pi-smart-compact 7.7.0 → 7.9.1

package/dist/index.js

@@ -1,6 +1,6 @@
 // @bun
 // src/constants.ts
-var VERSION = "7.7.0";
+var VERSION = "7.9.1";
 var CHARS_PER_TOKEN = 3.8;
 var COMPACT_SYSTEM_PREFIX = "You are an expert conversation summarizer for a coding agent. " + "Produce structured markdown summaries. " + "Follow output format exactly. " + "Use EXACT names \u2014 never paraphrase code identifiers. " + "Trust deterministic extraction data over intuition.";
 var PROFILES = {
@@ -161,6 +161,11 @@ var SESSION_TYPE_INSTRUCTIONS = {
   review: "Focus on: files read, issues found, recommendations, approval status. Prioritize findings over changes. Read-only tool calls = REVIEW, not implementation.",
   discussion: "Focus on: decisions made, trade-offs discussed, consensus reached. Prioritize rationale over implementation details."
 };
+var LOG_PREFIX = "[smart-compact]";
+var MIN_TOKEN_THRESHOLD = 5000;
+var MAX_EXPLORATION_ROUNDS = 8;
+var CONFIG_KEY = "smartCompact";
+var CONFIG_KEY_ALT = "semanticCompact";
 var EXPLORER_SYSTEM_PROMPT = `You are a conversation analyst. You have deterministic extraction data and can query the raw conversation using tools.
 
 ` + `Your job:
@@ -179,6 +184,19 @@ var EXPLORER_SYSTEM_PROMPT = `You are a conversation analyst. You have determini
 import fs from "fs";
 import path from "path";
 import crypto from "crypto";
+
+// src/utils/logger.ts
+var DEBUG = process.env.DEBUG?.includes("smart-compact") ?? false;
+function warn(msg, err) {
+  const detail = err instanceof Error ? err.message : err ?? "";
+  console.error(LOG_PREFIX + " " + msg + (detail ? ": " + detail : ""));
+}
+function debug(msg, ...args) {
+  if (DEBUG)
+    console.error(LOG_PREFIX + " [debug] " + msg, ...args);
+}
+
+// src/utils/helpers.ts
 var _cfg = null;
 var _cfgMtime = 0;
 function loadConfig() {
@@ -188,7 +206,7 @@ function loadConfig() {
     if (_cfg && stat.mtimeMs === _cfgMtime)
       return _cfg;
     const raw = JSON.parse(fs.readFileSync(p, "utf-8"));
-    const sc = raw.smartCompact ?? raw.semanticCompact ?? {};
+    const sc = raw[CONFIG_KEY] ?? raw[CONFIG_KEY_ALT] ?? {};
     const merged = { ...DEFAULT_CONFIG, ...sc };
     if (sc.profiles)
       merged.profiles = { ...PROFILES, ...sc.profiles };
@@ -197,8 +215,11 @@ function loadConfig() {
     _cfg = merged;
     _cfgMtime = stat.mtimeMs;
     return _cfg;
-  } catch {
-    return { ...DEFAULT_CONFIG, backupDir: path.join(process.env.HOME ?? "/tmp", ".pi/agent/compact-backups") };
+  } catch (e) {
+    warn("loadConfig failed, using defaults", e);
+    const fallback = { ...DEFAULT_CONFIG, backupDir: path.join(process.env.HOME ?? "/tmp", ".pi/agent/compact-backups") };
+    _cfg = fallback;
+    return fallback;
   }
 }
 function backupConversation(convText, sessionId) {
@@ -217,7 +238,8 @@ function backupConversation(convText, sessionId) {
 
 ` + convText);
     return fp;
-  } catch {
+  } catch (e) {
+    warn("backupConversation failed", e);
     return null;
   }
 }
@@ -238,8 +260,23 @@ function smartKeepBoundary(msgs, keepFromIndex) {
   const last = msgs[keepFromIndex - 1];
   const first = msgs[keepFromIndex];
   if (last && first) {
-    const lastText = JSON.stringify(last.message).toLowerCase();
-    const keptText = JSON.stringify(first.message).toLowerCase();
+    const getText = (msg) => {
+      const m = msg;
+      const c = m?.content;
+      if (typeof c === "string")
+        return c;
+      if (Array.isArray(c))
+        return c.map((b) => {
+          if (typeof b === "string")
+            return b;
+          if (typeof b === "object" && b !== null && b.type === "text")
+            return b.text ?? "";
+          return "";
+        }).join("");
+      return "";
+    };
+    const lastText = getText(last.message).toLowerCase();
+    const keptText = getText(first.message).toLowerCase();
     const fileRe = /(?:path|file)=["']([^"']+)["']/g;
     const lastFiles = new Set([...lastText.matchAll(fileRe)].map((m) => m[1].split("/").pop()));
     fileRe.lastIndex = 0;
@@ -249,6 +286,12 @@ function smartKeepBoundary(msgs, keepFromIndex) {
   }
   return keepFromIndex;
 }
+function extractUserNote(args) {
+  const SKIP = new Set(["verbose", "debug", "dry-run", "light", "balanced", "aggressive"]);
+  const tokens = args.trim().split(/\s+/).filter(Boolean);
+  const nonFlags = tokens.filter((t) => !t.includes("/") && !SKIP.has(t.toLowerCase()));
+  return nonFlags.length > 0 ? nonFlags.join(" ") : undefined;
+}
 function createBatches(chunks, maxTokens) {
   const batches = [];
   let batch = [], bt = 0;
@@ -359,6 +402,42 @@ var PROVIDER_MAP = {
     concurrencyLimit: 3,
     cacheStrategy: "anthropic"
   },
+  anthropic: {
+    maxOutputTokens: 8192,
+    supportsTools: true,
+    jsonReliability: "high",
+    instructionFollowing: "high",
+    tokenRatioEstimate: 3.5,
+    concurrencyLimit: 3,
+    cacheStrategy: "anthropic"
+  },
+  openai: {
+    maxOutputTokens: 16384,
+    supportsTools: true,
+    jsonReliability: "high",
+    instructionFollowing: "high",
+    tokenRatioEstimate: 4,
+    concurrencyLimit: 5,
+    cacheStrategy: "openai"
+  },
+  google: {
+    maxOutputTokens: 8192,
+    supportsTools: true,
+    jsonReliability: "high",
+    instructionFollowing: "high",
+    tokenRatioEstimate: 3.8,
+    concurrencyLimit: 3,
+    cacheStrategy: "openai"
+  },
+  deepseek: {
+    maxOutputTokens: 8192,
+    supportsTools: true,
+    jsonReliability: "medium",
+    instructionFollowing: "medium",
+    tokenRatioEstimate: 3.6,
+    concurrencyLimit: 2,
+    cacheStrategy: "none"
+  },
   minimax: {
     maxOutputTokens: 4096,
     supportsTools: "probe",
@@ -377,26 +456,54 @@ var PROVIDER_MAP = {
     concurrencyLimit: 2,
     cacheStrategy: "openai"
   },
-  openai: {
-    maxOutputTokens: 16384,
+  mistral: {
+    maxOutputTokens: 8192,
     supportsTools: true,
     jsonReliability: "high",
     instructionFollowing: "high",
-    tokenRatioEstimate: 4,
-    concurrencyLimit: 5,
+    tokenRatioEstimate: 3.5,
+    concurrencyLimit: 3,
     cacheStrategy: "openai"
-  }
-};
-function getProviderCaps(provider) {
-  return PROVIDER_MAP[provider] ?? {
+  },
+  xai: {
     maxOutputTokens: 8192,
-    supportsTools: "probe",
+    supportsTools: true,
     jsonReliability: "medium",
-    instructionFollowing: "medium",
+    instructionFollowing: "high",
     tokenRatioEstimate: 3.8,
-    concurrencyLimit: 2,
