autopilot-ai 1.13.0 → 1.14.0

package/dist/index.js

@@ -1832,9 +1832,14 @@ function shouldCompact(opts2) {
   const { turns } = groupIntoTurns(opts2.messages);
   return tokens > tokenThreshold || turns.length > turnThreshold;
 }
+function isCompiledStateMessage(m) {
+  return m.role === "system" && typeof m.content === "string" && m.content.startsWith("[compiled session state]");
+}
 function compactMessagesViaArtifacts(opts2) {
   const keepLastTurns = opts2.keepLastTurns ?? 4;
-  const { prefix, turns } = groupIntoTurns(opts2.messages);
+  const grouped = groupIntoTurns(opts2.messages);
+  const turns = grouped.turns;
+  const prefix = grouped.prefix.filter((m) => !isCompiledStateMessage(m));
   const tokensBefore = estimatePromptTokens(opts2.messages);
   if (turns.length <= keepLastTurns) {
     return {
@@ -4530,12 +4535,7 @@ ${opts2.lspContext}` : "";
 The following MCP server information was pre-computed by the coordinator.
 
 ${opts2.mcpContext}` : "";
-  const delegationBlock = opts2.delegationDirective ? `
-
-## Subagent policy for this turn
-
-${opts2.delegationDirective}` : "";
-  return identity + "\n\n" + env2 + "\n\n" + tools + lspBlock + userShellBlock + contextBlock + modeBlock + skillsBlock + memoryBlock + lspContextBlock + mcpContextBlock + delegationBlock;
+  return identity + "\n\n" + env2 + "\n\n" + tools + lspBlock + userShellBlock + contextBlock + modeBlock + skillsBlock + memoryBlock + lspContextBlock + mcpContextBlock;
 }
 function buildSystemPrompt(opts2) {
   return buildStaticPrefix(opts2) + "\n\n" + buildSessionPrefix(opts2);
@@ -4557,6 +4557,139 @@ var init_system_prompt = __esm({
   }
 });
 
+// src/agent/permission-gate.ts
+function createPermissionGate(ask, calls) {
+  const decisions = /* @__PURE__ */ new Map();
+  let tail = Promise.resolve();
+  return (req) => {
+    const existing = decisions.get(req.sessionKey);
+    if (existing) return existing;
+    const sameTool = calls.filter((call) => call.name === req.tool.name);
+    const request = sameTool.length > 1 ? { ...req, args: { ...req.args, batch: sameTool.map((call) => call.args) } } : req;
+    const decision = tail.then(() => ask(request));
+    tail = decision.catch(() => void 0);
+    decisions.set(req.sessionKey, decision);
+    return decision;
+  };
+}
+var init_permission_gate = __esm({
+  "src/agent/permission-gate.ts"() {
+    "use strict";
+  }
+});
+
+// src/agent/context-budget.ts
+function effectiveInputBudget(model, maxCompletionTokens) {
+  const contextWindow = getModelOrInfer(model).contextWindow;
+  const completion = maxCompletionTokens ?? DEFAULT_MAX_COMPLETION_TOKENS;
+  return Math.max(1, contextWindow - completion - BUDGET_SAFETY_MARGIN_TOKENS);
+}
+function proactiveCompactionThreshold(model, maxCompletionTokens) {
+  return Math.min(
+    PROACTIVE_COMPACTION_CAP_TOKENS,
+    Math.floor(effectiveInputBudget(model, maxCompletionTokens) * PROACTIVE_SHARE_OF_BUDGET)
+  );
+}
+function compactToFit(messages, budgetTokens, target) {
+  const tokensBefore = estimatePromptTokens(messages);
+  for (const keepLastTurns of KEEP_TURNS_LADDER) {
+    const scratch = new ArtifactStore();
+    const attempt = compactMessagesViaArtifacts({ messages, state: target.getState(), store: scratch, keepLastTurns });
+    if (attempt.metrics.rawTurnsRemoved === 0) continue;
+    const isLast = keepLastTurns === KEEP_TURNS_LADDER[KEEP_TURNS_LADDER.length - 1];
+    if (attempt.metrics.estimatedTokensAfter > budgetTokens && !isLast) continue;
+    const store = target.getStore();
+    for (const artifact of scratch.list()) store.add(artifact);
+    target.setState(attempt.newState);
+    target.onChange?.();
+    return {
+      messages: attempt.newMessages,
+      tokensBefore,
+      tokensAfter: attempt.metrics.estimatedTokensAfter,
+      turnsRemoved: attempt.metrics.rawTurnsRemoved,
+      artifactsArchived: attempt.metrics.archivedArtifacts
+    };
+  }
+  return null;
+}
+function sessionFileCompactionTarget(file) {
+  let store = null;
+  return {
+    getState: () => file.sessionState ?? emptySessionState(),
+    setState: (state) => {
+      file.sessionState = state;
+    },
+    getStore: () => store ??= file.artifactStore ? deserializeArtifactStore(file.artifactStore) : new ArtifactStore(),
+    onChange: () => {
+      if (store) file.artifactStore = serializeArtifactStore(store);
+    }
+  };
+}
+var DEFAULT_MAX_COMPLETION_TOKENS, BUDGET_SAFETY_MARGIN_TOKENS, PROACTIVE_COMPACTION_CAP_TOKENS, PROACTIVE_SHARE_OF_BUDGET, KEEP_TURNS_LADDER, ContextBudgetError;
+var init_context_budget = __esm({
+  "src/agent/context-budget.ts"() {
+    "use strict";
+    init_registry();
+    init_artifact_compaction();
+    init_session_state();
+    DEFAULT_MAX_COMPLETION_TOKENS = 16384;
+    BUDGET_SAFETY_MARGIN_TOKENS = 8192;
+    PROACTIVE_COMPACTION_CAP_TOKENS = 8e4;
+    PROACTIVE_SHARE_OF_BUDGET = 0.75;
+    KEEP_TURNS_LADDER = [4, 2, 1];
+    ContextBudgetError = class extends Error {
+      constructor(estimatedTokens, budgetTokens, contextWindow) {
+        super(
+          `kimiflare: context window exceeded (~${estimatedTokens.toLocaleString()} tokens needed, ${budgetTokens.toLocaleString()} available of ${contextWindow.toLocaleString()}). Older turns were already compacted; the current turn alone is too large. Run /clear to start fresh, pick a model with a larger context window, or reduce the size of the request (e.g. fewer or smaller attachments).`
+        );
+        this.estimatedTokens = estimatedTokens;
+        this.budgetTokens = budgetTokens;
+        this.contextWindow = contextWindow;
+        this.name = "ContextBudgetError";
+      }
+      estimatedTokens;
+      budgetTokens;
+      contextWindow;
+    };
+  }
+});
+
+// src/agent/turn-directive.ts
+function withTurnDirective(messages, directive, turnUserMessage) {
+  let out = messages;
+  if (messages.some(hasPersistedBlock)) {
+    out = messages.map(
+      (m) => hasPersistedBlock(m) ? { ...m, content: m.content.replace(PERSISTED_BLOCK, "") } : m
+    );
+  }
+  if (!directive) return out;
+  let anchor = turnUserMessage ? out.indexOf(turnUserMessage) : -1;
+  if (anchor === -1) {
+    for (let i = out.length - 1; i >= 0; i--) {
+      if (out[i].role === "user") {
+        anchor = i;
+        break;
+      }
+    }
+  }
+  const directiveMessage = { role: "system", content: `${HEADING}
+
+${directive}` };
+  const insertAt = anchor === -1 ? out.length : anchor + 1;
+  return [...out.slice(0, insertAt), directiveMessage, ...out.slice(insertAt)];
+}
+function hasPersistedBlock(m) {
+  return m.role === "system" && typeof m.content === "string" && m.content.includes(HEADING);
+}
+var HEADING, PERSISTED_BLOCK;
+var init_turn_directive = __esm({
+  "src/agent/turn-directive.ts"() {
+    "use strict";
+    HEADING = "## Subagent policy for this turn";
+    PERSISTED_BLOCK = /\n*## Subagent policy for this turn\n[\s\S]*?(?=\n## |$)/;
+  }
+});
+
 // src/agent/loop.ts
 function getSessionWebFetchHistory(sessionId) {
   const key = sessionId ?? "default";
@@ -4621,6 +4754,7 @@ async function runAgentTurn(opts2) {
   let memoryRecalledCount = 0;
   let skillResult;
   const lastUserPrompt = extractLastUserText(opts2.messages);
+  const turnUserMessage = [...opts2.messages].reverse().find((m) => m.role === "user");
   const userPromptPreview = lastUserPrompt.slice(0, 200);
   const skipSkillRouting = opts2.intentClassification?.tier === "light" && lastUserPrompt.length < 40;
   const recallPromise = opts2.sessionStartRecall && opts2.memoryManager && !hasRecalledMemory(opts2.messages) ? (async () => {
@@ -4670,8 +4804,7 @@ async function runAgentTurn(opts2) {
           model: opts2.model,
           mode: opts2.mode,
           skillContext: skillResult.skillContext,
-          preferPullRequests: opts2.preferPullRequests,
-          delegationDirective: opts2.delegationDirective
+          preferPullRequests: opts2.preferPullRequests
         })
       };
     } else {
@@ -4683,8 +4816,7 @@ async function runAgentTurn(opts2) {
           model: opts2.model,
           mode: opts2.mode,
           skillContext: skillResult.skillContext,
-          preferPullRequests: opts2.preferPullRequests,
-          delegationDirective: opts2.delegationDirective
+          preferPullRequests: opts2.preferPullRequests
         })
       };
     }
@@ -4752,7 +4884,6 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
   let webFetchesThisTurn = 0;
   let codeModeRedirects = 0;
   const MAX_WEB_FETCH_PER_TURN = 5;
-  const WEB_FETCH_DOMAIN_THRESHOLD = 2;
   let cumulativePromptTokens = 0;
   let iter = 0;
   let totalIter = 0;
