Provider API Reference

A trace-it-fast reference for every AI provider / coding-CLI that Cyberine integrates directly. Each provider has its own page with the official API/CLI surface (endpoints, auth, models, gotchas, doc links). This index maps those providers onto how the Cyberine codebase wires them so you can jump from "which env var / adapter / flag" to "the upstream API doc" in one hop.

For "anything new upstream?" jump straight to CHANGELOG-WATCH.md — one table mapping every vendor to its live changelog/releases URL. For the MCP SDK 1.x -> 2.x package split, see mcp-2.0-audit.md -- but read its correction banner first: the migration landed 2026-09-05 and three of that audit's predictions did not survive contact with the stable 2.0.0 packages.

For "which of these models is strongest", see model-strength.md: a dated, sourced 1 to 5 rank per model id that the router reads and fleet.json can override.

Source of truth for the wiring section is the code, not this doc — if they drift, trust: src/core/config.ts, src/orchestrator/buildAdapters.ts, src/adapters/LoopAgentAdapter.ts, src/adapters/CliAgentAdapter.ts, .env.example. Last synced: 2026-06-07.

A provider page here is not a promise that its worker dispatches today. Seven worker ids were RETIRED (2026-08-27 onward) for account/credential reasons, not for lack of support — the adapters and these reference pages stay, the workers are skipped. See Retired workers below and RETIRED_WORKERS in src/core/workerProfiles.ts, which is the authority.

Provider pages

Provider	Kind	Cyberine adapter	Page
DeepSeek (direct)	raw API (OpenAI-compatible)	LoopAgentAdapter	deepseek.md
OpenRouter	raw API (OpenAI-compatible aggregator)	LoopAgentAdapter	openrouter.md
Moonshot AI / Kimi	raw API (OpenAI-compatible)	LoopAgentAdapter	moonshot-kimi.md
LiteLLM gateway	OpenAI-compatible proxy	LoopAgentAdapter	litellm.md
OpenAI	raw API (behind gateway + Codex)	LoopAgentAdapter / Codex	openai.md
Anthropic Claude	raw API (behind gateway + Claude Code)	LoopAgentAdapter / Claude Code	anthropic.md
Claude Code CLI	headless coding agent	CliAgentAdapter	claude-code-cli.md
OpenAI Codex CLI	headless coding agent	CliAgentAdapter	codex-cli.md
Google Gemini CLI	headless coding agent	CliAgentAdapter	gemini-cli.md
OpenCode CLI	headless coding agent (multi-provider)	CliAgentAdapter	opencode-cli.md
OpenCode 2.0 beta (opencode2)	headless coding agent (beta, no GA date)	CliAgentAdapter (makeOpenCode2Adapter)	opencode2-v2-beta.md
Muse Code (Meta)	headless coding agent (flat-rate subscription)	CliAgentAdapter (makeMuseAdapter)	muse-cli.md
Cursor Agent CLI	headless coding agent (subscription)	CliAgentAdapter (makeCursorAgentAdapter)	cursor-agent.md
CLI Coding Agents (survey)	headless coding agents (integration reference)	CliAgentAdapter	cli-agents.md
Ollama (local)	local inference server (OpenAI-compatible + native)	LoopAgentAdapter	ollama.md
Ollama Cloud	hosted inference (OpenAI-compatible)	LoopAgentAdapter	ollama-cloud.md
MiniMax	raw API (OpenAI-compatible, api.minimax.io)	LoopAgentAdapter	minimax.md
Command Code	coding CLI (cmd) + raw API (OpenAI-compatible Provider API, api.commandcode.ai)	LoopAgentAdapter (loop-commandcode-build, loop-commandcode-free-build) + CliAgentAdapter (command-code)	command-code.md

Cyberine integration map

Three adapter families implement the same Agent port; the core never special-cases a provider. (LoopAgentAdapter — OpenAI-compatible loop; CliAgentAdapter — headless coding CLI; and AnthropicAgentAdapter — the native Anthropic Messages API, the key-only path to Claude with no gateway/CLI, see anthropic.md and the build-fleet table below.)

1. LoopAgentAdapter — raw OpenAI-compatible Chat Completions

A hand-run tool-calling loop against any OpenAI-compatible endpoint, using the openai npm SDK (chat.completions.create, stream: true, stream_options.include_usage). Provider independence is pure constructor config: swap baseURL + apiKey + model, no code change.

