@vibgrate/cli 2026.801.1 → 2026.803.3

package/dist/{chunk-IRCYOTRG.js → chunk-WGPS3NZQ.js}

@@ -1,13 +1,14 @@
 import { secretEgressRefusal, isSecretPath, secretRefusal, redactText } from './chunk-543E6WDJ.js';
-import { buildAgentMessages, redactSecrets, unifiedDiff, applyEdit } from './chunk-G25STLE7.js';
-import { localGraphBackend } from './chunk-A43I2AJ4.js';
+import { buildAgentMessages, redactSecrets, unifiedDiff, applyEdit } from './chunk-SJFBXCII.js';
+import { localGraphBackend } from './chunk-TM5XQY3L.js';
 import { buildIdentifierTrieFromGraph, KvBlockRegistry, graphDraftCandidates, extractIdentifiers, scanIdentifiersAgainstTrie } from './chunk-5HYCB55P.js';
-import { summarizeCapsule, CAPSULE_RANKING_VERSION, buildCodeContext, buildTaskCapsule, capsuleToCodeContext, recordCliCall, CLI_TOOL_ALIASES } from './chunk-MPP4S4OE.js';
-import { resolve } from './chunk-IPICO2RV.js';
-import { searchSymbols, resolveLib, loadCatalog, readDoc, localPackageDocs, resolveVersion } from './chunk-I6RBAIS3.js';
+import { summarizeCapsule, CAPSULE_RANKING_VERSION, buildCodeContext, buildTaskCapsule, capsuleToCodeContext, recordCliCall, CLI_TOOL_ALIASES } from './chunk-K6Q6HTFD.js';
+import { resolve } from './chunk-VIWZ4QT7.js';
+import { searchSymbols, resolveLib, loadCatalog, readDoc, localPackageDocs, resolveVersion } from './chunk-XK5LKOBF.js';
 import { resolveOne } from './chunk-J7LKEB2R.js';
-import { isLocateOnlyInstruction, extractLiteralNeedles, impactOf } from './chunk-SDG242R4.js';
+import { isLocateOnlyInstruction, extractLiteralNeedles, impactOf } from './chunk-HC44772H.js';
 import { parseSource, langForExtension } from './chunk-UTNNGKKX.js';
+import { analyzeQuestion, loadTopicTags } from './chunk-3Q3ZXJZP.js';
 import { repositoryIdFromRoot } from './chunk-BWPBB45M.js';
 import { hashString, canonicalize } from './chunk-JFGYT6BI.js';
 import { spawnSync } from 'child_process';
@@ -1992,12 +1993,124 @@ function sanitizeAgentDisplayText(text) {
   return t;
 }
 
+// src/code/model-capabilities.ts
+var VISION_YES = [
+  /gpt-4o/,
+  /gpt-4\.1/,
+  /gpt-5/,
+  /^o[34]/,
+  // OpenAI multimodal families
+  /claude/,
+  // Claude 3+ are all vision-capable
+  /gemini/,
+  /gemma-?3/,
+  // Google
+  /llava/,
+  /vision/,
+  /-vl\b/,
+  /-vl-/,
+  /qwen.*vl/,
+  // open vision families / naming markers
+  /pixtral/,
+  /moondream/,
+  /minicpm-v/,
+  /bakllava/,
+  /cogvlm/,
+  /internvl/,
+  /phi-?4-multimodal/
+];
+var VISION_NO = [
+  /deepseek/,
+  // DeepSeek chat/coder/r1 lines are text-only
+  /codestral/,
+  /starcoder/,
+  /codegemma/,
+  /codellama/,
+  /code-?llama/,
+  /qwen.*coder/,
+  /devstral/,
+  /^(mistral|mixtral)(?!.*(pixtral|medium-3|small-3\.1))/,
+  // classic Mistral text models
+  /phi-?[123]\b/,
+  /granite(?!.*vision)/,
+  /command-r(?!.*vision)/,
+  /stablelm/,
+  /tinyllama/,
+  /smollm(?!.*v)/,
+  /^llama-?[12]/,
+  /llama-?3(\.[01])?(?!.*vision)/
+  // Llama ≤3.1 text-only; 3.2-vision matches VISION_YES first
+];
+var TOOLS_YES = [
+  /gpt-4/,
+  /gpt-5/,
+  /^o[134]/,
+  /claude/,
+  /gemini/,
+  /grok/,
+  /deepseek-(chat|v3)/,
+  /kimi/,
+  /qwen-?[23]/,
+  /qwen.*(coder|instruct)/,
+  /llama-?3\.[13]/,
+  /llama-?4/,
+  /mistral-(small|medium|large)/,
+  /devstral/,