-    cacheStrategy: "none"
-  };
+    concurrencyLimit: 3,
+    cacheStrategy: "openai"
+  }
+};
+var PROVIDER_ALIASES = [
+  { pattern: /anthropic/i, provider: "anthropic" },
+  { pattern: /zai/i, provider: "zai-anthropic" },
+  { pattern: /openai/i, provider: "openai" },
+  { pattern: /gpt/i, provider: "openai" },
+  { pattern: /google|gemini/i, provider: "google" },
+  { pattern: /deepseek/i, provider: "deepseek" },
+  { pattern: /minimax/i, provider: "minimax" },
+  { pattern: /xiaomi/i, provider: "xiaomi-token-plan" },
+  { pattern: /mistral/i, provider: "mistral" },
+  { pattern: /xai|grok/i, provider: "xai" }
+];
+var DEFAULT_CAPS = {
+  maxOutputTokens: 8192,
+  supportsTools: "probe",
+  jsonReliability: "medium",
+  instructionFollowing: "medium",
+  tokenRatioEstimate: 3.8,
+  concurrencyLimit: 2,
+  cacheStrategy: "none"
+};
+function getProviderCaps(provider) {
+  if (PROVIDER_MAP[provider])
+    return PROVIDER_MAP[provider];
+  for (const { pattern, provider: key } of PROVIDER_ALIASES) {
+    if (pattern.test(provider))
+      return PROVIDER_MAP[key] ?? DEFAULT_CAPS;
+  }
+  return DEFAULT_CAPS;
 }
 var _calibrationFactors = new Map;
 function getCalibrationFactor(provider) {
@@ -435,12 +542,13 @@ function getCompactSessionId() {
 function resetCompactSessionId() {
   _compactSessionId = null;
 }
-function cacheOpts(opts) {
-  const retention = opts.cacheRetention ?? "short";
+function cacheOpts(opts, provider) {
+  const strategy = provider ? getProviderCaps(provider).cacheStrategy : "none";
+  const retention = strategy === "none" ? "none" : opts.cacheRetention ?? "short";
   if (retention === "none") {
     return { ...opts, cacheRetention: "none" };
   }
-  return { ...opts, sessionId: getCompactSessionId(), cacheRetention: "short" };
+  return { ...opts, sessionId: getCompactSessionId(), cacheRetention: retention };
 }
 var _metrics = [];
 function resetMetrics() {
@@ -486,9 +594,13 @@ async function trackedComplete(phase, model, reqBody, opts) {
       latencyMs: latency,
       success: true
     });
-    if (inputT > 0 && "messages" in reqBody) {
-      const rawText = JSON.stringify(reqBody.messages);
-      calibrateFromResponse(estimateTokens(rawText), inputT, model.provider);
+    try {
+      if (inputT > 0 && "messages" in reqBody) {
+        const rawText = JSON.stringify(reqBody.messages);
+        calibrateFromResponse(estimateTokens(rawText), inputT, model.provider);
+      }
+    } catch (e) {
+      debug("token calibration failed", e);
     }
     return resp;
   } catch (err) {
@@ -518,7 +630,9 @@ function saveCachedExtraction(sessionId, extraction, msgCount) {
       timestamp: Date.now()
     };
     fs2.writeFileSync(getCachePath(sessionId), JSON.stringify(cached));
-  } catch {}
+  } catch (e) {
+    warn("saveCachedExtraction failed", e);
+  }
 }
 function loadCachedExtraction(sessionId) {
   try {
@@ -529,13 +643,14 @@ function loadCachedExtraction(sessionId) {
     if (Date.now() - cached.timestamp > 3600000)
       return null;
     return cached;
-  } catch {
+  } catch (e) {
+    warn("loadCachedExtraction failed", e);
     return null;
   }
 }
 function mergeExtractions(base, delta, baseMsgCount) {
   return {
-    modifiedFiles: [...base.modifiedFiles, ...delta.modifiedFiles],
+    modifiedFiles: [...new Map([...base.modifiedFiles, ...delta.modifiedFiles].map((f) => [f.path, f])).values()],
     readFiles: [...new Set([...base.readFiles, ...delta.readFiles])],
     deletedFiles: [...new Set([...base.deletedFiles, ...delta.deletedFiles])],
     errors: [...base.errors, ...delta.errors],
@@ -557,13 +672,15 @@ function appendMetricsLog(sessionId) {
     const entry = { ts: new Date().toISOString(), sessionId, ...getMetricsSummary() };
     fs2.appendFileSync(logPath, JSON.stringify(entry) + `
 `);
-  } catch {}
+  } catch (e) {
+    warn("appendMetricsLog failed", e);
+  }
 }
 
 // src/utils/extraction.ts
 import path3 from "path";
 
-// src/types.ts
+// src/utils/type-guards.ts
 function isToolCallBlock(c) {
   return typeof c === "object" && c !== null && c.type === "toolCall" && typeof c.name === "string";
 }
@@ -607,8 +724,8 @@ function buildToolCallIndex(msgs) {
   }
   return idx;
 }
-function trackFileOps(msgs) {
-  const tcIdx = buildToolCallIndex(msgs);
+function trackFileOps(msgs, _tcIdx) {
+  const tcIdx = _tcIdx ?? buildToolCallIndex(msgs);
   const modMap = new Map;
   const readSet = new Set;
   const delSet = new Set;
@@ -642,8 +759,8 @@ function trackFileOps(msgs) {
     deleted: [...delSet]
   };
 }
-function catalogErrors(msgs) {
-  const tcIdx = buildToolCallIndex(msgs);
+function catalogErrors(msgs, _tcIdx) {
+  const tcIdx = _tcIdx ?? buildToolCallIndex(msgs);
   const errors = [];
   for (let i = 0;i < msgs.length; i++) {
     const m = msgs[i];
@@ -665,7 +782,8 @@ function catalogErrors(msgs) {
   for (const err of errors) {
     for (let j = err.index + 1;j < Math.min(msgs.length, err.index + 6); j++) {
       if (msgs[j]?.role === "assistant") {
-        const blocks = Array.isArray(msgs[j]?.content) ? msgs[j].content : [];
+        const rawBlocks = msgs[j]?.content;
+        const blocks = Array.isArray(rawBlocks) ? rawBlocks : [];
         for (const b of blocks) {
           if (isToolCallBlock(b) && b.name === err.tool) {
             err.retryAttempted = true;
@@ -685,8 +803,8 @@ function catalogErrors(msgs) {
   }
   return errors;
 }
-function extractDecisions(msgs) {
-  const tcIdx = buildToolCallIndex(msgs);
+function extractDecisions(msgs, _tcIdx) {
+  const tcIdx = _tcIdx ?? buildToolCallIndex(msgs);
   const decisions = [];
   for (const [id, tc] of tcIdx) {
     if (tc.name !== "ask_user")
@@ -737,10 +855,10 @@ function mineConstraints(msgs) {
   }
   return constraints;
 }
-function segmentTopicsHeuristic(msgs, pc, maxSegs = 20) {
+function segmentTopicsHeuristic(msgs, pc, maxSegs = 20, _tcIdx) {
   const topics = [];
   let startIdx = 0, tokenAcc = 0, lastFile = null, errAcc = 0;
-  const tcIdx = buildToolCallIndex(msgs);
+  const tcIdx = _tcIdx ?? buildToolCallIndex(msgs);
   for (let i = 0;i < msgs.length; i++) {
     const m = msgs[i];
     const txt = extractText(m.content);
@@ -896,11 +1014,12 @@ function extractOpenLoops(msgs, extraction) {
   return loops;
 }
 function extractStructured(msgs, pc) {
-  const { modified, read, deleted } = trackFileOps(msgs);
-  const errors = catalogErrors(msgs);
