workbuddy-cli-model-bridge
Install, audit, repair, and manage a loopback-only CLIProxyAPI
zjp1997720 · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add zjp1997720/zhijian-skills --skill workbuddy-cli-model-bridgeAbout this skill
name: workbuddy-cli-model-bridge description: Install, audit, repair, and manage a loopback-only CLIProxyAPI bridge that registers subscription-backed Codex, Grok, Antigravity/Gemini, GLM Coding Plan, or newly added CLI models in WorkBuddy. Use whenever the user asks to connect a local CLI or agent model to WorkBuddy, install CLIProxyAPI, add GLM-5.3 from a Coding Plan key, add or repair a WorkBuddy custom model, preserve image/tool/reasoning/Fast capabilities, onboard another CLI provider, or diagnose a WorkBuddy model that stopped working. This Skill is macOS-first and should be used even when the user names only the model or says their local WorkBuddy API has expired. disable-model-invocation: true
WorkBuddy CLI Model Bridge
Turn supported CLI subscriptions into verified WorkBuddy custom-model entries through CLIProxyAPI. Use the bundled script for discovery, installation, OAuth handoff, capability probes, and idempotent WorkBuddy updates. Keep model claims tied to real probes; a model name or marketing page is not evidence that a particular proxy route preserves the capability.
Resolve the Skill directory
Resolve this loaded Skill's directory and use its absolute path as <skill-dir>. Do not assume a specific installation root.
The deterministic entry point is:
python3 <skill-dir>/scripts/bridge.py
Default workflow
1. Audit before changing anything
python3 <skill-dir>/scripts/bridge.py audit
Read the JSON findings. The audit redacts secrets and detects:
- Homebrew, CLIProxyAPI, active config, bind address, and proxy reachability
- WorkBuddy initialization and existing model count
- bundled and local Provider manifests
- CLI executables and login signals without reading token contents
- CLIProxyAPI auth-file counts without exposing account names
Treat an existing CLI login as a discovery signal. CLIProxyAPI may still require its own OAuth grant; do not copy another CLI's token file.
2. Bootstrap CLIProxyAPI when required
Preview:
python3 <skill-dir>/scripts/bridge.py bootstrap
Apply the plan when the user asked to install, configure, repair, or converge the bridge:
python3 <skill-dir>/scripts/bridge.py bootstrap --apply
The request to set up the bridge authorizes reversible Homebrew installation, loopback service startup, a dedicated random proxy client key, secure local state, and timestamped backups. Pause only when:
- Homebrew itself is missing and its installer requires administrator interaction
- an existing proxy listens beyond loopback; explain the remote-client impact before using
--allow-rebind-local - an overlapping deployment or config conflict makes ownership ambiguous
Keep a healthy existing installation in place. Prefer official Homebrew installation for a new macOS setup.
3. Authorize only relevant Providers
Bundled OAuth Provider IDs are codex, xai-grok, and antigravity. Authorize one when the user requested it or the audit found its CLI/login signal and CLIProxyAPI has no matching auth/model route.
python3 <skill-dir>/scripts/bridge.py authorize codex
python3 <skill-dir>/scripts/bridge.py authorize xai-grok
python3 <skill-dir>/scripts/bridge.py authorize antigravity
The command delegates to CLIProxyAPI's native OAuth flag. Tell the user to approve the browser page, then continue automatically. Never paste, print, transform, or reuse OAuth tokens. Secure resulting auth JSON files to owner-only permissions.
Bundled glm-coding is a Coding Plan API-key Provider, not OAuth. Do not run authorize glm-coding. Follow glm-coding-plan.md: add the official coding /paas/v4 route to CLIProxyAPI, then sync --providers glm-coding. Do not use 智谱清言 website login.
Do not authorize unrelated Providers merely because they are bundled.
4. Probe and synchronize WorkBuddy
Pass only the relevant Provider IDs:
python3 <skill-dir>/scripts/bridge.py sync --providers codex,xai-grok,glm-coding --apply
The sync command:
- fetches the live
/v1/modelslist - selects recommended chat/agent models from Provider manifests
- probes text, SSE streaming, tools, images, and reasoning controls where declared
- resolves token limits by exact model ID, using Provider-declared catalogs before sourced manifest metadata
- refuses to register a model with an unknown input/output limit or a rejected output-limit probe
- skips models whose text or streaming probes fail
- downgrades optional capability flags when their probes fail
- backs up and atomically updates
~/.workbuddy/models.json - preserves manual entries and stale managed entries
- refuses to overwrite a manual entry with the same model ID
- records ownership separately under
~/.config/workbuddy-cli-model-bridge/
Do not use --skip-probes for a live setup. That option exists for deterministic offline tests only.
When the user explicitly requests Fast mode and no Fast model appears, read the Fast-model section in troubleshooting.md. A Fast alias needs both a distinct CLIProxyAPI route and verified priority semantics; never relabel an ordinary model as Fast.
A reasoning probe verifies control compatibility. It cannot expose private chain of thought. For teaching or screen recording, ask the model to emit a deliberate problem decomposition as normal answer content.
WorkBuddy represents each custom model as a separate entry. This Skill registers verified entries; it does not change the user's currently selected conversation model.
5. Verify the application consumed the change
Run the audit again. If WorkBuddy is already open, inspect its current main-thread log for a models.json changed or Loaded custom models config event. If no reload event appears, ask the user to reopen model settings or start a new conversation; do not rewrite the same config repeatedly.
Report:
- CLIProxyAPI version/path and local endp