Adapter instances built in buildDefaultAdapters() (selection order favours paths needing NO gateway):

Adapter id	Provider	Base URL (env)	Model (env)	When registered
loop-deepseek-direct	DeepSeek	DEEPSEEK_BASE_URL	DEEPSEEK_MODEL	DEEPSEEK_API_KEY set
loop-deepseek-openrouter	OpenRouter	OPENROUTER_BASE_URL	OPENROUTER_MODEL	no DeepSeek key, OPENROUTER_API_KEY set
loop-gateway-a	LiteLLM → LOOP_MODEL_A	GATEWAY_BASE_URL	LOOP_MODEL_A	any GATEWAY_*/LOOP_MODEL_* env set (incl. defaults)
loop-gateway-b	LiteLLM → LOOP_MODEL_B	GATEWAY_BASE_URL	LOOP_MODEL_B	any GATEWAY_*/LOOP_MODEL_* env set (incl. defaults)
loop-ollama	Ollama (local, keyless)	OLLAMA_BASE_URL	OLLAMA_MODEL	OLLAMA_BASE_URL or OLLAMA_MODEL set (env opt-in; static bearer ollama)

Reasoning models stream chain-of-thought in a separate reasoning_content delta field (DeepSeek extension). Cyberine surfaces it in the peek view but never folds it into the answer, and replays it on the assistant tool-call turn when present (required by Kimi k2.6).

2. CliAgentAdapter — headless coding-CLI subprocess

Spawns a coding CLI as a black-box subprocess, captures JSONL/stdout, normalises into a Result. Each runs under a host sandbox (hostSandbox.ts: seatbelt/bubblewrap) and with auto-approve flags so it never stalls on a permission prompt headless. The real work product is the file diff left in task.cwd.

Factory	CLI	Headless invocation (argv)	Auth dir
makeClaudeAdapter	claude	-p <prompt> --output-format stream-json --verbose --model <m> --dangerously-skip-permissions --add-dir <cwd> (opt-in: --effort, --forward-subagent-text, --json-schema; unconditional on the default-prompt path: --exclude-dynamic-system-prompt-sections)	~/.claude
makeCodexHeadlessAdapter	codex	exec --skip-git-repo-check --cd <cwd> -m <m> [--output-schema <path>] --dangerously-bypass-approvals-and-sandbox --dangerously-bypass-hook-trust <prompt>	~/.config/codex
makeGeminiAdapter	gemini	-p <prompt> -o json --approval-mode yolo --skip-trust -m <m>	~/.gemini
makeOpenCodeAdapter	opencode	run --model <prov/model> [--variant <v>] [--agent plan|build] --dir <cwd> --format json [--dangerously-skip-permissions] <prompt>	~/.local/share/opencode (XDG)
makeCodexAdapter	codex	exec --json [--output-schema <path>] <prompt> (LIBRARY-ONLY read-only variant, T656 — no production call site; the build-fleet codex worker is makeCodexHeadlessAdapter above)	~/.config/codex

CLI workers are gated on install and headless login (cliAvailability.ts); an installed but logged-out CLI is skipped with a note. opencode is treated as trusted-if-installed (its own stored auth / free-model path).

Build fleet (file-writing workers, opt-in via CYBERINE_BUILD_FLEET)

Comma-separated worker ids registered by buildFleetAdapters():