-  const decisions = extractDecisions(msgs);
+  const tcIdx = buildToolCallIndex(msgs);
+  const { modified, read, deleted } = trackFileOps(msgs, tcIdx);
+  const errors = catalogErrors(msgs, tcIdx);
+  const decisions = extractDecisions(msgs, tcIdx);
   const constraints = mineConstraints(msgs);
-  const topics = segmentTopicsHeuristic(msgs, pc);
+  const topics = segmentTopicsHeuristic(msgs, pc, 20, tcIdx);
   const timeline = buildTimeline(msgs, errors);
   const mainGoal = extractMainGoal(msgs);
   const lastUserMessages = msgs.filter((m) => m.role === "user").slice(-5).map((m) => extractText(m.content));
@@ -933,7 +1052,9 @@ function saveCompactionState(projectId, state) {
     if (!fs3.existsSync(STATE_DIR))
       fs3.mkdirSync(STATE_DIR, { recursive: true });
     fs3.writeFileSync(getStatePath(projectId), JSON.stringify(state, null, 2));
-  } catch {}
+  } catch (e) {
+    warn("saveCompactionState failed", e);
+  }
 }
 function loadCompactionState(projectId) {
   try {
@@ -941,9 +1062,22 @@ function loadCompactionState(projectId) {
     if (!fs3.existsSync(fp))
       return null;
     const data = JSON.parse(fs3.readFileSync(fp, "utf8"));
-    if (data.compactionVersion && Date.now() - 0 > 7 * 24 * 60 * 60 * 1000) {}
+    if (data.compactionVersion) {
+      let updatedAt = data.updatedAt;
+      if (!updatedAt) {
+        try {
+          updatedAt = fs3.statSync(fp).mtimeMs;
+        } catch (e) {
+          debug("statSync failed for state file", e);
+          updatedAt = 0;
+        }
+      }
+      if (Date.now() - updatedAt > 7 * 24 * 60 * 60 * 1000)
+        return null;
+    }
     return data;
-  } catch {
+  } catch (e) {
+    warn("loadCompactionState failed", e);
     return null;
   }
 }
@@ -991,7 +1125,8 @@ function buildCompactionState(extraction, openLoops, report, nextActions, critic
     nextActions,
     criticalContext,
     sessionType: report?.sessionType ?? "implementation",
-    compactionVersion: VERSION
+    compactionVersion: VERSION,
+    updatedAt: Date.now()
   };
 }
 function injectOpenLoopsSection(summary, openLoops) {
@@ -1178,7 +1313,8 @@ function pruneRedundant(msgs) {
   for (let idx = 0;idx < msgs.length; idx++) {
     if (msgs[idx].role !== "assistant")
       continue;
-    const blocks = Array.isArray(msgs[idx].content) ? msgs[idx].content : [];
+    const rawBlocks = msgs[idx].content;
+    const blocks = Array.isArray(rawBlocks) ? rawBlocks : [];
     const hasToolCall = blocks.some((b) => isToolCallBlock(b));
     if (hasToolCall)
       continue;
@@ -1220,6 +1356,7 @@ function pruneRedundant(msgs) {
 // src/utils/fingerprint.ts
 import fs4 from "fs";
 import path5 from "path";
+import crypto3 from "crypto";
 var FINGERPRINT_DIR = path5.join(process.env.HOME ?? "/tmp", ".pi", "agent", ".cache", "smart-compact", "projects");
 var LANG_MAP = {
   ".ts": "typescript",
@@ -1269,11 +1406,8 @@ function deriveProjectId(extraction) {
     roots.set(root, (roots.get(root) ?? 0) + 1);
   }
   const topRoot = [...roots.entries()].sort((a, b) => b[1] - a[1])[0]?.[0] ?? "unknown";
-  let hash = 0;
-  for (let i = 0;i < topRoot.length; i++) {
-    hash = (hash << 5) - hash + topRoot.charCodeAt(i) | 0;
-  }
-  return "proj-" + Math.abs(hash).toString(36);
+  const hash = crypto3.createHash("sha256").update(topRoot).digest("hex").slice(0, 12);
+  return "proj-" + hash;
 }
 function detectLanguage(extraction) {
   const extCounts = new Map;
@@ -1321,7 +1455,8 @@ function loadProjectFingerprint(projectId) {
     if (Date.now() - data.updatedAt > 30 * 24 * 60 * 60 * 1000)
       return null;
     return data;
-  } catch {
+  } catch (e) {
+    warn("loadProjectFingerprint failed", e);
     return null;
   }
 }
@@ -1345,7 +1480,9 @@ function saveProjectFingerprint(projectId, extraction) {
       updatedAt: Date.now()
     };
     fs4.writeFileSync(getFingerprintPath(projectId), JSON.stringify(fingerprint, null, 2));
-  } catch {}
+  } catch (e) {
+    warn("saveProjectFingerprint failed", e);
+  }
 }
 function buildProjectContext(fingerprint) {
   if (!fingerprint)
@@ -1459,7 +1596,9 @@ function logDamageReport(sessionId, report, details) {
     };
     fs5.appendFileSync(logPath, JSON.stringify(entry) + `
 `);
-  } catch {}
+  } catch (e) {
+    warn("logDamageReport failed", e);
+  }
 }
 
 // src/phases/explore.ts
@@ -1525,7 +1664,13 @@ function executeExplorationTool(call, llmMessages) {
     }
     case "search_conversation": {
       const q = (args.query ?? "").toLowerCase();
-      return JSON.stringify(llmMessages.filter((m) => JSON.stringify(m).toLowerCase().includes(q)).slice(0, 10).map((m) => ({
+      return JSON.stringify(llmMessages.filter((m) => {
+        const text = extractText(m?.content).toLowerCase();
+        if (text.includes(q))
+          return true;
+        const tcs = filterToolCalls(m?.content);
+        return tcs.some((tc) => JSON.stringify(tc.arguments).toLowerCase().includes(q));
+      }).slice(0, 10).map((m) => ({
         idx: llmMessages.indexOf(m),
         role: m?.role,
         preview: extractText(m?.content).slice(0, 150)
@@ -1589,11 +1734,15 @@ function parseExplorationReport(text, llmMessages) {
   let rawJson = json.slice(s, e + 1);
   try {
     return buildExplorationReportFromParsed(JSON.parse(rawJson), llmMessages);
-  } catch {}
+  } catch (e2) {
+    debug("JSON parse attempt 1 failed", e2);
+  }
   const cleaned = rawJson.replace(/,\s*([}\]])/g, "$1").replace(/'/g, '"').replace(/\/\/.*$/gm, "").replace(/\/\*[\s\S]*?\*\//g, "");
   try {
     return buildExplorationReportFromParsed(JSON.parse(cleaned), llmMessages);
-  } catch {}
+  } catch (e2) {
+    debug("JSON parse attempt 2 (cleaned) failed", e2);
+  }
   const boundaryMatch = rawJson.match(/"boundaries"\s*:\s*\[([\s\S]*?)\]/);
   if (boundaryMatch) {
     try {
@@ -1604,7 +1753,9 @@ function parseExplorationReport(text, llmMessages) {
         priority: ["critical", "high", "normal", "low"].includes(b.priority) ? b.priority : "normal",
         confidence: Math.min(1, Math.max(0, b.confidence ?? 0.5))
       })) };
-    } catch {}
+    } catch (e2) {
+      debug("Boundary JSON parse failed", e2);
+    }
   }
   return fallbackExplorationReport(llmMessages);
 }
@@ -1641,7 +1792,7 @@ function fallbackExplorationReport(llmMessages) {
     keyDecisions: []
   };
 }
-async function exploreConversation(llmMessages, extraction, model, auth, prevSummary, userNote, signal, maxRounds = 8, notify) {
