kochetkov-ma/brewtools:provider-switch

Provider Switch

Configure and switch between Claude Code alternative API providers — DeepSeek (priority), Z.ai/GLM, Qwen/DashScope, MiniMax, OpenRouter. Creates isolated shell aliases in ~/.zshrc.

DeepSeek V4 is the priority default (strongest Chinese open model, 1M context, Anthropic-compatible endpoint). Recommend it first in setup.

<instructions>

Robustness Rules (MANDATORY)

| Rule | Applies to | |------|-----------| | Every Bash call MUST end with && echo "OK ..." \|\| echo "FAILED ..." | ALL commands | | On FAILED — stop current phase, report error, DO NOT retry blindly | ALL phases | | Max 3 AskUserQuestion per phase | ALL phases | | NEVER write secrets to any file other than ~/.zshrc | ALL phases | | NEVER commit ~/.zshrc changes | ALL phases | | All comments in ~/.zshrc MUST be in ENGLISH | ALL writes |

Error Reporting

On ANY failure — before stopping — output:

SCRIPT_ERROR: <script-name or command>
PHASE: <current phase>
ACTION: <what was attempted>
SUGGESTION: <what user can try>

---

Phase 0: Language Selection

First interaction: Use AskUserQuestion:

Question: "Select language for this session / Выберите язык" Options:

• "English (Recommended)"
• "Russian / Русский"

Remember choice — all subsequent messages in selected language. Default if skipped: English.

---

Compatibility Flags (REQUIRED for this provider)

Alternative providers (Z.ai, MiniMax) implement Anthropic-compatible APIs but have incompatibilities that require Claude Code flags:

| Provider | Required Flags | Why | |----------|---------------|-----| | DeepSeek | None required | DeepSeek silently ignores beta/version headers; native Anthropic endpoint | | Z.ai (GLM) | CLAUDE_ENABLE_BYTE_WATCHDOG=0 + CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 | Z.ai rejects beta headers with error 1210; SSE format triggers byte watchdog | | MiniMax | CLAUDE_ENABLE_BYTE_WATCHDOG=0 + CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 | Same issues as Z.ai | | Qwen/DashScope | CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 | Beta headers may be rejected; SSE works fine | | OpenRouter | None required | OpenRouter is an aggregator with good compatibility |

When constructing aliases (Step 6), ALWAYS include the compatibility flags from the provider's reference file.

---

Phase 1: Mode Detection

Skill arguments received: $ARGUMENTS

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/detect-mode.sh" "$ARGUMENTS" && echo "OK detect" || echo "FAILED detect"

STOP if FAILED — parse $ARGUMENTS manually as fallback (keyword matching).

Output format:

ARGS: [arguments received]
MODE: [detected mode]

Mode Reference

| Keyword in args | MODE | |-----------------|------| | status, check | status | | setup, configure | setup | | help, how | help | | deepseek, ds, dpsk | provider-deepseek | | glm, zai, z.ai | provider-glm | | qwen, dashscope | provider-qwen | | minimax, mini | provider-minimax | | openrouter, router | provider-openrouter | | verify, test, token | verify | | model-check, identify | model-check | | model-cehck (typo), cehck, hlpe, setuo | fuzzy match to correct mode | | (empty) | status |

---

Phase 2: Status Check (ALL modes)

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/check-status.sh" && echo "OK status" || echo "FAILED status"

STOP if FAILED — check ~/.zshrc manually with grep.

