Trellis Engine

Trellis is an agentic harness that lets a language model work over a corpus far larger than its context window, and keeps what the model works out. It is OpenCnid's Recursive Language Model runtime: the model reaches its data by writing code in a persistent Python REPL and calling itself over slices, instead of reading everything into context, where context rot sets in. On a twenty-query benchmark it scored F1 1.000 on every query, twice over in independent runs: the opening queries paid for classification, the rest answered from cached facts with no sub-LLM calls at all, for $0.81–$0.87 across the whole run.

Kept is not the same as trusted. Source bytes are immutable content-addressed blocks; derived beliefs carry the exact block hashes they came from, so any claim traces to the bytes behind it. Nothing moves up a trust tier without a human running a gated promotion command, and when a source changes, a Merkle diff marks every belief that depended on the dead bytes contested — quarantined and auditable rather than quietly wrong.

The same rule turns on the system itself. Trellis writes its own extensions as modules, and each one lands the way a belief does: carrying research provenance, gated by an operator, and contested by that same sweep when the evidence under it dies. Two modules are active today, one has been retired, and one is contested — a capability here is held to the standard its knowledge is.

Agents, LLMs, and coding harnesses start at AGENTS.md. It carries the project basis, the annotated file tree, and an index that fans out to AMBIENT.md, which binds every session whatever the task, and to .claude/rules/, where each file carries the rules that fire on one kind of work.

System shape

1. PostgreSQL stores immutable content-addressed AST nodes and document-version membership.
2. Neo4j stores semantic beliefs carrying sourceNodeIds back to exact source blocks.
3. Redis and BullMQ isolate asynchronous extraction, verification, invalidation, resolution, RLM, and orchestration work.
4. The Python RLM harness exposes bounded database, workspace, MCP, editing, and by-reference answer surfaces inside a persistent REPL.
5. Promotion is the only Tier-3-to-Tier-1 bridge; provenance is enforced at the write path.
6. The model never counts and never copies: locations are engine-computed and existing bytes move through guarded code operations.

Fast path

Prerequisites are Node.js 22.18+ or 24.11+, Python 3.11+, and Docker Compose v2.

npm ci
cp .env.example .env
docker compose up -d postgres neo4j redis
npm run db:init:dev
npm run dev

Verification, all of it offline — no database, no containers, no network:

npm run check:repo-surface
npm run wiki:check -- --verify
npm test
npm run build
npm run python:check
npm run test:textedit
docker compose --profile test config --quiet

python:check and test:textedit run against the bare-host Python runtime, which the operator manual below installs.

Where to go next

You want to	Open
orient at the shallowest depth that answers you	docs/ORIENTATION.md — the whole system five times over, D0 one sentence through D4 the index
settle what a load-bearing term means	docs/GLOSSARY.md
find the record that governs a topic	docs/README.md
check a measured claim, and read what it does not prove	docs/benchmarks/ — dated reports, with CRITIQUE_AND_FUTURE.md as the honest ledger
deploy, operate, or recover the stack	docs/operations/OPERATOR_MANUAL.md and docs/operations/RUNBOOK.md
call the HTTP and SSE surfaces	docs/reference/API_REFERENCE.md
contribute code	.github/CONTRIBUTING.md

The project is licensed under the MIT License.