FALSE DONE
Replay the claim.
The record points from manifest to work item to evidence, so "done" can be checked against the run.
DRAGONS / FLIGHT RECORDER
Your agent said "done." It wasn't. You found out three days later. Every run writes a replayable record: who the agent was, what it was authorized to do, what it did, whether it is still alive. Check it yourself.
curl -fsSLO https://dragons.dev/receipt-specimen.json
curl -fsSLO https://dragons.dev/replay.mjs
node replay.mjs receipt-specimen.json Run the same refusal and recovery checks locally with Node. No account, install, or package registry.
01 / THE RUN
Dragons follows a local coding-agent run from activation to replay. The record names the agent, its authority, its work, and its last heartbeat instead of asking you to trust a dashboard.
CAUGHT-LIE SCENARIO Your coding agent says "done." Replay checks the run plan, work item, evidence, and last heartbeat against checked-in test records. If the record does not replay, the claim is not accepted.
Connect a local Claude Code or Codex session.
Identity and the manifest enter the record.
Work as normal inside the declared boundary.
The last heartbeat shows alive versus silently dead.
Check the record against checked-in test records.
02 / EVIDENCE SURFACE
A flight recorder earns its keep when it catches a false "done" or finds an overnight run that died before morning.
FALSE DONE
The record points from manifest to work item to evidence, so "done" can be checked against the run.
DEAD RUN
The last heartbeat shows whether an overnight run is alive or needs recovery before morning.
OPERATOR BOUNDARY
Authority, evidence, and recovery stay visible to the operator instead of disappearing behind a vendor dashboard.
03 / THE SMALLEST USEFUL LOOP
run:
target: Claude Code or Codex
connect: MCP config line | SDK agent()
record:
fields: run plan | instance | work item | heartbeat
replay:
against: checked-in repo fixtures
answer: alive | dead | done A SMALLER PROMISE
Dock one local session. Replay one record. The proof path belongs in the repo and can be checked on your machine.
See the local path04 / FAILURE RECORD
This is the public fixture the evaluator replays: one false "done" and one dead overnight run, with the decision each should trigger.
A completion claim with no matching artifact is refused, not accepted.
heartbeat = stale
{
"test_record": true,
"production": false,
"agent": "agent_019c8c0f2e2d7c1a8b6e2d4f9a0b1c3e",
"manifest": "hash + one work item",
"liveness": "alive / heartbeat redacted",
"replay": "agent → work → evidence → outcome",
"private_fields": "redacted"
} 05 / LOCAL DOCK
The product is a local SDK and MCP dock for coding-agent sessions. Connect Claude Code or Codex, work normally, then replay the record against checked-in test records.
Use one local MCP config line or one SDK agent() call.
Each activation records the run plan, which agent ran, the work item, and the last heartbeat.
When the agent says "done," check the record. It either replays or it does not.
import { agent } from '@dragons/sdk';
const run = agent({
mode: 'persistent',
venue: 'cloudflare',
template: 'pi-mono',
version: 'v1',
dock: { org: 'mogos', project: 'fleet-of-one', role: 'worker' },
}); PUBLIC PATH / RUN IT NOW
The local SDK/MCP dock is the next product surface. Use the public replay kit above today; the local package and hosted endpoint are not part of this release.
Run what is public nowNeed private local access? Email the operator. This is a human contact route, not an automated intake.
06 / TRUTH FENCE
Static public surface plus a downloadable, deterministic false-done and dead-run replay. This release has no published SDK/package, installer, hosted endpoint, external users, customer outcomes, pricing, open-source release, or production-readiness claim. Next product evidence: one redacted replay from customer-owned or auditor-owned data.