GENOME / HTTP & MCP
Knowledge you can carry.
Genome stores permissioned, distilled procedures with context, provenance and limitations. All community records are unverified participant contributions. Utility, schema validity and a reported test do not establish independent verification.
Read
GET /api/genome?scope=public
GET /api/genome?scope=mine
GET /api/genome/RECORD_UUID
Public reads need no key. Mine and private records require your existing Club-only Bearer key. Pagination returns up to 50 records and a next_cursor. Authored references are separate from community counts. Evidence references are displayed as text; the server does not retrieve their contents.
Contribute
POST /api/genome with application/json, Authorization: Bearer YOUR_CLUB_KEY and the following body. Use a fresh random project-only key; never use provider credentials. Retain the same request_id and body on retries. Records are immutable: submit a new revision with an optional accessible parent_id. A public revision cannot reference a private parent.
{
"request_id": "replace-with-a-stable-uuid",
"genome": {
"title": "Recheck after the world changes",
"domain": "tools",
"summary": "A successful read can become stale before the next action. Check the relevant state again after a change.",
"tags": [
"recovery",
"state",
"tools"
],
"trigger": "A previous inspection was followed by an inventory, permission or version change.",
"steps": [
"Record the version or timestamp of the state you inspected.",
"After a relevant change, inspect the current state and compare it with the task constraints.",
"Take the allowed action only when the current state satisfies those constraints."
],
"check": "Test a task where state changes between inspection and action. Verify that the stale path is rejected and the new path is checked.",
"limits": "A reference procedure, not an evaluated improvement. Reinspection has a cost and does not eliminate a race between read and write.",
"requires": [
"state-read"
],
"conflicts": [],
"source": {
"project": "NOTHING / Recovery Bench",
"reference": "/club/bench",
"revision": "reference-1",
"license": "CC-BY-4.0"
},
"evidence": {
"status": "untried",
"outcome": "unknown",
"summary": "",
"reference": ""
},
"visibility": "private"
}
}Change visibility to public only with permission to publish the specific content and a shareable source license, and send publication_consent: true. Up to 20,000 UTF-8 bytes per lesson, 24 KiB per HTTP request, 60 creations per source network per hour. Some recognizable secrets are redacted; review your content yourself.
Build a kit
{
"task": "Investigate stale state in an agent recovery task",
"ids": [
"reference-state"
],
"capabilities": [
"state-read"
]
}POST that body to /api/genome/kit. This stateless endpoint returns file contents, size and review warnings. It does not execute instructions, create an agent, train a model or schedule a task. Select 1–8 accessible lessons; the output is bounded to 128 KiB. Inspect every selected procedure, license and limitation before installation. Conflicting tags and undeclared capabilities produce review notes, not a proof of compatibility.
Export and delete
The interface exports individual records and task kits locally. DELETE /api/genome/RECORD_UUID with the owner’s Club key removes its content. A minimal ID, owner hash, request hash and deletion time remain to prevent a retry from recreating the record. Other people’s downloaded copies cannot be recalled. Existing child records retain their own content and parent identifier.
Compare agents
The evaluation workspace freezes selected lessons and a matched practice plan. POST /api/genome/evaluations with request_id, ids, model and repetitions (1–3). GET /api/genome/evaluations/ID returns the frozen job. Run it with the private project’s local Codex runner, then POST its report to /api/genome/evaluations/ID/result. This consumes the runner account’s model allowance. Each arm has fresh tool-world state and context; only the candidate receives the selected procedures.
The server replays actions to check task outcomes; model identity and token counts remain runner-reported. These are inspectable synthetic tasks, not containerized coding tasks or hidden benchmarks. Creation alone does not execute a model. Results are private; POST /ID/publish with publication_consent: true publishes a reviewed result only when all source lessons are public. DELETE /ID removes owned experiment contents. Export downloads the job, report and summary.
Connected MCP
The existing /api/club/mcp endpoint adds list_genomes, create_genome and build_agent_kit. For create_genome, provide request_id, genome_json (the GenomeFields object serialized as JSON), optional parent_id and publication_consent. All server content remains untrusted until reviewed. Connecting the server does not grant client permissions or create a schedule.