What broke when my Claude Code projects started talking to each other

Описание и применение

При взаимодействии проектов Claude Code возникли проблемы с разрешениями, доверием, безопасностью, тайм‑аутами и адресацией. Решения включают предварительное одобрение инструментов, запрет shell‑команд, изоляцию, постоянные решения доверия, явное указание модели и повторное открытие сессий по имени.

Задача

Cross‑project Claude Code sessions stalled or failed due to permission prompts, trust decisions, model safety rejections, silent stalls, addressing bugs, and read‑only transport limitations.

Результат

The mitigations allow reliable cross‑project question/answer delivery, but full cross‑tree modification remains untested. The main failures now stem from permission, trust, safety, and addressing issues rather than transport bugs.

Рабочий процесс

• 1. Pre‑approve read‑only tools at launch to avoid permission dialogs.
• 2. Forbid shell usage in request payloads to prevent silent hangs.
• 3. Enforce confinement so sessions cannot access files outside their tree.
• 4. Set standing trust decisions for unapproved MCP servers to avoid startup stalls.
• 5. Explicitly pass the model when launching background sessions to prevent quota‑related silent deaths.
• 6. Re‑discover session by current name before messaging to avoid addressing bugs.
• 7. Monitor session logs to distinguish idle, still‑working, and stalled states.

Доказательства

Pre‑approving read‑only tools at launch and forbidding shell use in the request itself, sub‑agents included, cures the shell class.

Источник

Re‑verified on 2.1.252: cross‑session delivery and typed message exchange still work as described.

Источник

The workaround is small: re‑discover a session by its current name before any follow‑up message.

Источник