Worker id	Backed by	Needs
loop-deepseek-build	DeepSeek API + implementer tools (sandbox-pinned, zero CLI)	DEEPSEEK_API_KEY
loop-openrouter-build	OpenRouter API + implementer tools	OPENROUTER_API_KEY
loop-kimi-build	Moonshot/Kimi API + implementer tools	MOONSHOT_API_KEY
loop-mimo-build	Xiaomi MiMo API + implementer tools (api.xiaomimimo.com)	XIAOMI_API_KEY
loop-minimax-build	MiniMax API + implementer tools (api.minimax.io)	MINIMAX_API_KEY
loop-openai-build	OpenAI direct API + implementer tools	OPENAI_API_KEY
loop-gemini-build	Google Gemini direct API + implementer tools	GEMINI_API_KEY
loop-xai-build	xAI Grok direct API + implementer tools	XAI_API_KEY
loop-groq-build	Groq direct API + implementer tools	GROQ_API_KEY
loop-mistral-build	Mistral direct API + implementer tools	MISTRAL_API_KEY
loop-together-build	Together AI direct API + implementer tools	TOGETHER_API_KEY
loop-fireworks-build	Fireworks AI direct API + implementer tools	FIREWORKS_API_KEY
loop-qwen-build	Qwen/DashScope direct API + implementer tools	QWEN_API_KEY
loop-zhipu-build	Zhipu GLM direct API + implementer tools	ZHIPU_API_KEY
loop-perplexity-build	Perplexity Sonar direct API + implementer tools	PERPLEXITY_API_KEY
loop-cerebras-build	Cerebras direct API + implementer tools	CEREBRAS_API_KEY
loop-nvidia-build	NVIDIA NIM direct API + implementer tools	NVIDIA_API_KEY
loop-cohere-build	Cohere direct API + implementer tools	COHERE_API_KEY
loop-azure-build	Azure OpenAI instance endpoint + implementer tools	AZURE_OPENAI_API_KEY + AZURE_OPENAI_ENDPOINT
loop-copilot-build	GitHub Copilot (OAuth device-code) + implementer tools	Copilot subscription; run cyberine auth login copilot
loop-claude-build	Anthropic Claude via LiteLLM gateway + implementer tools	gateway configured (LOOP_MODEL_A)
loop-gpt-build	OpenAI GPT via LiteLLM gateway + implementer tools	gateway configured (LOOP_MODEL_B)
loop-claude-native-build	Anthropic Claude via the NATIVE Messages API + implementer tools (key-only, no gateway/CLI)	ANTHROPIC_API_KEY
loop-ollama-build	local Ollama server + implementer tools (keyless, free, private)	OLLAMA_MODEL or OLLAMA_BASE_URL set; model pulled
loop-commandcode-build	Command Code Provider API + implementer tools, paid catalog model drawing the plan's credit windows	CMD_API_KEY; CMD_MODEL (default deepseek/deepseek-v4-flash)
loop-commandcode-free-build	Command Code Provider API + implementer tools on a zero-credit model resolved from the live catalog (the free lane)	CMD_API_KEY and at least one :free / -free id in GET /provider/v1/models; CMD_FREE_MODEL pins one
loop-bedrock-build	AWS Bedrock Converse API + implementer tools	BEDROCK_ACCESS_KEY_ID + BEDROCK_SECRET_ACCESS_KEY, or BEDROCK_PROFILE
loop-vertex-build	Google Vertex AI (Gemini / Claude-on-Vertex) + implementer tools	VERTEX_PROJECT_ID; ADC or VERTEX_SERVICE_ACCOUNT_KEY_JSON
opencode-deepseek	OpenCode CLI (--agent build), DeepSeek model	opencode installed; CYBERINE_OPENCODE_DEEPSEEK_MODEL (default deepseek/deepseek-v4-pro)
opencode-kimi	OpenCode CLI, Kimi model	DEAD 2026-08-25, self-skipping — opencode installed AND a moonshotai credential (opencode auth login) or a funded MOONSHOT_API_KEY; today neither holds and the account is suspended. CYBERINE_OPENCODE_KIMI_MODEL (default moonshotai/kimi-k2.7-code). See opencode-cli.md
claude	Claude Code CLI	claude installed + logged in
codex	Codex CLI	codex installed + logged in
gemini	Gemini CLI	gemini installed + logged in
aider	Aider CLI (litellm backend, any provider)	aider installed + provider API key; CYBERINE_AIDER_MODEL selects model
qwen-code	Qwen Code CLI (@qwen-code/qwen-code)	DASHSCOPE_API_KEY; CYBERINE_QWEN_MODEL (default qwen3-coder-plus)
crush	Crush CLI (charmbracelet, openai-compat)	crush installed + provider key; CYBERINE_CRUSH_MODEL selects provider/model
cursor-agent	Cursor Agent CLI (subscription)	cursor-agent installed + Copilot/Cursor subscription
amp	Amp CLI (ampcode.com)	AMP_API_KEY
goose	Goose CLI (Block, 40+ providers)	goose installed + metered key; CYBERINE_GOOSE_PROVIDER / CYBERINE_GOOSE_MODEL
openhands	OpenHands CLI (All-Hands AI, model-agnostic)	openhands installed; LLM configured at first run; CYBERINE_OPENHANDS_MODEL overrides per-run
command-code	Command Code CLI (command-code -p --yolo, subscription credit pool)	command-code installed + ~/.commandcode/auth.json apiKey or COMMAND_CODE_API_KEY; CYBERINE_COMMANDCODE_MODEL (default deepseek/deepseek-v4-flash), CYBERINE_COMMANDCODE_MAX_TURNS (default 60)
mmx-text-build	MiniMax mmx CLI (text chat)	mmx installed + MMX_TOKEN_KEY; CYBERINE_MMX_MODEL (default MiniMax-Text-01)

