Agent-Skills.md

Agent Skills: Copilot BYOK Provider Configuration

Configure and switch between BYOK (Bring Your Own Key) LLM providers for both GitHub Copilot CLI and VS Code Chat. Use when setting up OpenAI, Azure OpenAI, Anthropic, Ollama, Moonshot, OpenCode Go, or other OpenAI-compatible endpoints; creating or switching reusable provider profiles for CLI; switching between multiple accounts (API keys) for the same provider; configuring chatLanguageModels.json for VS Code; calculating max prompt or output token overrides; configuring wire API and reasoning effort; or troubleshooting COPILOT_PROVIDER_BASE_URL, COPILOT_PROVIDER_TYPE, COPILOT_PROVIDER_API_KEY, COPILOT_MODEL, COPILOT_PROVIDER_WIRE_API, COPILOT_PROVIDER_MAX_PROMPT_TOKENS, COPILOT_PROVIDER_MAX_OUTPUT_TOKENS, COPILOT_OFFLINE, and VS Code language model settings.

UncategorizedID: arisng/github-copilot-fc/copilot-byok

Repository

arisng/github-copilot-fc
arisng
4

Install this agent skill to your local

pnpm dlx add-skill https://github.com/arisng/github-copilot-fc/tree/HEAD/skills/copilot-byok

Skill Files

Browse the full folder contents for copilot-byok.

Download Skill

Loading file tree…

skills/copilot-byok/SKILL.md

Skill Metadata

Name
copilot-byok
Description
Configure and switch between BYOK (Bring Your Own Key) LLM providers for both GitHub Copilot CLI and VS Code Chat. Use when setting up OpenAI, Azure OpenAI, Anthropic, Ollama, Moonshot, OpenCode Go, or other OpenAI-compatible endpoints; creating or switching reusable provider profiles for CLI; switching between multiple accounts (API keys) for the same provider; configuring chatLanguageModels.json for VS Code; calculating max prompt or output token overrides; configuring wire API and reasoning effort; or troubleshooting COPILOT_PROVIDER_BASE_URL, COPILOT_PROVIDER_TYPE, COPILOT_PROVIDER_API_KEY, COPILOT_MODEL, COPILOT_PROVIDER_WIRE_API, COPILOT_PROVIDER_MAX_PROMPT_TOKENS, COPILOT_PROVIDER_MAX_OUTPUT_TOKENS, COPILOT_OFFLINE, and VS Code language model settings.

Copilot BYOK Provider Configuration

Use this skill to configure BYOK (Bring Your Own Key) LLM providers for both GitHub Copilot CLI and VS Code Chat. Manage repeatable CLI provider profiles and VS Code chatLanguageModels.json from a single source of truth.

Follow this workflow

  1. Determine whether the user needs a CLI profile-based setup, VS Code BYOK setup, a one-off manual setup, or troubleshooting.
  2. For CLI, prefer scripts/byok-profile.ps1 for repeated use or quick switching between providers.
  3. For VS Code, read references/copilot-vscode-providers.md and use the Chat: Manage Language Models UI command.
  4. Read only the reference file that matches the current need.
  5. Keep secrets out of files. Prefer ${ENV_VAR} placeholders and user-scoped environment variables.

Choose the path

CLI paths

  • Reusable profile workflow: Use scripts/byok-profile.ps1.
  • Manual one-off environment setup: Read references/copilot-cli-providers.md.
  • API key storage or rotation: Read references/api-key-storage.md.
  • Token-limit sizing: Read references/copilot-cli-providers.md first, then calculate conservative prompt and output limits.
  • Reasoning-level configuration: Read the grounded per-model lookup references/reasoning-effort-lookup.md, then apply --reasoning-effort per invocation. Cross-surface mapping (CLI flag ↔ env var ↔ SDK API ↔ VS Code) lives in the copilot-cli-subsession skill's copilot-sdk-parity-matrix.md.

VS Code path

  • VS Code BYOK setup: Read references/copilot-vscode-providers.md and use the Chat: Manage Language Models UI command to add models. VS Code uses chatLanguageModels.json and ignores COPILOT_PROVIDER_* env vars.

