promptpilot 0.1.2 → 0.1.3

package/dist/cli.js

@@ -834,6 +834,9 @@ var DEFAULT_PROVIDER = "ollama";
 var DEFAULT_MAX_INPUT_TOKENS = 1200;
 var DEFAULT_MAX_CONTEXT_TOKENS = 800;
 var DEFAULT_MAX_TOTAL_TOKENS = 2200;
+var DEFAULT_ROUTING_PRIORITY = "cheapest_adequate";
+var DEFAULT_ROUTING_TOP_K = 3;
+var DEFAULT_WORKLOAD_BIAS = "code_first";
 var PromptOptimizer = class {
   config;
   logger;
@@ -871,6 +874,10 @@ var PromptOptimizer = class {
     const maxInputTokens = input.maxInputTokens ?? this.config.maxInputTokens ?? DEFAULT_MAX_INPUT_TOKENS;
     const maxContextTokens = input.maxContextTokens ?? this.config.maxContextTokens ?? DEFAULT_MAX_CONTEXT_TOKENS;
     const maxTotalTokens = input.maxTotalTokens ?? this.config.maxTotalTokens ?? DEFAULT_MAX_TOTAL_TOKENS;
+    const routingEnabled = input.routingEnabled !== false;
+    const routingPriority = input.routingPriority ?? DEFAULT_ROUTING_PRIORITY;
+    const routingTopK = input.routingTopK ?? DEFAULT_ROUTING_TOP_K;
+    const workloadBias = input.workloadBias ?? DEFAULT_WORKLOAD_BIAS;
     const warnings = [];
     const changes = [];
     const useContext = input.useContext !== false && Boolean(input.sessionId);
@@ -959,6 +966,22 @@ var PromptOptimizer = class {
     }
     warnings.push(...providerWarnings);
     changes.push(...providerChanges);
+    const routingDecision = await this.routeDownstreamTargets({
+      input: {
+        ...input,
+        prompt: originalPrompt,
+        mode,
+        preset,
+        routingPriority,
+        routingTopK,
+        workloadBias
+      },
+      routingEnabled,
+      routingPriority,
+      routingTopK,
+      workloadBias
+    });
+    warnings.push(...routingDecision.routingWarnings);
     let finalPrompt = composeFinalPrompt({
       optimizedPrompt,
       input: {
@@ -967,7 +990,8 @@ var PromptOptimizer = class {
         mode,
         preset
       },
-      context: relevantContext
+      context: relevantContext,
+      routingDecision
     });
     let estimatedTokensAfter = {
       prompt: this.estimator.estimateText(optimizedPrompt),
@@ -984,6 +1008,7 @@ var PromptOptimizer = class {
         },
         optimizedPrompt,
         context: relevantContext,
+        routingDecision,
         maxTotalTokens
       });
       finalPrompt = reduced.finalPrompt;
@@ -1019,6 +1044,11 @@ var PromptOptimizer = class {
       mode,
       provider,
       model,
+      selectedTarget: routingDecision.selectedTarget,
+      rankedTargets: routingDecision.rankedTargets,
+      routingReason: routingDecision.routingReason,
+      routingWarnings: routingDecision.routingWarnings,
+      routingProvider: routingDecision.routingProvider,
       warnings,
       changes,
       debugInfo: input.debug ? {
@@ -1027,7 +1057,8 @@ var PromptOptimizer = class {
         estimatedTokensAfter,
         extractedConstraints,
         preset,
-        selectedModel: model
+        selectedModel: model,
+        routingDecision
       } : void 0
     };
   }
@@ -1243,6 +1274,137 @@ var PromptOptimizer = class {
       };
     }
   }
+  async routeDownstreamTargets(options) {
+    const availableTargets = normalizeAvailableTargets(options.input.availableTargets ?? []);
+    if (!options.routingEnabled || availableTargets.length === 0) {
+      return {
+        selectedTarget: null,
+        rankedTargets: [],
+        routingReason: null,
+        routingWarnings: [],
+        routingProvider: null
+      };
+    }
+    if (availableTargets.length === 1) {
+      return {
+        selectedTarget: stripInternalTargetFields(availableTargets[0]),
+        rankedTargets: [
+          {
+            ...stripInternalTargetFields(availableTargets[0]),
+            rank: 1,
+            reason: "Only one downstream target was supplied."