loop-*-build workers are sandbox-pinned (cannot leave their worktree); CLI workers run with full host privileges (worktrees isolate concurrency, not the host) — enable only for trusted tasks.

Environment variables

From src/core/config.ts (defaults) and .env.example.

Env var	Default	Purpose
GATEWAY_BASE_URL	http://localhost:41400/v1	LiteLLM proxy base URL (uncommon on-demand host port)
GATEWAY_API_KEY	sk-litellm-master-key	LiteLLM master key
LOOP_MODEL_A	anthropic/claude	gateway model alias A
LOOP_MODEL_B	openai/gpt	gateway model alias B (proves provider independence)
OPENAI_API_KEY	(empty)	OpenAI direct loop adapter / loop-openai-build; also consumed by gateway to route openai/*
OPENAI_BASE_URL	https://api.openai.com/v1	OpenAI direct base URL
OPENAI_MODEL	gpt-4o	OpenAI model id (loop-openai-build)
GEMINI_API_KEY	(empty)	Google Gemini direct loop adapter / loop-gemini-build
GEMINI_BASE_URL	https://generativelanguage.googleapis.com/v1beta/openai	Gemini OpenAI-compatible base URL
GEMINI_MODEL	gemini-2.5-pro	Gemini model id (loop-gemini-build)
XAI_API_KEY	(empty)	xAI Grok direct loop adapter / loop-xai-build
XAI_BASE_URL	https://api.x.ai/v1	xAI base URL
XAI_MODEL	grok-3	xAI Grok model id (loop-xai-build)
GROQ_API_KEY	(empty)	Groq direct loop adapter / loop-groq-build
GROQ_BASE_URL	https://api.groq.com/openai/v1	Groq base URL
GROQ_MODEL	llama-3.3-70b-versatile	Groq model id (loop-groq-build)
MISTRAL_API_KEY	(empty)	Mistral direct loop adapter / loop-mistral-build
MISTRAL_BASE_URL	https://api.mistral.ai/v1	Mistral base URL
MISTRAL_MODEL	mistral-large-latest	Mistral model id (loop-mistral-build)
TOGETHER_API_KEY	(empty)	Together AI direct loop adapter / loop-together-build
TOGETHER_BASE_URL	https://api.together.xyz/v1	Together AI base URL
TOGETHER_MODEL	meta-llama/Llama-3.3-70B-Instruct-Turbo	Together AI model id (loop-together-build)
FIREWORKS_API_KEY	(empty)	Fireworks AI direct loop adapter / loop-fireworks-build
FIREWORKS_BASE_URL	https://api.fireworks.ai/inference/v1	Fireworks AI base URL
FIREWORKS_MODEL	accounts/fireworks/models/deepseek-v3	Fireworks AI model id (loop-fireworks-build)
QWEN_API_KEY	(empty)	Qwen/DashScope direct loop adapter / loop-qwen-build
QWEN_BASE_URL	https://dashscope-intl.aliyuncs.com/compatible-mode/v1	Qwen international base URL (mainland: dashscope.aliyuncs.com/compatible-mode/v1)
QWEN_MODEL	qwen3-coder-plus	Qwen model id (loop-qwen-build)
ZHIPU_API_KEY	(empty)	Zhipu GLM direct loop adapter / loop-zhipu-build
ZHIPU_BASE_URL	https://open.bigmodel.cn/api/paas/v4	Zhipu base URL
ZHIPU_MODEL	glm-4.6	Zhipu GLM model id (loop-zhipu-build)
PERPLEXITY_API_KEY	(empty)	Perplexity direct loop adapter / loop-perplexity-build
PERPLEXITY_BASE_URL	https://api.perplexity.ai	Perplexity base URL
PERPLEXITY_MODEL	sonar-pro	Perplexity model id (loop-perplexity-build)
CEREBRAS_API_KEY	(empty)	Cerebras direct loop adapter / loop-cerebras-build
CEREBRAS_BASE_URL	https://api.cerebras.ai/v1	Cerebras base URL