@@ -4770,6 +4901,11 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
     mutatingToolsSinceLastTasksSet = 0;
     originalOnTasks?.(tasks);
   };
+  const appendPendingInput = () => {
+    const pending = opts2.drainPendingInput?.() ?? [];
+    for (const msg of pending) opts2.messages.push(msg);
+    return pending.length;
+  };
   while (true) {
     if (budgetExhausted) {
       opts2.messages.push({
@@ -4827,46 +4963,53 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
     const shadowStrip = process.env.KIMIFLARE_SHADOW_STRIP === "1";
     const keepLastRaw = process.env.KIMIFLARE_REASONING_KEEP_LAST;
     const keepLast = keepLastRaw ? parseInt(keepLastRaw, 10) : 1;
-    let apiMessages = opts2.messages;
     let shadowStripMetrics;
-    if (stripReasoning || shadowStrip) {
-      const stripped = stripHistoricalReasoning(opts2.messages, {
-        keepLast: Number.isNaN(keepLast) ? 1 : keepLast
-      });
-      if (shadowStrip) {
-        const originalSections = analyzePrompt(opts2.messages);
-        const strippedSections = analyzePrompt(stripped);
-        const originalApproxTokens = originalSections.reduce(
-          (sum, s) => sum + s.approxTokens,
-          0
-        );
-        const strippedApproxTokens = strippedSections.reduce(
-          (sum, s) => sum + s.approxTokens,
-          0
-        );
-        shadowStripMetrics = {
-          originalApproxTokens,
-          strippedApproxTokens,
-          savingsPct: originalApproxTokens > 0 ? Math.round(
-            (originalApproxTokens - strippedApproxTokens) / originalApproxTokens * 100
-          ) : 0
-        };
+    const buildApiMessages = () => {
+      let out = opts2.messages;
+      if (stripReasoning || shadowStrip) {
+        const stripped = stripHistoricalReasoning(opts2.messages, {
+          keepLast: Number.isNaN(keepLast) ? 1 : keepLast
+        });
+        if (shadowStrip) {
+          const originalApproxTokens = analyzePrompt(opts2.messages).reduce((sum, sec) => sum + sec.approxTokens, 0);
+          const strippedApproxTokens = analyzePrompt(stripped).reduce((sum, sec) => sum + sec.approxTokens, 0);
+          shadowStripMetrics = {
+            originalApproxTokens,
+            strippedApproxTokens,
+            savingsPct: originalApproxTokens > 0 ? Math.round((originalApproxTokens - strippedApproxTokens) / originalApproxTokens * 100) : 0
+          };
+        }
+        if (stripReasoning) out = stripped;
       }
-      if (stripReasoning) {
-        apiMessages = stripped;
+      if (opts2.keepLastImageTurns !== void 0) {
+        out = stripOldImages(out, opts2.keepLastImageTurns);
+      }
+      return withTurnDirective(out, opts2.delegationDirective, turnUserMessage);
+    };
+    let apiMessages = buildApiMessages();
+    let promptTokens = estimatePromptTokens(apiMessages);
+    const maxPromptTokens = effectiveInputBudget(opts2.model, opts2.maxCompletionTokens);
+    if (promptTokens > maxPromptTokens && opts2.compaction) {
+      const requestOverhead = Math.max(0, promptTokens - estimatePromptTokens(opts2.messages));
+      const outcome = compactToFit(opts2.messages, maxPromptTokens - requestOverhead, opts2.compaction);
+      if (outcome) {
+        opts2.messages.splice(0, opts2.messages.length, ...outcome.messages);
+        apiMessages = buildApiMessages();
+        promptTokens = estimatePromptTokens(apiMessages);
+        const info = {
+          tokensBefore: outcome.tokensBefore,
+          tokensAfter: outcome.tokensAfter,
+          turnsRemoved: outcome.turnsRemoved,
+          artifactsArchived: outcome.artifactsArchived,
+          budgetTokens: maxPromptTokens
+        };
+        logger.info("context:compacted", { sessionId: opts2.sessionId, ...info });
+        opts2.callbacks.onCompacted?.(info);
       }
     }
-    if (opts2.keepLastImageTurns !== void 0) {
-      apiMessages = stripOldImages(apiMessages, opts2.keepLastImageTurns);
-    }
-    const promptTokens = estimatePromptTokens(apiMessages);
-    const ctxWindow = getModelOrInfer(opts2.model).contextWindow;
-    const completionBudget = opts2.maxCompletionTokens ?? DEFAULT_MAX_COMPLETION_TOKENS;
-    const maxPromptTokens = ctxWindow - completionBudget - BUDGET_SAFETY_MARGIN_TOKENS;
     if (promptTokens > maxPromptTokens) {
-      throw new Error(
-        `kimiflare: context window exceeded (~${promptTokens.toLocaleString()} / ${ctxWindow.toLocaleString()} tokens). Run /compact to summarize older turns, or /clear to start fresh.`
-      );
+      logger.warn("context:irreducible", { sessionId: opts2.sessionId, promptTokens, budgetTokens: maxPromptTokens });
+      throw new ContextBudgetError(promptTokens, maxPromptTokens, getModelOrInfer(opts2.model).contextWindow);
     }
     logger.debug("turn:api_request", { sessionId: opts2.sessionId, messageCount: apiMessages.length });
     const events = runKimi({
@@ -4989,6 +5132,7 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
       if (budgetExhausted) {
         throw new BudgetExhaustedError();
       }
+      if (appendPendingInput() > 0) continue;
       logger.info("turn:complete", { sessionId: opts2.sessionId, durationMs: Math.round(performance.now() - turnStart) });
       await fireStopHook();
       return;
@@ -4996,12 +5140,20 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
     let blockedCount = 0;
     const availableToolNames = new Set(opts2.tools.map((tool) => tool.name));
     if (codeMode) availableToolNames.add("execute_code");
-    const allReadOnly = toolCalls.length > 1 && toolCalls.every((tc) => {
-      const tool = opts2.executor.list().find((t) => t.name === tc.function.name);
-      return availableToolNames.has(tc.function.name) && tool?.isReadOnly === true;
+    const executorTools = opts2.executor.list();
+    const allConcurrent = toolCalls.length > 1 && toolCalls.every((tc) => {
+      const tool = executorTools.find((t) => t.name === tc.function.name);
+      return availableToolNames.has(tc.function.name) && (tool?.isReadOnly === true || tool?.concurrent === true);
     });
-    if (allReadOnly) {
-      opts2.callbacks.onInfo?.(`${toolCalls.length} read-only tools running in parallel`);
+    if (allConcurrent) {
+      const workerCount = toolCalls.filter((tc) => executorTools.find((t) => t.name === tc.function.name)?.concurrent === true).length;
+      opts2.callbacks.onInfo?.(
+        workerCount > 0 ? `${toolCalls.length} tool calls running in parallel (${workerCount} isolated worker${workerCount === 1 ? "" : "s"})` : `${toolCalls.length} read-only tools running in parallel`
+      );
+      const batchAskPermission = createPermissionGate(
+        opts2.callbacks.askPermission,
+        toolCalls.map((tc) => ({ name: tc.function.name, args: parseArgsObject(tc.function.arguments) }))
+      );
       const items = [];
       for (const tc of toolCalls) {
         if (opts2.signal.aborted) throw new DOMException("aborted", "AbortError");
@@ -5016,7 +5168,8 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
               tool_call_id: tc.id,
               name: tc.function.name,
               content: `Loop detected: you have called ${tc.function.name} with the same arguments multiple times in a row. Consider a different approach.`,
-              ok: false
+              ok: false,
+              guardrail: true
             }
           });
           continue;
@@ -5035,27 +5188,19 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
           const args = JSON.parse(tc.function.arguments || "{}");
           const url = args.url || "";
           try {
-            const domain = new URL(url).hostname;
-            const domainCount = webFetchHistory.filter((h) => h.domain === domain).length;
+            new URL(url);
             const totalSessionFetches = webFetchHistory.length;
-            if (webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN || totalSessionFetches >= SESSION_WEB_FETCH_CAP || domainCount >= WEB_FETCH_DOMAIN_THRESHOLD) {
-              let warning;
-              if (webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN) {
-                warning = `Research budget exceeded: you have already made ${MAX_WEB_FETCH_PER_TURN} web requests this turn. Synthesize what you have learned instead of fetching more pages.`;
-              } else if (totalSessionFetches >= SESSION_WEB_FETCH_CAP) {
-                warning = `Session research budget exceeded: ${totalSessionFetches} web fetches across this session. Synthesize what you have learned from prior fetches instead of starting another page.`;
-              } else {
