feat(model,capacity,vision,io): per-persona 路由 + 本地三级(默认关) + L3 分类 + OCR 独立模型 + image-crop 工具（裁决 03.4.1/03.4.3/14.4.1/14.4.4/14.4.5）

- 03.4.1：per-persona 路由载体选 models.json（组内新增 personaModels/personaProviders + 校验 + PersonaRoute()）， Router/RuntimeLLM 快照绑定、loop 流式路径按该 turn 的 persona 路由；主步用路由结果、降级步钉自己的模型 - 03.4.3：本地第三级降级链（LLM_LOCAL_*，Configured() 含 Enabled ⇒ **默认关**），warn 标注来源 - 14.4.4：capacity 接入 L3（分类器接口 + TTL 缓存 + 严格 JSON 解析），CAPACITY_L3_ENABLED **默认 false** - 14.4.1：语义修正 —— group.multimodal 才是「主模型能否收图」（env 路径不再臆断），LLM_VISION_MODEL 只配视觉工具； vision/OCR 可各用不同模型；runtime 不再因有图抢路由（vision.Router 是唯一分层权威） - 14.4.5：新增 image-crop 工具（工具数 5→6，策略默认 owner-only），schema 支持按区域裁切 - 组合根：wire_runtime 读 group.Multimodal；capacity 装配拆出 wire_capacity.go（守 500 行） - 复核：`go build`/`go vet`/各包 `go test -count=1` 全绿；**6 条变异反证**先红后还原（含作者自查出的 2 条假阳性变异） - 未落：按 provider 的 thinking 开关（需 openai_encode.go/openai.go 文件面）、persona awakening（需 loader 文件面）
2026-10-04 12:44:52 +08:00
1 parent 63e5b70e3d
commit fa33c79e11
@@ -0,0 +1,99 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"log/slog"
+	"strings"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
+)
+
+// 本文件拥有唤醒式能力（capacity）决策层的装配：判定器构造、工具集 JSON 解析、
+// L3 缓存 TTL 解析与 capacity_awaken 事件出口。它从 wire_runtime.go 拆出
+// （AGENTS §10 文件 ≤500 行；wire.go 拆分建议同样适用于本文件）。
+
+// buildCapacity constructs the awakened-capacity decider when enabled, or nil
+// (which keeps the loop's default all-tools behavior). It is opt-in via
+// CAPACITY_ENABLED; the allowed-set mapping comes from CAPACITY_TOOL_SETS and
+// falls back to built-in defaults on a parse error.
+//
+// The model-assisted L3 layer (裁决 14.4.4) is a second, independent switch
+// (CAPACITY_L3_ENABLED, default false): even with capacity on, L3 stays silent
+// unless it was explicitly asked for, and its verdict cache TTL
+// (CAPACITY_L3_CACHE_TTL, default 5m) is parsed here because the config loader
+// only carries scalars.
+func (a *app) buildCapacity() *capacity.Decider {
+	if !a.cfg.Capacity.Enabled {
+		return nil
+	}
+	sets, err := parseCapacityToolSets(a.cfg.Capacity.ToolSetsJSON)
+	if err != nil {
+		a.log.Warn("capacity tool sets parse failed, using built-in defaults", "error", err)
+		sets = nil
+	}
+	cfg := capacity.Config{
+		Enabled:       true,
+		MinConfidence: a.cfg.Capacity.MinConfidence,
+		ToolSets:      sets,
+	}
+	if a.cfg.Capacity.L3Enabled {
+		cfg.L3Enabled = true
+		cfg.L3Classifier = capacity.NewLLMClassifier(a.modelAdapter, a.log)
+		cfg.L3CacheTTL = parseL3CacheTTL(a.cfg.Capacity.L3CacheTTL, a.log)
+	}
+	return capacity.New(cfg, capacity.WithLogger(a.log))
+}
+
+// parseL3CacheTTL parses the CAPACITY_L3_CACHE_TTL duration string. An empty
+// value selects the code default (5m, the landing suggestion of 裁决 14.4.4); an
+// unparsable value is reported and falls back to the same default rather than
+// silently disabling the cache.
+func parseL3CacheTTL(raw string, log *slog.Logger) time.Duration {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return capacity.DefaultL3CacheTTL
+	}
+	d, err := time.ParseDuration(raw)
+	if err != nil {
+		logger.WithSource(log, "cmd/cyrene").Warn("capacity L3 cache ttl unparsable, using the default",
+			"value", raw, "default", capacity.DefaultL3CacheTTL.String(), "error", err)
+		return capacity.DefaultL3CacheTTL
+	}
+	return d
+}
+
+// parseCapacityToolSets decodes the CAPACITY_TOOL_SETS JSON map of
+// state→allowed tool name list. A blank string yields nil (built-in defaults); a
+// null value for a state means unrestricted (all tools).
+func parseCapacityToolSets(raw string) (map[capacity.State][]string, error) {
+	if strings.TrimSpace(raw) == "" {
+		return nil, nil
+	}
+	var m map[string][]string
+	if err := json.Unmarshal([]byte(raw), &m); err != nil {
+		return nil, fmt.Errorf("capacity tool sets: %w", err)
+	}
+	out := make(map[capacity.State][]string, len(m))
+	for k, v := range m {
+		out[capacity.State(k)] = v
+	}
+	return out, nil
+}
+
+// onCapacity is the loop's OnCapacity sink: it observes a capacity_awaken
+// transition and pushes it to the Web UI (design/14 §5) as a `capacity_awaken`
+// frame. The loop only emits on a real state transition (never every round), and
+// the event carries the originating session id so it routes to that session, not
+// globally. The payload is the frontend AwakenPayload {fromState,toState,tool}.
+func (a *app) onCapacity(ev capacity.Event) {
+	logger.WithSource(a.log, "cmd/cyrene").Debug("capacity event",
+		"type", ev.Type, "from", ev.FromState, "to", ev.ToState,
+		"tool", ev.Tool, "at", ev.At, "session", ev.SessionID)
+	if a.capacitySink == nil {
+		return
+	}
+	a.capacitySink.BroadcastCapacityAwaken(ev.SessionID, string(ev.FromState), string(ev.ToState), ev.Tool)
+}
@@ -2,7 +2,6 @@ package main

