Codex Mission Control: Production-ready mission orchestration dashboard (Next.js + local persistence + OpenAI planning/execution)

We will deliver a real, end-to-end MVP of Codex Mission Control that plans and executes missions with OpenAI and persists all outputs locally for a reliable judge demo. First, we lock the MVP acceptance criteria and demo script so every engineering decision supports a crisp narrative: goal in, live plan out, specialists execute, artifacts/logs/risks/next steps returned. Next, we lay foundations in a greenfield Next.js App Router codebase with consistent structure and environment handling. We then define the domain model (Mission/Plan/TaskRun/Artifact/LogEvent/Risk/NextStep) and implement local persistence (preferably SQLite via an ORM) to ensure the demo survives refreshes and feels production-oriented. With persistence in place, we build an OpenAI server-only wrapper and strict prompt contracts using structured outputs and validation. This enables a reliable planner that converts the user’s goal into an execution plan, and an executor that performs each specialist task and emits artifacts and logs. We then implement the mission lifecycle API: create mission, generate plan, run tasks, stream events, retry failures, and fetch state. On top of that, we build a simple but real orchestration engine with a state machine, dependency resolution, concurrency limits, and retry semantics. The engine persists incremental log events and artifacts so the UI can show live progress. The UI is designed to be judge-facing and practical: a Mission Builder to start runs, a Plan View to inspect dependencies, and a Live Run View with task statuses, streaming logs, and an artifacts panel with previews and downloads. We prioritize clarity and control (run/cancel/retry/export) to demonstrate real orchestration rather than a static mock. Finally, we harden for demo reliability: observability, guardrails, graceful error handling, and curated mission templates that still execute real calls. We close with tests, performance polish (especially for logs), and a shareable final mission report that aggregates artifacts, risks, and next steps into a compelling output suitable for judges.

Defines judge-facing MVP: goal->plan->execute->stream logs->persist artifacts/risks/next steps->report. Includes non-negotiables (SQLite persistence, structured OpenAI outputs w/ validation, SSE, error handling, retries) and demo checklist.

Prisma SQLite schema covers Mission/Plan/Task/TaskRun/Artifact/LogEvent/Risk/NextStep with timestamps and traceId. Zod contracts enforce MissionInput, Plan shape (phases/tasks/deps/artifacts/riskChecks), and ExecutorResult (summary/logs/artifacts/risks/nextSteps).

Server-only OpenAI client uses abortable timeouts, retries, response_format=json_object, JSON.parse + Zod parsing, and secret redaction. Prompts define planner (phases/tasks with unique keys/deps) and executor (structured task results with artifacts, risks, next steps).

Top risks: (1) SSE + frequent SQLite writes can degrade performance—mitigate by throttling UI, batching inserts, limiting log history, WAL mode. (2) OpenAI structured output drift—mitigate with strict Zod parsing, repair prompts, short prompts, low temperature, cap task count. Remaining work: wire specs into real Next.js routes/pages and run an end-to-end mission.