home/roleplay/phase-c-e2e-verification

Phase c E2e Verification

GPTClaudeGemini··356 copies·updated 2026-07-14
phase-c-e2e-verification.prompt
# Phase C E2E verification: lessons.sh extract + review real run

## Goal

Phase C lessons loop 截至目前只 detection + count/list 跑過。**extract(headless `claude -p`)+ review(互動 promote/delete)** 從沒實測 — 如果 claude CLI 介面變了、prompt 格式錯了、或 review 流邏輯有 bug,整條 self-improving loop 死掉。本輪做 end-to-end 驗證,**不留任何 test 殘檔**。

## Execution profile

- model: sonnet
- effort: medium(多 step verify + cleanup 要乾淨)
- 0 deliverable commits(純 verification round)+ archive = 1 commit
- **本輪不應該對 Tandem repo 有任何 code 改動** — 全部 test 動作在 user-level state(shared memory + staging)+ 跑完還原

## Background context

- `lessons.sh extract` 用 `claude -p "..."` headless 餵 raw signals → AI 寫 candidate(含 frontmatter + Why/How body)
- `lessons.sh review` 互動 walk candidates,default action `[p]romote-shared` 把 candidate 寫進 `~/.claude-work/_shared/memory/<slug>.md` 並 append 索引到 `~/.claude-work/_shared/memory/MEMORY.md`
- 測試需考慮:claude CLI 可能需要 auth / API key 在 env / 不同回應格式

## Two-sub-test 策略

| Sub-test | 目標 | 失敗 means |
|---|---|---|
| **A. extract test** | 驗 `claude -p` 真能跑 + 輸出可被 lessons.sh parse 成 candidate state | claude CLI 介面破 / prompt 格式不被 model 理解 / parsing 邏輯錯 |
| **B. review test (manual candidate)** | 跳過 extract 直接 inject 已知 candidate,驗 review 互動 + promote 真寫入 shared memory | review 邏輯錯 / slug 提取錯 / shared MEMORY.md append 錯 |

A 失敗不擋 B(兩條獨立路徑)。本輪結束 Result block 分開報告兩個 sub-test 結果。

## Step-by-step

### Step 1: Pre-flight + snapshot

when to use it

Community prompt sourced from the open-source GitHub repo woodylin0920-bit/Tandem (MIT). A "Phase c E2e Verification" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.

tags

roleplaycommunitygeneral

source

woodylin0920-bit/Tandem · MIT