feat(aidd-context): a stale memory fact has neither a marker nor a trigger

未關閉
#942 2 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
3/5
預估耗時
1-2 天
新手友好度
68/100
Issue 類型
功能
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
javascript
領域
tooling

研究方向

Start with memory-rules.md, the 03-check flow, and scripts/check-context-reference-form.js; inspect how the memory files are selected and how existing format checks report failures. Add the volatility rule and the numeric/date-claim check, then verify that claims without derivation markers are flagged while valid memory content and existing checks remain green.

由索引模型根據 Issue 內容生成。

描述

Problem to solve

The rule already exists, in two places:

Flag each of these, with a reason and a location:

  • A claim the code contradicts, or that has gone stale.
    — plugins/aidd-context/skills/02-project-memory/references/review-protocol.md

Reflects the current state only: no unused section, no placeholder.
— memory-rules.md

So this is not a missing rule. What is missing is that nothing marks which facts are the
kind that go stale, and nothing runs the review often enough to catch one before it is cited.

Two specifics, both in the current tree:

  • memory-rules.md has exactly one volatility rule — "Names a tech without its version
    (React, not React 19)"
    — and it covers versions only. A count, a file total, a date, a
    "currently N branches" line is the same class of fact and has no rule.
  • 03-check step 1 matches the bank against memory-destinations.md and structure.md; the
    review in step 2 is what catches a stale claim, and it runs when a human asks for it. Between
    two runs, any agent may cite the stale line, and the citation will be faithful.

Observed on a project I run AIDD on. aidd_docs/memory/vcs.md recorded, on 2026-09-29:

The project has no branch convention yet. All 41 commits sit directly on main, unpushed.
This is a fact, not a rule.

The repository held 134 commits with 0 unpushed at the moment the line was written, and 137
with 1 unpushed when I looked. Both numbers were wrong when written. The session report that
produced the file repeats it and labels it "Vrai fact, consigné comme tel dans memory/vcs.md".
Its own section "Méthode — ce qui a failli mentir" lists six measurement near-misses that were
caught and re-derived; this one was not among them.

An agent then read the file, cited it verbatim, and was right about what the file said and wrong
about the repository. Nothing in the chain can tell the difference: update_memory.js emits
references and never reads content, check-context-reference-form.js reads TARGET_FILES from
the writer itself, and both are green.

Proposed solution

Two changes, both small, neither a verification pipeline.

  1. A volatility rule in memory-rules.md, in the spirit of the existing version rule: a
    memory line that states a count, a total, a date or a size carries the date it was derived on,
    or is written so it stays true without one. "One local branch, main" holds; "41 commits,
    unpushed" does not.
  2. A check, in the shape of scripts/check-context-reference-form.js, that reads the memory
    files and flags a numeric or dated claim with no derivation marker. It judges format only — it
    does not try to decide whether the number is true, because a script that guessed truth would be
    worse than no check.

I am deliberately not proposing that memory be re-derived on read. memory-rules.md wants
"the macro and the non-derivable", and a memory file that is cheap to re-derive was the wrong file.

Content type

  • Rule
  • Other — the accompanying script check

Target tool(s)

  • Other — no tool involved; aidd-context writes and checks the bank.

Additional context

  • Adjacent, not the same, and I checked the bodies rather than the titles. #864 audits code
    conformance against project Rules and states in its own scope that it must not depend on
    aidd-context, because "aidd-dev evaluates code conformance; aidd-context owns Rule content
    and context health"
    . This is a aidd-context question about the bank's own content. #793
    audits context architecture and ownership; #887 is about when the check loop should stop.
  • The precedent for the shape of the answer is #751. There, the rule existed and a machine
    drifted back into the wrong form; the fix restored it and added the check that notices when it
    drifts back
    . That is the standard I am aiming at, and why I am proposing a marker plus a check
    rather than another paragraph of doctrine.
  • Disclosure, because it cuts against me. The false fact is mine, in my own repository, written
    while following the framework's rules and missed by the report section whose whole purpose is
    catching exactly this. I have corrected the file and the report. I file this because the
    discipline failed, not because the author did.
  • I am not asking for a PR to be opened pre-emptively. Per CONTRIBUTING.md an issue is validated
    first — a Certified Member or Maintainer moves it to Todo — and after that anyone may pick it up,
    including me.
主要語言
TypeScript
星號
481
分支
45
平均合併
1 天 13 小時
30 天內合併 PR
91

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

ai-driven-dev/framework 的其他 Issue

查看 ai-driven-dev/framework 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。