Add tuned routing evaluation levers

cmd/router/main.go

@@ -537,6 +537,8 @@ func main() {
 	// below can be overridden per deployment.
 	plannerEnabled := config.GetOr("ROUTER_PLANNER_ENABLED", "true") == "true"
 	scoreToolResultTurns := config.GetOr("ROUTER_SCORE_TOOL_RESULT_TURNS", "true") == "true"
+	// Keep ToolResult candidates within the requested tier for safer evaluation.
+	toolResultTierCeiling := config.GetOr("ROUTER_TOOL_RESULT_TIER_CEILING", "false") == "true"
 	// Defensive backstop for Anthropic's cyber safety classifier: re-pin a
 	// session off a model that returned a safety refusal. Default off; enabled
 	// on the defensive deploy. Fallback target when the pin has no runner-up.
@@ -771,6 +773,7 @@ func main() {
 		WithSubAgentOverride(subAgentProvider, subAgentModel).
 		WithPlannerEnabled(plannerEnabled).
 		WithScoreToolResultTurns(scoreToolResultTurns).
+		WithToolResultTierCeiling(toolResultTierCeiling).
 		WithCyberRefusalRepin(cyberRefusalRepin).
 		WithCyberRefusalFallbackModel(cyberRefusalFallbackModel).
 		WithPrefixTrimFreeSwitch(prefixTrimFreeSwitch).
@@ -846,6 +849,18 @@ func main() {
 		logger.Info("Provider exclusion override active", "excluded_providers", cleaned)
 	}
 
+	if respectRaw := strings.TrimSpace(config.GetOr("ROUTER_RESPECT_REQUESTED_MODEL", "")); respectRaw != "" {
+		parts := strings.Split(respectRaw, ",")
+		cleaned := make([]string, 0, len(parts))
+		for _, p := range parts {
+			if trimmed := strings.TrimSpace(p); trimmed != "" {
+				cleaned = append(cleaned, trimmed)
+			}
+		}
+		proxySvc = proxySvc.WithRespectRequestedModel(cleaned)
+		logger.Info("Requested-model passthrough active", "honored_models", cleaned)
+	}
+
 	// The usage observer is always wired (cheap, side-effect-free) even though
 	// the cost discount below is env-gated: it also feeds the per-installation
 	// usage-bypass gate, which is DB-gated and can't know the env flag's state.
@@ -1070,6 +1085,13 @@ func buildClusterScorer(availableProviders map[string]struct{}) (router.Router,
 			logger.Info("Cluster top_p overridden", "top_p", n)
 		}
 	}
+	if v := strings.TrimSpace(config.GetOr("ROUTER_STATIC_CLUSTER_PIN", "")); v != "" {
+		pins := parseStaticClusterPins(v, logger)
+		if len(pins) > 0 {
+			cfg.StaticClusterPin = pins
+			logger.Info("Static cluster pin active", "pins", pins)
+		}
+	}
 	scorers := make(map[string]*cluster.Scorer, len(versions))
 	warmed := make(map[string]cluster.Embedder)
 	var defaultEmbedderID string

cmd/router/static_cluster_pin.go

@@ -0,0 +1,38 @@
+package main
+
+import (
+	"fmt"
+	"log/slog"
+	"strconv"
+	"strings"
+
+	"workweave/router/internal/router/catalog"
+)
+
+func parseStaticClusterPins(raw string, logger *slog.Logger) map[int]string {
+	pins := make(map[int]string)
+	for _, pair := range strings.Split(raw, ",") {
+		pair = strings.TrimSpace(pair)
+		if pair == "" {
+			continue
+		}
+		clusterID, model, found := strings.Cut(pair, ":")
+		if !found || strings.TrimSpace(model) == "" {
+			panic(fmt.Sprintf("invalid ROUTER_STATIC_CLUSTER_PIN entry %q: expected cluster_id:model", pair))
+		}
+		id, err := strconv.Atoi(strings.TrimSpace(clusterID))
+		if err != nil {
+			panic(fmt.Sprintf("invalid ROUTER_STATIC_CLUSTER_PIN cluster ID in %q", pair))
+		}
+		model = strings.TrimSpace(model)
+		if _, known := catalog.ByID(model); !known {
+			logger.Warn("Static cluster pin model is not in catalog; skipping",
+				"cluster_id", id,
+				"model", model,
+			)
+			continue
+		}
+		pins[id] = model
+	}
+	return pins
+}

