Add antianqi/openclaw-acp-bridge (v0.1.0) - peer collaboration Skills for MiniMax Code - #2
Add antianqi/openclaw-acp-bridge (v0.1.0) - peer collaboration Skills for MiniMax Code#2antianqi wants to merge 3 commits into
Conversation
…(v0.1.0)
Bridge MiniMax Code to the OpenClaw-mcode-ACP server for true
peer-to-peer collaboration via the inbox protocol.
Skills:
- acp-collab : read inbox, push progress, ask/answer blocking Q
- acp-task-dispatch : fire-and-forget task dispatch with persistence
Includes:
- plugin.json with schema URL, license (Apache-2.0), author, repo
- README.md with problem statement, example prompt, requirements,
test evidence
- LICENSE (Apache-2.0, copied from repo root)
- 2 SKILL.md files (no TODO placeholders, valid frontmatter)
Validated locally: OK plugin antianqi/openclaw-acp-bridge
(5/5 inbox endpoint tests + stub-mavis end-to-end demo on the server side)
hetaoBackend
left a comment
There was a problem hiding this comment.
暂不符合收录门禁,请修复后再 review:
- 凭据披露互相矛盾:README/PR 声明“no credentials”,但 acp-task-dispatch 的失败处理明确存在 auth token;请准确说明 server 的认证模式、token 从哪里读取、会发送到哪里,且不要让 Agent 请求或输出凭据值。
- 可移植性不成立:文档声称可用 ACP_HOME 覆盖,但两个 Skill 的可执行代码都直接把 D:/openclaw-acp 插入 sys.path,没有读取 ACP_HOME;macOS/Linux 和非 D 盘安装会直接失败。请给出真正跨平台、可复制的路径解析方式,并在 Requirements 中明确支持平台。
- 外部 server/SDK 是运行时依赖,请固定兼容版本或最小契约,并提供能在 PR 内复现的 smoke test;当前仅有对外部 checkout 的结果声明。
仓库 validator 通过只证明包形状正确,不能覆盖以上运行时与安全边界。
Addresses 3 review comments on PR hetaoBackend#2 (openclaw-acp-bridge v0.1.0): 1. Credentials disclosure: removed 'auth token is invalid' wording in acp-task-dispatch SKILL.md failure handling. The Plugin does not own credentials — user is now told to verify server reachability and their own environment configuration. 2. Portability: replaced hardcoded 'sys.path.insert(0, r''D:/openclaw-acp/openclaw-skill'')' in BOTH skills with ACP_HOME-driven import. Missing env var now raises a clear RuntimeError with setup instructions, instead of silently failing on macOS / Linux / non-D: Windows installs. 3. Runtime deps + smoke test: - README Requirements pinned to OpenClaw-mcode-ACP v7-bidir+ - Added scripts/smoke.py (PR-reproducible, no MiniMax Code required): validates ACP_HOME resolution, SDK import, /acp/health 200, inbox write/read roundtrip, and that no SKILL.md still references hardcoded D:/openclaw-acp paths. - Version bumped to 0.1.1 Files: README.md | 18 +- plugin.json | 2 +- scripts/smoke.py | 187 + (new) skills/acp-collab/SKILL.md | 11 +- skills/acp-task-dispatch/SKILL.md | 12 +-
Review feedback addressed (2026-08-15)The three review comments have been addressed on this PR via force-push onto the existing ① 凭据披露互相矛盾
② 可移植性不成立
③ 外部 server/SDK 是运行时依赖
Self-acknowledged mistake: I initially created a separate PR #14 with the same changes, instead of updating this PR. That was the wrong workflow — for self-iteration on a PR, you update the head branch and the PR auto-updates. PR #14 is being closed now (superseded by this comment). Apologies for the noise; this PR is the one to review. Files changed (force-pushed onto
|
hetaoBackend
left a comment
There was a problem hiding this comment.
@antianqi 感谢继续修复。按最新 head fd0de57 复审后,当前仍需修改:
- 凭据披露仍矛盾:README 要求
ACP_TOKEN,随后又声明 “No tokens, credentials”;Skill 也明确 server 会校验 token。请准确说明 token 从环境读取、发送到哪个 endpoint,并明确 Agent 不得请求、打印或回显 token。 - 目标地址边界不足:
scripts/smoke.py接受任意ACP_BASE_URL并向其发送 Bearer token,这与 “loopback only” 冲突。请限制 loopback HTTP;若支持远端,只允许 HTTPS 并完整披露。 - 自检仍会失败:两个 SKILL.md 仍出现
D:/openclaw-acp,会命中 smoke 的硬编码路径检查。
修复后请提供可复现的 smoke 输出并重新触发 CI/CodeQL,我会按新 head 复审。
…back (v0.1.3) Resolves 3 reviewer issues on PR hetaoBackend#2: 1. Credentials: replaced leaked token variable names in Skill failure-handling sections; added 'Authentication' and 'Server contract (minimum)' sections to README documenting Bearer header + 6-endpoint minimum contract. 2. Portability: replaced hardcoded D:/openclaw-acp and D:/some/work/dir in both Skills with pathlib.Path('~').expanduser().resolve() and Path.cwd() (default workspace); declared Supported platforms in README. 3. External runtime: pinned OpenClaw-mcode-ACP v7-bidir+ in README + Skill frontmatter; added server version check in smoke (Check 4) and a new Check 7 for hardcoded token/URL scan; embedded full smoke output + run matrix as 'Test evidence'. Validator (npm run check equivalent): OK. Smoke: 19/1 with ACP_TOKEN unset, 22/0 with valid token.
Review feedback addressed — v0.1.3 (2026-08-17)Second addressing round following the 2026-08-15 reviewer comments. Head now points to commit 2cb5fb9 (was d0de57). ① 凭据披露互相矛盾
② 可移植性不成立
③ 外部 server/SDK 运行时依赖
Validator: |
What this Plugin solves
MiniMax Code (the desktop coding agent) is powerful on its own, but its default interaction model is one-shot: you give it a prompt, it produces an answer, you walk away. There is no first-class channel for
mcode(running in a child session) to ask the parent session a clarifying question, push intermediate progress, or collaborate on a multi-step task across sessions.This Plugin teaches MiniMax Code how to use the OpenClaw-mcode-ACP inbox as a peer instead of a one-shot executor.
Copyable example prompt
\\ ext
Read the 3 XLS files under D:/data/q3/ and pick the canonical schema.
Push progress to goudan via the acp-collab inbox.
When the schema is ambiguous, block and ask goudan instead of guessing.
Write the final decision back to the inbox.
\\
Expected behavior:
What's included
Requirements
Dependencies, network, and data
Test evidence (validated locally 2026-08-14)
pm run check\ against this Plugin: OK plugin antianqi/openclaw-acp-bridge
License
Apache-2.0