+async function exploreConversation(llmMessages, extraction, model, auth, prevSummary, userNote, signal, maxRounds = MAX_EXPLORATION_ROUNDS, notify) {
   const extractionContext = [
     "## Deterministic Extraction (verified facts)",
     "Message count: " + extraction.messageCount,
@@ -1679,13 +1830,13 @@ async function exploreConversation(llmMessages, extraction, model, auth, prevSum
         if (retried.boundaries.length)
           return { report: retried, rounds: 1, toolSupported: false };
       }
-      return { report: report2, rounds: 0, toolSupported: false };
+      return { report: report2, rounds: 1, toolSupported: false };
     }
     const probeResp = await trackedComplete("explore", model, {
       systemPrompt: COMPACT_SYSTEM_PREFIX,
-      messages: [{ role: "user", content: [{ type: "text", text: userContent }] }],
+      messages: [{ role: "user", content: [{ type: "text", text: userContent }], timestamp: Date.now() }],
       tools: EXPLORATION_TOOLS
-    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, signal }));
+    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, signal }, model.provider));
     const toolCalls = probeResp.content.filter((c) => c.type === "toolCall");
     if (toolCalls.length > 0) {
       supportsTools = true;
@@ -1709,9 +1860,9 @@ async function exploreConversation(llmMessages, extraction, model, auth, prevSum
 ` + EXPLORER_SYSTEM_PROMPT,
             messages,
             tools: EXPLORATION_TOOLS
-          }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, signal }));
+          }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, signal }, model.provider));
         } catch (err) {
-          console.error("[smart-compact] Explore loop error:", err instanceof Error ? err.message : err);
+          warn("Explore loop error", err);
           break;
         }
         const nextToolCalls = response.content.filter((c) => c.type === "toolCall");
@@ -1749,8 +1900,8 @@ async function exploreConversation(llmMessages, extraction, model, auth, prevSum
       }
       return { report: report2, rounds: 1, toolSupported: true };
     }
-  } catch {
-    console.error("[smart-compact] Tool calling probe failed for " + cacheKey);
+  } catch (e) {
+    warn("Tool calling probe failed for " + cacheKey, e);
     _toolSupportCache.set(cacheKey, { result: false, timestamp: Date.now() });
     if (notify)
       notify("Tool calling not supported, using direct exploration", "warning");
@@ -1761,7 +1912,7 @@ async function exploreConversation(llmMessages, extraction, model, auth, prevSum
     if (retried.boundaries.length)
       return { report: retried, rounds: 1, toolSupported: false };
   }
-  return { report, rounds: 0, toolSupported: supportsTools };
+  return { report, rounds: 1, toolSupported: supportsTools };
 }
 async function explorationRetry(model, auth, llmMessages, extraction, prevSummary, userNote, signal) {
   const last5 = llmMessages.slice(-5).map((m) => "[" + m?.role + "] " + extractText(m?.content).slice(0, 150)).join(`
@@ -1780,11 +1931,12 @@ User steering: ` + userNote : "");
   try {
     const resp = await trackedComplete("explore-retry", model, {
       systemPrompt: COMPACT_SYSTEM_PREFIX,
-      messages: [{ role: "user", content: [{ type: "text", text: retryPrompt }] }]
-    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: 4096, signal }));
+      messages: [{ role: "user", content: [{ type: "text", text: retryPrompt }], timestamp: Date.now() }]
+    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: Math.min(4096, getProviderCaps(model.provider).maxOutputTokens), signal }, model.provider));
     const text = resp.content.filter((c) => c.type === "text").map((c) => c.text).join("").trim();
     return parseExplorationReport(text, llmMessages);
-  } catch {
+  } catch (e) {
+    debug("explorationRetry failed", e);
     return fallbackExplorationReport(llmMessages);
   }
 }
@@ -1817,17 +1969,21 @@ Output ONLY JSON: {"mainGoal":"...","sessionType":"implementation|review|debuggi
   try {
     const resp = await trackedComplete("explore-direct", model, {
       systemPrompt: COMPACT_SYSTEM_PREFIX,
-      messages: [{ role: "user", content: [{ type: "text", text: prompt }] }]
-    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: 4096, signal }));
+      messages: [{ role: "user", content: [{ type: "text", text: prompt }], timestamp: Date.now() }]
+    }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: Math.min(4096, getProviderCaps(model.provider).maxOutputTokens), signal }, model.provider));
     const text = resp.content.filter((c) => c.type === "text").map((c) => c.text).join(`
 `).trim();
     return parseExplorationReport(text, llmMessages);
-  } catch {
+  } catch (e) {
+    debug("directExploration failed", e);
     return fallbackExplorationReport(llmMessages);
   }
 }
 
 // src/phases/synthesize.ts
+function estimateChunkTokens(msgs) {
+  return estimateTokens(msgs.map((m) => extractText(m.content)).join(""));
+}
 function chunkLlmMessages(msgs, boundaries, pc) {
   if (!msgs.length)
     return [];
@@ -1835,7 +1991,7 @@ function chunkLlmMessages(msgs, boundaries, pc) {
     return [{
       startIndex: 0,
       endIndex: msgs.length - 1,
-      tokenEstimate: estimateTokens(JSON.stringify(msgs)),
+      tokenEstimate: estimateChunkTokens(msgs),
       topic: "Full conversation",
       priority: "normal",
       messages: msgs
@@ -1851,7 +2007,7 @@ function chunkLlmMessages(msgs, boundaries, pc) {
       chunks.push({
         startIndex: start,
         endIndex: end - 1,
-        tokenEstimate: estimateTokens(JSON.stringify(slice)),
+        tokenEstimate: estimateChunkTokens(slice),
         topic: bp.topic || "Segment " + (chunks.length + 1),
         priority: bp.priority,
         messages: slice
@@ -1865,7 +2021,7 @@ function chunkLlmMessages(msgs, boundaries, pc) {
     chunks.push({
       startIndex: start,
       endIndex: msgs.length - 1,
-      tokenEstimate: estimateTokens(JSON.stringify(slice)),
+      tokenEstimate: estimateChunkTokens(slice),
       topic: lastTopic,
       priority: "normal",
       messages: slice
@@ -1897,10 +2053,10 @@ Session-specific instructions:
   const resp = await trackedComplete("single-pass", model, {
     systemPrompt: COMPACT_SYSTEM_PREFIX,
     messages: [
-      { role: "user", content: [{ type: "text", text: adaptedPrefix }] },
