Documentation
¶
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ( OLLAMA = LLMProvider{Name: "ollama", ApiUrl: "http://localhost:11434", Models: []Model{QWEN_3_6}} ANTHROPIC = LLMProvider{Name: "anthropic", ApiUrl: "https://api.anthropic.com", Models: []Model{SONNET_4_6, OPUS_4_8}} DEEPSEEK = LLMProvider{Name: "deepseek", ApiUrl: "https://api.deepseek.com", Models: []Model{DEEPSEEK_V4_FLASH, DEEPSEEK_V4_PRO}} // OPENAI — all currently listed models are reasoning-class (gpt-5 / o-series). // If a non-reasoning model (gpt-4*, gpt-3.5*) is added here later, update // pkg/llm/openai/client.go isReasoningModel to match. OPENAI = LLMProvider{Name: "openai", ApiUrl: "https://api.openai.com", Models: []Model{GPT_5_4_MINI, GPT_5_5}} // GLM — Zhipu AI / z.ai, reached over its Anthropic-compatible endpoint // (same wire format as ANTHROPIC; pkg/llm/glm copies that engine and swaps // auth to Authorization: Bearer). ApiUrl is the z.ai Anthropic gateway, so // pkg/llm/glm hits ApiUrl + "/v1/messages". GLM = LLMProvider{Name: "glm", ApiUrl: "https://api.z.ai/api/anthropic", Models: []Model{GLM_4_6, GLM_5_2}} // QWEN — Alibaba Cloud Tongyi/Qwen via the DashScope OpenAI-compatible // endpoint (same wire format as OPENAI/DEEPSEEK; pkg/llm/qwen copies the // deepseek engine, Bearer auth, thinking via enable_thinking). ApiUrl is the // international (Singapore) gateway; pkg/llm/qwen hits ApiUrl + "/chat/completions". // China users override api_url to https://dashscope.aliyuncs.com/compatible-mode/v1. QWEN = LLMProvider{Name: "qwen", ApiUrl: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", Models: []Model{QWEN3_7_PLUS, QWEN3_7_MAX}} )
Use var for struct "constants" Models are ordered by price
var MODEL_CONTEXT_SIZE = map[Model]int{ QWEN_3_6: 262_144, SONNET_4_6: 500_000, OPUS_4_8: 500_000, DEEPSEEK_V4_FLASH: 1_000_000, DEEPSEEK_V4_PRO: 1_000_000, GPT_5_4_MINI: 400_000, GPT_5_5: 1_050_000, GLM_4_6: 200_000, GLM_5_2: 1_000_000, QWEN3_7_PLUS: 1_000_000, QWEN3_7_MAX: 1_000_000, }
for completion
var MODEL_PRICING = map[Model]Pricing{ QWEN_3_6: {}, SONNET_4_6: {Input: 3.00, Output: 15.00, CacheRead: 0.30, CacheWrite: 3.75}, OPUS_4_8: {Input: 5.00, Output: 25.00, CacheRead: 0.50, CacheWrite: 6.25}, DEEPSEEK_V4_FLASH: {Input: 0.14, Output: 0.28, CacheRead: 0.14, CacheWrite: 0.14}, DEEPSEEK_V4_PRO: {Input: 1.74, Output: 3.48, CacheRead: 1.74, CacheWrite: 1.74}, GPT_5_4_MINI: {Input: 0.75, Output: 4.50, CacheRead: 0.75, CacheWrite: 0.75}, GPT_5_5: {Input: 5.00, Output: 30.00, CacheRead: 5.00, CacheWrite: 5.00}, GLM_4_6: {Input: 0.43, Output: 1.74, CacheRead: 0.11, CacheWrite: 0.43}, GLM_5_2: {Input: 1.40, Output: 4.40, CacheRead: 0.26, CacheWrite: 1.40}, QWEN3_7_PLUS: {Input: 0.40, Output: 1.60, CacheRead: 0.08, CacheWrite: 0.40}, QWEN3_7_MAX: {Input: 2.50, Output: 7.50, CacheRead: 2.50, CacheWrite: 2.50}, }
MODEL_PRICING is the per-model USD rate card, keyed like MODEL_CONTEXT_SIZE. Rates are USD per 1M tokens and were verified against each provider's published list price via web search in June 2026 — treat them as sane defaults for the cost HUD, not invoices; an operator on negotiated rates can edit them here. A model absent from this map is "unpriced" and the HUD hides its dollar figure rather than guessing.