-                warning = `Loop detected: you have fetched from ${domain} multiple times. Consider a different approach or synthesize existing findings.`;
-              }
+            if (webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN || totalSessionFetches >= SESSION_WEB_FETCH_CAP) {
+              const warning = webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN ? `Research budget exceeded: you have already made ${MAX_WEB_FETCH_PER_TURN} web requests this turn. Synthesize what you have learned instead of fetching more pages.` : `Session research budget exceeded: ${totalSessionFetches} web fetches across this session. Synthesize what you have learned from prior fetches instead of starting another page.`;
               items.push({
                 kind: "blocked",
                 tc,
                 loopSignature,
-                result: { tool_call_id: tc.id, name: "web_fetch", content: warning, ok: false }
+                result: { tool_call_id: tc.id, name: "web_fetch", content: warning, ok: false, guardrail: true }
               });
               continue;
             }
-            webFetchHistory.push({ url, domain });
+            webFetchHistory.push(url);
             webFetchesThisTurn++;
           } catch {
           }
@@ -5069,7 +5214,7 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
           logger.debug("turn:tool_start", { sessionId: opts2.sessionId, tool: it.tc.function.name, toolCallId: it.tc.id });
           const result2 = await opts2.executor.run(
             { id: it.tc.id, name: it.tc.function.name, arguments: it.tc.function.arguments },
-            opts2.callbacks.askPermission,
+            batchAskPermission,
             {
               cwd: opts2.cwd,
               signal: opts2.signal,
@@ -5085,6 +5230,7 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
               model: opts2.model,
               allowDirectPush: opts2.allowDirectPush,
               runId: opts2.runId,
+              onRunBudgetExceeded: opts2.callbacks.onRunBudgetExceeded,
               runsDbPath: opts2.runsDbPath
             },
             opts2.onFileChange
@@ -5267,7 +5413,8 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
             tool_call_id: tc.id,
             name: tc.function.name,
             content: warning,
-            ok: false
+            ok: false,
+            guardrail: true
           };
           toolResults.push(loopResult);
           opts2.messages.push({
@@ -5286,8 +5433,7 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
           const args = JSON.parse(tc.function.arguments || "{}");
           const url = args.url || "";
           try {
-            const domain = new URL(url).hostname;
-            const domainCount = webFetchHistory.filter((h) => h.domain === domain).length;
+            new URL(url);
             const totalSessionFetches = webFetchHistory.length;
             if (webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN) {
               const warning = `Research budget exceeded: you have already made ${MAX_WEB_FETCH_PER_TURN} web requests this turn. Synthesize what you have learned instead of fetching more pages.`;
@@ -5295,7 +5441,8 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
                 tool_call_id: tc.id,
                 name: "web_fetch",
                 content: warning,
-                ok: false
+                ok: false,
+                guardrail: true
               };
               toolResults.push(budgetResult);
               opts2.messages.push({
@@ -5316,7 +5463,8 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
                 tool_call_id: tc.id,
                 name: "web_fetch",
                 content: warning,
-                ok: false
+                ok: false,
+                guardrail: true
               };
               toolResults.push(sessionCapResult);
               opts2.messages.push({
@@ -5331,28 +5479,7 @@ IMPORTANT \u2014 explore through code, not direct tools: to read files, run shel
               blockedCount++;
               continue;
             }
-            if (domainCount >= WEB_FETCH_DOMAIN_THRESHOLD) {
-              const warning = `Loop detected: you have fetched from ${domain} multiple times. Consider a different approach or synthesize existing findings.`;
-              const loopResult = {
-                tool_call_id: tc.id,
-                name: "web_fetch",
-                content: warning,
-                ok: false
-              };
-              toolResults.push(loopResult);
-              opts2.messages.push({
-                role: "tool",
-                tool_call_id: tc.id,
-                content: sanitizeString(warning),
-                name: "web_fetch"
-              });
-              opts2.callbacks.onToolResult?.(loopResult);
-              recentToolCalls.push(loopSignature);
-              if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
-              blockedCount++;
-              continue;
-            }
-            webFetchHistory.push({ url, domain });
+            webFetchHistory.push(url);
             webFetchesThisTurn++;
           } catch {
           }
@@ -5490,7 +5617,7 @@ ${sandboxResult.output}` : sandboxResult.output;
           opts2.callbacks.onToolResult?.(result2);
           if (result2.waitRequest) {
             if (opts2.onIterationEnd) {
-              opts2.messages = await opts2.onIterationEnd(opts2.messages, opts2.signal);
+              replaceMessagesInPlace(opts2.messages, await opts2.onIterationEnd(opts2.messages, opts2.signal));
             }
             logger.info("run:yielded", {
               runId: result2.waitRequest.runId,
@@ -5621,8 +5748,9 @@ ${sandboxResult.output}` : sandboxResult.output;
     if (blockedCount === toolCalls.length && toolCalls.length > 0) {
       loopExhausted = true;
     }
+    appendPendingInput();
     if (opts2.onIterationEnd) {
-      opts2.messages = await opts2.onIterationEnd(opts2.messages, opts2.signal);
+      replaceMessagesInPlace(opts2.messages, await opts2.onIterationEnd(opts2.messages, opts2.signal));
       if (opts2.signal.aborted) throw new DOMException("aborted", "AbortError");
     }
     if (opts2.sessionId && lastUsage) {
@@ -5674,6 +5802,18 @@ ${sandboxResult.output}` : sandboxResult.output;
     }
   }
 }
+function parseArgsObject(raw) {
+  try {
+    const parsed = JSON.parse(raw || "{}");
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : {};
+  } catch {
+    return {};
+  }
+}
+function replaceMessagesInPlace(target, next) {
+  if (next === target) return;
+  target.splice(0, target.length, ...next);
+}
 function validateToolArguments(raw) {
   if (!raw || !raw.trim()) return "{}";
   try {
@@ -5691,7 +5831,7 @@ function llmAuthOf(opts2) {
     ...opts2.provider ? { provider: opts2.provider } : {}
   };
 }
-var MAX_LOOP_RECOVERIES, BudgetExhaustedError, AgentLoopError, codeModeApiCache, driftEvents, DRIFT_WINDOW, DRIFT_THRESHOLD, memoryExtractionErrorCounts, sessionWebFetchHistory, SESSION_WEB_FETCH_CAP, DEFAULT_MAX_COMPLETION_TOKENS, BUDGET_SAFETY_MARGIN_TOKENS, MAX_TOOL_CONTENT_CHARS, CODE_MODE_REDIRECT_TOOLS, MAX_CODE_MODE_REDIRECTS;
+var MAX_LOOP_RECOVERIES, BudgetExhaustedError, AgentLoopError, codeModeApiCache, driftEvents, DRIFT_WINDOW, DRIFT_THRESHOLD, memoryExtractionErrorCounts, sessionWebFetchHistory, SESSION_WEB_FETCH_CAP, MAX_TOOL_CONTENT_CHARS, CODE_MODE_REDIRECT_TOOLS, MAX_CODE_MODE_REDIRECTS;
 var init_loop = __esm({
   "src/agent/loop.ts"() {
     "use strict";
@@ -5708,6 +5848,9 @@ var init_loop = __esm({
     init_router();
     init_system_prompt();
     init_registry();
+    init_permission_gate();
+    init_context_budget();
+    init_turn_directive();
     MAX_LOOP_RECOVERIES = 1;
     BudgetExhaustedError = class extends Error {
       constructor(message2 = "Cumulative input token budget exhausted") {
@@ -5728,8 +5871,6 @@ var init_loop = __esm({
     memoryExtractionErrorCounts = /* @__PURE__ */ new Map();
     sessionWebFetchHistory = /* @__PURE__ */ new Map();
     SESSION_WEB_FETCH_CAP = 25;
-    DEFAULT_MAX_COMPLETION_TOKENS = 16384;
-    BUDGET_SAFETY_MARGIN_TOKENS = 8192;
     MAX_TOOL_CONTENT_CHARS = 1e4;
     CODE_MODE_REDIRECT_TOOLS = /* @__PURE__ */ new Set(["read", "bash", "grep", "glob"]);
     MAX_CODE_MODE_REDIRECTS = 4;
@@ -6664,47 +6805,64 @@ function parsePushTarget(command) {
   }
   return { kind: "ref", ref: last.replace(/^refs\/heads\//, "") };
 }
+function buildCoauthorTrailer(coauthor) {
+  const unsafe = /[\u0000-\u001f\u007f\u2028\u2029]/;
+  if (unsafe.test(coauthor.name) || unsafe.test(coauthor.email)) return null;
