@kmugalkhod/lightcode 0.5.0 → 0.6.0

package/server.js

@@ -747,9 +747,10 @@ var init_schema = __esm(() => {
 
 // packages/ai/src/edit-file/schema.ts
 import { z as z6 } from "zod";
-var editFileDescription = "Edit a file via literal search and replace. Use this for focused, small edits.", editFileInputSchema, editFileProviderInputSchema, editFileOutputSchema;
+var editFileDescription, editFileInputSchema, editFileProviderInputSchema, editFileOutputSchema;
 var init_schema2 = __esm(() => {
   init_base_schemas();
+  editFileDescription = "Edit a file via search and replace. Use this for focused, small edits. " + "Line endings (CRLF/LF) and surrounding whitespace are matched tolerantly, so " + "copy the lines as shown by read_file. The search text must match exactly one " + "place \u2014 include enough surrounding context to make it unique, or set " + "replaceAll: true to change every occurrence.";
   editFileInputSchema = z6.object({
     path: boundedPathSchema,
     search: z6.string().min(1),
@@ -2225,126 +2226,7 @@ function selectCodingAgentIntentTools({
   if (!normalizedText || casualPromptPattern.test(normalizedText)) {
     return [];
   }
-  const selectedTools = [];
-  const hasContinuationIntent = includesAny(normalizedText, [
-    "continue",
-    "proceed",
-    "go on",
-    "go ahead",
-    "keep going",
-    "carry on",
-    "resume",
-    "finish",
-    "next step"
-  ]);
-  const hasCodeIntent = includesAny(normalizedText, [
-    "file",
-    "folder",
-    "code",
-    "repo",
-    "project",
-    "implement",
-    "fix",
-    "add",
-    "build",
-    "refactor",
-    "change",
-    "create",
-    "update",
-    "edit",
-    "delete",
-    "run",
-    "test",
-    "typecheck",
-    "debug",
-    "error",
-    "issue",
-    "problem",
-    "fails",
-    "failed",
-    "failing",
-    "bug",
-    "crash",
-    "slow",
-    "slowness",
-    "not working",
-    "epic",
-    "ticket"
-  ]);
-  const hasReadIntent = includesAny(normalizedText, [
-    "read",
-    "list",
-    "find",
-    "search",
-    "grep",
-    "glob",
-    "inspect",
-    "analyse",
-    "analyze",
-    "check"
-  ]);
-  const workspaceReferents = [
-    "this",
-    "these",
-    "that",
-    "here",
-    "code",
-    "codebase",
-    "repo",
-    "repository",
-    "project",
-    "file",
-    "files",
-    "directory",
-    "folder",
-    "app",
-    "application",
-    "module",
-    "my"
-  ];
-  const strongReviewIntent = includesAny(normalizedText, [
-    "review",
-    "audit",
-    "walk through",
-    "walkthrough",
-    "code review"
-  ]);
-  const softReviewIntent = includesAny(normalizedText, [
-    "explain",
-    "understand",
-    "summarize",
-    "summarise",
-    "overview",
-    "describe",
-    "explore",
-    "look at"
-  ]) && includesAny(normalizedText, workspaceReferents);
-  const hasReviewIntent = strongReviewIntent || softReviewIntent;
-  const hasWriteIntent = mode === "build" && (hasContinuationIntent || includesAny(normalizedText, [
-    "write",
-    "edit",
-    "change",
-    "create",
-    "implement",
-    "fix",
-    "update",
-    "delete",
-    "remove",
-    "rename"
-  ]));
-  const hasShellIntent = mode === "build" && (hasContinuationIntent || includesAny(normalizedText, [
-    "run",
-    "command",
-    "terminal",
-    "shell",
-    "bash",
-    "powershell",
-    "bun",
-    "tests",
-    "run test",
-    "run tests",
-    "typecheck"
-  ]));
+  const situational = [];
   const hasGitIntent = includesAny(normalizedText, [
     "git",
     "diff",
@@ -2355,95 +2237,64 @@ function selectCodingAgentIntentTools({
     "revision",
     "changes"
   ]);
-  const hasTodoIntent = includesAny(normalizedText, [
-    "todo",
-    "tasks",
-    "checklist",
-    "steps",
-    "epic",
-    "ticket"
-  ]);
-  const hasWebIntent = includesAny(normalizedText, [
-    "http://",
-    "https://",
-    "url",
-    "fetch",
-    "website",
-    "web",
-    "online",
-    "search internet",
-    "latest"
-  ]);
-  const explicitToolSearch = includesAny(normalizedText, [
-    "tool_search",
-    "tool search",
-    "find tool",
-    "available tool"
-  ]);
-  const hasSkillIntent = includesAny(normalizedText, [
-    "skill",
-    "skills",
-    "load skill"
-  ]);
-  const hasMcpIntent = includesAny(normalizedText, [
-    "mcp",
-    "resource",
-    "resources",
-    "mcp tool",
-    "mcp server"
-  ]);
-  if (hasCodeIntent || hasContinuationIntent || hasReviewIntent || hasReadIntent && !hasGitIntent && !hasWebIntent && !explicitToolSearch) {