Use the profile manager first

Use scripts/byok-profile.ps1 when the user wants repeatable setup, named profiles, or quick switching.

Run the following commands from the installed copilot-byok skill folder (the folder that contains this SKILL.md).

Common commands:

# List profiles
.\scripts\byok-profile.ps1 list

# Add a profile interactively
.\scripts\byok-profile.ps1 add

# Inspect a stored profile
.\scripts\byok-profile.ps1 show openai

# Run Copilot CLI with a profile for one session
.\scripts\byok-profile.ps1 run ollama

# Apply a profile to the current shell
. .\scripts\byok-profile.ps1 set-env openai

Pass extra Copilot CLI arguments through run (do not pass --model; model is sourced from the profile):

.\scripts\byok-profile.ps1 run openai --help

Profiles are stored in ~/.copilot/byok-profiles.json or $env:COPILOT_HOME\byok-profiles.json.

Switch between multiple provider accounts

When you have multiple subscriptions for the same provider (for example, two OpenCode Zen accounts with separate API keys), register the accounts once and switch per session — no profile edits needed.

1. Register accounts in ~/.copilot/byok-profiles.json

{
  "accounts": {
    "opencode-home": { "keyEnv": "OPENCODE_API_KEY_HOME", "label": "OpenCode Zen (Home)" },
    "opencode-work": { "keyEnv": "OPENCODE_API_KEY_WORK", "label": "OpenCode Zen (Work)" }
  },
  "activeAccount": "opencode-home"
}

keyEnv holds the name of an environment variable with that account's key — never the raw key. activeAccount sets the default.

2. Opt profiles in with accountGroup

Add "accountGroup": "opencode" to each profile that should use the registry (the add wizard sets it automatically for the OpenCode Go preset). Profiles without accountGroup never participate.

3. Manage and switch accounts

.\scripts\byok-profile.ps1 accounts
.\scripts\byok-profile.ps1 use opencode-work
.\scripts\byok-profile.ps1 run opencode-go-deepseek-v4-flash
.\scripts\byok-profile.ps1 run opencode-go-deepseek-v4-flash --account opencode-work

Resolution order: --account flag → profile account pin → activeAccount. If nothing resolves, the profile falls back to its legacy apiKey with a warning. For sub-sessions, pass -ByokAccount opencode-work to Invoke-CopilotCliSubSession.ps1.

Read references on demand

  • references/copilot-cli-providers.md
    • Read when you need CLI-specific provider environment variables, examples, model requirements, or offline-mode notes.
  • references/copilot-vscode-providers.md
    • Read when the user wants to configure BYOK models in VS Code Chat. VS Code uses a completely different mechanism (chatLanguageModels.json) and ignores COPILOT_PROVIDER_* env vars. This reference covers multi-provider setup, per-agent model pinning via .agent.md frontmatter, agent-specific model settings, and the full model-to-VS-Code mapping from byok-profiles.json.
  • references/api-key-storage.md
    • Read when the user needs secure key storage, persistent Windows environment variables, key rotation, or ${ENV_VAR} placeholder guidance.

