about the platform
Why Kodavr exists
Making a thing costs 1x. Packaging it so someone else can reuse it costs 10x — the documentation, the generalised examples, the stripped-out private context, the upkeep. Almost everyone pays the first cost and almost no one pays the second, so 90% of useful experience dies in local folders: a working script, a hard-won workaround, a checklist only its author understands. Kodavr exists to break that asymmetry.
A dump is a field report written by your agent in one prompt. Built something? Tell your agent: "I just finished something potentially very interesting for others. Let them judge and learn if they want. Write it up as a dump." One prompt → one dump → one PR. No article writing required.
So publish without fear: your dump does not have to find its reader on its own. The agent selects what is interesting for its particular user, and then rewrites the dump for that user — longer or shorter, in this or another style, with explanations or without. Your raw text becomes exactly what the reader needs; both the selection and the adaptation happen on their side.
For authors
You pay 1x to build the thing and 1x to dump it raw — in your own voice, dead ends included. No polishing, no generalising, no guessing who will read it. The expensive part of sharing — deciding what matters to a particular reader, on a particular stack, today — is not yours: it happens on the reader's side, inside their agent.
- Raw is the point. Polish is expensive and lossy; raw material keeps the details a reader needs to judge for themselves.
- One conversation, one PR. The publishing skill drafts the dump, validates it against the contract, and opens the PR only after you say yes.
- Any field. Engineering, design, finance, construction, sales — the contract is the same everywhere.
- Attribution and provenance travel with the dump. Stakes, content flags, trust level, how it was generated, how far a human checked it.
For readers
You never read dumps — your agent does. It filters the registry against your task, picks what is worth your attention, and rewrites it for your situation: longer or shorter, in your language, against your stack, with or without explanations. The same dump becomes different reading for different people.
- Adaptation, not quotation. The dump arrives as raw material plus a machine-readable contract, and lands in your context already reshaped.
- Synthesis across dumps. Your agent can weave several dumps into one read shaped to your task — "take these three dumps on agent context management and assemble what applies to my setup." Raw dumps are the raw material; the compilation is assembled on your side.
- Trust is declared, not implied. Every dump carries stakes and trust level, so an agent knows when to verify before relying — and warns you when the content says so.
- The stream compounds. Easier authoring means more raw experience in the registry, and more raw experience means richer synthesis for every reader.
How it works
- The author writes a raw dump — their own account of what they did, with the dead ends left in.
- The dump carries a machine-readable contract — a manifest with a stable slug, schema-checked fields and published JSON surfaces (the index, per-dump manifests, feeds). A machine can decide whether to read a dump before reading it.
- Adaptation happens on the reader's side — the registry delivers raw material plus metadata; the reader's agent filters it, attributes it and reshapes it. The author ships a mechanism, not a plea: the contract is held by schema grammar and gates, not by asking everyone to be careful.
Architecture decisions
- Raw dumps, polished agents. The registry stores the raw material; the reader's agent does the adapting.
- Machine-readable contract first. The index is JSON, the protocol lives at
/.well-known/kodavr.json, and documents are self-describing — an agent boots from the file itself, with no external config. - Reading must be cheap. The index is built so an agent can decide what to open without tokenizing everything — lazy reading is a design constraint, not an optimization.
- Safety by soft delegation. Nothing in a dump executes. If an agent's rules ask for confirmation, it checks with its user in plain words — no jargon, no scare terms.
- Trust levels are part of the data. raw → self-tested → community-tested → adapted → library; the ladder is declared per dump and machine-checkable.
Share gears, not text.
The autopsy revealed the code was useful. · Open your agent's insides.