-    for (const toolName of baseReadTools) {
-      pushUniqueTool(selectedTools, toolName);
-    }
-  }
   if (hasGitIntent) {
     if (includesAny(normalizedText, ["status", "state", "dirty", "working tree"]) || !includesAny(normalizedText, ["diff", "log", "show", "history", "commit"])) {
-      pushUniqueTool(selectedTools, "git_status");
+      pushUniqueTool(situational, "git_status");
     }
     if (includesAny(normalizedText, ["diff", "changes", "patch"])) {
-      pushUniqueTool(selectedTools, "git_diff");
+      pushUniqueTool(situational, "git_diff");
     }
     if (includesAny(normalizedText, ["log", "history", "commits"])) {
-      pushUniqueTool(selectedTools, "git_log");
+      pushUniqueTool(situational, "git_log");
     }
     if (includesAny(normalizedText, ["show", "revision", "commit"])) {
-      pushUniqueTool(selectedTools, "git_show");
+      pushUniqueTool(situational, "git_show");
     }
   }
-  if (hasTodoIntent && mode === "build") {
-    pushUniqueTool(selectedTools, "todo_write");
+  if (mode === "build" && includesAny(normalizedText, ["todo", "tasks", "checklist", "steps", "epic", "ticket"])) {
+    pushUniqueTool(situational, "todo_write");
   }
-  if (explicitToolSearch) {
-    pushUniqueTool(selectedTools, "tool_search");
+  if (includesAny(normalizedText, ["tool_search", "tool search", "find tool", "available tool"])) {
+    pushUniqueTool(situational, "tool_search");
   }
-  if (hasSkillIntent) {
-    pushUniqueTool(selectedTools, "skill");
+  if (includesAny(normalizedText, ["skill", "skills", "load skill"])) {
+    pushUniqueTool(situational, "skill");
   }
-  if (hasMcpIntent) {
-    pushUniqueTool(selectedTools, "list_mcp_resources");
+  if (includesAny(normalizedText, ["mcp", "resource", "resources", "mcp tool", "mcp server"])) {
+    pushUniqueTool(situational, "list_mcp_resources");
     if (includesAny(normalizedText, ["read", "resource", "resources"])) {
-      pushUniqueTool(selectedTools, "read_mcp_resource");
+      pushUniqueTool(situational, "read_mcp_resource");
     }
     if (mode === "build" && includesAny(normalizedText, ["call", "execute", "run"])) {
-      pushUniqueTool(selectedTools, "call_mcp_tool");
+      pushUniqueTool(situational, "call_mcp_tool");
     }
   }
-  if (hasWriteIntent) {
-    pushUniqueTool(selectedTools, "write_file");
-    pushUniqueTool(selectedTools, "edit_file");
+  if (includesAny(normalizedText, [
+    "http://",
+    "https://",
+    "url",
+    "fetch",
+    "website",
+    "web",
+    "online",
+    "search internet",
+    "latest"
+  ])) {
+    pushUniqueTool(situational, "web_fetch");
+    pushUniqueTool(situational, "web_search");
   }
-  if (hasShellIntent) {
-    pushUniqueTool(selectedTools, "bash");
+  if (mode === "plan" && includesAny(normalizedText, ["question", "ask", "clarify"])) {
+    pushUniqueTool(situational, "request_user_input");
   }
-  if (hasWebIntent) {
-    pushUniqueTool(selectedTools, "web_fetch");
-    pushUniqueTool(selectedTools, "web_search");
+  const ordered = [...baseReadTools, ...situational];
+  if (mode === "build") {
+    ordered.push("write_file", "edit_file", "bash");
   }
-  if (mode === "plan" && includesAny(normalizedText, ["question", "ask", "clarify"])) {
-    pushUniqueTool(selectedTools, "request_user_input");
+  const selected = [];
+  for (const toolName of ordered) {
+    pushUniqueTool(selected, toolName);
   }
-  return selectedTools;
+  return selected.slice(0, maxProviderActiveTools);
 }
 function limitProviderActiveTools(tools, maxTools = maxProviderActiveTools) {
   if (tools.length <= maxTools) {
@@ -2714,17 +2565,33 @@ function normalizeContextOptimizerConfig(config) {
 }
 function resolveContextWindowTokens({
   config,
-  modelContextWindow
+  modelContextWindow,
+  endpointLimitTokens
+}) {
+  const resolved = config.contextWindowOverride ?? modelContextWindow ?? DEFAULT_CONTEXT_WINDOW_TOKENS;
+  if (typeof endpointLimitTokens === "number" && Number.isFinite(endpointLimitTokens) && endpointLimitTokens > 0) {
+    return Math.min(resolved, endpointLimitTokens);
+  }
+  return resolved;
+}
+function resolveInputBudgetTokens({
+  contextWindow,
+  reservedOutputTokens,
+  safetyFraction = DEFAULT_INPUT_BUDGET_SAFETY_FRACTION
 }) {
-  return config.contextWindowOverride ?? modelContextWindow ?? DEFAULT_CONTEXT_WINDOW_TOKENS;
+  const reserved = typeof reservedOutputTokens === "number" && Number.isFinite(reservedOutputTokens) && reservedOutputTokens > 0 ? reservedOutputTokens : 0;
+  const clampedFraction = Math.min(1, Math.max(0.1, safetyFraction));