CEREBRAS_MODEL	llama-3.3-70b	Cerebras model id (loop-cerebras-build)
NVIDIA_API_KEY	(empty)	NVIDIA NIM direct loop adapter / loop-nvidia-build
NVIDIA_BASE_URL	https://integrate.api.nvidia.com/v1	NVIDIA NIM base URL
NVIDIA_MODEL	meta/llama-3.3-70b-instruct	NVIDIA NIM model id (loop-nvidia-build)
COHERE_API_KEY	(empty)	Cohere direct loop adapter / loop-cohere-build
COHERE_BASE_URL	https://api.cohere.ai/compatibility/v1	Cohere OpenAI-compatible base URL
COHERE_MODEL	command-r-plus	Cohere model id (loop-cohere-build)
AZURE_OPENAI_API_KEY	(empty)	Azure OpenAI key / loop-azure-build
AZURE_OPENAI_ENDPOINT	(empty)	Azure instance endpoint (e.g. https://<name>.openai.azure.com/openai/deployments/<deployment>/)
AZURE_OPENAI_DEPLOYMENT	gpt-4o	Azure deployment name — also the worker model id
AZURE_OPENAI_API_VERSION	2024-12-01-preview	Azure API version query parameter
COPILOT_MODEL	gpt-4o	GitHub Copilot model id (loop-copilot-build); requires active Copilot subscription + cyberine auth login copilot
ANTHROPIC_API_KEY	(empty)	consumed by the gateway to route anthropic/*; ALSO enables the native loop-claude-native-build worker (no gateway needed)
ANTHROPIC_BASE_URL	https://api.anthropic.com	native Anthropic adapter base URL (override only for a proxy)
ANTHROPIC_MODEL	claude-sonnet-4-6	native Anthropic adapter model id (loop-claude-native-build)
BEDROCK_REGION	us-east-1 (falls back to AWS_REGION / AWS_DEFAULT_REGION)	AWS region for Bedrock calls
BEDROCK_MODEL	anthropic.claude-3-5-sonnet-20241022-v2:0	Bedrock model id or inference-profile ARN (loop-bedrock-build)
BEDROCK_ACCESS_KEY_ID	(empty, falls back to AWS_ACCESS_KEY_ID)	AWS access key for Bedrock (optional if using ADC profile)
BEDROCK_SECRET_ACCESS_KEY	(empty, falls back to AWS_SECRET_ACCESS_KEY)	AWS secret access key for Bedrock
BEDROCK_SESSION_TOKEN	(empty, falls back to AWS_SESSION_TOKEN)	AWS session token for temporary credentials
BEDROCK_PROFILE	(empty, falls back to AWS_PROFILE)	Named AWS profile for Bedrock
VERTEX_PROJECT_ID	(empty)	GCP project id for Vertex AI (loop-vertex-build)
VERTEX_LOCATION	us-central1	Vertex AI region
VERTEX_MODEL	gemini-2.5-pro	Vertex AI model id — supports Gemini and Claude-on-Vertex
VERTEX_SERVICE_ACCOUNT_KEY_JSON	(empty)	Service account JSON key string; when absent uses ADC
XIAOMI_API_KEY	(empty)	Xiaomi MiMo direct loop adapter / loop-mimo-build
XIAOMI_BASE_URL	https://api.xiaomimimo.com/v1	MiMo base URL (regional: token-plan-cn/ams/sgp.xiaomimimo.com/v1)
XIAOMI_MODEL	mimo-v2.5-pro	Xiaomi MiMo model id (loop-mimo-build)
MINIMAX_API_KEY	(empty)	MiniMax direct loop adapter / loop-minimax-build
MINIMAX_BASE_URL	https://api.minimax.io/v1	MiniMax base URL (international platform)
MINIMAX_MODEL	MiniMax-M3	MiniMax model id (loop-minimax-build)
MMX_TOKEN_KEY	(empty)	MiniMax mmx CLI Token Plan key — required for mmx-text-build and multimedia generation tools
CYBERINE_OPENCODE_DEEPSEEK_MODEL	deepseek/deepseek-v4-pro	model for opencode-deepseek worker (pinned independently of OPENCODE_MODEL)