cmd/router/static_cluster_pin_test.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"io"
+	"log/slog"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestParseStaticClusterPinsSkipsUnknownModels(t *testing.T) {
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+
+	got := parseStaticClusterPins("0:claude-haiku-4-5,1:not-in-catalog", logger)
+
+	require.Equal(t, map[int]string{0: "claude-haiku-4-5"}, got)
+}
+
+func TestParseStaticClusterPinsPanicsOnMalformedEntry(t *testing.T) {
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+
+	assert.Panics(t, func() {
+		parseStaticClusterPins("not-a-pair", logger)
+	})
+}

docs/CONFIGURATION.md

@@ -150,6 +150,9 @@ Set `DATABASE_URL` directly, or compose it from the individual vars:
 | `ROUTER_HARD_PIN_EXPLORE`         | `true`                       | Pin Claude Code Task-tool sub-agent turns to `ROUTER_HARD_PIN_MODEL`/`ROUTER_HARD_PIN_PROVIDER` (or the cheapest deployed model, if those are unset). Set `false` to route sub-agents through the scorer like any other turn. |
 | `ROUTER_SUBAGENT_MODEL`           | *(none)*                     | Route Claude Code Task-tool sub-agent turns to a distinct model, independent of `ROUTER_HARD_PIN_MODEL` — e.g. a local/self-hosted OpenAI-compatible model (point `OPENROUTER_BASE_URL` at your local server) while the main loop keeps using Anthropic/whatever the scorer picks. Requires `ROUTER_SUBAGENT_PROVIDER`; either alone is ignored. Takes effect regardless of `ROUTER_HARD_PIN_EXPLORE`, but the HMM strategy keeps its own sub-agent handling and isn't affected. |
 | `ROUTER_SUBAGENT_PROVIDER`        | *(none)*                     | Pair with `ROUTER_SUBAGENT_MODEL`. |
+| `ROUTER_RESPECT_REQUESTED_MODEL`  | *(none)*                     | Comma-separated models served verbatim when requested, bypassing planner and scorer. Automatic hard pins still win; an explicit force-model override also wins over this list. Excluded or ineligible honored models fall back to automatic routing. |
+| `ROUTER_STATIC_CLUSTER_PIN`       | *(none)*                     | `clusterID:model` pairs, comma-separated. A nearest-cluster match skips `blendScoresV2`; an ineligible pinned model falls through to normal routing. Malformed pairs or cluster IDs abort startup; valid pins naming models absent from the catalog are warned and skipped. `Reason` is tagged `cluster-pin:`. |
+| `ROUTER_TOOL_RESULT_TIER_CEILING` | `false`                      | On `ToolResult` turns, excludes models above the requested model's tier on Messages, OpenAI Chat Completions, and Gemini surfaces. It is a no-op when no at-or-below-tier model survives, and a session-pin fit-check cannot lift its exclusions. Unknown requested-model tiers are a no-op. |
 | `ROUTER_TRANSLATION_COMPATIBILITY_MODE` | `shadow` | Translation representability rollout: `off` disables broad filtering, `shadow` records candidate exclusions without changing routes, and `enforce` makes declared semantic requirements hard routing constraints. Native-only safety paths (such as unsupported Responses tool unions and native Gemini ingress) remain protected unless mode is `off`. |
 | `ROUTER_COMPACTION_PCT`           | `0.85`                       | Fraction of the largest eligible model's context window at which the proactive compaction cascade engages (clear old tool results → structured summary → trim). Range `(0,1]`; `0` disables compaction (over-window requests then 413). Mirrors Claude Code's ~0.85 auto-compact trigger. |
 | `ROUTER_ONNX_ASSETS_DIR`          | `/opt/router/assets`         | Directory containing `model.onnx` + `tokenizer.json`. |

internal/proxy/force_model_tier_fallback_internal_test.go

@@ -24,6 +24,7 @@ import (
 type tierProbeRouter struct {
 	available map[string]struct{}
 	captured  []router.Request
+	provider  string
 }
 
 func (r *tierProbeRouter) Route(_ context.Context, req router.Request) (router.Decision, error) {
@@ -41,7 +42,11 @@ func (r *tierProbeRouter) Route(_ context.Context, req router.Request) (router.D
 	if best == "" {
 		return router.Decision{}, errors.New("no eligible candidate")
 	}
-	return router.Decision{Provider: providers.ProviderAnthropic, Model: best, Reason: "fake"}, nil
+	provider := r.provider
+	if provider == "" {
+		provider = providers.ProviderAnthropic
+	}
+	return router.Decision{Provider: provider, Model: best, Reason: "fake"}, nil
 }
 
 // forcedPinStore returns a single user-forced pin for every lookup.