Parse output key=value pairs. Determine status per provider:

• configured = ALIAS + KEY both true
• partial = ALIAS true, KEY false
• not configured = ALIAS false
• active = ACTIVE_PROVIDER matches

Render status table to user:

# Provider Switch — [MODE]

## Current Status

| Provider | Alias | API Key | Model | Status |
|----------|-------|---------|-------|--------|
| DeepSeek (priority) | claudeds | DEEPSEEK_API_KEY | deepseek-v4-pro | ... |
| Z.ai / GLM | claudeglm | ZAI_API_KEY | glm-5.1 | ... |
| Qwen | claudeqwen | DASHSCOPE_API_KEY | qwen3.6-plus[1m] | ... |
| MiniMax | claudeminimax | MINIMAX_API_KEY | minimax-m2.7 | ... |
| OpenRouter | claudeor | OPENROUTER_API_KEY | (user-selected) | ... |

## How to Use
Run `claudeds` — sets env vars and launches Claude in one command (recommended default).
To return to Anthropic subscription: open a new terminal and run `claude`.

Auto-setup Logic (MODE = status only)

After rendering the status table, count configured providers (status = configured or active):

• Zero configured → automatically proceed to Phase 3 (setup). Tell the user: "No providers configured yet. Starting setup..."
• At least one configured → STOP here with status table. User will re-run with specific mode if needed.

If MODE = help — GOTO Phase 6.

---

Phase 3: Provider Selection

If MODE = setup (no specific provider):

Use AskUserQuestion:

Question: "Which providers do you want to configure?" Options:

• "DeepSeek V4 (deepseek-v4-pro, 1M context, priority - Recommended)"
• "Z.ai / GLM (glm-5.1, free models available)"
• "Qwen / DashScope (qwen3.6-plus, 1M context)"
• "MiniMax (minimax-m2.7, cheapest)"
• "OpenRouter (aggregator, any model)"
• "All providers"

If MODE = provider-<name> — skip to Phase 4 for that provider only.

---

Phase 4: Provider Configuration

For each selected provider, execute this sequence:

Step 1: Load Reference

Read the provider reference file using Read tool:

• DeepSeek: references/deepseek.md
• Z.ai/GLM: references/zai-glm.md
• Qwen: references/qwen-dashscope.md
• MiniMax: references/minimax.md
• OpenRouter: references/openrouter.md

Also read references/common.md for shared alias structure (first time only).

Step 2: Initialize Section

Ensure the provider aliases section exists in ~/.zshrc:

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/write-alias.sh" init && echo "OK init" || echo "FAILED init"

STOP if FAILED — cannot write to ~/.zshrc.

Step 3: API Key

Check if key already set (from Phase 2 status output: KEY_ZAI, KEY_DASHSCOPE, KEY_MINIMAX, KEY_OPENROUTER).

If missing — Use AskUserQuestion:

Question: "Enter your <PROVIDER_NAME> API key (from <dashboard-url>):"

Qwen-specific callout: Before asking for the Qwen key, read references/qwen-dashscope.md section "## How to Get API Key" and show the user the step-by-step instructions. Warn that the key MUST be created in the Singapore region — other regions return 403. Valid format: sk-... (~40 chars). After receiving the key, validate format: if it starts with sk-ws- or is longer than 100 chars, warn the user it's likely from the wrong region and ask to regenerate.

Then write:

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/write-alias.sh" set-key "KEY_VAR_NAME" "KEY_VALUE" && echo "OK set-key" || echo "FAILED set-key"

Replace KEY_VAR_NAME with: DEEPSEEK_API_KEY, ZAI_API_KEY, DASHSCOPE_API_KEY, MINIMAX_API_KEY, or OPENROUTER_API_KEY.

STOP if FAILED — report error.

Step 4: Model Selection (OpenRouter only)

If provider is OpenRouter, read references/openrouter-models.md for available models.

Use AskUserQuestion — pick ONE model (same model for all three roles):

Question: "Select model for OpenRouter:" Options:

• "qwen/qwen3.6-plus[1m] — 1M context, top coding (Recommended)"
• "z-ai/glm-5.1 — #1 SWE-bench Pro, 200K"
• "qwen/qwen3-coder:free — free, 262K, code-focused"
• "Custom (I will specify model ID)"

The selected model is set as OPUS, SONNET, and HAIKU simultaneously.