Apply these operating rules

  • Prefer ${ENV_VAR} placeholders over raw API keys in JSON.
  • Treat openai as the default provider type for OpenAI-compatible endpoints such as Ollama, vLLM, Foundry Local, and Moonshot.
  • Set COPILOT_PROVIDER_TYPE=azure only for Azure OpenAI and anthropic only for Anthropic.
  • OpenCode Go uses a shared base URL with two possible provider types: openai (append /v1) for DeepSeek, GLM, Kimi, and MiMo models; anthropic (no /v1 suffix — SDK adds it) for MiniMax and Qwen models. Store the personal OpenCode key as OPENCODE_API_KEY_HOME and the work key as OPENCODE_API_KEY_WORK — both at User scope (never Machine scope).
  • CRITICAL: COPILOT_MODEL must use the bare model ID (e.g., deepseek-v4-flash), not the opencode-go/ prefix. The prefix is only used in OpenCode TUI config and in Copilot CLI profile names — never in COPILOT_MODEL.
  • For GPT-5 class OpenAI models, prefer COPILOT_PROVIDER_WIRE_API=responses.
  • Use COPILOT_OFFLINE=true only when the user explicitly wants Copilot CLI isolated from GitHub services; note that full isolation still depends on the provider endpoint being local or private.
  • If the model is not in Copilot CLI's built-in catalog, set explicit prompt and output token overrides instead of assuming Copilot will infer them correctly.
  • Profile proxyPort field: For Kimi models from https://api.moonshot.ai (which require top_p=0.95), add "proxyPort": 443 to the profile and the run command will auto-start the local proxy and route through https://moonshot.local/v1. The proxy strips top_p to 0.95 before forwarding to Moonshot.

Configure reasoning effort correctly

Use Copilot CLI's --reasoning-effort option for model reasoning level control. The authoritative per-model lookup — which levels a specific model supports and the recommended default — is references/reasoning-effort-lookup.md (OpenCode Go focus; method applies to any BYOK model).

  • Supported levels: none, minimal, low, medium, high, xhigh, max (per-model subset may vary — e.g., DeepSeek V4 models on OpenCode Go only support low, medium, high; minimal is newer, verified in CLI 1.0.77).
  • Apply it per run, for example:
.\scripts\byok-profile.ps1 run dprocess-openai-gpt-54 --reasoning-effort medium

For OpenAI models, you may also enable summaries:

.\scripts\byok-profile.ps1 run dprocess-openai-gpt-54 --reasoning-effort high --enable-reasoning-summaries

Model compatibility warning

Not all models support Copilot CLI's --reasoning-effort levels. If you get:

Model "glm-5.2" does not support reasoning effort configuration (requested: "high").

it means the model's API does not expose controllable reasoning effort levels. Known families without reasoning-effort support:

  • GLM (glm-5, glm-5.1, glm-5.2) — Zhipu AI / OpenCode Go
  • MiMo (mimo-v2.5, mimo-v2.5-pro, mimo-v2-pro, mimo-v2-omni) — Xiaomi / OpenCode Go
  • Kimi K2.x (kimi-k2.7-code, kimi-k2.6, kimi-k2.5) — Moonshot AI / OpenCode Go (thinking is implicit / always-on)
  • Qwen3.x (qwen3.7-plus, qwen3.7-max, qwen3.6-plus, qwen3.5-plus) — Alibaba / OpenCode Go (Anthropic-type, implicit thinking)
  • MiniMax (minimax-m3, minimax-m2.7, minimax-m2.5) — MiniMax / OpenCode Go (Anthropic-type, implicit thinking)

When using these models, omit --reasoning-effort entirely. The model will use its built-in default reasoning behavior.

To check whether a model supports it, try --reasoning-effort none (the least demanding level). If that also fails, the model simply doesn't support the feature.

The profile system tracks this per model. Profiles for models that do not support reasoning effort set "reasoningEffortSupported": false. The run command detects incompatible --reasoning-effort/--effort arguments, strips them before forwarding to Copilot CLI, and displays a clear notice. This prevents the API error. Support is derived from the shared no-support model list when the profile flag is absent (hand-added profiles), so run/set-env/show and Invoke-CopilotCliSubSession.ps1 stay consistent.

Do not claim a dedicated COPILOT_* environment variable exists for reasoning effort unless copilot help environment in the user's installed CLI version explicitly lists one.

Grounding and evidence standard

When answering questions in this domain, always separate grounded facts from inference:

  1. Cite authoritative evidence used (for example, copilot --help, copilot help environment, provider model docs).
  2. State what is directly evidenced versus inferred operational guidance.
  3. End with an explicit conclusion:
  • Grounding status: evidence-backed when all key claims are directly supported.
  • Grounding status: mixed (evidence + inference) when any recommendation is inferred.

Do not present inferred workarounds (for example wrapper aliases for sticky defaults) as first-class documented product features.

