如何在 Codex 上构建一个自我优化的外呼系统

1codex-self-improving-outbound/
2  AGENTS.md
3  README.md
4  config/
5    scoring.yaml
6    plays.yaml
7  prompts/
8    improve_scoring.md
9    improve_prompt.md
10    pr_summary.md
11  memory/
12    outcomes.jsonl
13  evals/
14    fixtures.yaml
15    score.py
16  scripts/
17    append_outcome.py
18    run_codex_step.sh
19    propose_improvement.py
20    open_pr.sh
21    weekly_tune.sh
22  examples/
23    outcomes.sample.jsonl
24    weekly-pr.md

1# 自我改进对外联系规则
2
3你根据结果改进对外联系系统。
4
5硬性规则：
6- 永远不要发送消息。
7- 永远不要抓取或丰富真实用户数据。
8- 永远不要自行合并。
9- 只编辑本仓库中的文件。
10- 一次只改动一个概念。
11- 每次改动都要引用 memory/outcomes.jsonl 中的结果。
12- 在改动成为 PR 之前，先改进 evals/score.py。
13- 如果评估没有改进，则撤销你的编辑并停止。
14
15允许的编辑范围：
16- config/scoring.yaml
17- config/plays.yaml
18- prompts/*.md
19
20必须输出的内容：
21- 修改过的文件
22- 每次改动的原因
23- 改动前的评分
24- 改动后的评分
25- pull request 摘要

1signals:
2  competitor_comparison:
3    weight: 8
4    reason: "买家正在比较替代方案"
5  implementation_page_visit:
6    weight: 6
7    reason: "买家在核实是否能安装此方案"
8  job_repost:
9    weight: 5
10    reason: "职位仍然开放且紧急"
11  funding_event:
12    weight: 5
13    reason: "预算或授权可能已变更"
14  generic_download:
15    weight: 1
16    reason: "内容兴趣，购买意图较弱"
17
18thresholds:
19  draft: 6
20  human_review: 10
21
22negative_signals:
23  student_research: -8
24  vendor_pitch: -6
25  competitor: -10

1{"date":"2026-07-01","account":"Northwind Finance","signal":"competitor_comparison","play":"migration_note","score":8,"outcome":"reply","reason":"asked for migration notes"}
2{"date":"2026-07-01","account":"Bluepeak Studio","signal":"generic_download","play":"resource_followup","score":1,"outcome":"no_reply","reason":"content-only intent"}
3{"date":"2026-07-02","account":"KiteOps","signal":"implementation_page_visit","play":"implementation_angle","score":6,"outcome":"reply","reason":"asked about implementation timeline"}
4{"date":"2026-07-02","account":"Atlas Recruiting","signal":"job_repost","play":"hiring_angle","score":5,"outcome":"bad_fit","reason":"student research request"}

1构建 scripts/append_outcome.py。
2
3它接受：
4- date
5- account
6- signal
7- play
8- score
9- outcome: reply | meeting | no_reply | bad_fit | bounced
10- reason
11
12它拒绝：
13- 缺少字段
14- 未知的 outcome
15- 空的 reason
16- 未来的日期
17
18将有效的行追加到 memory/outcomes.jsonl。
19打印追加的行。

1cases:
2  - account: Northwind Finance
3    signals: [competitor_comparison, implementation_page_visit]
4    expected: human_review
5    note: "一个账号有两个强信号"
6
7  - account: Bluepeak Studio
8    signals: [generic_download]
9    expected: ignore
10    note: "纯内容意图"
11
12  - account: KiteOps
13    signals: [implementation_page_visit]
14    expected: draft
15    note: "实现意图应超过草稿阈值"
16
17  - account: Atlas Recruiting
18    signals: [job_repost, student_research]
19    expected: ignore
20    note: "不匹配标记抵消了信号"

1构建 evals/score.py。
2
3读取 config/scoring.yaml 和 evals/fixtures.yaml。
4
5对每个案例：
61. 对每个信号累加权重。
72. 加上负面信号的惩罚。
83. 对账号进行路由：
9   - score >= thresholds.human_review => human_review
10   - score >= thresholds.draft => draft
11   - otherwise => ignore
124. 将路由结果与期望结果比较。
13
14打印每个预测结果。
15打印最终准确率，格式为 score=0.00 到 score=1.00。
16仅在准确率为 1.00 时退出 0。

1Northwind Finance: predicted=human_review expected=human_review
2Bluepeak Studio: predicted=ignore expected=ignore
3KiteOps: predicted=ignore expected=draft
4Atlas Recruiting: predicted=ignore expected=ignore
5score=0.75

