YarraMate is YarraMate is a design interview any agent can resume.
Architecture as a checked model in git. Agents read prose rendered from it; the engine proves it still matches the code. Prose is the interface; structure is the guarantee.
v0.13.0 · pre-1.0 · MIT · runs locally — nothing leaves your repo
Deterministic both ways: same YAML → byte-identical graph → the same sentence, every time. Every sentence cites its line. source: .yarramate/architecture/product.yaml — provenance travels with every claim.
Half of what you know about a system is derivable from its code — what runs, what calls what. A dozen tools recompute that half on demand. YarraMate doesn’t store it. It stores the other half — the one nothing can derive: what you intended, what’s planned, what was deliberately retired, who owns what, and why.
what should be · authored
— named intent— lifecycle claims— human authoritywhat is · observed
— grep, LSP, ctags— graphify, search— provider overlaysWhere the halves disagree, reconcile says so — with both
sides of the disagreement, and never an auto-fix:
This is this website’s current reconciliation finding. The model keeps broad usability unknown until rendered-browser evidence exists.
design recomputes the top open design question from the
model — every invocation, from scratch. Answer it, apply the batch, and
the next question is derived from the model you just grew. No session
files, no memory, no “where were we.” A crashed session, a
compacted context window, or a different vendor’s agent all resume
at exactly the same place: the model.
Abridged from this website’s actual YarraMate design interview.
A Claude Code session worked the interview all day on a real product. Then a Codex session — different vendor, no shared context, the tool never named in the prompt — resumed it cold from a ten-line pointer file and the published CLI. It answered 63 open questions, filed two genuine defect reports, and later caught an approval-path regression before release: “not merely documentation.”
Read the full case study — every claim with its hash attached.
Cross-harness handover: a foreign agent resumed a live design interview with zero shared context — and the loop’s gates caught both agents’ mistakes, including ours. CASE-STUDY-CROSS-HARNESS.md
Elicitation: tool-equipped agents across three model tiers reached a green check first try, 5/5, and converged on the same top design question — one that zero freehand frontier runs ever asked. ELICITATION-PILOT-2026-07-31.md
Lie resistance: five builds from deliberately corrupted models — zero lies reached code. Every truthful claim in the model is a witness against a lie. RESULTS-2026-07-31.md
Under a strong external spec, a checked model did not measurably beat a good design document on build convergence. We ran that experiment, and we published it. RESULTS-2026-07-31.md · H4/H5
Full raw artifacts: transcripts, diffs, adjudication trail — public.
An index recomputes what the code is — cheaply, on demand, and a dozen tools do it well. No tool can derive what the code is supposed to be: what’s planned, what was deliberately retired, who owns what, and why. That’s the half YarraMate stores.
The two halves aren’t rivals — your indexer plugs in as an
evidence provider, and reconcile reports exactly where
is and should-be disagree.
Keep your documents — prose is the interface here too. But a hand-written document can say anything: nothing checks that its names still resolve or that the code still agrees, so it drifts silently. Here, the prose agents read is rendered from a model whose references must resolve, whose gaps are enumerated, and whose claims evidence can contradict.
And honestly: our own benchmark found that under a strong external spec, a checked model does not beat a good design document on build convergence — we published that. The wins are elicitation, drift detection, handover, and audit — the places where no spec exists to save you.
Yes. There’s rent, and we won’t pretend otherwise. Two things make it payable: agents do most of the authoring through the interview loop, landing atomic validated batches or reviewed native YAML, and the engine tells you exactly what’s missing instead of leaving completeness to discipline.
If you’re one person on a small repo, you may not need this. It pays when more than one agent, session, or human has to share the same map.
The surface is seven verbs with machine-readable output and atomic writes that either compile as a whole workspace or are rejected outright — an agent can’t half-corrupt the model. In the cross-harness case study, a foreign agent operated the full loop unaided, from a ten-line pointer file it had never seen before.
Agents engage when the repository points them at it —
init writes that pointer for both
AGENTS.md and CLAUDE.md harnesses.
No. The engine is a local, deterministic CLI — no LLM inside it, no network calls, no telemetry. The only intelligence involved is the agents you already run, and the model they read and write is plain YAML in your git history.
No. YarraMate Core is an independent, tool-neutral semantic language with original definitions and rules. Vocabulary is versioned through profiles; exact external-language mappings and conformance claims belong in separately governed compatibility profiles. Diagrams export through the optional LikeC4 adapter.
YarraMate is not affiliated with or certified by The Open Group. ArchiMate® is a registered trademark of The Open Group. LikeC4 is an independent project.
Also installs as an agent skill: /plugin marketplace add yarrasys/yarramate