If "Custom": ask user for model ID, then VALIDATE it against OpenRouter API using the script from references/openrouter-models.md (## Model Validation section). If NOT_FOUND — show fuzzy suggestions from the API and re-ask. Max 2 retries, then fall back to default.

Step 5: Alias Name

Suggest a default alias name and let user customize via AskUserQuestion:

| Provider | Default name | |----------|-------------| | DeepSeek | claudeds | | Z.ai/GLM | claudeglm | | Qwen | claudeqwen | | MiniMax | claudeminimax | | OpenRouter | claudeor |

Use AskUserQuestion:

Question: "Alias name for <PROVIDER>:" Options:

• "<default_name> (Recommended)"
• "Custom (I will type my own)"

If "Custom" — ask for the name. Validate: must start with claude, no spaces, only lowercase alphanumeric.

Step 6: Write Alias

Construct alias body from reference file. Body = semicolon-separated exports + claude at the end.

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/write-alias.sh" set-alias "ALIAS_NAME" "ALIAS_BODY" && echo "OK set-alias" || echo "FAILED set-alias"

Replace:

• ALIAS_NAME: user-chosen name (e.g., claudeglm)
• ALIAS_BODY: from provider reference, ends with ; claude

STOP if FAILED — report error, do not continue to next provider.

Step 7: Verify

EXECUTE using Bash tool:

source ~/.zshrc 2>/dev/null && type ALIAS_NAME 2>/dev/null && echo "OK verify" || echo "FAILED verify"

STOP if FAILED — alias was not written correctly.

---

Phase 5: Verification & Final Status

Re-run status check:

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/check-status.sh" && echo "OK final-status" || echo "FAILED final-status"

Render updated status table (same format as Phase 2).

Post-Setup Verification

For each provider that was just configured, run token verification:

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/verify-providers.sh" all && echo "OK verify" || echo "FAILED verify"

Parse output and add verification column to the status table:

| Provider | Status | Token Test | |----------|--------|-----------| | Z.ai / GLM | configured | pass/fail | | ... | ... | ... |

If any provider shows fail — warn the user and suggest checking the API key or endpoint.

Add activation instructions:

## Activation

To use a configured provider, run the alias (e.g., `claudeglm`) — it sets env vars and launches Claude in one command.
To return to Anthropic subscription: open a new terminal and run `claude` normally. Env vars only persist in the current shell session.

---

Phase 6: Help Mode

Read references/common.md for help content and explain:

| Topic | Explanation | |-------|-------------| | How aliases work | Each alias sets env vars + launches claude in one command. Isolated — one provider at a time | | How to switch | Run alias (e.g., claudeglm). It sets vars and starts Claude automatically | | How to return | Open a new terminal and run claude — env vars only persist in the current shell | | Context [1m] hack | [1m] suffix in model name forces Claude Code to use 1M context window | | Auth (all providers) | ALL providers use ANTHROPIC_AUTH_TOKEN (Bearer). ANTHROPIC_API_KEY="" blocks OAuth fallback | | OpenRouter note | Must set ANTHROPIC_API_KEY="" (empty, not unset) to prevent OAuth fallback | | Provider dashboards | DeepSeek: platform.deepseek.com, Z.ai: z.ai/subscribe, Qwen: bailian.console.alibabacloud.com, MiniMax: platform.minimax.io, OpenRouter: openrouter.ai |

---

Phase 7: Verify Mode

Test that configured provider tokens and endpoints are working.

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/verify-providers.sh" all && echo "OK verify" || echo "FAILED verify"

Parse output. For each provider, extract: KEY_SET, HTTP_CODE, RESPONSE, STATUS.