+          }
+        ],
+        routingReason: "Only one downstream target was supplied, so it was selected directly.",
+        routingWarnings: [],
+        routingProvider: "direct"
+      };
+    }
+    if (!this.client.listModels) {
+      return {
+        selectedTarget: null,
+        rankedTargets: [],
+        routingReason: null,
+        routingWarnings: [
+          "Downstream target routing requires local Ollama model discovery so a Qwen router can run."
+        ],
+        routingProvider: null
+      };
+    }
+    try {
+      const installedModels = await this.client.listModels();
+      const routerModel = getQwenRouterModel(installedModels, this.config.routerModel);
+      if (!routerModel) {
+        return {
+          selectedTarget: null,
+          rankedTargets: [],
+          routingReason: null,
+          routingWarnings: [
+            "Downstream target routing could not run because no suitable local Qwen router model is installed."
+          ],
+          routingProvider: null
+        };
+      }
+      const response = await this.client.generateJson({
+        model: routerModel,
+        timeoutMs: options.input.timeoutMs ?? this.config.timeoutMs,
+        temperature: 0,
+        format: "json",
+        systemPrompt: buildDownstreamRoutingSystemPrompt(options.routingPriority, options.workloadBias),
+        prompt: JSON.stringify(
+          {
+            objective: "Rank the caller-supplied downstream targets for this prompt and choose the best top target.",
+            prompt: options.input.prompt,
+            task: options.input.task ?? null,
+            mode: options.input.mode,
+            preset: options.input.preset,
+            tone: options.input.tone ?? null,
+            targetHints: options.input.targetHints ?? [],
+            workloadBias: options.workloadBias,
+            routingPriority: options.routingPriority,
+            candidateTargets: availableTargets.map((target) => ({
+              id: target.id,
+              provider: target.provider,
+              model: target.model,
+              label: target.label ?? null,
+              costRank: target.costRank,
+              latencyRank: target.latencyRank,
+              capabilities: target.capabilities,
+              profile: describeDownstreamTarget(target)
+            }))
+          },
+          null,
+          2
+        )
+      });
+      const rankedTargetIds = Array.from(
+        new Set((response.rankedTargetIds ?? []).map((value) => value.trim()).filter(Boolean))
+      ).slice(0, Math.max(1, options.routingTopK));
+      const rankedTargets = rankedTargetIds.map((id, index) => {
+        const target = availableTargets.find((candidate) => candidate.id === id);
+        if (!target) {
+          return null;
+        }
+        return {
+          ...stripInternalTargetFields(target),
+          rank: index + 1,
+          reason: index === 0 ? response.reason?.trim() || "Selected by the local Qwen downstream router." : `Ranked #${index + 1} by the local Qwen downstream router.`
+        };
+      }).filter((value) => value !== null);
+      const selectedTargetId = response.selectedTargetId?.trim();
+      const selectedTargetCandidate = (selectedTargetId && availableTargets.find((candidate) => candidate.id === selectedTargetId)) ?? (rankedTargets[0] ? availableTargets.find(
+        (candidate) => candidate.provider === rankedTargets[0].provider && candidate.model === rankedTargets[0].model && candidate.label === rankedTargets[0].label
+      ) ?? null : null);
+      if (!selectedTargetCandidate || rankedTargets.length === 0) {
+        return {
+          selectedTarget: null,
+          rankedTargets: [],
+          routingReason: null,
+          routingWarnings: [
+            "Downstream target routing returned an invalid selection, so no downstream target was chosen."
