@kmugalkhod/lightcode 0.4.0 → 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),
@@ -1549,7 +1550,8 @@ var init_agent_tools = __esm(() => {
     permissionMode: permissionModeSchema.optional(),
     allowedTools: z21.array(codingAgentToolNameSchema).optional(),
     permissionRules: permissionRulesSchema.optional(),
-    sandbox: sandboxConfigSchema.optional()
+    sandbox: sandboxConfigSchema.optional(),
+    environmentContext: z21.string().optional()
   });
 });
 
@@ -1883,7 +1885,8 @@ function buildCodingAgentSystemPrompt({
   cwd,
   override,
   mode = defaultCodingAgentMode,
-  includeToolDiscipline = false
+  includeToolDiscipline = false,
+  environmentContext
 }) {
   const promptTemplate = override?.trim() ? override : defaultCodingAgentSystemPrompt;
   const normalizedPrompt = normalizePromptTemplate(promptTemplate, cwd);
@@ -1894,12 +1897,14 @@ function buildCodingAgentSystemPrompt({
 ` + `Mode purpose: ${modeDefinition.purpose}
 ` + `${modeDefinition.instructions}` + (includeToolDiscipline ? `
 
-${toolCallingDisciplineAddendum}` : "");
+${toolCallingDisciplineAddendum}` : "") + (environmentContext?.trim() ? `
+
+${environmentContext.trim()}` : "");
 }
 var cwdPromptPlaceholder = "{cwd}", defaultCodingAgentSystemPrompt, toolCallingDisciplineAddendum;
 var init_coding_agent_prompt = __esm(() => {
   init_coding_agent_modes();
-  defaultCodingAgentSystemPrompt = "You are a coding agent. Use tools for filesystem and codebase tasks instead of guessing, and only interact with files under this working directory: {cwd}. " + "Prefer dedicated tools \u2014 glob_search for paths, grep for text search, structured git tools for repository inspection \u2014 and bash only when those are not enough. " + `Use todo_write to keep multi-step work visible and current, inspect context before risky operations, and emit a brief, natural progress note before major tool actions and after important findings.
+  defaultCodingAgentSystemPrompt = "You are a coding agent working inside the user's project at this working directory: {cwd}. " + 'You can already see the project \u2014 proactively read it. When the user asks you to review, explain, or work on "my code"/"this", inspect the working directory yourself with list_files, glob_search, grep, and read_file; NEVER ask the user to paste or share code that lives in this directory \u2014 open it. ' + "Use tools for filesystem and codebase tasks instead of guessing, and only interact with files under the working directory. " + "Prefer dedicated tools \u2014 glob_search for paths, grep for text search, structured git tools for repository inspection \u2014 and bash only when those are not enough. " + `Use todo_write to keep multi-step work visible and current, inspect context before risky operations, and emit a brief, natural progress note before major tool actions and after important findings.
 
 ` + `You are an agent \u2014 keep working until the request is fully resolved. Never stop with a partial answer; if you say you will do something, do it in the same turn by calling a tool. Do not finish while any todo is pending or in_progress \u2014 end your turn only when the task is complete or you are blocked on input that only the user can provide.