+  /command-r/,
+  /glm-?4/,
+  /minimax/,
+  /hermes/
+  // Hermes models are tool-call trained
+];
+var TOOLS_NO = [
+  /phi-?[12]\b/,
+  /tinyllama/,
+  /stablelm/,
+  /smollm/,
+  /starcoder/,
+  /codellama/,
+  /codegemma/,
+  /gemma\b/,
+  /gemma-?[12]\b/,
+  // Gemma 1/2 have no tool support; Gemma 3 is unreliable but attempts it
+  /moondream/,
+  /llava/,
+  // pure-vision chat models
+  /-base\b/,
+  /-base-/
+  // base (non-instruct) checkpoints never tool-call
+];
+function match(rules, id) {
+  return rules.some((r) => r.test(id));
+}
+function lookupModelCapabilities(model, overrides) {
+  const id = (model ?? "").trim().toLowerCase();
+  const result = { nativeTools: null, vision: null };
+  if (!id) return result;
+  if (match(VISION_YES, id)) result.vision = true;
+  else if (match(VISION_NO, id)) result.vision = false;
+  if (match(TOOLS_NO, id)) result.nativeTools = false;
+  else if (match(TOOLS_YES, id)) result.nativeTools = true;
+  if (overrides) {
+    for (const [key, caps] of Object.entries(overrides)) {
+      const k = key.trim().toLowerCase();
+      if (!k) continue;
+      if (id === k || id.includes(k)) {
+        if (caps.nativeTools !== void 0) result.nativeTools = caps.nativeTools;
+        if (caps.vision !== void 0) result.vision = caps.vision;
+      }
+    }
+  }
+  return result;
+}
+
 // src/code/agent.ts
 var DEFAULT_MAX_STEPS = 24;
 var DEFAULT_CONTEXT_BUDGET = 16e3;
 var KEEP_ROUNDS = 8;
 var NUDGE_AT = 3;
 var STOP_AT = 5;
+var EMPTY_REPLY_RETRIES = 2;
 async function runAgent(options) {
   const { root, instruction, providers } = options;
   let graph = options.graph;
@@ -2038,11 +2151,27 @@ async function runAgent(options) {
   const advancedMode = options.advancedMode ?? true;
   const discoveryGate = createDiscoveryGateState();
   const allTools = [...AGENT_TOOLS, ...options.externalTools?.specs ?? []];
-  const built = buildAgentContext(graph, instruction, { ...options, fsImpl, budget });
+  const relevance = await analyzeQuestion(instruction);
+  const topicTags = relevance ? await loadTopicTags(graph, options.root) : null;
+  const built = buildAgentContext(graph, instruction, { ...options, fsImpl, budget, relevance, topicTags });
   const context = built.context;
   let capsule = built.capsule;
   const useLadder = options.verifyLadder ?? !!capsule;
   const messages = buildAgentMessages(context);
+  if (options.images?.length) {
+    const caps = lookupModelCapabilities(providers[0]?.model);
+    if (caps.vision === false) {
+      onEvent({
+        type: "notice",
+        text: `${providers[0]?.model ?? "the selected model"} cannot view images \u2014 attached image(s) are referenced by name/path only. Pick a vision-capable model to have it read the pixels.`
+      });
+    } else {
+      const last = messages[messages.length - 1];
+      if (last?.role === "user") {
+        messages[messages.length - 1] = { ...last, images: options.images };
+      }
+    }
+  }
   if (options.projectRules) {
     messages.push({ role: "user", content: options.projectRules });
   }
@@ -2055,6 +2184,7 @@ async function runAgent(options) {
   const spans = buildSpanIndex(graph);
   const changes = [];
   const repeats = /* @__PURE__ */ new Map();
+  let emptyReplies = 0;