+  const name = coauthor.name.trim();
+  const email = coauthor.email.trim();
+  if (!name || !email || /[<>]/.test(email)) return null;
+  return `Co-authored-by: ${name} <${email}>`;
+}
 function injectCoauthor(command, coauthor) {
-  if (!coauthor) return command;
-  const trailer = `Co-authored-by: ${coauthor.name} <${coauthor.email}>`;
+  const unchanged = { command, env: {} };
+  if (!coauthor) return unchanged;
+  const trailer = buildCoauthorTrailer(coauthor);
+  if (!trailer) {
+    logger.warn("bash:coauthor_rejected", { reason: "empty value, control character, or line break in co-author name/email" });
+    return unchanged;
+  }
   const trimmed = command.trim();
-  if (command.includes(trailer)) return command;
+  if (command.includes(trailer)) return unchanged;
   const createsCommit = /\bgit\s+(commit|merge|revert|cherry-pick)\b/.test(trimmed);
   const isRebaseContinue = /\bgit\s+rebase\b/.test(trimmed) && !/\b--abort\b|\b--skip\b/.test(trimmed);
   const movesHeadOnly = /\bgit\s+(reset|checkout|switch)\b/.test(trimmed);
   const mentionsGit = /\bgit\b/.test(trimmed);
-  if (!createsCommit && !isRebaseContinue && !mentionsGit) return command;
-  if (movesHeadOnly) return command;
-  const tmpFile = join14(tmpdir(), `kf-coauthor-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`);
-  const amendBlock = `
-    if ! git log -1 --pretty=%B 2>/dev/null | grep -qF "${trailer}"; then
-      git log -1 --pretty=%B | git interpret-trailers --trailer "${trailer}" > "${tmpFile}" && git commit --amend -F "${tmpFile}" --no-edit && rm -f "${tmpFile}"
-    fi
-  `.trim();
+  if (!createsCommit && !isRebaseContinue && !mentionsGit) return unchanged;
+  if (movesHeadOnly) return unchanged;
+  const env2 = {
+    [COAUTHOR_TRAILER_ENV]: trailer,
+    [COAUTHOR_MSG_ENV]: join14(tmpdir(), `kf-coauthor-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`)
+  };
   if (createsCommit || isRebaseContinue) {
-    return `(${command}) && { ${amendBlock}; }`;
+    return { command: `(${command}) && { ${AMEND_BLOCK}
+}`, env: env2 };
   }
   const beforeHead = `git rev-parse HEAD 2>/dev/null || echo "NO_HEAD"`;
-  const afterCheck = `
-    _KF_AFTER_HEAD=$(git rev-parse HEAD 2>/dev/null || echo "NO_HEAD")
-    if [ "$_KF_BEFORE_HEAD" != "$_KF_AFTER_HEAD" ] && [ "$_KF_AFTER_HEAD" != "NO_HEAD" ] && git merge-base --is-ancestor "$_KF_BEFORE_HEAD" "$_KF_AFTER_HEAD" 2>/dev/null; then
-      ${amendBlock}
-    fi
-  `.trim();
-  return `_KF_BEFORE_HEAD=$(${beforeHead}); (${command}); _KF_EXIT=$?; [ $_KF_EXIT -eq 0 ] && { ${afterCheck}; }; exit $_KF_EXIT`;
+  const afterCheck = [
+    `_KF_AFTER_HEAD=$(git rev-parse HEAD 2>/dev/null || echo "NO_HEAD")`,
+    `if [ "$_KF_BEFORE_HEAD" != "$_KF_AFTER_HEAD" ] && [ "$_KF_AFTER_HEAD" != "NO_HEAD" ] && git merge-base --is-ancestor "$_KF_BEFORE_HEAD" "$_KF_AFTER_HEAD" 2>/dev/null; then`,
+    AMEND_BLOCK,
+    `fi`
+  ].join("\n");
+  return {
+    command: `_KF_BEFORE_HEAD=$(${beforeHead}); (${command}); _KF_EXIT=$?; [ $_KF_EXIT -eq 0 ] && { ${afterCheck}
+}; exit $_KF_EXIT`,
+    env: env2
+  };
 }
 async function runBash(args, ctx) {
   const timeout = Math.min(Math.max(1e3, args.timeout_ms ?? DEFAULT_TIMEOUT), MAX_TIMEOUT);
   const pushGuard = await guardGitPush(args.command, ctx);
   if (pushGuard) return pushGuard;
   const { shell, args: shellArgs, isPosix } = getShellCommand(ctx.shell);
-  const command = isPosix ? injectCoauthor(args.command, ctx.coauthor) : args.command;
+  const { command, env: coauthorEnv } = isPosix ? injectCoauthor(args.command, ctx.coauthor) : { command: args.command, env: {} };
   return new Promise((resolve15, reject) => {
     logger.debug("bash:spawn", { command: args.command.slice(0, 200), cwd: ctx.cwd, shell });
     const child = spawn(shell, [...shellArgs, command], {
       cwd: ctx.cwd,
       env: {
         ...process.env,
+        ...coauthorEnv,
         GIT_EDITOR: "true"
       }
     });
@@ -6771,7 +6929,7 @@ ${stderr}`
     });
   });
 }
-var DEFAULT_TIMEOUT, MAX_TIMEOUT, bashTool;
+var DEFAULT_TIMEOUT, MAX_TIMEOUT, bashTool, COAUTHOR_TRAILER_ENV, COAUTHOR_MSG_ENV, AMEND_BLOCK;
 var init_bash = __esm({
   "src/tools/bash.ts"() {
     "use strict";
@@ -6801,6 +6959,14 @@ var init_bash = __esm({
       render: (args) => ({ title: formatBashTitle(String(args.command ?? "")) }),
       run: (args, ctx) => runBash(args, ctx)
     };
+    COAUTHOR_TRAILER_ENV = "KF_COAUTHOR_TRAILER";
+    COAUTHOR_MSG_ENV = "KF_COAUTHOR_MSG";
+    AMEND_BLOCK = [
+      `if ! git log -1 --pretty=%B 2>/dev/null | grep -qF -- "$${COAUTHOR_TRAILER_ENV}"; then`,
+      `  git log -1 --pretty=%B | git interpret-trailers --trailer "$${COAUTHOR_TRAILER_ENV}" > "$${COAUTHOR_MSG_ENV}" && git commit --amend -F "$${COAUTHOR_MSG_ENV}" --no-edit; _KF_RC=$?;`,
+      `  rm -f "$${COAUTHOR_MSG_ENV}"; [ $_KF_RC -eq 0 ];`,
+      `fi`
+    ].join("\n");
   }
 });
 
@@ -8436,6 +8602,8 @@ async function runHotcellWorker(options) {
     throw new Error("Hotcell worker spend cap must be a positive number.");
   }
   if (options.signal?.aborted) return terminalResult(workerId, options.task, "cancelled", "Cancelled before worker setup.");
+  const workerPackage = resolveWorkerPackage(options.workerPackage);
+  const maxInputTokens = resolveWorkerInputBudget(model, options.budgetUsd, options.maxInputTokens);
   await acquireSlot(options.maxParallel ?? MAX_PARALLEL_WORKERS, options.signal);
   let cellId;
   let result2;
@@ -8487,9 +8655,43 @@ async function runHotcellWorker(options) {
       result2 = terminalResult(workerId, options.task, "failed", "Hotcell has no OpenRouter gateway route. Run `hotcell keys add openrouter` on the daemon host, then retry.", model);
       return result2;
     }
+    const setupCommand = [
+      ...FIND_REPO_ROOT,
+      `git -C "$REPO_ROOT" fetch --quiet origin ${shellQuote(repo.commit)}`,
+      `git -C "$REPO_ROOT" checkout --quiet --detach ${shellQuote(repo.commit)}`,
+      `npm install --prefix ${WORKER_PREFIX} --no-audit --no-fund --omit=optional --loglevel=error ${shellQuote(workerPackage)} 1>&2`
+    ].join(" && ");
+    const setup = await execute(command, ["exec", cellId, setupCommand, "--cwd", "/workspace"], {
+      cwd,
+      signal: options.signal,
+      timeoutMs: options.setupTimeoutMs ?? DEFAULT_SETUP_TIMEOUT_MS
+    });
+    if (setup.timedOut || setup.aborted) {
+      result2 = terminalResult(
+        workerId,
+        options.task,
+        options.signal?.aborted ? "cancelled" : "timed_out",
+        options.signal?.aborted ? "Cancelled during worker setup." : `Worker setup (checkout and ${workerPackage} install) timed out.`,
+        model
+      );
+      return result2;
+    }
+    if (setup.code !== 0) {
+      const detail = sanitizeHotcellDiagnostic(`${setup.stderr}
+${setup.stdout}`);
+      result2 = terminalResult(
+        workerId,
+        options.task,
+        "failed",
+        `Worker setup failed (exit ${setup.code}) while checking out the commit or installing ${workerPackage}${detail ? `: ${detail}` : "."}`,
+        model
+      );
+      return result2;
+    }
     const prompt = [
       "You are an isolated, read-only research worker. You cannot edit files, run shell commands, push, publish, or call write-capable integrations.",
       "Use the exact requested model and return concise findings, recommendations, and relevant file paths. Do not claim to have changed files.",
+      WORKER_REPORT_INSTRUCTIONS,
       `Mission:
 ${options.task}`,
       options.context ? `Coordinator context:
@@ -8497,16 +8699,11 @@ ${options.context}` : ""
     ].filter(Boolean).join("\n\n");
     const encodedPrompt = Buffer.from(prompt, "utf8").toString("base64");
     const shellCommand = [
-      'GIT_DIR="$(find /workspace -mindepth 2 -maxdepth 5 -name .git -print -quit)"',
-      'test -n "$GIT_DIR"',
-      'REPO_ROOT="${GIT_DIR%/.git}"',
-      `git -C "$REPO_ROOT" fetch --quiet origin ${shellQuote(repo.commit)}`,
-      `git -C "$REPO_ROOT" checkout --quiet --detach ${shellQuote(repo.commit)}`,
+      ...FIND_REPO_ROOT,
       'cd "$REPO_ROOT"',
-      "npm ci --no-audit --no-fund --loglevel=error 1>&2",
       'export OPENROUTER_BASE_URL="${OPENROUTER_BASE_URL%/}/v1"',
       `PROMPT="$(printf '%s' '${encodedPrompt}' | base64 -d)"`,
-      `node --import tsx "$REPO_ROOT/src/index.tsx" --format json --max-input-tokens 14000 --model ${shellQuote(model)} --worker-profile research -p "$PROMPT"`
+      `${WORKER_BIN} --format json --max-input-tokens ${maxInputTokens} --model ${shellQuote(model)} --worker-profile research -p "$PROMPT"`
     ].join(" && ");
