跳到正文
DEV Community· Siwar Grami·· 17 天前AI 評分56

Gitmore 如何讓編碼 Agent 讀取程式碼庫歷史並接續工作脈絡

Your coding agent starts every session with amnesia.

AI 導讀

Gitmore 透過 MCP 以唯讀方式向編碼 Agent 提供程式碼庫活動資訊,使用 API key 且不提供原始碼,讓 Agent 查詢本週交付內容及 PR 描述、提交訊息中的改動原因。Claude Code 的 Stop hook 會在工作階段結束時提交簡短自述,記錄工作內容、原因、分支、提交、PR 和修改檔案,供下一個 Agent 及工程師讀取。網址為 gitmore.io。

正文

Siwar Grami

It can read your entire codebase. It cannot read why any of it is that way.
The reasoning lives somewhere it can't reach in PR descriptions, in commit messages, in a review thread from March, and mostly in sessions that already ended.
So it re-derives decisions your team made six months ago. Sometimes it re-derives them wrong.
We gave the history back.
Gitmore exposes your repo's activity to the agent over MCP read-only, API key, no source code. Before it edits a file, the agent can ask what shipped this week, or search every PR description and commit message for why a thing changed.
Then the other direction, which matters more.
When a session ends, a Claude Code Stop hook posts a short self-report: what it did, why, the branch, the commit, the PR, the files it touched. Four thousand characters of reasoning that would otherwise have died with the session.
The next agent reads it. So does the next engineer.
Here's the part I underestimated. When one person used an agent occasionally, losing the "why" was an annoyance. When agents write a third of your merged work, losing the "why" is an institutional memory problem and it compounds quietly, commit by commit, until nobody can explain a subsystem.
Attribution tells you how much of your code agents wrote.
This decides whether any of that work accumulates into something your team actually knows.

gitmore.io

來源:DEV Community · dev.to