   let recordedSearch = false;
   let commandCount = 0;
   let verifyRounds = verifyConfig?.maxRounds ?? 2;
@@ -2126,6 +2256,7 @@ async function runAgent(options) {
     return "file" in action && action.file ? [action.file] : [];
   };
   const approveAndCheckpoint = async (action) => {
+    if (options.plan) return false;
     const approved = await options.approve(action);
     if (!approved || !options.checkpoint) return approved;
     const files = actionFiles(action);
@@ -2174,6 +2305,9 @@ async function runAgent(options) {
           const nextCapsule = buildTaskCapsule(session.graph, instruction, {
             budget,
             files: options.files,
+            relevance,
+            topicTags,
+            priorInstruction: options.priorInstruction,
             readFile: (rel) => fsImpl.read(rel),
             repositoryId: repositoryIdFromRoot(root),
             provenance: {
@@ -2353,7 +2487,12 @@ Fix the failing tests, then call finish again.${verifyRounds === 0 ? " (last ver
     }
     currentStep = step;
     onEvent({ type: "step", n: step });
-    const compacted = compact(messages, contextBudget, changes);
+    const compacted = options.llmCompaction ? await compactWithModel(
+      messages,
+      contextBudget,
+      changes,
+      (transcript) => summarizeForCompaction(providers, transcript)
+    ) : compact(messages, contextBudget, changes);
     if (compacted.droppedRounds > 0) {
       messages.length = 0;
       messages.push(...compacted.messages);
@@ -2405,8 +2544,21 @@ Fix the failing tests, then call finish again.${verifyRounds === 0 ? " (last ver
       if (isUserFacingProse(displayText)) {
         return finish("finished", displayText, step);
       }
-      return finish("no-tools", displayText || "done", step);
+      emptyReplies++;
+      if (emptyReplies <= EMPTY_REPLY_RETRIES && step < maxSteps) {
+        messages.push({
+          role: "user",
+          content: displayText.trim().length === 0 ? "Your last reply was empty. Either call a tool to keep working, or answer the task in plain Markdown now." : "Your last reply was neither a tool call nor a readable answer. Do not emit raw JSON or templates \u2014 call a tool, or answer the task in plain Markdown."
+        });
+        continue;
+      }
+      return finish(
+        "no-tools",
+        displayText || `The model (${providerInfo.model}) returned no usable output after ${emptyReplies} attempt(s). Try a stronger model, or re-ask with a more specific instruction.`,
+        step
+      );
     }
+    emptyReplies = 0;
     messages.push({ role: "assistant", content: result.text, toolCalls });
     for (const call of toolCalls) {
       onEvent({ type: "tool-call", name: call.name, args: call.arguments });
@@ -2510,15 +2662,90 @@ function compact(messages, budget, changes) {
   };
   return { messages: [...head, note, ...kept], droppedRounds: dropped.length };
 }
+var COMPACTION_PROMPT = [
+  "Summarize the following coding-agent transcript into a compact checkpoint the agent can resume from.",
+  "Cover, as short labelled lines: Objective; Key findings/decisions; Files read or changed (with paths);",
+  "Current state; Next steps. Under 250 words. Plain text only \u2014 no code blocks, no tool syntax."