     const execution = await execute(command, ["exec", cellId, shellCommand, "--cwd", "/workspace"], {
       cwd,
@@ -8524,14 +8721,16 @@ ${execution.stderr}`.toLowerCase();
     const executionDiagnostic = sanitizeHotcellDiagnostic(execution.stderr);
     const spendExhausted = /(?:http\s*)?402|spend cap|spending limit/.test(diagnostic);
     const status = execution.timedOut ? "timed_out" : execution.aborted ? options.signal?.aborted ? "cancelled" : "failed" : spendExhausted ? "spend_exhausted" : budgetExhausted ? "budget_exhausted" : execution.code === 0 ? "completed" : "failed";
-    const findings = summary ? [{ topic: "Research findings", summary: summary.slice(0, 4e4), confidence: "medium", sources: [], relevance: "high" }] : [];
+    const report = parseWorkerReport(summary);
+    const findings = report ? report.findings : summary ? [{ topic: "Research findings", summary: summary.slice(0, 4e4), confidence: "medium", sources: [], relevance: "high" }] : [];
     result2 = {
       workerId,
       status,
       task: options.task,
       findings,
       recommendations: [],
-      filesRead: [],
+      filesRead: report?.filesRead ?? [],
+      ...report ? { structured: true, openQuestions: report.openQuestions } : {},
       webSources: [],
       costUsd: metrics.costUsd ?? 0,
       tokensUsed: metrics.tokensUsed ?? tokensUsed,
@@ -8582,8 +8781,60 @@ async function getCleanRepository(cwd) {
   if (!/^(https:\/\/|git@|ssh:\/\/)/i.test(url) || /:\/\/[^/]*@/.test(url)) {
     throw new Error("Hotcell workers require a credential-free HTTPS or SSH origin URL that the Hotcell daemon can clone.");
   }
+  const pushed = (await execFileAsync("git", ["branch", "-r", "--contains", commit], { cwd: root })).stdout.trim();
+  if (!pushed) {
+    throw new Error(
+      `Hotcell workers clone from origin, but HEAD (${commit.slice(0, 8)}) is not on any remote-tracking branch. Push it first (git push -u origin ${ref}), or run git fetch if it was pushed elsewhere, then retry.`
+    );
+  }
   return { url, commit, ref };
 }
+function resolveWorkerInputBudget(model, budgetUsd, override) {
+  const clamp = (n) => Math.min(MAX_WORKER_INPUT_TOKENS, Math.max(MIN_WORKER_INPUT_TOKENS, Math.floor(n)));
+  const envOverride = Number(process.env.KIMIFLARE_WORKER_MAX_INPUT_TOKENS);
+  const explicit = override ?? (Number.isFinite(envOverride) && envOverride > 0 ? envOverride : void 0);
+  if (explicit !== void 0 && Number.isFinite(explicit) && explicit > 0) return clamp(explicit);
+  const perMillion = getModelOrInfer(model).pricing.inputPerMtok;
+  if (!(perMillion > 0) || !(budgetUsd > 0)) return DEFAULT_WORKER_INPUT_TOKENS;
+  return clamp(budgetUsd * INPUT_SHARE_OF_BUDGET * 1e6 / perMillion);
+}
+function parseWorkerReport(text) {
+  const blocks = [...text.matchAll(/```json\s*\n([\s\S]*?)```/g)];
+  const raw = blocks.at(-1)?.[1];
+  if (!raw) return null;
+  let parsed;
+  try {
+    parsed = JSON.parse(raw);
+  } catch {
+    return null;
+  }
+  if (!parsed || typeof parsed !== "object" || !Array.isArray(parsed.findings)) return null;
+  const obj = parsed;
+  const strings = (value, maxChars = 300) => Array.isArray(value) ? value.filter((v) => typeof v === "string" && v.trim().length > 0).slice(0, MAX_REPORT_LIST_ITEMS).map((v) => v.trim().slice(0, maxChars)) : [];
+  const findings = [];
+  for (const item of obj.findings.slice(0, MAX_REPORT_FINDINGS)) {
+    if (!item || typeof item !== "object") continue;
+    const f = item;
+    if (typeof f.summary !== "string" || !f.summary.trim()) continue;
+    const confidence = f.confidence === "high" || f.confidence === "low" ? f.confidence : "medium";
+    findings.push({
+      topic: typeof f.topic === "string" && f.topic.trim() ? f.topic.trim().slice(0, 120) : "Finding",
+      summary: f.summary.trim().slice(0, MAX_REPORT_SUMMARY_CHARS),
+      confidence,
+      sources: strings(f.files),
+      relevance: "high"
+    });
+  }
+  if (findings.length === 0) return null;
+  return { findings, openQuestions: strings(obj.openQuestions, 500), filesRead: strings(obj.filesRead) };
+}
+function resolveWorkerPackage(override) {
+  const spec = (override ?? process.env.KIMIFLARE_WORKER_PACKAGE ?? `${PACKAGE_NAME}@${getAppVersion()}`).trim();
+  if (!spec || !PACKAGE_SPEC.test(spec)) {
+    throw new Error(`Invalid Hotcell worker package spec: ${JSON.stringify(spec)}.`);
+  }
+  return spec;
+}
 function parseCellId(output) {
   const uuid = output.match(/\b[a-f0-9]{8}(?:-[a-f0-9]{4,}){0,4}\b/gi)?.at(-1);
   if (uuid) return uuid;
@@ -8734,22 +8985,65 @@ function releaseSlot() {
   runningWorkers = Math.max(0, runningWorkers - 1);
   for (const wake of workerQueue.slice()) wake();
 }
-var execFileAsync, MAX_OUTPUT_BYTES, MAX_PROMPT_CHARS, DEFAULT_CELL_TIMEOUT_MS, MAX_PARALLEL_WORKERS, runningWorkers, workerQueue;
+var execFileAsync, MAX_OUTPUT_BYTES, MAX_PROMPT_CHARS, DEFAULT_CELL_TIMEOUT_MS, DEFAULT_SETUP_TIMEOUT_MS, WORKER_PREFIX, WORKER_BIN, PACKAGE_SPEC, FIND_REPO_ROOT, MAX_PARALLEL_WORKERS, MIN_WORKER_INPUT_TOKENS, MAX_WORKER_INPUT_TOKENS, DEFAULT_WORKER_INPUT_TOKENS, INPUT_SHARE_OF_BUDGET, MAX_REPORT_FINDINGS, MAX_REPORT_SUMMARY_CHARS, MAX_REPORT_LIST_ITEMS, runningWorkers, workerQueue, WORKER_REPORT_INSTRUCTIONS;
 var init_hotcell_worker = __esm({
   "src/tools/hotcell-worker.ts"() {
     "use strict";
     init_client();
+    init_version();
+    init_registry();
     execFileAsync = promisify2(execFile2);
     MAX_OUTPUT_BYTES = 1e6;
     MAX_PROMPT_CHARS = 4e4;
     DEFAULT_CELL_TIMEOUT_MS = 3e5;
+    DEFAULT_SETUP_TIMEOUT_MS = 3e5;
+    WORKER_PREFIX = "/tmp/autopilot-worker";
+    WORKER_BIN = `${WORKER_PREFIX}/node_modules/.bin/${CLI_NAME}`;
+    PACKAGE_SPEC = /^[\w@./:+~^=<>#-]+$/;
+    FIND_REPO_ROOT = [
+      'GIT_DIR="$(find /workspace -mindepth 2 -maxdepth 5 -name .git -print -quit)"',
+      'test -n "$GIT_DIR"',
+      'REPO_ROOT="${GIT_DIR%/.git}"'
+    ];
     MAX_PARALLEL_WORKERS = 3;
+    MIN_WORKER_INPUT_TOKENS = 6e4;
+    MAX_WORKER_INPUT_TOKENS = 3e6;
+    DEFAULT_WORKER_INPUT_TOKENS = 4e5;
+    INPUT_SHARE_OF_BUDGET = 0.8;
+    MAX_REPORT_FINDINGS = 12;
+    MAX_REPORT_SUMMARY_CHARS = 2e3;
+    MAX_REPORT_LIST_ITEMS = 25;
     runningWorkers = 0;
     workerQueue = [];
+    WORKER_REPORT_INSTRUCTIONS = [
+      "Finish your answer with one fenced ```json block, and nothing after it, of exactly this shape:",
+      '{"findings":[{"topic":"short title","summary":"what you found and why it matters","files":["path/to/file.ts:123"],"confidence":"high|medium|low"}],"openQuestions":["what you could not determine"],"filesRead":["path/to/file.ts"]}',
+      "Cite repository-relative file paths (with line numbers where useful) for every finding. Use low confidence for anything you did not verify in the code."
+    ].join("\n");
   }
 });
 
 // src/tools/spawn-worker.ts
+function formatWorkerResult(result2, model) {
+  const lines = [
+    "Local Hotcell worker " + result2.status + (result2.status === "budget_exhausted" ? " (partial result)" : "") + ".",
+    "Model: " + (result2.model ?? model) + " \xB7 Tokens: " + result2.tokensUsed.toLocaleString() + " \xB7 Cost: " + (result2.costUsd > 0 ? result2.costUsd.toFixed(4) + " total Hotcell cost" : "unavailable from Hotcell stats") + ".",
+    ...result2.findings.map((finding) => "\n## " + finding.topic + (result2.structured ? " (" + finding.confidence + " confidence)" : "") + "\n" + finding.summary + (finding.sources.length ? "\nFiles: " + finding.sources.join(", ") : ""))
+  ];
+  if (result2.openQuestions?.length) {
+    lines.push("\n## Open questions\n" + result2.openQuestions.map((q) => "- " + q).join("\n"));
+  }
+  if (result2.filesRead.length) {
+    lines.push("\nFiles read: " + result2.filesRead.join(", "));
+  }
+  if (!result2.structured && result2.findings.length) {
+    lines.push("\n(Worker did not return a structured report; findings above are its raw answer and are unverified.)");
+  }
+  if (result2.status === "budget_exhausted") {
+    lines.push("\nWorker input-token budget was exhausted; review partial findings before relying on them.");
+  }
+  return lines.join("\n");
+}
 function textOutput(content) {
   const bytes = Buffer.byteLength(content, "utf8");
   return { content, rawBytes: bytes, reducedBytes: bytes };
@@ -8771,7 +9065,8 @@ var init_spawn_worker = __esm({
       name: "spawn_worker",
       description: [
         "Spawn a read-only research worker in a local Hotcell sandbox. Hotcell is the only worker backend.",
-        "Workers inspect the committed repository snapshot and cannot edit files; implement any changes in the coordinator."