-      { role: "user", content: [{ type: "text", text: dynamicSuffix }] }
+      { role: "user", content: [{ type: "text", text: adaptedPrefix }], timestamp: Date.now() },
+      { role: "user", content: [{ type: "text", text: dynamicSuffix }], timestamp: Date.now() }
     ]
-  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: 8192, signal }));
+  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: getProviderCaps(model.provider).maxOutputTokens, signal }, model.provider));
   const summary = resp.content.filter((c) => c.type === "text").map((c) => c.text).join(`
 `).trim();
   if (!summary.startsWith("##"))
@@ -1928,10 +2084,10 @@ async function summarizeBatch(batch, extraction, model, auth, signal) {
   const resp = await trackedComplete("batch", model, {
     systemPrompt: COMPACT_SYSTEM_PREFIX,
     messages: [
-      { role: "user", content: [{ type: "text", text: BATCH_PROMPT_PREFIX }] },
-      { role: "user", content: [{ type: "text", text: dynamicSuffix }] }
+      { role: "user", content: [{ type: "text", text: BATCH_PROMPT_PREFIX }], timestamp: Date.now() },
+      { role: "user", content: [{ type: "text", text: dynamicSuffix }], timestamp: Date.now() }
     ]
-  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: 4096, signal }));
+  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: Math.min(4096, getProviderCaps(model.provider).maxOutputTokens), signal }, model.provider));
   const output = resp.content.filter((c) => c.type === "text").map((c) => c.text).join(`
 `);
   const sections = output.split(/^### /m).filter((s) => s.trim());
@@ -1969,10 +2125,10 @@ async function assembleLLM(summaries, extraction, report, model, auth, budget, p
   const resp = await trackedComplete("assemble", model, {
     systemPrompt: COMPACT_SYSTEM_PREFIX,
     messages: [
-      { role: "user", content: [{ type: "text", text: ASSEMBLY_PROMPT_PREFIX }] },
-      { role: "user", content: [{ type: "text", text: dynamicSuffix }] }
+      { role: "user", content: [{ type: "text", text: ASSEMBLY_PROMPT_PREFIX }], timestamp: Date.now() },
+      { role: "user", content: [{ type: "text", text: dynamicSuffix }], timestamp: Date.now() }
     ]
-  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: Math.min(budget, 8192), signal }));
+  }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: Math.min(budget, getProviderCaps(model.provider).maxOutputTokens), signal }, model.provider));
   return resp.content.filter((c) => c.type === "text").map((c) => c.text).join(`
 `).trim();
 }
@@ -2120,10 +2276,14 @@ function patchDeterministic(summary, gaps, extraction) {
   const constraintGaps = gaps.filter((g) => g.startsWith("Missing constraint:"));
   const decisionGaps = gaps.filter((g) => g.startsWith("Missing decision:"));
   const otherGaps = gaps.filter((g) => !g.startsWith("Missing modified file:") && !g.startsWith("Missing error:") && !g.startsWith("Missing constraint:") && !g.startsWith("Missing decision:") && !g.startsWith("Potentially fabricated") && !g.startsWith("Inconsistency"));
+  const findSectionInsert = (header) => {
+    const re = new RegExp(header.replace(/[.*+?^${}()|[\]\\]/g, "\\$&") + "\\s*\\n", "i");
+    const m = patched.match(re);
+    return m?.index != null ? m.index + m[0].length : null;
+  };
   if (fileGaps.length > 0) {
-    const filesSection = patched.match(/## Files Modified\n/);
-    if (filesSection) {
-      const insertPos = filesSection.index + filesSection[0].length;
+    const insertPos = findSectionInsert("## Files Modified");
+    if (insertPos != null) {
       const entries = fileGaps.map((g) => "- " + g.replace("Missing modified file: ", "")).join(`
 `) + `
 `;
@@ -2131,9 +2291,8 @@ function patchDeterministic(summary, gaps, extraction) {
     }
   }
   if (errorGaps.length > 0) {
-    const ctxSection = patched.match(/## Critical Context\n/);
-    if (ctxSection) {
-      const insertPos = ctxSection.index + ctxSection[0].length;
+    const insertPos = findSectionInsert("## Critical Context");
+    if (insertPos != null) {
       const entries = errorGaps.map((g) => "- " + g).join(`
 `) + `
 `;
@@ -2141,9 +2300,8 @@ function patchDeterministic(summary, gaps, extraction) {
     }
   }
   if (constraintGaps.length > 0) {
-    const constrSection = patched.match(/## Constraints & Preferences\n/);
-    if (constrSection) {
-      const insertPos = constrSection.index + constrSection[0].length;
+    const insertPos = findSectionInsert("## Constraints & Preferences");
+    if (insertPos != null) {
       const entries = constraintGaps.map((g) => "- " + g).join(`
 `) + `
 `;
@@ -2151,9 +2309,8 @@ function patchDeterministic(summary, gaps, extraction) {
     }
   }
   if (decisionGaps.length > 0) {
-    const decSection = patched.match(/## Key Decisions\n/);
-    if (decSection) {
-      const insertPos = decSection.index + decSection[0].length;
+    const insertPos = findSectionInsert("## Key Decisions");
+    if (insertPos != null) {
       const entries = decisionGaps.map((g) => "- **" + g.replace("Missing decision: ", "") + "**").join(`
 `) + `
 `;
@@ -2182,12 +2339,13 @@ Return the COMPLETE updated summary with missing items integrated. Keep the same
   try {
     const resp = await trackedComplete("patch", model, {
       systemPrompt: COMPACT_SYSTEM_PREFIX,
-      messages: [{ role: "user", content: [{ type: "text", text: patchPrompt }] }]
+      messages: [{ role: "user", content: [{ type: "text", text: patchPrompt }], timestamp: Date.now() }]
     }, cacheOpts({ apiKey: auth.apiKey, headers: auth.headers, maxTokens: 8192, signal }));
     const patched = resp.content.filter((c) => c.type === "text").map((c) => c.text).join(`
 `).trim();
     return patched.startsWith("##") ? patched : summary;
-  } catch {
+  } catch (e) {
+    debug("patchSummary LLM failed", e);
     return summary;
   }
 }
@@ -2201,7 +2359,8 @@ function renderContextBar(theme, pct, tokens, barLen = 24) {
   const filled = Math.min(barLen, Math.round(clamped / 100 * barLen));
   const bar = "\u2588".repeat(filled) + "\u2591".repeat(barLen - filled);
   const color = clamped > 80 ? "error" : clamped > 50 ? "warning" : "success";
-  return theme.fg("text", "  Context: ") + theme.fg(color, bar) + theme.fg("text", " " + clamped + "%") + theme.fg("dim", " (" + (tokens ?? 0).toLocaleString() + "t)");
+  const fg = theme.fg;
