Cyrene_v2/internal/runtime/model/reasoning.go

package model

import (
	"io"
	"log/slog"
	"strings"

	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
)

// discardLogger returns a logger that drops every record but still carries the
// adapter's module source, so an adapter built without an injected logger can
// never emit an unsourced record.
func discardLogger(source string) *slog.Logger {
	return logger.WithSource(slog.New(slog.NewTextHandler(io.Discard, nil)), source)
}

// Anthropic extended-thinking protocol bounds (design/04 §3.2).
const (
	// anthropicMinThinkingBudget is the smallest budget_tokens the Messages API
	// accepts for thinking.type=enabled; anything below it is rejected with a
	// 400, so a lower non-zero budget is raised and reported rather than sent.
	anthropicMinThinkingBudget = 1024
	// anthropicDefaultThinkingBudget is used when thinking is switched on without
	// an explicit budget.
	anthropicDefaultThinkingBudget = 4096
	// anthropicThinkingBodyReserve is the body room kept on top of the thinking
	// budget: max_tokens is shared by thinking and the visible answer, so the
	// budget must never be allowed to consume it all (design/04 §3.2 red line:
	// "预算与正文不得互相吃光").
	anthropicThinkingBodyReserve = 2048
)

// reasoningVendor identifies the vendor behind an OpenAI-compatible endpoint.
// The wire protocol is the same for all of them, but the accepted
// reasoning_effort values are not, so a value must be checked against the right
// whitelist instead of being forwarded blindly.
type reasoningVendor string

const (
	vendorOpenAI   reasoningVendor = "openai"
	vendorDeepSeek reasoningVendor = "deepseek"
	vendorQwen     reasoningVendor = "qwen"
	vendorUnknown  reasoningVendor = "unknown"
)

// reasoningEffortLadder is the canonical ascending order of reasoning_effort
// tiers. It is used to round an unsupported value down to the nearest tier a
// vendor accepts.
var reasoningEffortLadder = []string{"none", "minimal", "low", "medium", "high", "xhigh", "max"}

// reasoningEffortAllowed is the per-vendor whitelist. A vendor absent from this
// table has no reasoning_effort support at all, which is reported rather than
// guessed.
var reasoningEffortAllowed = map[reasoningVendor]map[string]bool{
	vendorOpenAI:   tierSet("none", "minimal", "low", "medium", "high", "xhigh", "max"),
	vendorDeepSeek: tierSet("low", "high", "max"),
}

// tierSet builds a membership set from the given tiers.
func tierSet(tiers ...string) map[string]bool {
	set := make(map[string]bool, len(tiers))
	for _, t := range tiers {
		set[t] = true
	}
	return set
}

// reasoningVendorFor classifies the vendor of an OpenAI-compatible request from
// its model name, falling back to the base URL when the name does not identify
// one. An unrecognized pair yields vendorUnknown so the caller reports "not
// supported" instead of assuming OpenAI semantics.
func reasoningVendorFor(model, baseURL string) reasoningVendor {
	m := strings.ToLower(strings.TrimSpace(model))
	switch {
	case strings.Contains(m, "deepseek"):
		return vendorDeepSeek
	case strings.Contains(m, "qwen"):
		return vendorQwen
	case strings.Contains(m, "gpt-"), strings.Contains(m, "chatgpt"),
		strings.HasPrefix(m, "o1"), strings.HasPrefix(m, "o3"), strings.HasPrefix(m, "o4"):
		return vendorOpenAI
	}
	u := strings.ToLower(strings.TrimSpace(baseURL))
	switch {
	case strings.Contains(u, "deepseek"):
		return vendorDeepSeek
	case strings.Contains(u, "dashscope"), strings.Contains(u, "aliyuncs"):
		return vendorQwen
	case strings.Contains(u, "openai"):
		return vendorOpenAI
	}
	return vendorUnknown
}

// resolveReasoningEffort validates effort against the vendor's whitelist and
// returns the value to send. An unsupported tier is rounded down to the nearest
// tier the vendor accepts; a value the vendor does not accept at all — or a
// vendor with no whitelist — yields "", meaning the field is dropped. Every
// decision is logged (provider, field, value), never silently applied.
func resolveReasoningEffort(log *slog.Logger, vendor reasoningVendor, effort, model string) string {
	effort = strings.TrimSpace(effort)
	if effort == "" {
		return ""
	}
	allowed, ok := reasoningEffortAllowed[vendor]
	if !ok {
		log.Warn("reasoning_effort ignored: provider does not support it",
			"provider", string(vendor), "field", "reasoning_effort", "value", effort, "model", model)
		return ""
	}
	if allowed[effort] {
		return effort
	}
	downgraded, ok := downgradeReasoningEffort(allowed, effort)
	if !ok {
		log.Warn("reasoning_effort dropped: unrecognized value",
			"provider", string(vendor), "field", "reasoning_effort", "value", effort, "model", model)
		return ""
	}
	log.Warn("reasoning_effort downgraded: value not accepted by provider",
		"provider", string(vendor), "field", "reasoning_effort",
		"value", effort, "sent", downgraded, "model", model)
	return downgraded
}

