跳到正文
DEV Community· OwlMem·· 8 小時前AI 評分50

OwlMem 如何讓 Claude 將未完成任務交給 Codex,毋須重新説明

Hand a half-finished task from Claude to Codex without re-explaining

AI 導讀

OwlMem 透過共享記憶和收件箱,讓 Claude 把未完成任務連同目標、範圍及驗證方式交給 Codex,毋須重新交代背景。啟用 auto-wake 後,Codex 有待辦任務時會由背景程式啟動;每個 AI 同一時間限處理一項任務,每項限時 30 分鐘,逾時會停止並標記為 declined。

正文

Hand a half-finished task from Claude to Codex without re-explaining

Claude gets halfway through a job — then you hit its context limit, or you want Codex to take it from here. Normally that means re-explaining everything: what the job is, what's done, what's left, where the files are.

OwlMem removes the re-explaining step. (Disclosure: OwlMem is our own product — owlmem.com.) Its AIs share an inbox: one AI leaves a note or hands over a task, and the other picks it up the next time it connects — with the notes attached.

OwlMem docs: Collaboration between AIs — one AI hands over a task, the other picks it up

How a handover works

There are two pieces: the shared memory and the task itself.

1. The work lives in one place. Both AIs connect to the same OwlMem space, so what Claude recorded — decisions, progress, gotchas — is already there for Codex to find. Nothing is trapped in Claude's chat history.

2. The task is handed over, not re-typed. Claude (or you) sends Codex a task with the goal, the scope, how to verify it, and the done standard. A task brief looks like this:

Title: Finish login-form validation in ~/app
Body:
Goal: finish the validation Claude started on the login form.
Scope: only the login form and its tests — don't touch anything else.
Verify: run the form's tests until they pass.
Done: validation works, tests pass, result names the branch.
cwd: ~/app

Codex reads it from its inbox (kb_inbox), takes it on (kb_task_update → doing), and reports back with done and the result in its note — or declined with the reason. Done and declined send the sender a note automatically, so Claude hears the outcome too.

Codex doesn't even need to be open

A cloud AI only runs when someone talks to it, so OwlMem wakes it instead: turn on auto-wake once, and a small background program starts Codex when a task is waiting. When the task is done or declined, the AI that handed it out is woken to read the result.

A few mechanics worth knowing before you hand work over:

  • One task at a time per AI, 30 minutes per task. A task that runs out of time is stopped and marked declined, for a person to take over.
  • In a git repo, the AI works in its own worktree and branch, commits there, and doesn't merge. Its result names the branch — you review and merge.
  • Put the working folder in the task (cwd: ~/Projects/app). Without a folder it works in an empty one.
  • Give tasks to AIs that check their inbox regularly — one on a schedule of its own, or one you open often — not one that's rarely used.

You can watch all of it on the task board: in the console, Messages → Task board shows every note and task with its history.

The setup, if you haven't done it yet

Both AIs need to be connected to the same OwlMem space. One address for every client:

https://mcp.owlmem.com/mcp
claude mcp add --transport http --scope user owlmem https://mcp.owlmem.com/mcp
codex mcp add owlmem --url https://mcp.owlmem.com/mcp

Or run npx -y owlmem install and let it find both apps. Then turn on auto-wake for Codex with npx -y owlmem wake enable, or switch it on in the console under Agents & tools. Full steps are in the docs. Free for two AIs, forever.

The habit that makes this work: when Claude finishes a chunk of a job, have it record what's done and what's left — one sentence in the shared memory. Then the handover is just the task itself. The next AI starts where the last one stopped, and you never type the same context twice.

來源:DEV Community · dev.to