+  return fg("text", "  Context: ") + fg(color, bar) + fg("text", " " + clamped + "%") + fg("dim", " (" + (tokens ?? 0).toLocaleString() + "t)");
 }
 function renderTokenBar(theme, before, after, label, barLen = 30) {
   const ratio = before > 0 ? after / before : 0;
@@ -2211,7 +2370,6 @@ function renderTokenBar(theme, before, after, label, barLen = 30) {
   const savedColor = savedPct >= 50 ? "success" : savedPct >= 25 ? "warning" : "error";
   return theme.fg("text", "  " + label + ": ") + theme.fg(savedColor, bar) + theme.fg("text", " " + (after ?? 0).toLocaleString() + "t") + theme.fg(savedColor, " (saved " + savedPct + "%)");
 }
-var _toolSupportCache2 = new Map;
 async function selectModel(ctx, opts) {
   const available = ctx.modelRegistry.getAvailable();
   const options = available.map((m) => ({
@@ -2243,7 +2401,7 @@ async function selectModel(ctx, opts) {
       scrollInfo: (t) => theme.fg("dim", t),
       noMatch: (t) => theme.fg("warning", t)
     });
-    sel.selectedIndex = opts.defaultModelIndex;
+    sel.setSelectedIndex(opts.defaultModelIndex);
     sel.onSelect = (item) => done(item.value);
     sel.onCancel = () => done(null);
     c.addChild(sel);
@@ -2292,7 +2450,7 @@ async function selectProfile(ctx, selectedModel, opts) {
       scrollInfo: (t) => theme.fg("dim", t),
       noMatch: (t) => theme.fg("warning", t)
     });
-    sel.selectedIndex = 1;
+    sel.setSelectedIndex(1);
     sel.onSelect = (item) => done(item.value);
     sel.onCancel = () => done(null);
     c.addChild(sel);
@@ -2317,8 +2475,7 @@ function showProgressOverlay(ctx, state) {
   const progress = Math.round(state.phase / 4 * 100);
   const name = phaseNames[state.phase - 1] ?? "?";
   const detail = state.detail ? " (" + state.detail + ")" : "";
-  const type = state.phase >= 4 ? "success" : "info";
-  ctx.ui.notify("EESV [" + progress + "%] Phase " + state.phase + "/4: " + name + detail, type);
+  ctx.ui.notify("EESV [" + progress + "%] Phase " + state.phase + "/4: " + name + detail, state.phase >= 4 ? "info" : "info");
 }
 async function showResultScreen(ctx, details, extraction) {
   await ctx.ui.custom((tui, theme, _kb, done) => {
@@ -2332,7 +2489,8 @@ async function showResultScreen(ctx, details, extraction) {
     c.addChild(new Text("", 0, 0));
     const methodColors = { eesv: "accent", "single-pass": "success", heuristic: "warning" };
     const methodColor = methodColors[details.method] ?? "text";
-    c.addChild(new Text(theme.fg("text", "  Method: ") + theme.fg(methodColor, details.method.toUpperCase()) + theme.fg("dim", " \u2022 " + details.llmCalls + " LLM call(s) \u2022 Profile: " + details.profile), 0, 0));
+    const fg = theme.fg;
+    c.addChild(new Text(theme.fg("text", "  Method: ") + fg(methodColor, details.method.toUpperCase()) + theme.fg("dim", " \u2022 " + details.llmCalls + " LLM call(s) \u2022 Profile: " + details.profile), 0, 0));
     if (details.model) {
       c.addChild(new Text(theme.fg("dim", "  Model: " + details.model), 0, 0));
     }
@@ -2435,7 +2593,8 @@ async function showCompactUI(ctx, opts) {
 }
 
 // src/core.ts
-async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dryRun, pendingRef, isRunning, autoTriggered, userNote, skipCompact) {
+async function runSmartCompact(opts) {
+  const { ctx, summaryModel, segModel, profile, verbose = false, dryRun = false, pendingRef, isRunning, autoTriggered = false, userNote, skipCompact } = opts;
   if (isRunning.value)
     return;
   isRunning.value = true;
@@ -2459,16 +2618,18 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         ctx.ui.notify("Auth failed", "error");
       return;
     }
+    const apiKey = auth.apiKey;
+    const apiHeaders = auth.headers;
     const usage = ctx.getContextUsage();
     const totalTokens = usage?.tokens ?? 0;
-    if (!totalTokens || totalTokens < 5000) {
+    if (!totalTokens || totalTokens < MIN_TOKEN_THRESHOLD) {
       isRunning.value = false;
       if (!autoTriggered)
         ctx.ui.notify("Context OK or unknown", "info");
       return;
     }
     const notify = (msg, type = "info") => {
-      ctx.ui.notify(msg, type);
+      ctx.ui.notify(msg, type === "success" ? "info" : type);
     };
     const ctrl = new AbortController;
     const signal = ctrl.signal;
@@ -2498,7 +2659,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
       isRunning.value = false;
       return;
     }
-    const firstKeptId = msgs[keepFrom]?.id ?? msgs[msgs.length - 1]?.id ?? "";
+    const firstKeptId = msgs[keepFrom]?.id ?? msgs[msgs.length - 1]?.id;
     if (!autoTriggered) {
       showProgressOverlay(ctx, { phase: 1, phaseName: "Extract", detail: "Preparing...", model: modelLabel, profile });
     }
@@ -2542,7 +2703,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
       if (!autoTriggered)
         showProgressOverlay(ctx, { phase: 2, phaseName: "Explore", detail: "Single-pass (" + convTokens.toLocaleString() + "t)", model: modelLabel, profile, extraction });
       try {
-        const r = await singlePassCompact(convText, extraction, null, prevContext + projectCtx, summaryModel, { apiKey: auth.apiKey, headers: auth.headers }, signal);
+        const r = await singlePassCompact(convText, extraction, null, prevContext + projectCtx, summaryModel, { apiKey, headers: apiHeaders }, signal);
         finalSummary = r.summary;
         method = "single-pass";
         llmCalls = r.llmCalls;
@@ -2558,7 +2719,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         if (!autoTriggered)
           showProgressOverlay(ctx, { phase: 2, phaseName: "Explore", detail: "Exploring...", model: modelLabel, profile, extraction });
         try {
-          const expResult = await exploreConversation(llmMessages, extraction, segModel, { apiKey: segAuth.apiKey, headers: segAuth.headers }, prevContext || undefined, userNote, signal, 8, notify);
+          const expResult = await exploreConversation(llmMessages, extraction, segModel, { apiKey: segAuth.apiKey, headers: segAuth.headers }, prevContext || undefined, userNote, signal, MAX_EXPLORATION_ROUNDS, notify);
           explorationReport = expResult.report;
           explorationRounds = expResult.rounds;
           notify("Phase 2 Explore: " + expResult.rounds + " rounds, " + explorationReport.boundaries.length + " boundaries" + (expResult.toolSupported ? "" : " (no tool support)"), "info");
@@ -2607,7 +2768,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
