What I learned building a native console for two competing coding agents

Описание и применение

Создано приложение Agent Console для наблюдения за двумя кодирующими агентами, обеспечивающее быстрый одобрение, восстановление состояния, просмотр diff и доказуемый аудит.

Задача

Developing a desktop app to supervise coding agents (Claude Code and Codex) required features beyond what terminal UIs offered, such as quick approval, state snapshots, diff viewing, and verifiable audit trails.

Результат

Agent Console enabled efficient, auditable supervision of two competing coding agents, revealing architectural insights (engine neutrality, permission store differences, convergence) and practical lessons on cross‑platform hooks, clipboard handling, and release pipelines.

Рабочий процесс

• Launch Agent Console which auto‑starts the chosen agent in a PTY
• Agent outputs commands and diffs
• Human reviews command and diff, then approves or rejects via a two‑keystroke modal
• If approved, the agent executes the command; otherwise, execution is halted
• Each turn is recorded in a hash‑chained ledger with prompts, approvals, tool results, and diffs
• Users can export signed proof packets (Ed25519, DSSE) for audit
• The same workflow is used for supervising the app’s own development PRs

Доказательства

The agents wrote the overwhelming majority of the code, supervised through the very app they were building.

Источник

Every session in Agent Console is witnessed: prompts, human approvals, tool results and per-turn diffs land in a hash-chained ledger.

Источник