 
@@ -2221,78 +2226,7 @@ function selectCodingAgentIntentTools({
   if (!normalizedText || casualPromptPattern.test(normalizedText)) {
     return [];
   }
-  const selectedTools = [];
-  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 hasWriteIntent = mode === "build" && includesAny(normalizedText, [
-    "write",
-    "edit",
-    "change",
-    "create",
-    "implement",
-    "fix",
-    "update",
-    "delete",
-    "remove",
-    "rename"
-  ]);
-  const hasShellIntent = mode === "build" && includesAny(normalizedText, [
-    "run",
-    "command",
-    "terminal",
-    "shell",
-    "bash",
-    "powershell",
-    "bun",
-    "tests",
-    "run test",
-    "run tests",
-    "typecheck"
-  ]);
+  const situational = [];
   const hasGitIntent = includesAny(normalizedText, [
     "git",
     "diff",
@@ -2303,95 +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 || 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) {
@@ -2594,7 +2497,8 @@ function createCodingAgent({
           cwd: options.cwd,
           override: promptOverride,
           mode,
-          includeToolDiscipline
+          includeToolDiscipline,
+          environmentContext: options.environmentContext
         })
       };
     }
@@ -2661,17 +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(),
@@ -2685,6 +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(),
@@ -2697,6 +2618,7 @@ var init_config3 = __esm(() => {
     compactAtFraction: 0.7,
     pruneAtFraction: 0.6,
     preserveRecentMessages: 6,
+    maxCoverageTokensPerCompaction: 12000,
     contextWindowOverride: null,
     summaryMaxChars: 4000,
     aggressivePruneWhenUncached: true,
@@ -2776,10 +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();
@@ -3129,6 +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;
@@ -3142,11 +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;
@@ -3170,7 +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,
@@ -3185,15 +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";
     }
   }
@@ -3201,6 +3300,7 @@ function buildProviderView(options) {
     providerMessages,
     estimate,
     contextWindow,
+    inputBudgetTokens,
     needsCompaction: overCompactThreshold && compactionBlockedReason === null,
     compactionBlockedReason,
     anchorResolved,
@@ -3212,6 +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();
@@ -3257,6 +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();
@@ -3486,8 +3588,8 @@ var init_lightcode_config = __esm(() => {
   defaultAutoContinueConfig = {
     enabled: true,
     maxAutoContinues: 50,
-    maxErrorRetries: 5,
-    stallTimeoutSeconds: 180
+    maxErrorRetries: 8,
+    stallTimeoutSeconds: 300
   };
   headroomConfigSchema = z28.object({
     enabled: z28.boolean().optional(),
@@ -3530,7 +3632,7 @@ var init_lightcode_config = __esm(() => {
     context: defaultContextOptimizerConfig,
     maxOutputTokens: 32768,
     maxSteps: 30,
-    maxRetries: 5,
+    maxRetries: 3,
     autoContinue: defaultAutoContinueConfig,
     headroom: defaultHeadroomConfig
   };
@@ -4044,6 +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) {
@@ -4131,7 +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 = [
@@ -4158,6 +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;
 });
@@ -5046,6 +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");
 }
@@ -5086,7 +5201,8 @@ function classifyChatError(error) {
         kind,
         statusCode: error.statusCode,
         retryable,
-        message: message2
+        message: message2,
+        contextLimitTokens: kind === "context_overflow" ? extractContextLimitTokens(message2) : undefined
       };
     }
     return {
@@ -5100,7 +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 };
@@ -5166,7 +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 = {
@@ -5185,6 +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
@@ -6080,9 +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";
@@ -6118,11 +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();
 });
 
@@ -6169,6 +6292,112 @@ var init_runtime8 = __esm(() => {
 // packages/ai/src/git/runtime.ts
 import { execFile as execFile2 } from "child_process";