// downgradeReasoningEffort rounds an unsupported tier down to the nearest tier
// the vendor accepts using the canonical ladder order; a tier below the vendor's
// minimum maps to that minimum. It reports false when effort is not on the
// ladder at all, so the caller drops the field instead of inventing a value.
func downgradeReasoningEffort(allowed map[string]bool, effort string) (string, bool) {
	start := -1
	for i, tier := range reasoningEffortLadder {
		if tier == effort {
			start = i
			break
		}
	}
	if start < 0 {
		return "", false
	}
	for i := start; i >= 0; i-- {
		if allowed[reasoningEffortLadder[i]] {
			return reasoningEffortLadder[i], true
		}
	}
	for _, tier := range reasoningEffortLadder {
		if allowed[tier] {
			return tier, true
		}
	}
	return "", false
}

// anthropicThinkingFor resolves the extended-thinking directive together with the
// max_tokens it requires. It returns the (possibly raised) max_tokens and the
// wire block; a nil block means "do not send thinking" (no directive, or an
// explicit off). Because max_tokens is shared by thinking and the visible answer,
// the returned budget always leaves anthropicThinkingBodyReserve for the body,
// and a request whose budget would eat the answer raises max_tokens instead — the
// adjustment is logged so it is never a silent rewrite.
func anthropicThinkingFor(th *Thinking, maxTokens int, log *slog.Logger) (int, *anthropicThinking) {
	if th == nil || (th.Enable != nil && !*th.Enable) {
		return maxTokens, nil
	}
	budget := th.BudgetTokens
	if budget == 0 {
		budget = anthropicDefaultThinkingBudget
	}
	if budget < anthropicMinThinkingBudget {
		log.Warn("thinking budget raised to the protocol minimum",
			"provider", "anthropic", "field", "thinking.budget_tokens",
			"budget_tokens", th.BudgetTokens, "sent", anthropicMinThinkingBudget)
		budget = anthropicMinThinkingBudget
	}
	if need := budget + anthropicThinkingBodyReserve; maxTokens < need {
		log.Warn("max_tokens raised to leave room for thinking plus the answer",
			"provider", "anthropic", "budget_tokens", budget,
			"max_tokens_requested", maxTokens, "max_tokens_sent", need,
			"body_reserve", anthropicThinkingBodyReserve)
		maxTokens = need
	}
	return maxTokens, &anthropicThinking{Type: anthropicThinkingTypeEnabled, BudgetTokens: budget}
}

// reportUnmappedThinking logs why a Thinking directive cannot be carried by an
// OpenAI-compatible request, naming the provider and the field that was ignored.
// Qwen3 does have enable_thinking, but the protocol has no portable field for
// it (vLLM wants chat_template_kwargs.enable_thinking, DashScope's compatible
// mode wants a top-level enable_thinking), and sending the wrong shape can 400
// the request — so this version reports it instead of guessing.
func reportUnmappedThinking(log *slog.Logger, vendor reasoningVendor, model string, th *Thinking) {
	if th == nil {
		return
	}
	if th.Enable != nil {
		switch vendor {
		case vendorQwen:
			log.Warn("enable_thinking not forwarded: the OpenAI-compatible protocol has no portable field for it",
				"provider", "qwen", "field", "extra_body.chat_template_kwargs.enable_thinking",
				"enable", *th.Enable, "model", model)
		case vendorDeepSeek:
			log.Warn("deepseek thinking switch not mapped in this version; only reasoning_effort is forwarded",
				"provider", "deepseek", "field", "thinking.type", "enable", *th.Enable, "model", model)
		default:
			log.Warn("thinking switch ignored: provider has no boolean thinking switch",
				"provider", string(vendor), "field", "thinking.enable", "enable", *th.Enable, "model", model)
		}
		return
	}
	if th.BudgetTokens > 0 {
		log.Warn("thinking budget ignored: OpenAI-compatible providers take reasoning_effort, not a token budget",
			"provider", string(vendor), "field", "thinking.budget_tokens",
			"budget_tokens", th.BudgetTokens, "model", model)
	}
}