* feat: auto-install WeChat article tools during agent-reach install
WeChat (微信公众号) deps are now auto-installed during 'agent-reach install':
- Python packages: camoufox[geoip], markdownify, beautifulsoup4, httpx, miku_ai
- Tool repo: wechat-article-for-ai cloned to ~/.agent-reach/tools/
No login or configuration needed — works out of the box like YouTube/RSS.
Updated README tables to reflect zero-config status.
Also covers safe mode and dry-run.
* feat: auto-install WeChat + add channel-setup reference to SKILL.md
1. WeChat auto-install during 'agent-reach install':
- pip install camoufox[geoip], markdownify, bs4, httpx, miku_ai
- clone wechat-article-for-ai to ~/.agent-reach/tools/
- covers safe mode and dry-run
- README tables: WeChat → zero-config
2. SKILL.md improvements:
- Added 'configure' triggers to description
- Added references/channel-setup.md with install steps for all
login-required platforms (Twitter, XHS, Douyin, LinkedIn, Boss)
- Principle: user only provides cookies, agent does everything else
* simplify: point to install.md URL instead of bundled reference
No need to maintain a separate channel-setup.md. Just tell agents
to fetch install.md when they need setup instructions.
---------
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
SKILL.md was mixing installation/configuration with usage commands.
Now it's a pure usage guide per skill-creator conventions:
- Description includes ALL platform trigger words (search, read,
browse for each of 13 channels) so agents load it when users
ask to search/read any platform
- Body is concise command reference only — no install/config
instructions (those belong in install.md)
- Removed duplicate workspace rules explanation
- Removed cookie setup, proxy config, and channel setup sections
- 324 → 160 lines (51% reduction in context tokens)
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
Two subprocess.run(["npm", ...]) calls in undici installation were
using bare command name, which fails on Windows where npm is
installed as npm.cmd. Now resolves full path via shutil.which()
first, consistent with all other subprocess calls.
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
Problem: After installing Agent Reach, agents may clone repos, create
files, or write output into the workspace directory, gradually polluting
the user's project. Over time this degrades the agent experience.
Fix:
- SKILL.md: add Workspace Rules section — never create files in workspace
- install.md: add Directory Rules table, explicit boundary about workspace
- install.md: all git clone commands now target ~/.agent-reach/tools/
- update.md: add workspace warning
- cli.py: install creates ~/.agent-reach/tools/ directory
All agent-facing docs now enforce: /tmp for temp, ~/.agent-reach/ for
persistent data, workspace stays untouched.
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
- wechat channel now supports both search and read
- Search: miku_ai (sogou weixin search) → article list with URLs
- Read: camoufox (stealth Firefox) → full Markdown content
- Doctor shows granular status (search only / read only / both)
- SKILL.md updated with search + read workflow examples
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
- Replace text=True with encoding='utf-8', errors='replace' in all
subprocess.run calls (channels + cli.py) to fix GBK decode errors
on Chinese Windows systems
- Use shutil.which() resolved paths in channel subprocess calls to
handle Windows .cmd wrappers (npm global installs)
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
- Remove dead '# ── read ──' placeholder comment in cli.py
- Reword SKILL.md to make it explicit that agent-reach is an
installer/config tool, not a content fetcher
- Addresses confusion from issue #58
Co-authored-by: Panniantong <panniantong@users.noreply.github.com>
When running 'agent-reach configure twitter-cookies', credentials are now
automatically written to ~/.config/xfetch/session.json so that 'xreach auth
check' works without manual intervention.
Preserves existing session.json fields if the file already exists.
Falls back gracefully if the sync fails (still saves to agent-reach config).
Fixes#50
yt-dlp now requires an external JS runtime (deno or Node.js) for YouTube.
Since agent-reach already installs Node.js, we configure yt-dlp to use it:
1. cli.py: install writes '--js-runtimes node' to ~/.config/yt-dlp/config
2. youtube.py: doctor checks JS runtime availability and config
3. SKILL.md: updated YouTube section with accurate troubleshooting
Bilibili note: 412 errors are IP-based (needs proxy), not cookie-related.
Fixes#47
YouTube may trigger 'Sign in to confirm you're not a bot' on server IPs.
Bilibili may return 412 on non-residential IPs.
Added --cookies-from-browser fallback examples for both channels.
Fixes#47
- Detect ARM64 architecture and auto-add --platform linux/amd64 to
docker run hints in doctor/check output
- Add ARM64 FAQ section to setup guide with two options:
Rosetta emulation (--platform linux/amd64) or build from source
- Closes#44
mcporter list probes all configured servers sequentially, which can hang
indefinitely when servers are slow/offline. mcporter config list reads
config without probing — sufficient for checking if a server is configured.
Affected: doctor checks (4 channels) and install command (2 calls).
Fixes#34
Fixes#29 — __version__ in __init__.py was still '1.0.0' while pyproject.toml
had already been bumped to '1.2.0'. This caused misleading output in
'agent-reach --version' and 'agent-reach doctor'.
- docs/README_en.md: update FAQ, Works Out of the Box, Design Philosophy
- docs/troubleshooting.md: Exa fallback uses mcporter directly
- agent_reach/guides/setup-exa.md: test with mcporter call
- agent_reach/guides/setup-wechat.md: test with Jina Reader
Adds a check in 'agent-reach doctor' that warns when ~/.agent-reach/config.yaml
is readable by other users (group/world). Suggests chmod 600 to fix.
Ref: #6
核心改动：
1. install.md 加入 Agent 指令：'You are a capable AI Agent. If a tool
fails, try another method. Don't give up after one attempt.'
2. install.md 加入常见问题自修复表（9 种场景 + 对应解法）
3. CLI 错误输出改为 Agent-friendly 语气，给出替代方案而非只说失败
4. install 结束时如果有未激活渠道，提示 Agent 自行修复