Calculate token overrides conservatively

When the user asks for COPILOT_PROVIDER_MAX_PROMPT_TOKENS or COPILOT_PROVIDER_MAX_OUTPUT_TOKENS:

  1. Get the model's documented context window.
  2. Pick a realistic max output budget for the workload.
  3. Reserve a safety buffer for tool calls, system instructions, and multi-turn variance.
  4. Compute:

maxPromptTokens = contextWindow - plannedMaxOutput - safetyBuffer

Prefer stable values over theoretical maximums. If the user reports context-limit failures, reduce prompt tokens by 5-10% and retry.

Provider-enforced limits vs theoretical context windows

Some gateway providers (notably OpenCode Go) enforce per-request token limits lower than the model's theoretical context window. The native DeepSeek V4 Flash model supports 1M context, but OpenCode Go's gateway enforces an effective limit around ~300K for prompt tokens. Using the theoretical 1M to calculate maxPromptTokens (840K) will cause compaction failures with 400 Error from provider (Console Go): Upstream request failed.

Empirical approach: Test compaction at increasing context-usage levels. When compaction fails, derive the provider's effective limit:

maxPromptTokens = successfulCompactionTokens / 0.78

For OpenCode Go's DeepSeek V4 models, the empirically validated safe values are:

  • maxPromptTokens: 325,000
  • maxOutputTokens: 64,000

This same limit discovery process applies to any provider whose gateway enforces stricter limits than the model's published context window.

Troubleshoot in this order

  1. Confirm the base URL, provider type, model name, and API key source.
  2. Confirm the model supports streaming and tool calling.
  3. If using a stored profile, run show or list to verify the saved values.
  4. If ${ENV_VAR} placeholders are used, confirm the environment variable actually exists.
  5. If long-context models fail, add or lower explicit max prompt and output token overrides.

Moonshot proxy (top_p workaround)

Problem: Kimi models from https://api.moonshot.ai only accepts top_p=0.95, but VS Code Copilot BYOK always sends top_p=1.0 and the chatLanguageModels.json schema doesn't support per-model parameter overrides.

Solution: A local HTTPS proxy that strips top_p before forwarding. Scripts live in the skill's scripts/ folder; runtime cert data is stored at ~/.copilot/moonshot-proxy/.

| File | Location | Purpose | |------|----------|---------| | proxy.js | scripts/proxy.js | HTTPS proxy server (Node.js) — dual-port: 3002 (always) + 443 (elevated) | | start-proxy.ps1 | scripts/start-proxy.ps1 | Auto-elevates admin, kills old proxy, starts via node proxy.js | | setup-dns.ps1 | scripts/setup-dns.ps1 | One-time admin setup: adds 127.0.0.1 moonshot.local to hosts + trusted cert | | Certs | ~/.copilot/moonshot-proxy/ | Runtime cert data (moonshot.pfx, cert.pfx) |

How to use

# One-time setup (run once, elevated):
.\scripts\setup-dns.ps1

# Start proxy (after every reboot):
.\scripts\start-proxy.ps1

# Or from published skill location:
pwsh -NoProfile "~\.copilot\skills\copilot-byok\scripts\start-proxy.ps1"

# Check status:
curl -s https://moonshot.local/health

Profile integration

Add "proxyPort": 443 to any byok-profiles.json profile that needs the proxy. The run command auto-starts the proxy and routes through https://moonshot.local/v1.

VS Code integration

Configure the Moonshot provider's url to https://moonshot.local/v1/chat/completions in chatLanguageModels.json. The .vscode/tasks.json background task with runOn: "folderOpen" auto-starts the proxy, or run Terminal → Run Task → "Moonshot Proxy".

Moonshot/Kimi AI credentials

Use MOONSHOT_API_KEY for the Kimi AI Platform (api.moonshot.ai/v1). All models use OpenAI-compatible format, 262K context, and support tool calling and streaming.

Related skill

For MCP server configuration rather than model-provider configuration, read ../copilot-cli-mcp-config/SKILL.md.