+  const usable = Math.max(0, contextWindow - reserved);
+  return Math.max(0, Math.floor(usable * clampedFraction));
 }
-var DEFAULT_CONTEXT_WINDOW_TOKENS = 128000, contextOptimizerConfigSchema, resolvedContextOptimizerConfigSchema, defaultContextOptimizerConfig;
+var DEFAULT_CONTEXT_WINDOW_TOKENS = 128000, contextOptimizerConfigSchema, resolvedContextOptimizerConfigSchema, defaultContextOptimizerConfig, DEFAULT_INPUT_BUDGET_SAFETY_FRACTION = 0.92;
 var init_config3 = __esm(() => {
   contextOptimizerConfigSchema = z26.object({
     autoCompact: z26.boolean().optional(),
     compactAtFraction: z26.number().min(0.3).max(0.95).optional(),
     pruneAtFraction: z26.number().min(0.2).max(0.9).optional(),
     preserveRecentMessages: z26.number().int().min(2).max(100).optional(),
+    maxCoverageTokensPerCompaction: z26.number().int().min(2000).max(200000).optional(),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).optional(),
     summaryMaxChars: z26.number().int().min(200).max(50000).optional(),
     aggressivePruneWhenUncached: z26.boolean().optional(),
@@ -2738,6 +2605,7 @@ var init_config3 = __esm(() => {
     compactAtFraction: z26.number().min(0.3).max(0.95),
     pruneAtFraction: z26.number().min(0.2).max(0.9),
     preserveRecentMessages: z26.number().int().min(2).max(100),
+    maxCoverageTokensPerCompaction: z26.number().int().min(2000).max(200000),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).nullable(),
     summaryMaxChars: z26.number().int().min(200).max(50000),
     aggressivePruneWhenUncached: z26.boolean(),
@@ -2750,6 +2618,7 @@ var init_config3 = __esm(() => {
     compactAtFraction: 0.7,
     pruneAtFraction: 0.6,
     preserveRecentMessages: 6,
+    maxCoverageTokensPerCompaction: 12000,
     contextWindowOverride: null,
     summaryMaxChars: 4000,
     aggressivePruneWhenUncached: true,
@@ -2829,10 +2698,44 @@ var init_estimate = __esm(() => {
       totalTokens: z27.number().nonnegative().optional()
     }),
     modelId: z27.string().optional(),
-    finishReason: languageModelFinishReasonSchema.optional()
+    finishReason: languageModelFinishReasonSchema.optional(),
+    context: z27.object({
+      inputTokens: z27.number().nonnegative(),
+      contextWindow: z27.number().positive(),
+      compactedMessages: z27.number().nonnegative().optional()
+    }).optional()
   }).loose();
 });
 
+// packages/ai/src/common/output-utils.ts
+function omissionMarker(omitted) {
+  return `
+
+... [${omitted} characters omitted; re-read a line range with read_file or raise maxChars] ...
+
+`;
+}
+function truncateText(text, maxChars = DEFAULT_TOOL_TEXT_OUTPUT_CHARS) {
+  if (text.length <= maxChars) {
+    return {
+      text,
+      truncated: false
+    };
+  }
+  const markerReserve = omissionMarker(text.length).length;
+  const usable = Math.max(0, maxChars - markerReserve);
+  const headChars = Math.floor(usable * 0.7);
+  const tailChars = usable - headChars;
+  const head = text.slice(0, headChars);
+  const tail = tailChars > 0 ? text.slice(text.length - tailChars) : "";
+  const omitted = text.length - head.length - tail.length;
+  return {
+    text: `${head}${omissionMarker(omitted)}${tail}`,
+    truncated: true
+  };
+}
+var init_output_utils = () => {};
+
 // packages/ai/src/context/heuristic-summary.ts
 function truncateInline(text, maxChars) {
   const normalized = text.replace(/\s+/g, " ").trim();
@@ -3182,6 +3085,127 @@ var init_tier1_prune = __esm(() => {
   init_message_parts();
 });
 
+// packages/ai/src/context/fit-to-budget.ts
+function estimateStructuralTokens(messages) {
+  return messages.reduce((total, message) => total + estimateMessageTokens(message), 0);
+}
+function getPartState2(part) {
+  const state = isRecord3(part) ? Reflect.get(part, "state") : undefined;
+  return typeof state === "string" ? state : null;
+}
+function elideOutput(part, originalChars) {
+  const stub = {
+    lightcodeElided: true,
+    note: "Tool output elided to fit the model's context budget. Re-run the tool if this output is needed again.",
+    originalChars
+  };
+  return { ...part, output: stub };
+}
+function elideToolOutputsInMessage(message) {
+  let elided = 0;
+  const parts = message.parts.map((part) => {
+    if (!getToolNameFromPart(part) || getPartState2(part) !== "output-available") {
+      return part;
+    }
+    const output = isRecord3(part) ? Reflect.get(part, "output") : undefined;
+    if (output === undefined || isElidedToolOutput(output)) {
+      return part;
+    }
+    elided += 1;