OLLAMA_BASE_URL	http://localhost:11434/v1	local Ollama OpenAI-compat endpoint; setting it (or OLLAMA_MODEL) opts the keyless adapters in
OLLAMA_MODEL	llama3.1:8b	Ollama model id (loop-ollama / loop-ollama-build); must be pulled already
CLI_AGENT_CMD	claude	binary for the generic CLI adapter
OPENROUTER_API_KEY	(empty)	OpenRouter loop adapter / loop-openrouter-build
OPENROUTER_BASE_URL	https://openrouter.ai/api/v1	OpenRouter base URL
OPENROUTER_MODEL	deepseek/deepseek-v4-pro-0813	OpenRouter model id (vendor/model). Dated on purpose — the undated alias is a frozen 0423 GA on this route; see openrouter.md
DEEPSEEK_API_KEY	(empty)	DeepSeek-direct loop adapter / loop-deepseek-build; preferred over OpenRouter
DEEPSEEK_BASE_URL	https://api.deepseek.com	DeepSeek base URL
DEEPSEEK_MODEL	deepseek-v4-pro (config) / deepseek-v4-flash (.env.example)	DeepSeek model in native naming
MOONSHOT_API_KEY	(empty)	Moonshot/Kimi loop worker loop-kimi-build
MOONSHOT_BASE_URL	https://api.moonshot.ai/v1	Moonshot base (mainland: https://api.moonshot.cn/v1)
MOONSHOT_MODEL	kimi-k2.7-code	Moonshot/Kimi model id
OPENCODE_MODEL	deepseek/deepseek-v4-pro	OpenCode provider/model (canonical id)
OPENCODE_VARIANT	high	reasoning effort: minimal/high/max
OPENCODE_AGENT	plan (config) / build (interactive launcher)	OpenCode --agent
OPENCODE_SKIP_PERMISSIONS	true	pass --dangerously-skip-permissions
OPENCODE_TIMEOUT_MS	1800000	OpenCode subprocess timeout (30 min)
CYBERINE_OPENCODE_MAX_RETRIES	2	retries on transient SQLite WAL-lock crash
CYBERINE_OPENCODE_KIMI_MODEL	moonshotai/kimi-k2.7-code	model for opencode-kimi worker (default provider uncredentialled + suspended as of 2026-08-25)
CYBERINE_CLAUDE_MODEL	claude-sonnet-4-6	model for the claude build worker
CYBERINE_CODEX_MODEL	(unset → codex default)	model for the codex build worker
CYBERINE_GEMINI_MODEL	gemini-2.5-flash	model for the gemini build worker
JUDGE_MODEL	deepseek/deepseek-v4-pro-0813	model for LLM/adversarial judges (OpenRouter path)
CYBERINE_JUDGE_BACKEND	auto	T651c: moonshot opts the LLM-judge panel into Moonshot-direct (requires MOONSHOT_API_KEY); any other value (default auto) keeps the DeepSeek-direct/OpenRouter precedence unchanged
CYBERINE_MOONSHOT_JUDGE_MODEL	kimi-k3	judge model used when CYBERINE_JUDGE_BACKEND=moonshot — deliberately distinct from MOONSHOT_MODEL (kimi-k2.7-code, the coding-specialized build-fleet default)
LOOP_MAX_TOKENS	16000	max completion tokens per loop request
LOOP_MAX_TOOL_ROUNDS	25	max tool-calling rounds in the loop
CYBERINE_CHAT_CONTEXT_WINDOW	65536	planner context window (chat TUI warning only)
CYBERINE_BUILD_FLEET	(empty)	comma-separated build-fleet worker ids
CYBERINE_ROUTE	auto	router mode for auto fleet selection
HOST_SANDBOX_MODE	auto	seatbelt/bubblewrap/none
HOST_SANDBOX_NETWORK	allow	network inside the sandbox
CYBERINE_TASK_TRUSTED	true	when false, CLI workers require a sandbox