Functions ¶
func CostOf ¶ added in v1.8.0
CostOf prices a usage slice for model m against MODEL_PRICING. ok is false when m has no rate card — the caller should then hide the cost rather than render a misleading $0.00. A model priced at zero (e.g. a local Ollama model) returns (0, true): genuinely free, not unknown.
Types ¶
type AgentStatus ¶
type AgentStatus string
const ( INIT AgentStatus = "init" DRAINING AgentStatus = "draining" THINKING AgentStatus = "thinking" TEXTING AgentStatus = "texting" EXECUTING AgentStatus = "executing" INTERRUPTED AgentStatus = "interrupted" IDLE AgentStatus = "idle" MAX_ITERS AgentStatus = "max_iters" SAVING AgentStatus = "saving" COMPACTING AgentStatus = "compacting" READY_REPORT AgentStatus = "ready_report" SHUTDOWN AgentStatus = "shutdown" CRUSHED AgentStatus = "crushed" )
func (AgentStatus) String ¶
func (a AgentStatus) String() string
type LLMProvider ¶
func GetAllProviders ¶
func GetAllProviders() []LLMProvider
func GetProvider ¶
func GetProvider(name string) (LLMProvider, bool)
func ProviderOfModel ¶ added in v1.4.3
func ProviderOfModel(m Model) (LLMProvider, bool)
ProviderOfModel returns the provider that lists m in its Models. Lets a caller holding only a model id (e.g. a per-member `model:` pin in a swarm profile.yml) derive the provider to route through.
func (LLMProvider) ModelForLevel ¶
func (p LLMProvider) ModelForLevel(level int) Model
ModelForLevel returns the model to use at a given capability tier:
- level 1 (default) → the smallest / cheapest model the provider lists (Models[0]). Suitable for routine work; "normal" tier.
- level 2 → the largest model the provider lists (Models[len-1]). Suitable for hard reasoning; "big" tier. More expensive — the LLM should only pick this when the task genuinely needs deeper thinking.
Providers with only one configured model collapse both tiers to that model. Levels outside {1, 2} clamp to the nearest valid tier.
type Model ¶
type Model string
const ( // OLLAMA QWEN_3_6 Model = "qwen3.6" // ANTHROPIC SONNET_4_6 Model = "claude-sonnet-4-6" OPUS_4_8 Model = "claude-opus-4-8" // DEEPSEEK DEEPSEEK_V4_FLASH Model = "deepseek-v4-flash" DEEPSEEK_V4_PRO Model = "deepseek-v4-pro" // OPENAI GPT_5_4_MINI Model = "gpt-5.4-mini" GPT_5_5 Model = "gpt-5.5" // GLM (Zhipu / z.ai). glm-4.6 is the cheaper coding model (normal tier), // glm-5.2 the flagship (big tier). Both are reached over the Anthropic- // compatible endpoint; vision input requires a vision-capable GLM model. GLM_4_6 Model = "glm-4.6" GLM_5_2 Model = "glm-5.2" // QWEN (Alibaba DashScope). qwen3.7-plus is the cheaper tier, qwen3.7-max the // flagship (1M ctx, native extended-thinking). OpenAI-compatible wire. QWEN3_7_PLUS Model = "qwen3.7-plus" QWEN3_7_MAX Model = "qwen3.7-max" )
type Pricing ¶ added in v1.8.0
type Pricing struct {
Input float64 // uncached input
Output float64 // output — hidden reasoning tokens are a subset, billed here
CacheRead float64 // cached-prefix re-read
CacheWrite float64 // cached-prefix first write
}
Pricing is a model's USD rate card, every rate expressed per 1,000,000 tokens. A zero-value Pricing means "genuinely free" (a locally hosted Ollama model costs nothing to run) — distinct from a model that is absent from MODEL_PRICING, which is "unpriced / unknown" (see CostOf).
CacheRead / CacheWrite price Anthropic-style prompt-cache tokens: a READ re-uses a previously cached prefix at a steep discount, while the first WRITE of that prefix carries a surcharge. Providers without prompt caching report zero cache tokens (see llm.Usage), so for them these rates are never exercised and simply mirror Input.
func (Pricing) CostUSD ¶ added in v1.8.0
CostUSD prices one usage slice. The arguments mirror llm.Usage: in is the FULL input-token count, of which cacheRead + cacheWrite are a subset (Anthropic's accounting), and out is the output-token count (hidden reasoning tokens are already inside out). Cached tokens are billed at their own rates and carved out of the uncached-input pool so they are never double-charged.
Source Files