+    return elideOutput(part, safeStringify(output).length);
+  });
+  return elided > 0 ? { message: { ...message, parts }, elided } : { message, elided: 0 };
+}
+function truncateTextPartsInMessage(message, capChars) {
+  let truncated = 0;
+  const parts = message.parts.map((part) => {
+    if (!isRecord3(part)) {
+      return part;
+    }
+    const text = Reflect.get(part, "text");
+    if (typeof text !== "string" || text.length <= capChars) {
+      return part;
+    }
+    const result = truncateText(text, capChars);
+    if (!result.truncated) {
+      return part;
+    }
+    truncated += 1;
+    return { ...part, text: result.text };
+  });
+  return truncated > 0 ? { message: { ...message, parts }, truncated } : { message, truncated: 0 };
+}
+function fitMessagesToBudget(messages, options) {
+  const budget = Math.max(0, Math.floor(options.inputBudgetTokens));
+  const preserveRecentMessages = Math.max(0, options.preserveRecentMessages);
+  let working = [...messages];
+  let elidedToolOutputs = 0;
+  let droppedMessages = 0;
+  let truncatedTextParts = 0;
+  const fits = () => estimateStructuralTokens(working) <= budget;
+  if (fits()) {
+    return {
+      messages: working,
+      elidedToolOutputs: 0,
+      droppedMessages: 0,
+      truncatedTextParts: 0,
+      estimatedTokens: estimateStructuralTokens(working),
+      fitted: false,
+      withinBudget: true
+    };
+  }
+  const preserveFrom = Math.max(0, working.length - preserveRecentMessages);
+  for (let index = 0;index < preserveFrom && !fits(); index += 1) {
+    const { message, elided } = elideToolOutputsInMessage(working[index]);
+    if (elided > 0) {
+      working[index] = message;
+      elidedToolOutputs += elided;
+    }
+  }
+  const firstIsSummary = isSummaryMessage(working[0]);
+  let dropFrom = firstIsSummary ? 1 : 0;
+  while (!fits() && working.length - dropFrom > preserveRecentMessages && dropFrom < working.length) {
+    working.splice(dropFrom, 1);
+    droppedMessages += 1;
+  }
+  if (!fits()) {
+    working = working.map((message) => {
+      const { message: elidedMessage, elided } = elideToolOutputsInMessage(message);
+      elidedToolOutputs += elided;
+      return elidedMessage;
+    });
+    let cap = MAX_TOOL_TEXT_OUTPUT_CHARS;
+    while (!fits() && cap >= MIN_TRUNCATION_CAP_CHARS) {
+      working = working.map((message) => {
+        const { message: truncatedMessage, truncated } = truncateTextPartsInMessage(message, cap);
+        truncatedTextParts += truncated;
+        return truncatedMessage;
+      });
+      cap = Math.floor(cap / 2);
+    }
+  }
+  return {
+    messages: working,
+    elidedToolOutputs,
+    droppedMessages,
+    truncatedTextParts,
+    estimatedTokens: estimateStructuralTokens(working),
+    fitted: elidedToolOutputs > 0 || droppedMessages > 0 || truncatedTextParts > 0,
+    withinBudget: fits()
+  };
+}
+function isSummaryMessage(message) {
+  return message?.id === contextSummaryMessageId;
+}
+var MIN_TRUNCATION_CAP_CHARS = 256;
+var init_fit_to_budget = __esm(() => {
+  init_output_utils();
+  init_estimate();
+  init_heuristic_summary();
+  init_message_parts();
+  init_tier1_prune();
+});
+
 // packages/ai/src/context/build-provider-view.ts
 function computeCompactionCutoffIndex(messages, preserveRecentMessages, afterAnchorIndex) {
   const candidate = messages.length - preserveRecentMessages;
@@ -3195,11 +3219,32 @@ function computeCompactionCutoffIndex(messages, preserveRecentMessages, afterAnc
   }
   return afterAnchorIndex;
 }
+function capCoverageToTokenBudget(coveredMessages, maxTokens) {
+  if (coveredMessages.length <= 1) {
+    return [...coveredMessages];
+  }
+  let total = 0;
+  const capped = [];
+  for (const message of coveredMessages) {
+    const messageTokens = estimateStructuralTokens([message]);
+    if (capped.length > 0 && total + messageTokens > maxTokens) {
+      break;
+    }
+    capped.push(message);
+    total += messageTokens;
+  }
+  return capped;
+}
 function buildProviderView(options) {
   const { messages, contextState, config } = options;
   const contextWindow = resolveContextWindowTokens({
     config,
-    modelContextWindow: options.modelContextWindow
+    modelContextWindow: options.modelContextWindow,
+    endpointLimitTokens: options.endpointLimitTokens
+  });
+  const inputBudgetTokens = resolveInputBudgetTokens({
+    contextWindow,
+    reservedOutputTokens: options.reservedOutputTokens
   });
   let anchorResolved = false;
   let afterAnchorIndex = 0;
@@ -3223,7 +3268,8 @@ function buildProviderView(options) {