+        "Workers inspect the committed repository snapshot and cannot edit files; implement any changes in the coordinator.",
+        "To run several workers in parallel, emit all of their spawn_worker calls in the same response; the user approves them together."
       ].join(" "),
       parameters: {
         type: "object",
@@ -8805,10 +9100,22 @@ var init_spawn_worker = __esm({
         additionalProperties: false
       },
       needsPermission: true,
-      render: (args) => ({
-        title: "spawn_worker (Hotcell plan)",
-        body: args.task.slice(0, 200)
-      }),
+      // Workers run in isolated Hotcell cells and never touch the workspace, so
+      // several can run at once; the Hotcell slot semaphore bounds concurrency.
+      concurrent: true,
+      render: (args) => {
+        const batch = args.batch;
+        if (Array.isArray(batch) && batch.length > 1) {
+          return {
+            title: `spawn_worker \xD7 ${batch.length} (Hotcell plan, run in parallel)`,
+            body: batch.map((item, i) => `${i + 1}. ${String(item.task ?? "").slice(0, 160)}`).join("\n")
+          };
+        }
+        return {
+          title: "spawn_worker (Hotcell plan)",
+          body: args.task.slice(0, 200)
+        };
+      },
       async run(args, ctx) {
         const cfg = await loadConfig().catch(() => null);
         if (!cfg?.openrouterApiKey) {
@@ -8836,6 +9143,7 @@ var init_spawn_worker = __esm({
             model,
             budgetUsd,
             timeoutMs,
+            setupTimeoutMs: readNumberEnv2("KIMIFLARE_WORKER_SETUP_TIMEOUT_MS"),
             maxParallel: cfg.workerMaxParallel,
             cwd: ctx.cwd,
             signal: ctx.signal
@@ -8846,15 +9154,7 @@ var init_spawn_worker = __esm({
         if (result2.status !== "completed" && result2.status !== "budget_exhausted") {
           throw new Error("Local Hotcell worker " + result2.status + ": " + (result2.error ?? "unknown error"));
         }
-        const lines = [
-          "Local Hotcell worker " + result2.status + (result2.status === "budget_exhausted" ? " (partial result)" : "") + ".",
-          "Model: " + (result2.model ?? model) + " \xB7 Tokens: " + result2.tokensUsed.toLocaleString() + " \xB7 Cost: " + (result2.costUsd > 0 ? result2.costUsd.toFixed(4) + " total Hotcell cost" : "unavailable from Hotcell stats") + ".",
-          ...result2.findings.map((finding) => "\n## " + finding.topic + "\n" + finding.summary)
-        ];
-        if (result2.status === "budget_exhausted") {
-          lines.push("\nWorker input-token budget was exhausted; review partial findings before relying on them.");
-        }
-        return textOutput(lines.join("\n"));
+        return textOutput(formatWorkerResult(result2, model));
       }
     };
   }
@@ -10206,9 +10506,9 @@ var init_store = __esm({
             throw new Error(`Invalid run transition: ${current.status} -> ${status}`);
           }
           const now2 = Date.now();
-          const startedAt = status === "running" ? current.startedAt ?? now2 : current.startedAt;
+          const startedAt2 = status === "running" ? current.startedAt ?? now2 : current.startedAt;
           const finishedAt = ["completed", "failed", "cancelled", "interrupted_unknown"].includes(status) ? now2 : null;
-          this.db.prepare(`UPDATE runs SET status = ?, stop_reason = ?, updated_at = ?, started_at = ?, finished_at = ? WHERE id = ?`).run(status, reason ?? null, now2, startedAt, finishedAt, id);
+          this.db.prepare(`UPDATE runs SET status = ?, stop_reason = ?, updated_at = ?, started_at = ?, finished_at = ? WHERE id = ?`).run(status, reason ?? null, now2, startedAt2, finishedAt, id);
           this.appendEvent(id, "state", null, null, { status, reason: reason ?? null }, now2);
           return this.getRun(id);
         });
@@ -11410,13 +11710,13 @@ function directiveFor(kind) {
     return "The user explicitly asked not to delegate. Do not call spawn_worker; handle the task in the main session and respect any stated dependencies.";
   }
   if (kind === "explicit-delegate") {
-    return "The user explicitly requested subagents. Before acting, split the request into bounded missions that can make progress independently; use spawn_worker in plan mode for useful parallel read-only research, wait for the workers, and synthesize their findings before dependent implementation. Do not delegate tightly coupled or sequential steps. The tool's normal permission prompt and all configured safety limits still apply.";
+    return "The user explicitly requested subagents. Before acting, split the request into bounded missions that can make progress independently; use spawn_worker in plan mode for useful parallel read-only research (emit all independent spawn_worker calls in the same response so they run concurrently), wait for the workers, and synthesize their findings before dependent implementation. Do not delegate tightly coupled or sequential steps. The tool's normal permission prompt and all configured safety limits still apply.";
   }
   if (kind === "suggest") {
     return "The harness sees a possible parallel-work opportunity, but the user did not request delegation. Use spawn_worker only if you can define bounded, independent missions that materially help. The tool's normal permission prompt is the required user confirmation; never assume approval. Keep dependent implementation in the coordinator.";
   }
   if (kind === "auto-delegate") {
-    return "Proactive subagent delegation is enabled for this substantial task. First assess its structure, dependencies, and coordination cost. If two or more bounded research tracks can make meaningful progress independently, delegate them with spawn_worker in plan mode in parallel, wait for their results, and synthesize the findings before doing dependent work in the coordinator. Keep trivial, tightly coupled, and sequential work local. Worker calls still require normal permission and local Hotcell with OpenRouter, a clean committed repository, spend, concurrency, timeout, and read-only restrictions; never bypass them. If a worker cannot run or fails, explain that plainly and continue locally without implying delegation succeeded.";
+    return "Proactive subagent delegation is enabled for this substantial task. First assess its structure, dependencies, and coordination cost. If two or more bounded research tracks can make meaningful progress independently, delegate them with spawn_worker in plan mode by emitting all of those calls in the same response so they run concurrently, wait for their results, and synthesize the findings before doing dependent work in the coordinator. Keep trivial, tightly coupled, and sequential work local. Worker calls still require normal permission approval (one prompt covers a batch) and local Hotcell with OpenRouter, a clean pushed repository, spend cap, concurrency, timeout, and read-only restrictions; never bypass them. If a worker cannot run or fails, explain that plainly and continue locally without implying delegation succeeded.";
   }
   return void 0;
 }
@@ -15273,6 +15573,7 @@ async function runAgentTurnForSession(active, config2, allowAll) {
       return "deny";
     }
   };
+  const compaction = sessionFileCompactionTarget(sessionFile);
   try {
     if (run && runStore) {
       const tools = ALL_TOOLS.filter((tool) => active.allowedTools?.has(tool.name));
@@ -15292,6 +15593,7 @@ async function runAgentTurnForSession(active, config2, allowAll) {
         codeMode: false,
         allowDirectPush: config2.allowDirectPush,
         preferPullRequests: config2.preferPullRequests,
+        compaction,
         callbacks,
         onIterationEnd: async (updatedMessages) => {
           sessionFile.messages = updatedMessages;
@@ -15318,6 +15620,7 @@ async function runAgentTurnForSession(active, config2, allowAll) {
         codeMode: config2.codeMode,
         allowDirectPush: config2.allowDirectPush,
         preferPullRequests: config2.preferPullRequests,
+        compaction,
         callbacks
       });
     }
@@ -15374,6 +15677,7 @@ var init_routes = __esm({
     init_system_prompt();
     init_executor();
     init_sessions();
+    init_context_budget();
     init_logger();
     init_sse2();
     init_openapi();
@@ -17913,6 +18217,7 @@ async function createAgentSession(opts2) {
     { role: "system", content: systemPrompt },
     ...sessionFile.messages.filter((m) => m.role !== "system")
   ];
+  const isNewSession = sessionFile.messages.length === 0;
   const session = new InternalSession({
     sessionFile,
     cwd,
@@ -17928,6 +18233,11 @@ async function createAgentSession(opts2) {
     hooks,
     provider: opts2.provider
   });
+  if (isNewSession) {
+    await session.save().catch((err) => {
+      logger.warn("sdk:initial_save_failed", { sessionId: session.sessionId, error: err.message });
+    });
+  }
   return { session };
 }
 function costLookup(config2, meta) {
@@ -17951,6 +18261,7 @@ var init_session = __esm({
     init_requesty_catalog();
     init_custom_endpoint();
     init_logger();
+    init_context_budget();
     init_config2();
     init_permissions();
     InternalSession = class {
@@ -17986,8 +18297,15 @@ var init_session = __esm({
         turnCount: 0
       };
       disposed = false;
+      createdAt;
+      /** Persisted compaction state (session state + archived artifacts). */
+      compactionState;
+      compaction;
       constructor(opts2) {
         this.sessionId = opts2.sessionFile.id;