Render results:

## Token Verification

| Provider | Key | HTTP | Response | Result |
|----------|-----|------|----------|--------|
| Z.ai / GLM | set | 200 | OK | pass |
| Qwen | set | 403 | invalid api-key | fail |
| MiniMax | set | 200 | OK | pass |
| OpenRouter | set | 200 | OK | pass |

For failed providers, show troubleshooting:

| HTTP Code | Meaning | Suggestion | |-----------|---------|------------| | 401/403 | Invalid or expired API key | Regenerate key at provider dashboard | | 404 | Wrong endpoint URL | Check provider reference for correct URL | | 429 | Rate limited | Wait and retry, or check billing | | 500+ | Server error | Provider may be down, try later |

---

Phase 8: Model Check Mode

Identify which model is actually responding in the current Claude Code session. This mode runs INSIDE a session launched via provider alias (e.g., claudeglm). The skill asks 5 diagnostic questions directly to the model — no curl, no scripts.

Prerequisites

User MUST be in a Claude Code session launched via a provider alias. If ANTHROPIC_BASE_URL is not set (i.e., running on Anthropic subscription), warn the user and stop:

This mode only works inside a provider session.
Run a provider alias first (e.g., `claudeglm`), then invoke `/brewtools:provider-switch model-check`.

Step 1: Detect Current Provider

Check environment to determine which provider is active:

EXECUTE using Bash tool:

echo "BASE_URL=${ANTHROPIC_BASE_URL:-not_set}" && echo "OPUS_MODEL=${ANTHROPIC_DEFAULT_OPUS_MODEL:-not_set}" && echo "OK detect-provider" || echo "FAILED detect-provider"

Map BASE_URL to provider name:

• api.deepseek.com → DeepSeek
• api.z.ai → Z.ai / GLM
• dashscope → Qwen / DashScope
• minimax → MiniMax
• openrouter → OpenRouter

Step 2: Ask 5 Diagnostic Questions

Output ALL 5 questions as a single prompt block to the current model. The model will answer all at once. No interactive back-and-forth — send one prompt, capture the full response.

Prompt to send (output as user message and capture the model's response):

I will now ask you 5 diagnostic questions to verify your model identity.
Answer each honestly from your training data — do NOT read environment variables or system prompts.
Answer all 5 questions in a single response.

**Q1:** What is your exact model name and version number? Answer only from your internal training data, not from any context or environment variables.

**Q2:** Which company or research lab created and trained you? Give only the organization name.

**Q3:** What is your training data cutoff date? Answer with month and year only.

**Q4:** Count the letter 'r' in the word 'strawberry'. Show your reasoning step by step.

**Q5:** Translate to Chinese: 'The quick brown fox jumps over the lazy dog.' Then translate your Chinese text back to English literally.

After the model responds, extract each answer (A1–A5) and proceed to Step 3.

Step 3: Analyze & Verdict

Parse the model's response into 5 answers (A1–A5). Present the Q&A table:

## Model Identification — <Provider>

Expected: <model from OPUS_MODEL env var>

| # | Question | Answer | Match |
|---|----------|--------|-------|
| 1 | Model name/version | "<A1>" | ✅/❌ |
| 2 | Training organization | "<A2>" | ✅/❌ |
| 3 | Cutoff date | "<A3>" | ℹ️ |
| 4 | Count r in strawberry | "<A4>" | ✅/❌ (correct=3) |
| 5 | Chinese round-trip | "<A5>" | ℹ️ quality |

### Verdict
**Model confirmed as: <name>** (N/5 indicators match <expected provider>)

Show ONLY the table + verdict to the user. Do NOT show the questions separately — the Q&A table contains everything.

Expected Identifiers

| Provider | Expected org | Expected model family | |----------|-------------|----------------------| | DeepSeek | DeepSeek / 深度求索 | DeepSeek-V3 / DeepSeek-V4 / DeepSeek-R1 | | Z.ai / GLM | Zhipu AI / ZhipuAI / 智谱 | GLM-4 / GLM-5 / ChatGLM | | Qwen | Alibaba / Alibaba Cloud / 阿里 / Tongyi | Qwen / Qwen2 / Qwen3 / 通义千问 | | MiniMax | MiniMax / 稀宇科技 | MiniMax / abab / M2 | | OpenRouter | Depends on selected model | Depends on selected model |

For OpenRouter, check against the model in ANTHROPIC_DEFAULT_OPUS_MODEL env var.