 import { promisify as promisify3 } from "util";
+function isRecord7(value) {
+  return typeof value === "object" && value !== null;
+}
+function getStringProperty(record, key) {
+  const value = record[key];
+  return typeof value === "string" ? value : "";
+}
+function getExitCode(error) {
+  const code = error.code;
+  return typeof code === "number" ? code : null;
+}
+async function runGit(args, workspaceContext, maxBuffer = 10 * 1024 * 1024) {
+  try {
+    const result = await execFileAsync2("git", args, {
+      cwd: workspaceContext.root,
+      maxBuffer,
+      windowsHide: true
+    });
+    return {
+      ok: true,
+      stdout: result.stdout,
+      stderr: result.stderr,
+      exitCode: 0
+    };
+  } catch (error) {
+    const record = isRecord7(error) ? error : {};
+    const code = record.code;
+    const codeText = typeof code === "string" ? code : "git_failed";
+    const message = getStringProperty(record, "message") || "Git command failed.";
+    return {
+      ok: false,
+      error: {
+        code: codeText,
+        message,
+        exitCode: getExitCode(record),
+        stdout: getStringProperty(record, "stdout"),
+        stderr: getStringProperty(record, "stderr")
+      }
+    };
+  }
+}
+function parseBranchLine(line) {
+  const branchLine = line.replace(/^##\s*/, "");
+  const [branchSegment, trackingSegment] = branchLine.split("...");
+  const branch = branchSegment && branchSegment !== "No commits yet on " ? branchSegment.replace(/^No commits yet on\s+/, "") : null;
+  const trackingMatch = trackingSegment?.match(/^([^\s]+)(?:\s+\[(.*)\])?/);
+  const upstream = trackingMatch?.[1] ?? null;
+  const trackingDetails = trackingMatch?.[2] ?? "";
+  const aheadMatch = trackingDetails.match(/ahead\s+(\d+)/);
+  const behindMatch = trackingDetails.match(/behind\s+(\d+)/);
+  return {
+    branch,
+    upstream,
+    ahead: aheadMatch ? Number(aheadMatch[1]) : 0,
+    behind: behindMatch ? Number(behindMatch[1]) : 0
+  };
+}
+function parseStatusEntries(stdout, maxEntries) {
+  const lines = stdout.split(/\r?\n/).filter((line) => line.length > 0);
+  const branchInfo = lines[0]?.startsWith("##") ? parseBranchLine(lines[0]) : { branch: null, upstream: null, ahead: 0, behind: 0 };
+  const statusLines = lines.filter((line) => !line.startsWith("##"));
+  const entries = statusLines.slice(0, maxEntries).map((line) => {
+    const indexStatus = line[0] === " " ? "" : line[0] ?? "";
+    const workingTreeStatus = line[1] === " " ? "" : line[1] ?? "";
+    const rawPath = line.slice(3);
+    const renameParts = rawPath.split(" -> ");
+    const originalPath = renameParts.length > 1 ? renameParts[0] : null;
+    const filePath = renameParts.length > 1 ? renameParts[1] : rawPath;
+    return {
+      path: filePath,
+      indexStatus,
+      workingTreeStatus,
+      originalPath
+    };
+  });
+  return {
+    ...branchInfo,
+    entries,
+    totalEntries: statusLines.length,
+    truncated: statusLines.length > entries.length
+  };
+}
+async function executeGitStatus(input, workspaceContext = getDefaultWorkspaceContext()) {
+  const parsedInput = gitStatusInputSchema.parse(input);
+  const result = await runGit(["status", "--porcelain=v1", "-b", "--untracked-files=all", "--", "."], workspaceContext);
+  if (!result.ok) {
+    return gitStatusOutputSchema.parse({
+      ok: false,
+      cwd: workspaceContext.root,
+      branch: null,
+      upstream: null,
+      ahead: 0,
+      behind: 0,
+      entries: [],
+      totalEntries: 0,
+      truncated: false,
+      error: result.error
+    });
+  }
+  return gitStatusOutputSchema.parse({
+    ok: true,
+    cwd: workspaceContext.root,
+    ...parseStatusEntries(result.stdout, parsedInput.maxEntries),
+    error: null
+  });
+}
 var execFileAsync2;
 var init_runtime9 = __esm(() => {
   init_resolve_within_workspace();
@@ -6563,6 +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;
@@ -6788,6 +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) {
@@ -6852,12 +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,
@@ -6874,6 +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."