       const concurrency = caps.concurrencyLimit;
       if (totalBatches <= 1) {
         try {
-          summaries.push(...await summarizeBatch(batches[0], extraction, summaryModel, { apiKey: auth.apiKey, headers: auth.headers }, signal));
+          summaries.push(...await summarizeBatch(batches[0], extraction, summaryModel, { apiKey, headers: apiHeaders }, signal));
         } catch (err) {
           summaries.push(...batches[0].map((ch) => ({
             topic: ch.topic,
@@ -2630,7 +2791,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
           const wavePromises = waveBatches.map(async (batch, i) => {
             const idx = wave + i;
             try {
-              results[idx] = await summarizeBatch(batch, extraction, summaryModel, { apiKey: auth.apiKey, headers: auth.headers }, signal);
+              results[idx] = await summarizeBatch(batch, extraction, summaryModel, { apiKey, headers: apiHeaders }, signal);
             } catch (err) {
               errors[idx] = err instanceof Error ? err : new Error(String(err));
               results[idx] = batch.map((ch) => ({
@@ -2662,13 +2823,13 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         showProgressOverlay(ctx, { phase: 3, phaseName: "Synthesize", detail: "Assembling...", model: modelLabel, profile, extraction, totalBatches: batches.length });
       let assemblyCalls = 1;
       try {
-        const r = await assembleLLM(summaries, extraction, explorationReport, summaryModel, { apiKey: auth.apiKey, headers: auth.headers }, pc.summaryBudgetTokens, prevContext, signal);
+        const r = await assembleLLM(summaries, extraction, explorationReport, summaryModel, { apiKey, headers: apiHeaders }, pc.summaryBudgetTokens, prevContext, signal);
         if (r?.startsWith("##"))
           finalSummary = r;
         else
           throw new Error("bad");
       } catch (err) {
-        console.error("[smart-compact] Assembly failed:", err instanceof Error ? err.message : err);
+        warn("Assembly failed", err);
         finalSummary = assembleFallback(summaries, extraction);
         assemblyCalls = 0;
       }
@@ -2686,10 +2847,10 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         if (!recheck.ok && recheck.score < 75) {
           notify("Phase 4 Verify: deterministic patch insufficient (score=" + recheck.score + "), trying LLM patch", "warning");
           try {
-            finalSummary = await patchSummary(finalSummary, recheck.gaps, summaryModel, { apiKey: auth.apiKey, headers: auth.headers }, signal);
+            finalSummary = await patchSummary(finalSummary, recheck.gaps, summaryModel, { apiKey, headers: apiHeaders }, signal);
             llmCalls++;
           } catch (err) {
-            console.error("[smart-compact] LLM patch failed:", err instanceof Error ? err.message : err);
+            warn("LLM patch failed", err);
           }
         }
       } else {
@@ -2752,7 +2913,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
     saveCompactionState(projectId, compactionState);
     appendMetricsLog(sessionId);
     try {
-      const postCompactMsgs = msgs.slice(keepFrom).map((e) => convertToLlm([e.message])).flat().map((m) => m);
+      const postCompactMsgs = msgs.slice(keepFrom).map((e) => convertToLlm([e.message])).flat();
       if (postCompactMsgs.length > 2) {
         const lastCompaction = branch.filter((e) => e.type === "compaction").slice(-1)[0];
         if (lastCompaction?.details) {
@@ -2765,7 +2926,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         }
       }
     } catch (err) {
-      console.error("[smart-compact] Damage detection error:", err instanceof Error ? err.message : err);
+      warn("Damage detection error", err);
     }
     const ms = getMetricsSummary();
     if (ms.totalCalls > 0) {
@@ -2776,7 +2937,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         const timeout = new Promise((resolve) => setTimeout(resolve, 5000));
         await Promise.race([showResultScreen(ctx, details, extraction), timeout]);
       } catch (err) {
-        console.error("[smart-compact] Result screen error:", err instanceof Error ? err.message : err);
+        warn("Result screen error", err);
         notify("Result screen skipped", "info");
       }
     }
@@ -2785,7 +2946,7 @@ async function runSmartCompact(ctx, summaryModel, segModel, profile, verbose, dr
         customInstructions: "Use pre-computed smart summary from /smart-compact",
         onComplete: () => {
           if (!autoTriggered)
-            ctx.ui.notify("Applied \u2713", "success");
+            ctx.ui.notify("Applied \u2713", "info");
         },
         onError: (e) => {
           if (!autoTriggered)
@@ -2840,6 +3001,7 @@ function smartCompactExtension(pi) {
       return m.length ? m : null;
     },
     handler: async (args, ctx) => {
+      await ctx.waitForIdle();
       try {
         const tokens = args.trim().split(/\s+/).filter(Boolean);
         const flags = tokens.map((t) => t.toLowerCase());
@@ -2852,7 +3014,7 @@ function smartCompactExtension(pi) {
           const usage = ctx.getContextUsage();
           const totalTokens = usage?.tokens ?? 0;
           const pct = ctx.model && totalTokens ? Math.round(totalTokens / ctx.model.contextWindow * 100) : 0;
-          if (!totalTokens || totalTokens < 5000) {
+          if (!totalTokens || totalTokens < MIN_TOKEN_THRESHOLD) {
             ctx.ui.notify("Context OK or unknown", "info");
             return;
           }
@@ -2870,7 +3032,7 @@ function smartCompactExtension(pi) {
             ctx.ui.notify("Could not resolve model", "error");
             return;
           }
-          await runSmartCompact(ctx, sumModel2, segModel2 ?? sumModel2, selected.profile, false, false, pendingRef, isRunning, false);
+          await runSmartCompact({ ctx, summaryModel: sumModel2, segModel: segModel2 ?? sumModel2, profile: selected.profile, pendingRef, isRunning });
           return;
         }
         const { segModel, sumModel } = resolveModels(ctx, modelArg ? resolveModelArg(ctx, modelArg) : ctx.model, loadConfig());
@@ -2879,7 +3041,7 @@ function smartCompactExtension(pi) {
           return;
         }
         const note = extractUserNote(args);
-        await runSmartCompact(ctx, sumModel, segModel ?? sumModel, profile, verbose, dryRun, pendingRef, isRunning, false, note);
