BriefOps Relay
Inspiration
Codex can make strong progress inside one well-scoped task. Long-running projects fail differently: requirements live in documents, decisions live in earlier sessions, and constraints live in code and tests. By the next task, there is no reliable way to show which project context still matters or whether an implementation actually honored it.
We built BriefOps Relay to make that continuity inspectable rather than merely summarized.
What it does
BriefOps Relay turns bounded repository evidence into a task-scoped execution contract, audits an implementation diff against that contract, and leaves a verified handoff for the next Codex session.
- Prepare — Collect only bounded, line-addressable repository evidence and create an Execution Contract for the task.
- Build — Codex can implement against the Contract instead of a vague repository summary.
- Audit — Relay examines the recorded Git diff and classifies every Contract item as met, at risk, violated, or unverified.
- Validate — Every semantic claim must cite a supplied evidence ID. Unknown evidence, missing findings, duplicate findings, and score tampering are rejected rather than silently accepted.
- Handoff — Relay renders the changes, checks, unresolved risks, and next action as a Markdown handoff and a self-contained static HTML report.
The result is not an AI memory summary. It is an evidence-backed execution loop with a deterministic integrity score and completion gate.
How we built it
Relay is an additive TypeScript and Node.js command group inside BriefOps. It uses Commander for the CLI, Zod for artifact schemas, Git diff collection for change evidence, and Vitest for regression coverage. The generated report is one self-contained HTML file with no server, React app, analytics, or required network access.
The default demo is deliberately offline and API-key-free. It creates a seeded context-drift scenario, then runs demo, audit, handoff, report, and evaluation locally. The evaluation covers six adversarial artifact-integrity scenarios: known violation, unknown evidence, duplicate findings, missing finding, score tampering, and an all-met control.
For live semantic analysis, Relay offers two explicit, opt-in providers. The
Codex provider invokes the locally authenticated Codex CLI; the optional OpenAI
provider uses a structured Responses API request with store: false. Both
require an explicit network flag before bounded repository evidence is sent.
How Codex was used
Codex audited the original repository, implemented Relay modules and tests, ran the authenticated local Contract and Audit workflow, generated adversarial fixtures, iterated on the report against screenshot criteria, and automated the repeatable video-production pipeline. The project owner made the product scope, evidence policy, scoring semantics, feature-cut, and submission-claim decisions.
Challenges we ran into
The hard part was not generating a summary. It was making generated output safe to trust. We separated semantic reasoning from deterministic validation, bounded evidence before transmission, excluded secret-bearing paths, rejected unknown citations in full, and made the score a code calculation rather than a model opinion.
Another challenge was preserving a useful no-key path while supporting a real live provider. The offline demo stays reproducible without credentials; live evidence transfer remains explicit, bounded, and reviewable.
What we learned
AI coding workflows need more than larger context windows. They need an execution contract that survives a task boundary and a way to prove what the implementation did against it. Keeping semantic reasoning evidence-constrained and making acceptance logic deterministic produced a smaller, clearer product than adding another general-purpose agent layer.
What's next for BriefOps Relay
We plan to expand the fixture corpus, improve rule-based evidence selection, and add repair/re-audit examples across more real repositories. The current offline evaluation is intentionally artifact-integrity coverage, not a claim about general model accuracy. A relevant file can still be omitted by the selection rules, which Relay should surface as risk rather than pretend to solve.
Log in or sign up for Devpost to join the conversation.