@@ -6939,6 +7193,7 @@ var init_provider_registry = __esm(() => {
 var exports_runtime_config = {};
 __export(exports_runtime_config, {
   resolvedProviderModel: () => resolvedProviderModel,
+  reloadRuntimeConfig: () => reloadRuntimeConfig,
   lightcodeConfigResult: () => lightcodeConfigResult,
   initializeModelCapabilities: () => initializeModelCapabilities,
   disableHeadroomRouting: () => disableHeadroomRouting,
@@ -6999,6 +7254,31 @@ function applyModelSelection({
   });
   return configStatus;
 }
+function reloadRuntimeConfig() {
+  const nextConfigResult = loadLightcodeConfig({
+    cwd: process.cwd(),
+    env: Bun.env
+  });
+  const nextResolved = resolveConfiguredProviderModel({
+    config: nextConfigResult.config,
+    env: Bun.env
+  });
+  const nextAgent = buildCodingAgent(nextConfigResult.config, nextResolved);
+  lightcodeConfigResult = nextConfigResult;
+  resolvedProviderModel = nextResolved;
+  chatModelId = nextResolved.resolvedModelId;
+  codingAgent = nextAgent;
+  configStatus = createConfigStatus({
+    config: nextConfigResult.config,
+    loadedFiles: nextConfigResult.loadedFiles,
+    resolvedProviderModel: nextResolved
+  });
+  logger5.info("runtime_config_reloaded", {
+    provider: nextConfigResult.config.provider,
+    model: nextConfigResult.config.model
+  });
+  return configStatus;
+}
 function disableHeadroomRouting() {
   if (!lightcodeConfigResult.config.headroom?.enabled) {
     return;
@@ -7134,11 +7414,122 @@ var init_session_auto_title = __esm(() => {
   logger6 = createLogger("session-auto-title");
 });
 
+// apps/server/src/lib/workspace-context.ts
+import { readFile as readFile2, readdir } from "fs/promises";
+import { platform as platform3 } from "os";
+import path11 from "path";
+function describePlatform() {
+  const map = {
+    darwin: "macOS",
+    win32: "Windows",
+    linux: "Linux"
+  };
+  return map[platform3()] ?? platform3();
+}
+async function buildDirectoryListing(cwd) {
+  try {
+    const dirents = await readdir(cwd, { withFileTypes: true });
+    const names = dirents.filter((entry) => entry.name !== ".git").map((entry) => entry.isDirectory() ? `${entry.name}/` : entry.name).sort((a, b) => a.localeCompare(b));
+    const shown = names.slice(0, maxListedEntries);
+    const suffix = names.length > shown.length ? `
+\u2026 and ${names.length - shown.length} more` : "";
+    return shown.length > 0 ? shown.join(`
+`) + suffix : "(empty directory)";
+  } catch (error) {
+    logger7.warn("workspace_listing_failed", { cwd, error: getErrorMessage(error) });
+    return "(unable to list directory)";
+  }
+}
+async function buildGitSummary(cwd) {
+  try {
+    const context2 = createWorkspaceContext(cwd);
+    const status = await executeGitStatus({}, context2);
+    if (!status.ok) {
+      return "Not a git repository.";
+    }
+    const header = [
+      `branch: ${status.branch ?? "(detached)"}`,
+      status.ahead || status.behind ? `ahead ${status.ahead}, behind ${status.behind}` : null
+    ].filter(Boolean).join(" | ");
+    if (status.totalEntries === 0) {
+      return `${header}
+working tree clean`;
+    }
+    const lines = status.entries.slice(0, maxListedEntries).map((entry) => `${entry.indexStatus}${entry.workingTreeStatus} ${entry.path}`.trim());
+    const more = status.totalEntries > lines.length ? `
+\u2026 and ${status.totalEntries - lines.length} more` : "";
+    return `${header}
+changes:
+${lines.join(`
+`)}${more}`;
+  } catch (error) {
+    logger7.warn("workspace_git_summary_failed", { cwd, error: getErrorMessage(error) });
+    return "Not a git repository.";
+  }
+}
+async function buildProjectDocs(cwd) {
+  const sections = [];
+  for (const name of projectDocNames) {
+    try {
+      const content = (await readFile2(path11.join(cwd, name), "utf8")).trim();
+      if (!content) {
+        continue;
+      }
+      const truncated = content.length > maxDocChars ? `${content.slice(0, maxDocChars)}
+\u2026 (truncated)` : content;
+      sections.push(`### ${name}
+${truncated}`);
+    } catch {}
+  }
+  return sections.join(`
+
+`);
+}
+async function buildWorkspaceContext({
+  cwd
+}) {
+  const [listing, git, docs] = await Promise.all([
+    buildDirectoryListing(cwd),
+    buildGitSummary(cwd),
+    buildProjectDocs(cwd)
+  ]);
+  const today = new Date().toISOString().slice(0, 10);
+  const parts = [
+    "<environment>",
+    `Working directory: ${cwd}`,