+        this.createdAt = opts2.sessionFile.createdAt;
+        this.compactionState = { sessionState: opts2.sessionFile.sessionState, artifactStore: opts2.sessionFile.artifactStore };
+        this.compaction = sessionFileCompactionTarget(this.compactionState);
         this.cwd = opts2.cwd;
         this.messages = opts2.messages;
         this.config = opts2.config;
@@ -18059,15 +18377,12 @@ var init_session = __esm({
         this.isStreaming = true;
         this.emit({ type: "status", status: "streaming" });
         this.abortController = new AbortController();
+        let succeeded = false;
         try {
           await this.runTurn(mode, options?.maxToolIterations);
-          for (const followUp of this.followUpQueue) {
-            this.messages.push({ role: "user", content: followUp });
-          }
-          this.followUpQueue = [];
+          succeeded = true;
           this.isStreaming = false;
           this.emit({ type: "status", status: "idle" });
-          await this.save();
         } catch (err) {
           this.isStreaming = false;
           if (err.name === "AbortError") {
@@ -18082,6 +18397,22 @@ var init_session = __esm({
             this.emit({ type: "status", status: "error" });
             throw err;
           }
+        } finally {
+          await this.save().catch((saveErr) => {
+            logger.warn("sdk:save_failed", { sessionId: this.sessionId, error: saveErr.message });
+          });
+        }
+        if (succeeded) await this.runFollowUps(mode);
+      }
+      /** Run queued follow-ups as their own turns, in order. Steers that arrived
+       *  after the turn's last chance to see them become follow-ups too. */
+      async runFollowUps(mode) {
+        if (this.steerQueue.length > 0) {
+          this.followUpQueue.unshift(...this.steerQueue.splice(0));
+        }
+        while (this.followUpQueue.length > 0 && !this.disposed) {
+          const next = this.followUpQueue.shift();
+          await this.prompt(next, { mode });
         }
       }
       async steer(text) {
@@ -18134,9 +18465,11 @@ var init_session = __esm({
           id: this.sessionId,
           cwd: this.cwd,
           model: this.model,
-          createdAt: (/* @__PURE__ */ new Date()).toISOString(),
+          createdAt: this.createdAt,
           updatedAt: (/* @__PURE__ */ new Date()).toISOString(),
-          messages: this.messages
+          messages: this.messages,
+          sessionState: this.compactionState.sessionState,
+          artifactStore: this.compactionState.artifactStore
         });
       }
       dispose() {
@@ -18155,6 +18488,15 @@ var init_session = __esm({
           email: this.config.coauthorEmail || "kimiflare@proton.me"
         } : void 0;
         const callbacks = {
+          onCompacted: (info) => {
+            this.emit({
+              type: "context.compacted",
+              tokensBefore: info.tokensBefore,
+              tokensAfter: info.tokensAfter,
+              turnsRemoved: info.turnsRemoved,
+              artifactsArchived: info.artifactsArchived
+            });
+          },
           onAssistantStart: () => {
             const messageId = `msg_${this.nextMessageId++}`;
             this.currentAssistantMessageId = messageId;
@@ -18286,14 +18628,19 @@ var init_session = __esm({
           coauthor,
           sessionId: this.sessionId,
           memoryManager: this.memoryManager,
+          compaction: this.compaction,
           allowDirectPush: this.config.allowDirectPush,
           preferPullRequests: this.config.preferPullRequests,
-          onIterationEnd: async (messages, _signal) => {
-            for (const steerText of this.steerQueue) {
-              messages.push({ role: "user", content: steerText });
-            }
-            this.steerQueue = [];
-            return messages;
+          // Steering lands at the next tool boundary, or before the final answer
+          // ends the turn (the loop continues so the model can respond to it).
+          drainPendingInput: () => {
+            const steers = this.steerQueue.splice(0);
+            return steers.map((text) => {
+              const messageId = `msg_${this.nextMessageId++}`;
+              this.emit({ type: "message.start", messageId, role: "user" });
+              this.emit({ type: "message.end", messageId });
+              return { role: "user", content: text };
+            });
           },
           onFileChange: (path, content) => {
             if (content) {
@@ -23910,14 +24257,23 @@ var init_abort_scope = __esm({
 });
 
 // src/ui/spinner.tsx
+import { performance as performance2 } from "perf_hooks";
 import { useSyncExternalStore } from "react";
 import { Text as Text7 } from "ink";
 import { jsx as jsx8 } from "react/jsx-runtime";
+function animationTickAt(start, now2) {
+  return Math.max(0, Math.floor((now2 - start) / ANIMATION_TICK_MS));
+}
 function subscribe(listener) {
   listeners.add(listener);
   if (!timer2) {
+    tick = 0;
+    startedAt = performance2.now();
     timer2 = setInterval(() => {
-      tick++;
+      if (startedAt === null) return;
+      const nextTick = animationTickAt(startedAt, performance2.now());
+      if (nextTick === tick) return;
+      tick = nextTick;
       for (const l of listeners) l();
     }, ANIMATION_TICK_MS);
     timer2.unref?.();
@@ -23927,6 +24283,7 @@ function subscribe(listener) {
     if (listeners.size === 0 && timer2) {
       clearInterval(timer2);
       timer2 = null;
+      startedAt = null;
     }
   };
 }
@@ -23935,7 +24292,7 @@ function Spinner({ type = "dots" }) {
   const frames = FRAMES[type] ?? FRAMES.dots;
   return /* @__PURE__ */ jsx8(Text7, { children: frames[t % frames.length] });
 }
-var FRAMES, ANIMATION_TICK_MS, tick, timer2, listeners, getTick;
+var FRAMES, ANIMATION_TICK_MS, tick, startedAt, timer2, listeners, getTick;
 var init_spinner = __esm({
   "src/ui/spinner.tsx"() {
     "use strict";
@@ -23946,6 +24303,7 @@ var init_spinner = __esm({
     };
     ANIMATION_TICK_MS = 100;
     tick = 0;
+    startedAt = null;
     timer2 = null;
     listeners = /* @__PURE__ */ new Set();
     getTick = () => tick;
@@ -24644,7 +25002,7 @@ var init_tool_view = __esm({
         const id = setInterval(() => setNow(Date.now()), 1e3);
         return () => clearInterval(id);
       }, [evt.status, evt.startedAt]);
-      const statusIcon = evt.status === "queued" ? /* @__PURE__ */ jsx10(Text9, { color: theme.muted?.color ?? theme.info.color, children: "[ ]" }) : evt.status === "running" ? /* @__PURE__ */ jsx10(Text9, { color: theme.info.color, children: /* @__PURE__ */ jsx10(Spinner, { type: "dots" }) }) : evt.status === "error" ? /* @__PURE__ */ jsx10(Text9, { color: theme.palette.error, children: "[err]" }) : evt.status === "cancelled" ? /* @__PURE__ */ jsx10(Text9, { color: theme.info.color, children: "[x]" }) : evt.status === "rejected" ? /* @__PURE__ */ jsx10(Text9, { color: theme.palette.error, children: "[!]" }) : /* @__PURE__ */ jsx10(Text9, { color: theme.palette.success, children: "[ok]" });
+      const statusIcon = evt.status === "queued" ? /* @__PURE__ */ jsx10(Text9, { color: theme.muted?.color ?? theme.info.color, children: "[ ]" }) : evt.status === "running" ? /* @__PURE__ */ jsx10(Text9, { color: theme.info.color, children: /* @__PURE__ */ jsx10(Spinner, { type: "dots" }) }) : evt.status === "blocked" ? /* @__PURE__ */ jsx10(Text9, { color: theme.muted?.color ?? theme.info.color, dimColor: true, children: "[skip]" }) : evt.status === "error" ? /* @__PURE__ */ jsx10(Text9, { color: theme.palette.error, children: "[err]" }) : evt.status === "cancelled" ? /* @__PURE__ */ jsx10(Text9, { color: theme.info.color, children: "[x]" }) : evt.status === "rejected" ? /* @__PURE__ */ jsx10(Text9, { color: theme.palette.error, children: "[!]" }) : /* @__PURE__ */ jsx10(Text9, { color: theme.palette.success, children: "[ok]" });
       const rawTitle = evt.render?.title ?? `${evt.name}(${compactArgs(evt.args)})`;
       let title = humanizeToolTitle(evt.name, rawTitle, intentTier);
       if (evt.startedAt !== void 0 && (evt.status === "running" || evt.status === "done" || evt.status === "error")) {
@@ -24653,7 +25011,8 @@ var init_tool_view = __esm({
       const statusLabel = evt.status === "rejected" ? /* @__PURE__ */ jsx10(Text9, { color: theme.palette.error, children: " rejected" }) : evt.status === "cancelled" ? /* @__PURE__ */ jsx10(Text9, { color: theme.info.color, children: " cancelled" }) : null;
       const showItalic = evt.status === "queued" || evt.status === "cancelled" || evt.status === "rejected";
       const expand = Boolean(evt.expanded || verbose);
-      const lines = evt.result ? evt.result.split("\n") : [];
+      const visibleResult = evt.status === "blocked" ? void 0 : evt.result;
+      const lines = visibleResult ? visibleResult.split("\n") : [];
       const showLimit = verbose ? 200 : 20;
       return /* @__PURE__ */ jsxs7(Box8, { flexDirection: "column", marginLeft: 2, children: [
         /* @__PURE__ */ jsxs7(Text9, { children: [
@@ -24664,7 +25023,7 @@ var init_tool_view = __esm({
           isRepeated ? /* @__PURE__ */ jsx10(Text9, { color: theme.warn, children: " [warn] repeated" }) : null
         ] }),
         evt.render?.diff ? /* @__PURE__ */ jsx10(Box8, { marginLeft: 2, children: /* @__PURE__ */ jsx10(DiffView, { ...evt.render.diff }) }) : null,
-        evt.result && expand ? /* @__PURE__ */ jsxs7(
+        visibleResult && expand ? /* @__PURE__ */ jsxs7(