+          ],
+          routingProvider: routerModel
+        };
+      }
+      return {
+        selectedTarget: stripInternalTargetFields(selectedTargetCandidate),
+        rankedTargets,
+        routingReason: response.reason?.trim() || "Selected by the local Qwen downstream router.",
+        routingWarnings: [],
+        routingProvider: routerModel
+      };
+    } catch {
+      return {
+        selectedTarget: null,
+        rankedTargets: [],
+        routingReason: null,
+        routingWarnings: [
+          "Downstream target routing could not complete, so no downstream target was selected."
+        ],
+        routingProvider: null
+      };
+    }
+  }
   heuristicOptimize(options) {
     const lines = [
       `Request: ${options.input.prompt}`,
@@ -1274,7 +1436,8 @@ var PromptOptimizer = class {
     const finalPrompt = composeFinalPrompt({
       optimizedPrompt: this.estimator.truncateToBudget(options.optimizedPrompt, Math.floor(options.maxTotalTokens * 0.5)),
       input: options.input,
-      context: compactContext
+      context: compactContext,
+      routingDecision: options.routingDecision
     });
     return {
       finalPrompt,
@@ -1335,7 +1498,7 @@ ${contextBlock}`);
 - ${constraints.join("\n- ")}`);
   }
   const desiredOutput = [
-    input.input.targetModel ? `Target model: ${input.input.targetModel}` : "Target model: claude",
+    input.routingDecision.selectedTarget ? `Selected target: ${formatTargetLabel(input.routingDecision.selectedTarget)}` : input.input.targetModel ? `Target model: ${input.input.targetModel}` : "Target model: claude",
     `Mode: ${input.input.mode}`,
     `Preset: ${input.input.preset}`
   ];
@@ -1361,6 +1524,77 @@ function emptyRelevantContext() {
     debugInfo: {}
   };
 }
+function normalizeAvailableTargets(targets) {
+  return targets.map((target, index) => ({
+    ...target,
+    id: `${target.provider}:${target.model}:${index}`,
+    label: target.label ?? `${target.provider}:${target.model}`,
+    capabilities: target.capabilities ?? inferCapabilities(target),
+    costRank: target.costRank ?? index + 1,
+    latencyRank: target.latencyRank ?? index + 1
+  }));
+}
+function stripInternalTargetFields(target) {
+  return {
+    provider: target.provider,
+    model: target.model,
+    label: target.label,
+    capabilities: target.capabilities,
+    costRank: target.costRank,
+    latencyRank: target.latencyRank
+  };
+}
+function buildDownstreamRoutingSystemPrompt(priority, workloadBias) {
+  return [
+    "You are a downstream model router for PromptPilot.",
+    "Return strict JSON only with this shape:",
+    '{"selectedTargetId":"string","rankedTargetIds":["string"],"reason":"string"}',
+    "Choose only from the supplied candidate target IDs.",
+    "Rank up to the requested top targets in best-first order.",
+    `Routing priority: ${priority}.`,
+    `Workload bias: ${workloadBias}.`,
+    "Code-first means ambiguous prompts should default toward coding-capable or agentic-capable targets.",
+    "Explicit email, support, chat, and lightweight writing prompts may prefer cheaper lighter targets.",
+    "Do not invent targets. Do not output prose outside JSON."
