CONTROL ROOM

Inspiration

AI coding agents can change a repository in seconds, but speed does not teach the human operating skills that make agent work safe: setting a boundary, defining what counts as done, demanding evidence, resolving conflicting reports, and asking an independent agent to verify the patch. CONTROL ROOM turns those skills into a short, visible practice loop instead of another document developers are expected to read.

What it does

CONTROL ROOM provides two paths through the same workflow.

Trial Mode is a five-minute guided incident involving a synthetic university enrollment failure. The learner briefs a read-only Scout using four plain fields: objective, boundary, completion condition, and required proof. The learner reviews Scout's file, test, and log evidence, encounters a deliberate regional conflict, decides whether Builder may continue, requests independent verification, and receives an evidence-linked coaching debrief.

Real Mode brings that workflow to a clean Git repository selected by the user. The official Codex SDK starts three separate GPT-5.6 Sol threads. Scout is read only. Builder writes inside a temporary detached worktree. Verifier is separately read only and reviews the actual patch. CONTROL ROOM shows the diff and refuses to apply it until the user explicitly approves. It never commits or pushes.

How we built it

The interface is React 19 and TypeScript with an authored visual system built around a tactile Directive Table. A deterministic scenario reducer owns every approval transition, so the model cannot approve its own work or skip a gate. Zod validates browser, server, and model boundaries.

The local Real Mode companion uses the official @openai/codex-sdk. Each role has a separate thread and sandbox policy, agent shell network access is disabled, and the Builder works in an isolated Git worktree. The companion pins the starting commit, requires a clean repository, validates loopback Host and Origin headers, checks the patch with git apply --check, and requires a random browser token.

An optional Cloudflare Worker route uses the OpenAI Responses API with GPT-5.6 Sol, role-specific tools, bounded tool loops, evidence allowlists, and structured output. It is disabled on the public deployment. The public Trial Mode is a verified static replay, so judges can test the complete learning experience without an account, API key, or paid model call.

Codex was the primary development collaborator from research through deployment. It helped challenge the product concept, design the interaction and motion system, implement the React and TypeScript application, build the local Codex SDK runner, review security and privacy boundaries, write tests, produce documentation, and run the release audit. Product choices and every consequential approval remained human-owned.

How GPT-5.6 is used

Real Mode launches separate GPT-5.6 Sol Scout, Builder, and Verifier threads through the Codex SDK. Their permissions, evidence, timing, measured usage, and thread IDs appear in a downloadable trace. The recorded public proof includes a real run where Verifier returned concern because dependencies were unavailable in its isolated environment. The product exposed that uncertainty instead of claiming a false test pass.

The optional Responses API path also uses GPT-5.6 Sol for role-specific analysis and structured evidence reports. A deterministic state machine remains in control of sequence and authorization around every model call.

Challenges

The hardest product problem was making the boundary between simulation and real work unmistakable. Trial Mode needed to be instantly testable, while Real Mode needed to prove that it runs actual agents without pretending a public static site can safely execute a visitor's repository.

The hardest engineering problem was preserving human control across three agents. We separated permissions by role, isolated Builder, pinned the repository state, materialized evidence, added an independent verification step, and kept final patch application outside every model's authority.

Accomplishments

• A complete, public, no-account Trial Mode that reaches an evidence-linked debrief.
• A local Real Mode with three real GPT-5.6 Sol Codex threads and a human apply gate.
• An honest trace that records a Verifier concern instead of inventing success.
• Twenty-two automated tests covering workflow, permissions, security, and replay.
• Zero known dependency vulnerabilities and verified package signatures.
• A privacy-reviewed repository, trace, screenshot, README, and sub-three-minute demo.

What we learned

Agent supervision works better when uncertainty is a visible state rather than a paragraph buried in an agent response. Separating investigation, implementation, and verification also makes responsibility legible: one agent cannot quietly expand its own authority or certify its own patch. Finally, deterministic software should control consequential transitions even when GPT-5.6 supplies the intelligence inside each role.

What's next

The next useful step is not more autonomy. It is more practice: additional incident scenarios, trace comparison across repeated drills, and clearer coaching on how a learner's directives improve. Real Mode can later add configurable verification commands while preserving the same clean-repository, isolated-worktree, and human-approval boundaries.

Built With

• cloudflare
• codex
• openai
• react
• typescript
• vite