按这个模板,可直接整段粘贴:

## Inspiration

Coding agents write code well, but real work still lives in email, Feishu/Lark, GitLab, SSH, and release checklists. I was tired of tab-switching and re-explaining context every time. I wanted Codex to use the same tools I already trust—locally, with clear configs, and with secrets never committed to the repo.

## What it does

**AI Productivity Plugins** is a local Codex plugin marketplace. One repository installs composable plugins for:

- IMAP/SMTP mail (QQ, 163/126, Exmail, Ali Mail, and custom mailboxes)
- Lark / Feishu CLI (docs, wikis, calendar, messages, Base, sheets)
- GitLab (projects, MRs, issues, CI, discussions, approvals, file reads)
- Product release gate (fail-closed submission → tests → approval → RELEASE_READY)
- R&D flywheel (new requirements → production-proven capability with visual decision gates)
- SSH (strict host-key checks, remote commands, SCP, agent lifecycle)
- WeCom messaging + local Codex usage summaries
- Daily vulnerability bulletin email (built on top of Lark + mail)

Open the repo in Codex App, install from the local marketplace, and the agent can run real workplace workflows—not just generate snippets.

## How we built it

We structured everything around Codex’s local marketplace entrypoint (`.agents/plugins/marketplace.json`), then packaged each workflow as a plugin with skills, example configs, and setup wizards where needed.

Design choices:

- **Local-first secrets:** mailbox auth codes, GitLab tokens, WeCom secrets, and SSH keys stay on the user’s machine (`~/.imap-smtp-mail`, env vars, protected config paths)—never in git.
- **Compose, don’t reinvent:** Lark wraps the existing `lark-cli` + skills; the vuln bulletin reuses Lark + IMAP/SMTP instead of duplicating APIs.
- **Explicit contracts:** release-gate adapters return JSON; missing files, bad signatures, failed scans, or SHA1 drift block the event.
- **Provider-aware setup:** mail wizard guides QQ/163/Exmail auth-code flows so users don’t paste web login passwords.

## Challenges we ran into

- Keeping plugins **easy to demo** while remaining **safe to publish** (no real credentials in the repo).
- Building a **fail-closed** release gate: soft warnings are not enough—invalid evidence must stop the pipeline.
- Handling **Chinese mail provider quirks** (IMAP enablement + authorization codes).
- Wrapping existing CLIs/skills without breaking auth or behavior when OpenAPI surfaces change.
- Making multiple plugins feel like one product, not a pile of unrelated scripts.

## Accomplishments that we're proud of

- A working **local marketplace** with multiple production-shaped plugins in one repo.
- End-to-end composition: e.g. **daily vuln bulletin** that verifies Feishu subscribers, MIME Subject/Message-ID readback, and recipient privacy—on top of shared mail/Lark plugins.
- A **fail-closed product release gate** with auditable reports and immutable submission manifests.
- Setup wizards that remove hand-editing JSON for common paths (mail, WeCom).
- A clear boundary: plugins prove readiness and automate workflows; they don’t hide deployment credentials or fake production access.

## What we learned

Agent productivity is less about one mega-prompt and more about **small, reliable tool plugins** with explicit contracts, local config, and auditable gates. When plugins compose cleanly, the agent can finish real workplace loops—mail, docs, GitLab, release checks—inside one session.

We also learned that “works on my machine” only scales if secrets, adapters, and failure modes are designed up front.

## What's next for ai-productivity-plugins

- Submit to the official public Codex marketplace.
- Add tighter end-to-end demos: mail → Feishu → GitLab → release gate as one repeatable loop.
- More enterprise adapters (additional Git forges, chat platforms, and evidence sources).
- Stronger CI for plugin contracts so marketplace updates stay fail-closed as the set grows.

Built With

Share this project:

Updates