   let tier1 = null;
   const uncachedActive = options.cacheActive === false && config.aggressivePruneWhenUncached;
   const pruneFraction = uncachedActive ? config.uncachedPruneAtFraction : config.pruneAtFraction;
-  if (estimate.tokens > pruneFraction * contextWindow) {
+  const triggerTokens = (currentEstimateTokens) => Math.max(currentEstimateTokens, estimateStructuralTokens(providerMessages), typeof options.measuredInputTokens === "number" ? options.measuredInputTokens : 0);
+  if (triggerTokens(estimate.tokens) > pruneFraction * inputBudgetTokens) {
     tier1 = pruneToolOutputs(providerMessages, {
       preserveRecentMessages: config.preserveRecentMessages,
       minOutputChars: uncachedActive ? config.uncachedPruneMinOutputChars : undefined,
@@ -3238,15 +3284,15 @@ function buildProviderView(options) {
     }
   }
   const compactionCutoffIndex = computeCompactionCutoffIndex(messages, config.preserveRecentMessages, afterAnchorIndex);
-  const coveredMessages = messages.slice(afterAnchorIndex, compactionCutoffIndex);
-  const overCompactThreshold = estimate.tokens > config.compactAtFraction * contextWindow;
+  const coveredMessages = capCoverageToTokenBudget(messages.slice(afterAnchorIndex, compactionCutoffIndex), config.maxCoverageTokensPerCompaction);
+  const overCompactThreshold = triggerTokens(estimate.tokens) > config.compactAtFraction * inputBudgetTokens;
   let compactionBlockedReason = null;
   if (overCompactThreshold) {
     if (coveredMessages.length === 0) {
       compactionBlockedReason = "not_enough_messages";
     } else if ((options.pendingInteractionCount ?? 0) > 0) {
       compactionBlockedReason = "pending_interactions";
-    } else if (hasUnresolvedToolWork(messages)) {
+    } else if (hasUnresolvedToolWork(coveredMessages)) {
       compactionBlockedReason = "unresolved_tool_work";
     }
   }
@@ -3254,6 +3300,7 @@ function buildProviderView(options) {
     providerMessages,
     estimate,
     contextWindow,
+    inputBudgetTokens,
     needsCompaction: overCompactThreshold && compactionBlockedReason === null,
     compactionBlockedReason,
     anchorResolved,
@@ -3265,6 +3312,7 @@ function buildProviderView(options) {
 var init_build_provider_view = __esm(() => {
   init_config3();
   init_estimate();
+  init_fit_to_budget();
   init_heuristic_summary();
   init_message_parts();
   init_tier1_prune();
@@ -3310,6 +3358,7 @@ var init_context = __esm(() => {
   init_config3();
   init_context_state();
   init_estimate();
+  init_fit_to_budget();
   init_heuristic_summary();
   init_message_parts();
   init_normalize_provider_messages();
@@ -3583,7 +3632,7 @@ var init_lightcode_config = __esm(() => {
     context: defaultContextOptimizerConfig,
     maxOutputTokens: 32768,
     maxSteps: 30,
-    maxRetries: 5,
+    maxRetries: 3,
     autoContinue: defaultAutoContinueConfig,
     headroom: defaultHeadroomConfig
   };
@@ -4097,6 +4146,10 @@ async function extractToolCalls(text, knownToolNames) {
   while ((match = DEEPSEEK_WRAPPER_RE.exec(text)) !== null) {
     segments.push({ start: match.index, end: match.index + match[0].length, call: null });
   }
+  FUNCTION_CALLS_WRAPPER_RE.lastIndex = 0;
+  while ((match = FUNCTION_CALLS_WRAPPER_RE.exec(text)) !== null) {
+    segments.push({ start: match.index, end: match.index + match[0].length, call: null });
+  }
   const covered = (idx) => segments.some((s) => idx >= s.start && idx < s.end);
   const orphanIdx = findOrphanMarker(text, covered);
   if (orphanIdx !== -1) {
@@ -4184,7 +4237,7 @@ async function parseBareTrailingJson(text, knownToolNames) {
     return null;
   return { start: idx, payload };
 }
-var TOOL_CALL_OPEN_MARKERS, MAX_MARKER_LENGTH, MARKER_CONFIRMATION_WINDOW = 256, BLOCK_CLOSE_MARKERS, CLOSED_BLOCK_RES, DEEPSEEK_WRAPPER_RE, PARAMETER_TAG_RE, PARAMETER_EQ_RE;
+var TOOL_CALL_OPEN_MARKERS, MAX_MARKER_LENGTH, MARKER_CONFIRMATION_WINDOW = 256, BLOCK_CLOSE_MARKERS, CLOSED_BLOCK_RES, DEEPSEEK_WRAPPER_RE, FUNCTION_CALLS_WRAPPER_RE, PARAMETER_TAG_RE, PARAMETER_EQ_RE;
 var init_xml_tool_call_parser = __esm(() => {
   init_tool_call_json_repair();
   TOOL_CALL_OPEN_MARKERS = [
@@ -4211,6 +4264,7 @@ var init_xml_tool_call_parser = __esm(() => {
     deepseek: /<\uFF5Ctool\u2581call\u2581begin\uFF5C>\s*(?:function)?\s*(?:<\uFF5Ctool\u2581sep\uFF5C>)?\s*([\w.-]+)\s*\n?([\s\S]*?)<\uFF5Ctool\u2581call\u2581end\uFF5C>/g
   };
   DEEPSEEK_WRAPPER_RE = /<\uFF5Ctool\u2581calls\u2581(?:begin|end)\uFF5C>/g;
+  FUNCTION_CALLS_WRAPPER_RE = /<\/?function_calls\s*>/gi;