1你改进对外联系评分系统。
2
3阅读：
4- AGENTS.md
5- config/scoring.yaml
6- memory/outcomes.jsonl
7- evals/fixtures.yaml
8
9你的任务：
101. 找出一条应该改变的评分规则。
112. 理由必须引用 memory/outcomes.jsonl。
123. 只修改 config/scoring.yaml。
134. 运行 python3 evals/score.py。
145. 如果评分提高，保留改动。
156. 如果评分持平或下降，撤销你的编辑并停止。
16
17输出：
18- 具体改动的行
19- 导致该改动的结果行
20- 改动前的评分
21- 改动后的评分
22- 该改动是否应该成为 PR
23
24不要编辑 prompts。
25不要添加新信号。
26不要触碰发送逻辑。

1scripts/run_codex_step.sh improve_scoring

1- implementation_page_visit: 4
2+ implementation_page_visit: 6

1Northwind Finance: predicted=human_review expected=human_review
2Bluepeak Studio: predicted=ignore expected=ignore
3KiteOps: predicted=draft expected=draft
4Atlas Recruiting: predicted=ignore expected=ignore
5score=1.00

1plays:
2  migration_note:
3    prompt_file: prompts/plays/migration_note.md
4    use_when:
5      - competitor_comparison
6    banned_lines:
7      - "thought this might be relevant"
8      - "quick question"
9
10  implementation_angle:
11    prompt_file: prompts/plays/implementation_angle.md
12    use_when:
13      - implementation_page_visit
14    banned_lines:
15      - "checking out our solution"
16      - "would love to chat"

1你改进一个对外联系话术。
2
3阅读：
4- AGENTS.md
5- config/plays.yaml
6- memory/outcomes.jsonl
7- 所选话术的提示词文件
8
9选择一个至少有 10 个结果的话术。
10
11找出：
12- 在积极结果中出现的行或结构
13- 在 no_reply 或 bad_fit 结果中出现的行或结构
14- 任何应该被禁止的短语
15
16对该话术的提示词做一次小编辑。
17
18规则：
19- 不要修改评分。
20- 不要创建新的话术。
21- 不要添加新的渠道。
22- 引用结果行。
23- 写出修改前后的指令。
24
25然后运行文案评估（如果有）。
26如果没有文案评估，则将 PR 标记为 review_required。

1为这次对外联系改进编写一个 pull request 摘要。
2
3包含：
41. 改了什么。
52. 为什么改，引用结果行。
63. 改动前的评分。
74. 改动后的评分。
85. 修改了哪些文件。
96. 风险。
107. 人工审查者应该检查什么。
11
12保持简短。
13不要声称改动已上线。

1#!/usr/bin/env bash
2set -euo pipefail
3
4branch="codex/weekly-tune-$(date +%Y-%m-%d)"
5
6git checkout -b "$branch"
7git add config prompts evals memory
8git commit -m "Codex weekly outbound tune"
9
10body="$(cat outputs/pr-summary.md)"
11
12python3 scripts/create_pr.py \
13  "$branch" \
14  "Codex weekly outbound tune" \
15  "$body"

1Changed:
2- Raised implementation_page_visit from 4 to 6.
3
4Why:
5- KiteOps had implementation-page intent and replied with implementation timing.
6- The previous score routed this account to ignore.
7
8Before:
9- eval score 0.75
10
11After:
12- eval score 1.00
13
14Reviewer check:
15- Make sure implementation intent is specific enough.
16- Keep generic downloads low.
17- Merge only if this matches actual sales judgment.

1#!/usr/bin/env bash
2set -euo pipefail
3
4cd "$(dirname "$0")/.."
5
6python3 evals/score.py || true
7scripts/run_codex_step.sh improve_scoring
8python3 evals/score.py
9scripts/run_codex_step.sh pr_summary > outputs/pr-summary.md
10scripts/open_pr.sh

10 8 * * MON cd ~/codex-self-improving-outbound && scripts/weekly_tune.sh

1name: weekly-outbound-tune
2
3on:
4  schedule:
5    - cron: "0 8 * * 1"
6  workflow_dispatch:
7
8jobs:
9  tune:
10    runs-on: ubuntu-latest
11    steps:
12      - uses: actions/checkout@v4
13      - uses: actions/setup-python@v5
14        with:
15          python-version: "3.11"
16      - run: pip install -r requirements.txt
17      - run: python3 evals/score.py || true
18      - run: scripts/weekly_tune.sh

1git clone <repo>
2cd codex-self-improving-outbound
3python3 -m venv .venv
4source .venv/bin/activate
5pip install -r requirements.txt
6python3 evals/score.py
7python3 scripts/propose_improvement.py

1score=0.75
2changed config/scoring.yaml
3implementation_page_visit: 4 -> 6
4score=1.00
5open PR for human review