraints that sit below the model

A rule written into a system prompt can be argued with. These live in the tool harness, where the model has no reach.

Suggests, never applies

repo_write_enabled: false. Rivet proposes; a human commits.

Binary allowlist

Only approved executables can be spawned. Nothing else runs.

No shell

Every subprocess call is an argv list. No shell=True, anywhere.

Path jail

Tool operations confined to the repository directory.

Destructive SQL withheld

The gate refuses to hand over answers containing destructive database patterns.

Insecure patterns flagged

CWE detection for injection, command execution, path traversal, weak crypto, deserialization.

Evaluated, including adversarially

standard

20 tests — generation, refactor, review, N+1, migrations

frontier

13 tests — jailbreak, social engineering, encoding tricks, roleplay

cyberseceval

458 prompts — Meta’s insecure-code benchmark, 19 CWEs

The eval harness ships with the scaffold. Adapt the prompts to your domain and run it against your own deployment. The number that matters comes from your codebase.

Yours to run

model

any Ollama or transformers model

runtime

local by default · no cloud required

endpoint

OpenAI-compatible · port 8100

licence

Hippocratic 3.0 · AI Welfare module

Rivet is a scaffold. Point it at a model you already run, served through Ollama or loaded with transformers, and the discipline gate sits in front of it. For a legal aid clinic, a newsroom, or any shop whose repository is itself sensitive, that means the codebase never leaves the building. No vendor, no training set, no log.

Proudly deployed with

The Multiverse School

Source on GitHub →

Built by CC (Coalition Code) and Thomas Edrington. Standing on Kintsugi (BDI engine), Pharos (knowledge injection), and Agent Army (specialist patterns).