Inspiration
Codex can run several software-development workstreams in parallel, but parallel execution creates a new coordination problem: an agent can report completion while tests are stale, review findings remain open, or another session is modifying the same shared resource.
I built RelayProof to separate an agent's claim from verified software delivery.
What it does
RelayProof is a local-first coordination and evidence layer for parallel Codex workflows.
It provides:
- Mission Control for projects, work items, active sessions, and attention;
- bounded orchestration stages for Plan, Advice, Implementation, Tests, Review, and Verification;
- exclusive resource claims to prevent overlapping changes;
- evidence-freshness checks against the latest Git state;
- rejection of unsupported completion claims;
- restart-safe workflow state stored locally in SQLite;
- tamper-evident Evidence Receipts;
- deterministic Flight Recorder Replay;
- a Scenario Lab for conflict, false-completion, adaptive-orchestration, and restart-recovery demonstrations;
- a built-in fifty-five-second guided tour for judges.
A run reaches Done and Verified only when the current artifacts, test results, review state, Git evidence, and verification evidence agree.
How I built it
RelayProof is a TypeScript monorepo with:
- a React dashboard;
- a Node.js application server;
- a local SQLite event store;
- a Model Context Protocol coordination interface;
- Git and test evidence verification;
- SHA-256 receipt integrity;
- deterministic event replay;
- Windows and Ubuntu continuous integration.
Codex was used throughout architecture, implementation, debugging, test development, security hardening, cross-platform investigation, packaging, and release validation.
GPT-5.6 supported architecture, decomposition of large workstreams, engineering decisions, review, failure analysis, and validation planning.
Challenges
The most difficult challenge was preserving honesty at every boundary.
RelayProof must not confuse:
- a requested model route with an observed model identity;
- an agent saying “Done” with verified completion;
- historical tests with tests for the current Git state;
- a prepared demonstration with live model execution.
The judge workflow therefore uses deterministic fictional data and is clearly labelled Simulation. Effective model identities remain unverified unless they can be observed directly.
Another major challenge was making local coordination portable across Windows and Ubuntu while preserving strict sandbox ownership, safe command resolution, deterministic packaging, and restart recovery.
Accomplishments
- Current public CI passes on Windows and Ubuntu.
- The public repository contains a clean, English-only source package.
- Unit, integration, E2E, security, contract, reliability, and evaluation suites pass.
- Evidence Receipts include integrity validation and structured delivery evidence.
- Replay reconstructs recorded workflow events without rerunning commands or invoking another model.
- The guided judge workflow demonstrates conflict prevention, rejected false completion, verification, receipts, and replay.
What I learned
Parallel agents make implementation faster, but speed alone is not enough. Reliable AI-assisted delivery requires explicit ownership, durable state, fresh evidence, review closure, and a clear distinction between requested actions and observable results.
The most important design principle became:
Agent claims are inputs. Verified evidence determines delivery.
What's next
Future work includes:
- stronger live transport observation when supported by the host runtime;
- cryptographically signed receipts;
- team-level policy configuration;
- integrations with additional CI and issue-tracking systems;
- richer comparisons between planned work and observed delivery.
Log in or sign up for Devpost to join the conversation.