           Box8,
           {
             marginLeft: 2,
@@ -24683,9 +25042,9 @@ var init_tool_view = __esm({
             ]
           }
         ) : null,
-        evt.result && !expand && evt.status !== "running" && evt.status !== "queued" && evt.status !== "cancelled" && evt.status !== "rejected" ? /* @__PURE__ */ jsxs7(Text9, { color: theme.info.color, children: [
+        visibleResult && !expand && evt.status !== "running" && evt.status !== "queued" && evt.status !== "cancelled" && evt.status !== "rejected" ? /* @__PURE__ */ jsxs7(Text9, { color: theme.info.color, children: [
           "  ",
-          firstLine(evt.result)
+          firstLine(visibleResult)
         ] }) : null
       ] });
     });
@@ -25891,7 +26250,7 @@ var init_queue_plan_picker = __esm({
 import { useEffect as useEffect6, useRef as useRef4, useState as useState14 } from "react";
 import { Box as Box18, Text as Text19 } from "ink";
 import { jsx as jsx20, jsxs as jsxs17 } from "react/jsx-runtime";
-function TaskList({ tasks, startedAt, tokensDelta }) {
+function TaskList({ tasks, startedAt: startedAt2, tokensDelta }) {
   const theme = useTheme();
   const [now2, setNow] = useState14(Date.now());
   const [celebrating, setCelebrating] = useState14(false);
@@ -25899,7 +26258,7 @@ function TaskList({ tasks, startedAt, tokensDelta }) {
   const prevAllDoneRef = useRef4(false);
   tasksRef.current = tasks;
   useEffect6(() => {
-    if (startedAt === null) return;
+    if (startedAt2 === null) return;
     const id = setInterval(() => {
       setNow(Date.now());
       const current = tasksRef.current;
@@ -25908,7 +26267,7 @@ function TaskList({ tasks, startedAt, tokensDelta }) {
       }
     }, 1e3);
     return () => clearInterval(id);
-  }, [startedAt]);
+  }, [startedAt2]);
   useEffect6(() => {
     const allDone2 = tasks.length > 0 && tasks.every((t) => t.status === "completed");
     if (allDone2 && !prevAllDoneRef.current) {
@@ -25924,7 +26283,7 @@ function TaskList({ tasks, startedAt, tokensDelta }) {
   const total = tasks.length;
   const allDone = done === total;
   const header = active ? active.title : allDone ? `${total} tasks done` : `${done}/${total}`;
-  const elapsed = startedAt ? formatElapsed4(now2 - startedAt) : null;
+  const elapsed = startedAt2 ? formatElapsed4(now2 - startedAt2) : null;
   const headerStats = [elapsed, tokensDelta > 0 ? `\u2191 ${formatTokens3(tokensDelta)} tokens` : null].filter(Boolean).join(" \xB7 ");
   const visibleTasks = tasks.slice(0, MAX_VISIBLE);
   const hiddenPending = Math.max(0, tasks.length - visibleTasks.length);
@@ -31896,7 +32255,7 @@ var init_slash_commands = __esm({
           text: `subagent policy: ${cfg?.subagentPolicy ?? "auto"}
 usage: /subagents off|suggest|auto
 
-auto (default) lets the coordinator assess substantial tasks and delegate independent work; it is conditional, not a promise to spawn. suggest surfaces possible opportunities; off disables automatic suggestions. Explicit user instructions still win. Worker calls use normal permission checks and respect spend, concurrency, timeout, and read-only limits. Hotcell is the only worker backend: install the local CLI/daemon, configure its OpenRouter route, and use a clean committed Git checkout; workers are read-only.`
+auto (default) lets the coordinator assess substantial tasks and delegate independent work; it is conditional, not a promise to spawn. suggest surfaces possible opportunities; off disables automatic suggestions. Explicit user instructions still win. Worker calls use normal permission checks and respect spend, concurrency, timeout, and read-only limits. Hotcell is the only worker backend: install the local CLI/daemon, configure its OpenRouter route, and use a clean, pushed Git checkout; workers are read-only.`
         }]);
         return true;
       }
@@ -35203,7 +35562,7 @@ ${wcagWarnings.join("\n")}` }
   const onIterationEnd = useCallback8(
     async (messages, signal) => {
       if (signal.aborted) return messages;
-      if (!shouldCompact({ messages })) return messages;
+      if (!shouldCompact({ messages, tokenThreshold: cfg ? proactiveCompactionThreshold(cfg.model) : void 0 })) return messages;
       if (hooksManagerRef.current.hasEnabledHooks("PreCompact")) {
         void hooksManagerRef.current.fire(
           "PreCompact",
@@ -36012,6 +36371,16 @@ ${result2.output ?? "(output not captured)"}`,
       const turnScope = sessionScopeRef.current.createChild();
       activeScopeRef.current = turnScope;
       const sharedCallbacks = {
+        onCompacted: (info) => {
+          setEvents((e) => [
+            ...e,
+            {
+              kind: "info",
+              key: mkKey(),
+              text: `context budget: compacted ${info.turnsRemoved} older turn${info.turnsRemoved === 1 ? "" : "s"} (${info.tokensBefore} \u2192 ${info.tokensAfter} tokens, ${info.artifactsArchived} artifacts archived)`
+            }
+          ]);
+        },
         onAssistantStart: () => {
           const id = mkAssistantId();
           activeAsstIdRef.current = id;
@@ -36085,8 +36454,8 @@ ${result2.output ?? "(output not captured)"}`,
             endTurn();
           }
           updateTool(r.tool_call_id, {
-            status: !r.ok && typeof r.content === "string" && r.content.startsWith("Permission denied") ? "rejected" : r.ok ? "done" : "error",
-            result: r.content
+            status: r.guardrail ? "blocked" : !r.ok && typeof r.content === "string" && r.content.startsWith("Permission denied") ? "rejected" : r.ok ? "done" : "error",
+            result: r.guardrail ? void 0 : r.content
           });
         },
         onUsage: (u) => {
@@ -36118,6 +36487,7 @@ ${result2.output ?? "(output not captured)"}`,
         },
         askPermission: askForPermission,
         onGuardrail: (ev) => {
+          if (ev.kind === "loop_recovery" || ev.kind === "loop_stopped") return;
           setEvents((e) => [...e, { kind: "info", key: mkKey(), text: ev.message }]);
         },
         onKimiMdStale: () => {
@@ -36214,6 +36584,15 @@ ${result2.output ?? "(output not captured)"}`,
           allowDirectPush: cfg.allowDirectPush,
           preferPullRequests: cfg.preferPullRequests,
           onIterationEnd,
+          // Request preflight archives older turns here when the next request
+          // would not fit the model's input budget.
+          compaction: {
+            getState: () => sessionStateRef.current,
+            setState: (state) => {
+              sessionStateRef.current = state;
+            },
+            getStore: () => artifactStoreRef.current
+          },
           intentClassification: classification,
           delegationDirective: delegationGuidance.directive,
           sessionStartRecall: sessionStartRecallRef.current ?? void 0,
@@ -36244,7 +36623,7 @@ ${result2.output ?? "(output not captured)"}`,
                 return;
               }
               let didCompact = false;
-              if (shouldCompact({ messages: messagesRef.current })) {
+              if (shouldCompact({ messages: messagesRef.current, tokenThreshold: proactiveCompactionThreshold(overrideModel ?? cfg.model) })) {
                 if (hooksManagerRef.current.hasEnabledHooks("PreCompact")) {
                   void hooksManagerRef.current.fire(
                     "PreCompact",
@@ -37187,6 +37566,7 @@ var init_app = __esm({
     init_system_prompt();
     init_llm_summarize();
     init_artifact_compaction();
+    init_context_budget();
     init_session_state();
     init_executor();
     init_manager5();
@@ -37381,6 +37761,7 @@ init_system_prompt();
 init_executor();
 init_errors();
 init_sessions();
+init_context_budget();
 init_image();
 init_glob();
 init_permissions_evaluator();
@@ -37514,13 +37895,15 @@ async function runPrintMode(opts2) {
   if (isNew || sessionFile.messages.length === 0) {
     messages.push({
       role: "system",
-      content: buildSystemPrompt({ cwd, tools: workerTools, model: opts2.model, preferPullRequests: opts2.preferPullRequests, delegationDirective: delegationGuidance.directive })
+      content: buildSystemPrompt({ cwd, tools: workerTools, model: opts2.model, preferPullRequests: opts2.preferPullRequests })
     });
   } else {
-    const nonSystem = sessionFile.messages.filter((m) => m.role !== "system");
+    const nonSystem = sessionFile.messages.filter(
+      (m) => m.role !== "system" || typeof m.content === "string" && m.content.startsWith("[compiled session state]")
+    );
     messages.push({
       role: "system",
-      content: buildSystemPrompt({ cwd, tools: workerTools, model: opts2.model, preferPullRequests: opts2.preferPullRequests, delegationDirective: delegationGuidance.directive })
+      content: buildSystemPrompt({ cwd, tools: workerTools, model: opts2.model, preferPullRequests: opts2.preferPullRequests })
     });
     messages.push(...nonSystem);
   }
@@ -37686,6 +38069,9 @@ async function runPrintMode(opts2) {
       coauthor: opts2.coauthor !== false ? { name: opts2.coauthorName || "autopilot", email: opts2.coauthorEmail || "kimiflare@proton.me" } : void 0,
       allowDirectPush: opts2.allowDirectPush,
       preferPullRequests: opts2.preferPullRequests,
+      // --continue sessions grow across runs; archive older turns onto the
+      // session file (saved below) instead of failing on context size.
+      compaction: sessionFileCompactionTarget(sessionFile),
       callbacks
     });
   } catch (err) {