+].join(" ");
+function renderDroppedRounds(rounds) {
+  const lines = [];
+  for (const round2 of rounds) {
+    for (const m of round2) {
+      if (m.role === "assistant") {
+        if (m.content?.trim()) lines.push(`assistant: ${m.content.trim().slice(0, 400)}`);
+        for (const t of m.toolCalls ?? []) {
+          lines.push(`tool-call: ${t.name}(${JSON.stringify(t.arguments).slice(0, 200)})`);
+        }
+      } else if (m.role === "tool") {
+        lines.push(`tool-result[${m.name}]: ${m.content.split("\n")[0].slice(0, 300)}`);
+      } else if (m.role === "user") {
+        lines.push(`user: ${m.content.trim().slice(0, 300)}`);
+      }
+    }
+  }
+  return lines.join("\n").slice(0, 24e3);
+}
+async function summarizeForCompaction(providers, transcript) {
+  const provider = providers[0];
+  if (!provider) return null;
+  const r = await provider.chat(
+    [
+      { role: "system", content: COMPACTION_PROMPT },
+      { role: "user", content: transcript }
+    ],
+    { temperature: 0, maxTokens: 512, timeoutMs: 6e4 }
+  );
+  return (r.text ?? "").trim() || null;
+}
+function isUsableSummary(text) {
+  const t = (text ?? "").trim();
+  return !!t && !t.startsWith("{") && !t.includes("<tool_call>");
+}
+async function compactWithModel(messages, budget, changes, summarize) {
+  const deterministic = compact(messages, budget, changes);
+  if (deterministic.droppedRounds === 0) return deterministic;
+  const head = messages.slice(0, 3);
+  const body = messages.slice(3);
+  const rounds = [];
+  for (const m of body) {
+    if (m.role === "assistant" || rounds.length === 0) rounds.push([m]);
+    else rounds[rounds.length - 1].push(m);
+  }
+  const dropped = rounds.slice(0, rounds.length - KEEP_ROUNDS);
+  const kept = rounds.slice(-KEEP_ROUNDS).flat();
+  let summary = null;
+  try {
+    summary = await summarize(renderDroppedRounds(dropped));
+  } catch {
+    summary = null;
+  }
+  if (!isUsableSummary(summary)) return deterministic;
+  const filesChanged = [...new Set(changes.map((c) => c.file))];
+  const note = {
+    role: "user",
+    content: `[Earlier steps were compacted to save context. Checkpoint summary:
+${summary.trim().slice(0, 4e3)}
+${filesChanged.length ? `Files changed so far: ${filesChanged.join(", ")}. ` : ""}Continue the task.]`
+  };
+  return { messages: [...head, note, ...kept], droppedRounds: dropped.length };
+}
 function buildAgentContext(graph, instruction, options) {
   const budget = options.budget ?? 3e3;
   const files = options.files;
+  const relevance = options.relevance;
+  const topicTags = options.topicTags;
   if (!options.capsule) {
-    return { context: buildCodeContext(graph, instruction, { budget, files }), capsule: null };
+    return {
+      context: buildCodeContext(graph, instruction, { budget, files, relevance, topicTags, priorInstruction: options.priorInstruction }),
+      capsule: null
+    };
   }
   const capsule = buildTaskCapsule(graph, instruction, {
     budget,
     files,
+    relevance,
+    topicTags,
+    priorInstruction: options.priorInstruction,
     readFile: (rel) => options.fsImpl.read(rel),
     repositoryId: repositoryIdFromRoot(options.root),
     extraPinnedFacts: options.extraPinnedFacts,
@@ -2609,6 +2836,6 @@ function buildSpanIndex(graph) {
   return map;
 }
 
-export { CONTEXT_POLICY_VERSION, DISCOVERY_TOOLS, MUTATION_TOOLS, compact, computeFcsReport, createTrajectoryCollector, renderFcsReportMarkdown, runAgent };
-//# sourceMappingURL=chunk-IRCYOTRG.js.map
-//# sourceMappingURL=chunk-IRCYOTRG.js.map
+export { CONTEXT_POLICY_VERSION, DISCOVERY_TOOLS, MUTATION_TOOLS, compact, compactWithModel, computeFcsReport, createTrajectoryCollector, lookupModelCapabilities, renderFcsReportMarkdown, runAgent };
+//# sourceMappingURL=chunk-WGPS3NZQ.js.map
+//# sourceMappingURL=chunk-WGPS3NZQ.js.map