Forked session inherits parent's "acp:" runtime binding — every tool call fails with "runtime ... does not exist in workspace" once the parent session ends

Summary

A session that was forked from a session running on an ACP runtime keeps the parent's runtime binding (acp:<parent-session-id>) instead of getting its own. As long as the parent session's ACP runtime is alive, the fork works fine. Once the parent session ends / its ACP runtime is unregistered, every tool call in the fork fails with:

runtime acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0 does not exist in workspace wd_eb1243_b29b0a783e63

and the turn aborts with internal — surfaced in the UI as "模型请求失败，本轮对话已中断" (model request failed, turn interrupted), which is misleading: the LLM request itself succeeds; the failure happens when resolving the runtime for tool execution.

There is no validation, rebind, or fallback to the local runtime, so the session is permanently bricked: even local-only management tools (e.g. CronDelete) cannot run, and scheduled cron tasks keep firing and failing.

Environment

• Kimi Code CLI: 0.39.0 (wire protocol 1.5)
• OS: Linux x64, Node.js 24.15.0
• Client in use: Kimi web UI (ACP), session also forked twice (Fork: Fork: ...)

Timeline (from the exported wire log, all times local UTC+8)

Expected behavior

1. Forking a session should rebind the fork to local (or to its own ACP runtime if the fork is created through an ACP client) — never keep a binding that references the parent session's runtime.
2. If the bound runtime is gone, tool execution should detect this, fall back to local (or rebind) and/or surface an actionable error ("session bound to a runtime that no longer exists — reattach or rebind"), instead of failing every turn with internal and a misleading "model request failed" message.

Actual behavior

• Every tool call throws RuntimeError: runtime acp:<parent-session> does not exist in workspace <ws>; the turn is aborted as failed with code internal, retryable: false.
• The user cannot recover from within the session: scheduled cron jobs cannot even be deleted (CronDelete also goes through runtime resolution), and cron keeps firing into the broken session.
• Pure-text replies still work, so from the user's perspective "the model answers, but goes crazy the moment it touches a tool".

Steps to reproduce (inferred from wire log)

1. Open/resume session A through an ACP client (web UI / IDE) so its binding becomes acp:session_A.
2. Fork session A → session B (/fork or fork from the client).
3. Use session B — everything works while session A's runtime is registered.
4. Close session A (or otherwise let its ACP runtime unregister), e.g. reload the web UI.
5. In session B, send any prompt that triggers a tool call → turn fails with the error above; session is unrecoverable.

Evidence

Redacted excerpts from the session's agents/main/wire.jsonl and logs/kimi-web.jsonl (exported via the debug-zip export; workspace path redacted, otherwise verbatim):

{"type": "runtime.set_binding", "workspaceId": "wd_eb1243_b29b0a783e63", "runtimeId": "local", "agentId": "main", "time": 1787369623562}

{"type": "runtime.set_binding", "workspaceId": "wd_eb1243_b29b0a783e63", "runtimeId": "acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0", "agentId": "main", "time": 1787819978190}

{"type": "forked", "agentId": "main", "time": 1787838905535}

{"ts": 1787882250745, "event": "app:load:start"}
{"ts": 1787882251038, "event": "app:load:complete", "sessionId": "session_0474e71d-b859-49c9-b08c-38c12fa0ea86", "status": "accepted", "durationMs": 293, "sessionCount": 26, "workspaceCount": 14}
{"ts": 1787882251063, "event": "ws:connection", "status": "connected"}

{"type": "turn.ended", "agentId": "main", "turnId": 203, "reason": "failed", "error": {"code": "internal", "message": "runtime acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0 does not exist in workspace wd_eb1243_b29b0a783e63", "name": "RuntimeError", "retryable": false}, "durationMs": 96558, "time": 1787882253457}
{"type": "turn.ended", "agentId": "main", "turnId": 204, "reason": "failed", "error": {"code": "internal", "message": "runtime acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0 does not exist in workspace wd_eb1243_b29b0a783e63", "name": "RuntimeError", "retryable": false}, "durationMs": 61615, "time": 1787882466680}
{"type": "turn.ended", "agentId": "main", "turnId": 205, "reason": "failed", "error": {"code": "internal", "message": "runtime acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0 does not exist in workspace wd_eb1243_b29b0a783e63", "name": "RuntimeError", "retryable": false}, "durationMs": 42913, "time": 1787882517681}
{"type": "turn.ended", "agentId": "main", "turnId": 206, "reason": "failed", "error": {"code": "internal", "message": "runtime acp:session_0730c2bd-1a3f-4dea-a326-64c3b98396c0 does not exist in workspace wd_eb1243_b29b0a783e63", "name": "RuntimeError", "retryable": false}, "durationMs": 14691, "time": 1787882540119}

{"type": "step.end", "uuid": "0e81879d-897c-42b0-875d-cb11652f1126", "turnId": "203", "step": 1, "finishReason": "error"}

{"id":"session_0474e71d-b859-49c9-b08c-38c12fa0ea86","version":2,"cwd":"<workspace>","archived":false,"lastTurnReason":"failed","title":"Fork: Fork: 当前编译验证","forkedFrom":"session_0730c2bd-1a3f-4dea-a326-64c3b98396c0","lastPrompt":"继续"}

The full (unredacted) debug ZIP contains internal/company data, so it is not attached here; I can provide it via code@moonshot.ai referencing this issue if needed.

Suggested fix direction

• On fork: strip/reset any acp:* runtime binding and rebind the new session to local (or to a fresh ACP runtime registered under the new session id).
• On tool dispatch: if the bound runtime id is not registered in the workspace, attempt rebind to local and only fail with a clear, actionable error if that is impossible.
• Reword the user-facing message: this is not a model request failure — report the actual runtime resolution error.

Trace the fork handling and the runtime.set_binding events first, using the wire.jsonl evidence to confirm how the child retains the parent's ACP binding. Then inspect runtime resolution during tool dispatch, including CronDelete; done means a fork no longer uses the parent's runtime and a missing binding produces recovery or an actionable error instead of an internal failure.