feat: accurate context window sizing from arm capabilities + prefix token baseline + tokenizer wiring

cmd/gnoma/main.go

@@ -22,6 +22,7 @@ import (
 	"somegit.dev/Owlibou/gnoma/internal/provider"
 	"somegit.dev/Owlibou/gnoma/internal/router"
 	"somegit.dev/Owlibou/gnoma/internal/security"
+	"somegit.dev/Owlibou/gnoma/internal/tokenizer"
 	anthropicprov "somegit.dev/Owlibou/gnoma/internal/provider/anthropic"
 	"somegit.dev/Owlibou/gnoma/internal/provider/mistral"
 	googleprov "somegit.dev/Owlibou/gnoma/internal/provider/google"
@@ -329,16 +330,32 @@ func main() {
 		logger.Debug("loaded project docs as context prefix", "file", name, "size", len(data))
 	}
 
+	// Derive context window size from registered arm capabilities (accurate) or fall back to heuristic
+	contextWindowSize := int64(cfg.Provider.MaxTokens) * 20
+	if arm, ok := rtr.LookupArm(armID); ok && arm.Capabilities.ContextWindow > 0 {
+		contextWindowSize = int64(arm.Capabilities.ContextWindow)
+		logger.Debug("context window from arm capabilities", "arm", armID, "context_window", contextWindowSize)
+	}
+
 	// Create context window with summarize strategy (falls back to truncation)
 	var compactStrategy gnomactx.Strategy
 	compactStrategy = gnomactx.NewSummarizeStrategy(prov)
 	ctxWindow := gnomactx.NewWindow(gnomactx.WindowConfig{
-		MaxTokens:      cfg.Provider.MaxTokens * 20, // rough: max_tokens is per-turn, context window ~20x
+		MaxTokens:      contextWindowSize,
 		Strategy:       compactStrategy,
 		PrefixMessages: prefixMsgs,
 		Logger:         logger,
 	})
 
+	// Wire tokenizer and seed tracker with prefix cost
+	tok := tokenizer.ForProvider(prov.Name())
+	ctxWindow.Tracker().SetTokenizer(tok)
+	if len(prefixMsgs) > 0 {
+		prefixTokens := ctxWindow.Tracker().CountMessages(prefixMsgs)
+		ctxWindow.Tracker().Set(prefixTokens)
+		logger.Debug("prefix token baseline set", "tokens", prefixTokens)
+	}
+
 	// Create engine
 	eng, err := engine.New(engine.Config{
 		Provider:    prov,

internal/engine/loop.go

@@ -76,7 +76,11 @@ func (e *Engine) runLoop(ctx context.Context, cb Callback) (*Turn, error) {
 				}
 			}
 			task := router.ClassifyTask(prompt)
-			task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+			if e.cfg.Context != nil {
+				task.EstimatedTokens = int(e.cfg.Context.Tracker().CountTokens(prompt))
+			} else {
+				task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+			}
 
 			e.logger.Debug("routing request",
 				"task_type", task.Type,
@@ -117,7 +121,11 @@ func (e *Engine) runLoop(ctx context.Context, cb Callback) (*Turn, error) {
 						}
 					}
 					task := router.ClassifyTask(prompt)
-					task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+					if e.cfg.Context != nil {
+						task.EstimatedTokens = int(e.cfg.Context.Tracker().CountTokens(prompt))
+					} else {
+						task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+					}
 					var retryDecision router.RoutingDecision
 					s, retryDecision, err = e.cfg.Router.Stream(ctx, task, req)
 					decision = retryDecision // adopt new reservation on retry
@@ -455,7 +463,11 @@ func (e *Engine) handleRequestTooLarge(ctx context.Context, origErr error, req p
 			}
 		}
 		task := router.ClassifyTask(prompt)
-		task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+		if e.cfg.Context != nil {
+			task.EstimatedTokens = int(e.cfg.Context.Tracker().CountTokens(prompt))
+		} else {
+			task.EstimatedTokens = int(gnomactx.EstimateTokens(prompt))
+		}
 		s, _, err := e.cfg.Router.Stream(ctx, task, req)
 		return s, err
 	}