+  ].join("\n");
+}
+function inferCapabilities(target) {
+  const lower = `${target.provider} ${target.model} ${target.label ?? ""}`.toLowerCase();
+  const capabilities = /* @__PURE__ */ new Set();
+  if (/code|codex|coder|agent|tool/.test(lower)) {
+    capabilities.add("coding");
+  }
+  if (/agent|tool/.test(lower)) {
+    capabilities.add("agentic");
+    capabilities.add("tool_use");
+  }
+  if (/refactor|coder|codex/.test(lower)) {
+    capabilities.add("refactor");
+  }
+  if (/debug|fix|ci/.test(lower)) {
+    capabilities.add("debugging");
+  }
+  if (/write|email|chat|sonnet|mini/.test(lower)) {
+    capabilities.add("writing");
+  }
+  if (/email/.test(lower)) {
+    capabilities.add("email");
+  }
+  return Array.from(capabilities);
+}
+function describeDownstreamTarget(target) {
+  return [
+    `provider=${target.provider}`,
+    `model=${target.model}`,
+    `label=${target.label}`,
+    `costRank=${target.costRank}`,
+    `latencyRank=${target.latencyRank}`,
+    `capabilities=${target.capabilities?.join(",") || "none"}`
+  ].join("; ");
+}
+function formatTargetLabel(target) {
+  return target.label ?? `${target.provider}:${target.model}`;
+}
 
 // src/index.ts
 function createOptimizer(config = {}) {
@@ -1433,6 +1667,12 @@ async function runCli(argv, io = { stdout: process.stdout, stderr: process.stder
       maxLength: parsed.maxLength,
       tags: parsed.tags,
       pinnedConstraints: parsed.pinnedConstraints,
+      availableTargets: parsed.targets,
+      routingEnabled: parsed.routingEnabled,
+      routingPriority: parsed.routingPriority,
+      routingTopK: parsed.routingTopK,
+      targetHints: parsed.targetHints,
+      workloadBias: parsed.workloadBias,
       debug: parsed.debug,
       plainOutput: parsed.plain,
       maxTotalTokens: parsed.maxTotalTokens,
@@ -1456,6 +1696,10 @@ async function runCli(argv, io = { stdout: process.stdout, stderr: process.stder
 `);
     io.stdout.write(`provider=${result.provider} model=${result.model} tokens=${result.estimatedTokensAfter.total} savings=${result.tokenSavings}
 `);
+    if (result.selectedTarget) {
+      io.stdout.write(`selected_target=${formatTarget(result.selectedTarget)}
+`);
+    }
     if (result.warnings.length > 0) {
       io.stdout.write(`warnings=${result.warnings.join(" | ")}
 `);
@@ -1478,7 +1722,10 @@ function parseOptimizeArgs(args) {
     bypassOptimization: false,
     help: false,
     tags: [],
-    pinnedConstraints: []
+    pinnedConstraints: [],
+    targets: [],
+    targetHints: [],
+    routingEnabled: true
   };
   const positionals = [];
   for (let index = 0; index < args.length; index += 1) {
@@ -1517,6 +1764,24 @@ function parseOptimizeArgs(args) {
       case "--pin-constraint":
         parsed.pinnedConstraints.push(args[++index]);
         break;
+      case "--target":
+        parsed.targets.push(parseTargetCandidate(args[++index], parsed.targets.length));
+        break;
+      case "--target-hint":
+        parsed.targetHints.push(args[++index]);
+        break;
+      case "--routing-priority":
+        parsed.routingPriority = args[++index];
+        break;
+      case "--routing-top-k":
+        parsed.routingTopK = Number(args[++index]);
+        break;
+      case "--workload-bias":
+        parsed.workloadBias = args[++index];
+        break;
+      case "--no-routing":
+        parsed.routingEnabled = false;
+        break;
       case "--host":
         parsed.host = args[++index];
         break;
@@ -1589,6 +1854,12 @@ function getHelpText() {
     "  --max-length <n>",
     "  --tag <value>             Repeatable",
     "  --pin-constraint <text>   Repeatable",
+    "  --target <provider:model> Repeatable",
+    "  --target-hint <value>     Repeatable",
+    "  --routing-priority <value>",
+    "  --routing-top-k <n>",
+    "  --workload-bias <code_first>",
+    "  --no-routing",
     "  --host <url>",
     "  --store <local|sqlite>",
     "  --storage-dir <path>",
@@ -1606,6 +1877,20 @@ function getHelpText() {
     "  --bypass-optimization"
   ].join("\n");
 }
+function parseTargetCandidate(raw, index) {
+  const [provider, ...modelParts] = raw.split(":");
+  const model = modelParts.join(":").trim();
+  return {
+    provider: provider.trim(),
+    model,
+    label: raw,
+    costRank: index + 1,
+    latencyRank: index + 1
+  };
+}
+function formatTarget(target) {
+  return target.label ?? `${target.provider}:${target.model}`;
+}
 async function readStdin(stdin = process.stdin) {
   if (!stdin || stdin.isTTY) {
     return "";