/projects/proofloop
Proofloop
A Bun CLI that runs Claude and Codex as a paired worker/reviewer — turning prompts into reviewed plans and refusing to finish until proof and an adversarial security gate pass.
Proofloop
Proof-driven paired agent loop with proof review
08
Project context05
Problem
AI coding agents can produce plausible changes without proof, review, or adversarial pressure. In real engineering work, completion should depend on evidence and quality gates, not just model confidence.
Solution
Proofloop pairs Claude and Codex as worker/reviewer roles in a Bun CLI, requiring plans, quality commands, proof artifacts, security review, resume manifests, and draft PR preparation.
Challenges
The system has to coordinate multiple agents, terminal sessions, git state, review artifacts, and failure recovery while keeping the loop usable as a developer tool rather than a research toy.
Innovation
Proofloop operationalizes proof-driven agent development: the workflow refuses to finish until review and verification artifacts exist, making evidence a first-class output.
Domain expertise
This shows Stefan's depth in agent harness design, multi-agent review loops, CLI ergonomics, git automation, security gates, proof discipline, and developer workflow architecture.
Case study evidence11
Outcomes
- Raises the completion bar for AI coding work from plausible diffs to reviewable proof artifacts.
- Creates a repeatable operator workflow for planning, implementing, reviewing, testing, and preparing PRs.
- Makes agentic development safer by pairing implementation with adversarial review and quality gates.
Architecture decisions
- Worker and reviewer roles coordinate through a CLI-managed loop instead of relying on one model pass.
- Detached sessions and resume manifests keep long-running agent work inspectable and recoverable.
- Git worktrees and quality commands make proof collection part of the development path.
Domain expertise signals
Agent harnessesProof-driven developmentCLI workflowsSecurity reviewGit automation
Technical deep dive09
Proofloop is a safety harness for AI coding work. It treats plans, review, terminal evidence, security critique, resume state, and PR packaging as part of the product, not optional ceremony.
Worker reviewer model
Pairing worker and reviewer roles creates deliberate friction. One model can implement while another challenges assumptions, asks for proof, and catches risks before the run claims completion.
Evidence gate
The workflow requires quality commands, artifacts, and review outputs before completion. That makes the agent accountable to observable evidence rather than self-reported confidence.
Session recovery
Detached sessions and resume manifests matter for real coding tasks because agent work can run long, fail midway, or need inspection after interruption.
Human merge path
Draft PR creation keeps the human in charge of shipping while still turning verified agent output into a reviewable software-delivery artifact.
What this proves
- Paired Claude and Codex roles
- Quality gates before completion
- Resume manifests for interrupted runs
- Security and performance review as explicit stages
2paired agents
20loop cap
4quality checks
3resume modes
1security gate
100%proof-gated
Technology stack08
Bun
Chosen for a fast single-binary CLI that can install, run, and update with little ceremony.
TypeScript
Keeps the run manifest, options, gates, and agent state explicit across a complex CLI flow.
Anthropic
Claude is useful as a reviewer or worker because it is strong at critique, plans, and long-context code review.
OpenAI
Codex gives the loop a second implementation and review path instead of relying on one model family.
tmux
Detached sessions make long agent runs inspectable and resumable without tying up the main terminal.
Git
Worktrees isolate each run so the agent can work without contaminating the user's current branch.
GitHub
Draft PR support turns a proven run into something a human can review and merge.
Biome
Keeps formatting and lint checks fast enough to run inside the proof loop.
Tools implemented08
loop CLI
Starts a paired worker/reviewer run from a prompt, file, or proof requirement.
Plan review
Has the non-primary model sanity-check PLAN.md before implementation begins.
Agent selector
Lets the operator choose Claude, Codex, paired mode, or single-agent execution.
Message relay
Lets the worker and reviewer coordinate directly instead of routing every critique through the user.
Quality commands
Runs check, typecheck, test, and build commands as completion evidence.
Adversarial review
Blocks completion until security and performance risks are pressure-tested.
Resume manifest
Persists run state and transcripts so interrupted work can continue.
Draft PR creation
Packages the verified change set for human review without silently shipping it.