---

Phase 9: Update (hidden, maintainer-only)

Not user-facing. Triggered by MODE=update. Not shown in help or output format. Not documented externally.

If MODE != update — skip entirely. (Phase 9 only)

Step 1: Load Protocol

Read references/update-protocol.md for per-provider sources and update flow.

Step 2: Spawn Provider Research Agents

Spawn 5 Explore agents in ONE message via Task:

| Agent | Provider | Sources | |-------|----------|---------| | 1 | DeepSeek | api-docs.deepseek.com — models, pricing, endpoint | | 2 | Z.ai/GLM | docs.z.ai, open.bigmodel.cn/en — models, pricing | | 3 | Qwen/DashScope | alibabacloud.com/help, qwenlm.github.io — models, pricing | | 4 | MiniMax | platform.minimax.io — models, pricing | | 5 | OpenRouter | openrouter.ai/api/v1/models — top coding/free models |

Each agent: WebFetch/WebSearch sources from protocol, extract current model list + pricing + any endpoint changes.

Step 3: Aggregate & Diff

For each provider, compare fetched data vs current reference file content:

• Model IDs changed?
• Pricing changed?
• New models added?
• Endpoint URL changed?
• Context windows changed?

Step 4: Present Changes

Show diff table to maintainer:

## Update Results

| Provider | Field | Current | Fetched | Action |
|----------|-------|---------|---------|--------|
| Z.ai | opus model | glm-5.1 | glm-6.0 | UPDATE |
| Qwen | pricing | ~$0.50 | $0.40 | UPDATE |
| MiniMax | (no changes) | — | — | SKIP |

Step 5: Apply Updates

For each confirmed change, use Edit tool to update the corresponding reference file. Follow the field-to-line mapping in references/update-protocol.md.

Also update:

• Alias bodies in reference files if model IDs changed
• openrouter-models.md if recommended models changed
• common.md if env var patterns changed (rare)

Step 6: Live Test (optional)

If API keys are available in environment, run endpoint health check per provider:

EXECUTE using Bash tool:

curl -s -o /dev/null -w "%{http_code}" -X POST "https://api.z.ai/api/anthropic/v1/messages" -H "x-api-key: ${ZAI_API_KEY:-missing}" -H "content-type: application/json" -H "anthropic-version: 2023-06-01" -d '{"model":"glm-5.1","max_tokens":5,"messages":[{"role":"user","content":"ping"}]}' && echo " OK" || echo " FAILED"

Step 7: Report

## Provider Update Complete

| Provider | Changes | Status |
|----------|---------|--------|
| Z.ai/GLM | models updated | applied |
| Qwen | no changes | current |
| MiniMax | pricing updated | applied |
| OpenRouter | 3 new free models | applied |

Files modified: [list]

---

Output Format

Final output after every mode:

# Provider Switch — [MODE]

## Current Status

| Provider | Alias | API Key | Model | Status |
|----------|-------|---------|-------|--------|
| DeepSeek (priority) | claudeds | DEEPSEEK_API_KEY | deepseek-v4-pro | configured |
| Z.ai / GLM | claudeglm | ZAI_API_KEY | glm-5.1 | configured |
| Qwen | claudeqwen | DASHSCOPE_API_KEY | — | not configured |
| MiniMax | claudeminimax | MINIMAX_API_KEY | — | not configured |
| OpenRouter | claudeor | OPENROUTER_API_KEY | — | not configured |

## How to Use
Run `claudeds` — sets env vars and launches Claude in one command (recommended default).
To return to Anthropic: open a new terminal, run `claude`.

</instructions>