Note the one intentional discrepancy: config.ts defaults DEEPSEEK_MODEL to deepseek-v4-pro, while .env.example ships deepseek-v4-flash (the cheaper model) as the suggested starter value.

Retired workers (2026-08-27)

Retirement is a cost/credential decision, not a capability one: the adapter, the provider page, and the WORKER_PROFILES entry all stay in place — buildAdapters simply skips the id, and cyberine doctor reports it as RETIRED rather than FAIL. Nothing is deleted, so re-enabling a worker is a one-line edit once its account is funded again. RETIRED_WORKERS in src/core/workerProfiles.ts is the authority; this table mirrors it.

Worker id	Provider	Why retired
loop-kimi-build	Moonshot / Kimi	Account out of balance — retired rather than funded (kn4vz)
opencode-kimi	Moonshot / Kimi (via OpenCode)	Same credential as loop-kimi-build (kn4vz)
loop-minimax-build	MiniMax	402 insufficient balance 1008 on every dispatch (kf2mx)
loop-minimax-direct	MiniMax	Same credential; adapter-only, gated in buildAdapters (kf2mx)
mmx-text-build	MiniMax (mmx CLI)	Same unfunded MiniMax account (kf2mx)
gemini	Google Gemini CLI	Google revoked the free-tier client (IneligibleTierError), no paid tier funded (kp7wq)
loop-gemini-build	Google Gemini API	Same unfunded Google account as the CLI worker (kp7wq)
loop-muse-build (2026-09-05)	Meta Model API	402 billing_not_configured — key valid, no payment method; the muse CLI subscription does not cover the raw API (kjx6p)
opencode2-deepseek (2026-09-06)	DeepSeek (via OpenCode 2.0)	Same exhausted DeepSeek account as loop-deepseek-direct — 402 Insufficient Balance (kzhyc)
opencode-deepseek, opencode (2026-09-07)	DeepSeek (via OpenCode)	Operator is dropping the opencode binaries; cyberine run is the one-command edit tool now (ks4sf)

Run cyberine route "<task>" and read its available: line for the fleet that can actually be dispatched right now — that is the only answer that accounts for credentials, CLI login state, and ~/.cyberine/fleet.json gating at the same time.

Integration mismatches

Surfaced while cross-checking the official docs against the Cyberine code (2026-06-07). Items 1-3 were verified against ground truth and FIXED 2026-06-07 (commit lands with this note); item 4 is informational only.

1. reasoning_content replay conflicts with DeepSeek — FIXED. LoopAgentAdapter used to replay buffered reasoning_content on the assistant tool-call turn for every provider. Kimi k2.6 requires it (thinking on by default → 400 if missing); DeepSeek forbids it (echoing it back → HTTP 400). The replay is now gated by needsReasoningReplay(provider, model) (matches moonshot/kimi), with a replayReasoning constructor override for gateway-routed thinking models. Verified: DeepSeek + Kimi tool-call loops both complete (no 400). See deepseek.md and moonshot-kimi.md.
2. Codex config-dir path — FIXED. CliAgentAdapter.cliConfigDirs() allow-listed only ~/.config/codex, but Codex's official CODEX_HOME defaults to ~/.codex (the live dir on this host; ~/.config/codex did not exist) — so a codex build worker under seatbelt could not write its config/auth. Now honors CODEX_HOME when set, else allow-lists BOTH ~/.codex and ~/.config/codex. Verified: the generated seatbelt profile grants write to ~/.codex. See codex-cli.md.
3. codex usage cost_usd — clarified. Cyberine's Codex parse reads usage.cost_usd, which is not in Codex's documented usage schema (input_tokens/cached_input_tokens/output_tokens/reasoning_output_tokens). The read is kept as a harmless forward-compat fallback (stays undefined); the misleading comment that presented cost_usd as the Codex shape was corrected. See codex-cli.md.
4. Model-id naming drift (informational). Repo uses claude-sonnet-4-6 (hyphen); OpenRouter's slug is anthropic/claude-sonnet-4.6 (dot). Repo's kimi-k2.7-code (bumped 2026-07-01, T517) is the live Moonshot/OpenCode id but is NOT an OpenRouter slug (moonshotai/kimi-k2-0905 there) — Cyberine doesn't route Kimi via OpenRouter, so this is informational only. See openrouter.md.