   PARAMETER_TAG_RE = /<parameter\s+name=["']?([\w.-]+)["']?\s*>([\s\S]*?)(?:<\/parameter>|(?=<parameter\s)|$)/gi;
   PARAMETER_EQ_RE = /<parameter=([\w.-]+)>([\s\S]*?)(?=<parameter=|<\/function>|$)/gi;
 });
@@ -5099,6 +5153,14 @@ function isProviderBillingOrQuotaError(error) {
 function isContextOverflowMessage(message) {
   return contextOverflowPattern.test(message);
 }
+function extractContextLimitTokens(message) {
+  const match = contextLimitPattern.exec(message);
+  if (!match) {
+    return;
+  }
+  const value = Number.parseInt(match[1], 10);
+  return Number.isFinite(value) && value > 0 ? value : undefined;
+}
 function isAbortError(error) {
   return error instanceof Error && error.name === "AbortError" || normalizeErrorMessage(getErrorMessage(error)).includes("aborted");
 }
@@ -5139,7 +5201,8 @@ function classifyChatError(error) {
         kind,
         statusCode: error.statusCode,
         retryable,
-        message: message2
+        message: message2,
+        contextLimitTokens: kind === "context_overflow" ? extractContextLimitTokens(message2) : undefined
       };
     }
     return {
@@ -5153,7 +5216,12 @@ function classifyChatError(error) {
     return { kind: "aborted", retryable: false, message };
   }
   if (isContextOverflowMessage(message)) {
-    return { kind: "context_overflow", retryable: true, message };
+    return {
+      kind: "context_overflow",
+      retryable: true,
+      message,
+      contextLimitTokens: extractContextLimitTokens(message)
+    };
   }
   if (isProviderBillingOrQuotaError(error)) {
     return { kind: "billing", retryable: false, message };
@@ -5219,7 +5287,7 @@ function logChatDisconnectEvent({
     message: getErrorMessage(error)
   });
 }
-var chatFailureCounters, logger2, retryableStatusCodes, contextOverflowPattern;
+var chatFailureCounters, logger2, retryableStatusCodes, contextOverflowPattern, contextLimitPattern;
 var init_chat_observability = __esm(() => {
   init_src();
   chatFailureCounters = {
@@ -5238,6 +5306,7 @@ var init_chat_observability = __esm(() => {
   logger2 = createLogger("chat");
   retryableStatusCodes = new Set([408, 409, 429, 500, 502, 503, 504]);
   contextOverflowPattern = /(context window|context length|context_length|maximum context|prompt is too long|input is too long|too many tokens|exceeds? the (?:maximum|model'?s) (?:context|token)|max(?:imum)? input tokens)/i;
+  contextLimitPattern = /(?:maximum context (?:length|window)|max(?:imum)? input (?:length|tokens)|context (?:length|window) of)\D{0,40}?(\d{4,})/i;
 });
 
 // packages/db/generated/prisma/internal/class.ts
@@ -6133,9 +6202,6 @@ var init_resolve_within_workspace = __esm(() => {
   WORKSPACE = defaultWorkspaceContext.root;
 });
 
-// packages/ai/src/common/output-utils.ts
-var init_output_utils = () => {};
-
 // packages/ai/src/bash/runtime.ts
 import { exec } from "child_process";
 import { promisify } from "util";
@@ -6171,11 +6237,15 @@ var init_runtime5 = __esm(() => {
 import { createTwoFilesPatch } from "diff";
 var init_unified_diff = () => {};
 
+// packages/ai/src/edit-file/match.ts
+var init_match = () => {};
+
 // packages/ai/src/edit-file/runtime.ts
 var init_runtime6 = __esm(() => {
   init_runtime5();
   init_resolve_within_workspace();
   init_unified_diff();
+  init_match();
   init_schema2();
 });
 
@@ -6722,6 +6792,24 @@ var init_overflow_recovery = __esm(() => {
   overflowRecoveryRounds = new Map;
 });
 
+// apps/server/src/lib/learned-context-limit.ts
+function noteLearnedContextLimit(modelId, limitTokens) {
+  if (!modelId || typeof limitTokens !== "number" || !Number.isFinite(limitTokens) || limitTokens <= 0) {
+    return;
+  }
+  const existing = learnedLimits.get(modelId);
+  if (existing === undefined || limitTokens < existing) {
+    learnedLimits.set(modelId, Math.floor(limitTokens));
+  }
+}
+function getLearnedContextLimit(modelId) {
+  return learnedLimits.get(modelId) ?? null;
+}
+var learnedLimits;
+var init_learned_context_limit = __esm(() => {
+  learnedLimits = new Map;
+});
+
 // apps/server/src/lib/sse-heartbeat.ts
 function endsWithEventBoundary(chunk) {
   const length = chunk.length;
@@ -6947,6 +7035,16 @@ function getAnthropicProviderOptions(modelId) {
     }
   };
 }
+function getOpenRouterProviderOptions() {
+  return {
+    openrouter: {
+      provider: {
+        allow_fallbacks: true,
+        require_parameters: true
+      }
+    }
+  };
+}
 function resolveHeadroomRouting(config5, realBaseUrl) {
   const headroom = config5.headroom;
   if (!headroom?.enabled) {