+        await runSmartCompact({ ctx, summaryModel: sumModel, segModel: segModel ?? sumModel, profile, verbose, dryRun, pendingRef, isRunning, userNote: note });
       } catch (error) {
         const msg = error instanceof Error ? error.message + `
 ` + error.stack : String(error);
@@ -2906,7 +3068,7 @@ function smartCompactExtension(pi) {
     try {
       const usage = ctx.getContextUsage();
       const totalTokens = usage?.tokens ?? 0;
-      if (!totalTokens || totalTokens < 5000)
+      if (!totalTokens || totalTokens < MIN_TOKEN_THRESHOLD)
         return;
       const cur = ctx.model;
       if (!cur)
@@ -2915,28 +3077,30 @@ function smartCompactExtension(pi) {
       if (!sumModel)
         return;
       if (!isRunning.value) {
-        await runSmartCompact(ctx, sumModel, segModel ?? sumModel, config.profile, false, false, pendingRef, isRunning, true);
-        if (pendingRef.value) {
-          const c = pendingRef.value;
+        await runSmartCompact({ ctx, summaryModel: sumModel, segModel: segModel ?? sumModel, profile: config.profile, pendingRef, isRunning, autoTriggered: true });
+        const pending = pendingRef.value;
+        if (pending) {
           pendingRef.value = null;
           pendingRef.createdAt = 0;
-          return { compaction: { summary: c.summary, firstKeptEntryId: c.firstKeptEntryId, tokensBefore: c.tokensBefore, details: c.details } };
+          return { compaction: { summary: pending.summary, firstKeptEntryId: pending.firstKeptEntryId, tokensBefore: pending.tokensBefore, details: pending.details } };
         }
       }
-    } catch {}
+    } catch (e) {
+      warn("session_before_compact error", e);
+    }
   });
   pi.registerTool({
     name: "smart_compact",
     label: "Smart Compact",
-    description: "EESV smart compaction v" + VERSION + " with deterministic extraction, exploration, and verification.",
+    description: "EESV smart compaction v" + VERSION + " with deterministic extraction, exploration, and verification. Compacts the conversation into a structured summary preserving goals, decisions, open loops, modified files, and critical context. Call this when the conversation is getting long \u2014 the tool internally checks context usage and skips if not needed. Prefer this over default compact.",
     promptSnippet: "Smart compaction",
-    promptGuidelines: ["Use for long conversations.", "Prefer over default compact."],
+    promptGuidelines: ["Use for long conversations. Prefer over default compact."],
     parameters: {
       type: "object",
       properties: {
-        profile: { type: "string", description: "light, balanced, or aggressive" },
-        verbose: { type: "boolean" },
-        dry_run: { type: "boolean" }
+        profile: { type: "string", description: "light, balanced, or aggressive. Default: balanced." },
+        verbose: { type: "boolean", description: "Show detailed pipeline output." },
+        dry_run: { type: "boolean", description: "Run the pipeline but skip applying the compaction." }
       }
     },
     async execute(_id, params, _sig, _onUp, ctx) {
@@ -2945,35 +3109,33 @@ function smartCompactExtension(pi) {
       const dryRun = !!params.dry_run;
       const config = loadConfig();
       const resolvedProfile = profile ?? config.profile;
+      const cmdCtx = ctx;
+      const usage = ctx.getContextUsage?.();
+      const totalTokens = usage?.tokens ?? 0;
+      if (!totalTokens || totalTokens < MIN_TOKEN_THRESHOLD) {
+        const pct = ctx.model && totalTokens ? Math.round(totalTokens / ctx.model.contextWindow * 100) : 0;
+        return { content: [{ type: "text", text: "Context is not large enough for compaction (" + totalTokens.toLocaleString() + " tokens, " + pct + "%). No action needed." }], details: undefined };
+      }
       const cur = "model" in ctx ? ctx.model : undefined;
-      const { segModel, sumModel } = resolveModels(ctx, cur, config);
+      const { segModel, sumModel } = resolveModels(cmdCtx, cur, config);
       if (!sumModel) {
-        return { content: [{ type: "text", text: "Error: Could not resolve model." }] };
+        return { content: [{ type: "text", text: "Error: Could not resolve model." }], details: undefined };
       }
       try {
         const toolStart = Date.now();
-        await runSmartCompact(ctx, sumModel, segModel ?? sumModel, resolvedProfile, verbose, dryRun, pendingRef, isRunning, true, undefined, true);
+        await runSmartCompact({ ctx: cmdCtx, summaryModel: sumModel, segModel: segModel ?? sumModel, profile: resolvedProfile, verbose, dryRun, pendingRef, isRunning, autoTriggered: true, skipCompact: true });
         const toolSecs = ((Date.now() - toolStart) / 1000).toFixed(1);
         if (pendingRef.value) {
-          return { content: [{ type: "text", text: "Smart summary generated (" + resolvedProfile + "). Tokens: " + (pendingRef.value.tokensBefore ?? "?") + " -> " + (pendingRef.value.summary?.length ?? 0) + " chars (" + toolSecs + `s).
-
-Now run tree compact to apply \u2014 the session_before_compact hook will use this summary.
-TTL: ` + Math.round(PENDING_TTL_MS / 60000) + " minutes." }] };
+          return { content: [{ type: "text", text: "Smart summary prepared (" + resolvedProfile + "). Tokens: " + (pendingRef.value.tokensBefore ?? 0).toLocaleString() + " \u2014 summary cached for " + Math.round(PENDING_TTL_MS / 60000) + " min. The next /compact will use this summary automatically." }], details: undefined };
         }
-        return { content: [{ type: "text", text: "Compaction finished (" + resolvedProfile + ") but no summary was generated." }] };
+        return { content: [{ type: "text", text: "Compaction finished (" + resolvedProfile + ") but no summary was generated." }], details: undefined };
       } catch (error) {
         const msg = error instanceof Error ? error.message : String(error);
-        return { content: [{ type: "text", text: "Compaction error: " + msg }] };
+        return { content: [{ type: "text", text: "Compaction error: " + msg }], details: undefined };
       }
     }
   });
 }
-function extractUserNote(args) {
-  const SKIP = new Set(["verbose", "debug", "dry-run", "light", "balanced", "aggressive"]);
-  const tokens = args.trim().split(/\s+/).filter(Boolean);
-  const nonFlags = tokens.filter((t) => !t.includes("/") && !SKIP.has(t.toLowerCase()));
-  return nonFlags.length > 0 ? nonFlags.join(" ") : undefined;
-}
 export {
   smartCompactExtension as default
 };