 import (
 	"context"
-	"encoding/json"
 	"fmt"
 	"path/filepath"
 	"strings"
@@ -15,7 +14,6 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/persona/loader"
 	personareg "git.yeij.top/AskaEth/Cyrene_v2/internal/persona/registry"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
-	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/embedding"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/guardrail"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
@@ -191,8 +189,14 @@ func (a *app) wireRuntime() error {
 	}

 	// Vision routing: the loop rewrites an image-bearing user message per the
-	// vision Router — a vision-capable main model sees the image directly,
-	// otherwise the vision/OCR tool's text description replaces it (design/06 §0).
+	// vision Router — a main model that really accepts image input sees the image
+	// directly, otherwise the vision/OCR tool's text description replaces it
+	// (design/06 §0)。
+	//
+	// SupportsVision means "主模型能否收图"（裁决 14.4.1 修语义，C-7）, read from the
+	// active models.json group's `multimodal` declaration — NOT from
+	// LLM_VISION_MODEL, which configures the vision *tool's* model. It is read per
+	// call so a panel save/reload takes effect on the next turn.
 	visionRouter := vision.NewRouter(
 		vision.WithVisionTool(visionToolName),
 		vision.WithOCRTool(ocrToolName),
@@ -200,7 +204,8 @@ func (a *app) wireRuntime() error {
 	visionResolver := vision.NewMessageResolver(
 		visionRouter,
 		func() vision.ModelInfo {
-			return vision.ModelInfo{Name: a.cfg.LLM.VisionModel, SupportsVision: a.cfg.LLM.VisionModel != ""}
+			group := a.llm.Group()
+			return vision.ModelInfo{Name: group.Model, SupportsVision: group.Multimodal}
 		},
 		a.dispatcher.Invoke,
 	)
@@ -374,40 +379,8 @@ func (a *app) approvalActor(ctx context.Context) string {
 	return pid
 }

-// buildCapacity constructs the awakened-capacity decider when enabled, or nil
-// (which keeps the loop's default all-tools behavior). It is opt-in via
-// CAPACITY_ENABLED; the allowed-set mapping comes from CAPACITY_TOOL_SETS and
-// falls back to built-in defaults on a parse error.
-func (a *app) buildCapacity() *capacity.Decider {
-	if !a.cfg.Capacity.Enabled {
-		return nil
-	}
-	sets, err := parseCapacityToolSets(a.cfg.Capacity.ToolSetsJSON)
-	if err != nil {
-		a.log.Warn("capacity tool sets parse failed, using built-in defaults", "error", err)
-		sets = nil
-	}
-	return capacity.New(capacity.Config{
-		Enabled:       true,
-		MinConfidence: a.cfg.Capacity.MinConfidence,
-		ToolSets:      sets,
-	}, capacity.WithLogger(a.log))
-}
-
-// onCapacity is the loop's OnCapacity sink: it observes a capacity_awaken
-// transition and pushes it to the Web UI (design/14 §5) as a `capacity_awaken`
-// frame. The loop only emits on a real state transition (never every round), and
-// the event carries the originating session id so it routes to that session, not
-// globally. The payload is the frontend AwakenPayload {fromState,toState,tool}.
-func (a *app) onCapacity(ev capacity.Event) {
-	logger.WithSource(a.log, "cmd/cyrene").Debug("capacity event",
-		"type", ev.Type, "from", ev.FromState, "to", ev.ToState,
-		"tool", ev.Tool, "at", ev.At, "session", ev.SessionID)
-	if a.capacitySink == nil {
-		return
-	}
-	a.capacitySink.BroadcastCapacityAwaken(ev.SessionID, string(ev.FromState), string(ev.ToState), ev.Tool)
-}
+// buildCapacity / onCapacity / capacity JSON+TTL parsing live in
+// wire_capacity.go (AGENTS §10 文件 ≤500 行).

 // onToolAudit is the loop's OnAudit sink: it forwards a tool call denied by the
 // design/11 §3 security gate in a group/stranger environment into the gateway
@@ -424,23 +397,7 @@ func (a *app) onToolAudit(ev loop.AuditEvent) {
 	a.gw.AuditEvent(ev.Who, ev.What, ev.Result)
 }

-// parseCapacityToolSets decodes the CAPACITY_TOOL_SETS JSON map of
-// state→allowed tool name list. A blank string yields nil (built-in defaults); a
-// null value for a state means unrestricted (all tools).
-func parseCapacityToolSets(raw string) (map[capacity.State][]string, error) {
-	if strings.TrimSpace(raw) == "" {
-		return nil, nil
-	}
-	var m map[string][]string
-	if err := json.Unmarshal([]byte(raw), &m); err != nil {
-		return nil, fmt.Errorf("capacity tool sets: %w", err)
-	}
-	out := make(map[capacity.State][]string, len(m))
-	for k, v := range m {
-		out[capacity.State(k)] = v
-	}
-	return out, nil
-}
+// parseCapacityToolSets / parseL3CacheTTL live in wire_capacity.go.

 // dbCfgForEmbed returns the DB config used by the vector index, which opens the
 // same SQLite file as the primary pool.