@@ -7011,12 +7109,8 @@ function resolveOpenAITransportModel({
   if (isOpenRouter) {
     const referer = getEnvValue(env, "OPENROUTER_HTTP_REFERER");
     const title = getEnvValue(env, "OPENROUTER_APP_TITLE", "LIGHTCODE_APP_TITLE");
-    if (referer) {
-      headers["HTTP-Referer"] = referer;
-    }
-    if (title) {
-      headers["X-OpenRouter-Title"] = title;
-    }
+    headers["HTTP-Referer"] = referer ?? "https://github.com/kmugalkhod/lightcode";
+    headers["X-Title"] = title ?? "Lightcode";
   }
   const provider = createOpenAICompatible({
     name: config5.provider,
@@ -7033,6 +7127,7 @@ function resolveOpenAITransportModel({
     headroomRouted,
     model: withXmlToolCallSupport(provider(configuredModel)),
     needsToolCallDiscipline: true,
+    providerOptions: isOpenRouter ? getOpenRouterProviderOptions() : undefined,
     contextWindow: capabilities?.contextLength ?? getModelContextWindow(config5.provider, configuredModel),
     missingCredentialHints: apiKey ? [] : isOpenCodeZen ? ["Set OPENCODE_API_KEY to call OpenCode Zen models."] : isOpenRouter ? ["Set OPENROUTER_API_KEY to call OpenRouter models."] : [
       "Set LIGHTCODE_OPENAI_COMPATIBLE_API_KEY or OPENAI_API_KEY if your OpenAI-compatible endpoint requires authentication."
@@ -7553,6 +7648,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
   }
   const contextConfig = lightcodeConfigResult.config.context;
   const cacheActive = resolvedProviderModel.provider === "anthropic";
+  const reservedOutputTokens = resolveMaxOutputTokens(lightcodeConfigResult.config.maxOutputTokens, lightcodeConfigDefaults.maxOutputTokens, resolvedProviderModel.maxCompletionTokens);
+  const endpointLimitTokens = getLearnedContextLimit(chatModelId);
   const pendingInteractionCount = await countPendingChatInteractions(sessionId);
   let contextState = await loadSessionContextStateSafe(sessionId);
   let view = buildProviderView({
@@ -7560,44 +7657,56 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     contextState,
     config: contextConfig,
     modelContextWindow: resolvedProviderModel.contextWindow,
+    reservedOutputTokens,
+    endpointLimitTokens,
     pendingInteractionCount,
     cacheActive
   });
-  if (view.needsCompaction && contextConfig.autoCompact) {
-    try {
-      const estimatedTokensBefore = view.estimate.tokens;
-      const compaction = await compactSessionContext({
-        sessionId,
-        coveredMessages: view.coveredMessages,
-        previousState: contextState,
-        model: resolvedProviderModel.model,
-        modelId: chatModelId,
-        cwd,
-        config: contextConfig,
-        estimatedTokens: estimatedTokensBefore
-      });
-      contextState = compaction.state;
-      view = buildProviderView({
-        messages: validatedMessages,
-        contextState,
-        config: contextConfig,
-        modelContextWindow: resolvedProviderModel.contextWindow,
-        pendingInteractionCount,
-        cacheActive
-      });
-      logger8.info("context_auto_compacted", {
-        sessionId,
-        tier: compaction.state.tier,
-        usedFallback: compaction.usedFallback,
-        coveredMessageCount: compaction.state.coveredMessageCount,
-        estimatedTokensBefore,
-        estimatedTokensAfter: view.estimate.tokens
-      });
-    } catch (error) {
-      logger8.error("context_compaction_failed", {
-        sessionId,
-        error: getErrorMessage(error)
-      });
+  if (contextConfig.autoCompact) {
+    let compactionRound = 0;
+    while (view.needsCompaction && compactionRound < MAX_COMPACTION_ROUNDS) {
+      compactionRound += 1;
+      try {
+        const estimatedTokensBefore = view.estimate.tokens;
+        const compaction = await compactSessionContext({
+          sessionId,
+          coveredMessages: view.coveredMessages,
+          previousState: contextState,
+          model: resolvedProviderModel.model,
+          modelId: chatModelId,
+          cwd,
+          config: contextConfig,
+          estimatedTokens: estimatedTokensBefore
+        });
+        contextState = compaction.state;
+        view = buildProviderView({
+          messages: validatedMessages,
+          contextState,
+          config: contextConfig,
+          modelContextWindow: resolvedProviderModel.contextWindow,
+          reservedOutputTokens,
+          endpointLimitTokens,
+          pendingInteractionCount,
+          cacheActive
+        });
+        logger8.info("context_auto_compacted", {
+          sessionId,
+          round: compactionRound,
+          tier: compaction.state.tier,
+          usedFallback: compaction.usedFallback,
+          coveredMessageCount: compaction.state.coveredMessageCount,
+          estimatedTokensBefore,
+          estimatedTokensAfter: view.estimate.tokens,
+          stillNeedsCompaction: view.needsCompaction
+        });
+      } catch (error) {
+        logger8.error("context_compaction_failed", {
+          sessionId,
+          round: compactionRound,
+          error: getErrorMessage(error)
+        });
+        break;
+      }
     }
   }
   const overflowPreserve = getOverflowPreserveRecentMessages(sessionId);
@@ -7612,6 +7721,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
+        reservedOutputTokens,
+        endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive
       });
@@ -7633,6 +7744,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
+        reservedOutputTokens,
+        endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive
       });
@@ -7651,7 +7764,24 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
       });
     }
   }
-  const providerMessages = normalizeProviderMessages(view.providerMessages);
+  const fitted = fitMessagesToBudget(view.providerMessages, {
+    inputBudgetTokens: view.inputBudgetTokens,
+    preserveRecentMessages: contextConfig.preserveRecentMessages
+  });
+  if (fitted.fitted) {
+    logger8.warn("context_fit_to_budget", {
+      sessionId,
+      inputBudgetTokens: view.inputBudgetTokens,
+      contextWindow: view.contextWindow,
+      estimatedTokensBefore: estimateStructuralTokens(view.providerMessages),
+      estimatedTokensAfter: fitted.estimatedTokens,
+      elidedToolOutputs: fitted.elidedToolOutputs,
+      droppedMessages: fitted.droppedMessages,
+      truncatedTextParts: fitted.truncatedTextParts,
+      withinBudget: fitted.withinBudget
+    });
+  }
+  const providerMessages = normalizeProviderMessages(fitted.messages);
   let overflowNotedThisRequest = false;
   const noteOverflowForSession = () => {
     if (overflowNotedThisRequest) {
@@ -7745,7 +7875,12 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         totalTokens: part.totalUsage.totalTokens
       },
       modelId: chatModelId,
-      finishReason: part.finishReason
+      finishReason: part.finishReason,
+      context: {
+        inputTokens: view.estimate.tokens,
+        contextWindow: view.contextWindow,
+        compactedMessages: contextState?.coveredMessageCount ?? 0
+      }
     };
   };
   const environmentContext = await buildWorkspaceContext({ cwd });
@@ -7793,6 +7928,7 @@ ${environmentContext}`,
         onError: (error) => {
           const classified = classifyChatError(error);
           if (classified.kind === "context_overflow") {
+            noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
             noteOverflowForSession();
           }
           incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), { sessionId, statusCode: classified.statusCode });
@@ -7810,6 +7946,7 @@ ${environmentContext}`,
     const mapStreamError = (error) => {
       const classified = classifyChatError(error);
       if (classified.kind === "context_overflow") {
+        noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
         noteOverflowForSession();
       }
       incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), {
@@ -7864,6 +8001,7 @@ ${environmentContext}`,
   } catch (error) {
     const classified = classifyChatError(error);
     if (classified.kind === "context_overflow") {
+      noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
       noteContextOverflow(sessionId);
     }
     incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), {
@@ -7894,7 +8032,7 @@ ${environmentContext}`,
     }, 500);
   }
 }
-var logger8, recoverableDisconnectMessage = "Connection interrupted. Please retry or regenerate your last message.", providerBillingOrQuotaMessage, fastChatSystemPrompt, fastChatMaxOutputTokens = 512, fastChatRecentMessageCount = 6;
+var logger8, recoverableDisconnectMessage = "Connection interrupted. Please retry or regenerate your last message.", providerBillingOrQuotaMessage, fastChatSystemPrompt, fastChatMaxOutputTokens = 512, fastChatRecentMessageCount = 6, MAX_COMPACTION_ROUNDS = 3;
 var init_chat_stream = __esm(() => {
   init_src2();
   init_src();
@@ -7903,6 +8041,7 @@ var init_chat_stream = __esm(() => {
   init_chat_store();
   init_context_compaction();
   init_overflow_recovery();
+  init_learned_context_limit();
   init_sse_heartbeat();
   init_context_state_store();
   init_session_auto_title();