+    `Platform: ${describePlatform()}`,
+    `Date: ${today}`,
+    "",
+    "Git:",
+    git,
+    "",
+    "Top-level entries:",
+    listing,
+    docs ? `
+Project instructions:
+${docs}` : "",
+    "</environment>"
+  ];
+  const block = parts.join(`
+`);
+  return block.length > maxBlockChars ? `${block.slice(0, maxBlockChars)}
+\u2026 (environment truncated)
+</environment>` : block;
+}
+var logger7, projectDocNames, maxListedEntries = 40, maxDocChars = 1500, maxBlockChars = 8000;
+var init_workspace_context = __esm(() => {
+  init_runtime17();
+  init_src();
+  logger7 = createLogger("workspace-context");
+  projectDocNames = ["AGENTS.md", "CLAUDE.md", "README.md"];
+});
+
 // apps/server/src/lib/chat-stream.ts
 import {
   createAgentUIStreamResponse,
   consumeStream,
   generateId as generateId3,
+  isToolUIPart as isToolUIPart2,
   safeValidateUIMessages as safeValidateUIMessages2,
   streamText
 } from "ai";
@@ -7158,6 +7549,9 @@ function buildFastChatModelMessages(messages) {
     content: collectMessageText(message)
   })).filter((message) => message.content.length > 0).slice(-fastChatRecentMessageCount);
 }
+function messagesHaveToolActivity(messages) {
+  return messages.some((message) => message.parts?.some((part) => isToolUIPart2(part)));
+}
 function shouldUseFastChatPath({
   messages,
   mode,
@@ -7166,6 +7560,9 @@ function shouldUseFastChatPath({
   if (allowedTools && allowedTools.length > 0) {
     return false;
   }
+  if (messagesHaveToolActivity(messages)) {
+    return false;
+  }
   return selectCodingAgentIntentTools({
     mode,
     prompt: "",
@@ -7180,7 +7577,7 @@ async function countPendingChatInteractions(sessionId) {
     });
     return pendingInteractions.interactions.length;
   } catch (error) {
-    logger7.warn("context_compaction_pending_interactions_unavailable", {
+    logger8.warn("context_compaction_pending_interactions_unavailable", {
       sessionId,
       error: getErrorMessage(error)
     });
@@ -7191,7 +7588,7 @@ async function loadSessionContextStateSafe(sessionId) {
   try {
     return await getSessionContextState(sessionId);
   } catch (error) {
-    logger7.warn("context_state_unavailable", {
+    logger8.warn("context_state_unavailable", {
       sessionId,
       error: getErrorMessage(error)
     });
@@ -7240,7 +7637,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
       staleSkip: false
     });
   } catch (error) {
-    logger7.error("chat_pre_stream_persist_failed", {
+    logger8.error("chat_pre_stream_persist_failed", {
       sessionId,
       error: getErrorMessage(error)
     });
@@ -7251,6 +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({
@@ -7258,44 +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
-      });
-      logger7.info("context_auto_compacted", {
-        sessionId,
-        tier: compaction.state.tier,
-        usedFallback: compaction.usedFallback,
-        coveredMessageCount: compaction.state.coveredMessageCount,
-        estimatedTokensBefore,
-        estimatedTokensAfter: view.estimate.tokens
-      });
-    } catch (error) {
-      logger7.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);
@@ -7310,6 +7721,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
+        reservedOutputTokens,
+        endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive
       });
@@ -7331,10 +7744,12 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
+        reservedOutputTokens,
+        endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive
       });
-      logger7.info("context_overflow_recovery", {
+      logger8.info("context_overflow_recovery", {
         sessionId,
         preserveRecentMessages: overflowConfig.preserveRecentMessages,
         estimatedTokens: view.estimate.tokens,
@@ -7343,13 +7758,30 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         compactedMessages: view.coveredMessages.length
       });
     } catch (error) {
-      logger7.error("context_overflow_recovery_failed", {
+      logger8.error("context_overflow_recovery_failed", {
         sessionId,
         error: getErrorMessage(error)
       });
     }
   }
-  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) {
@@ -7379,14 +7811,14 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     });
     const normalizedMessages = removeTrailingEmptyAssistantMessages(fullHistory);
     if (normalizedMessages.length !== fullHistory.length) {
-      logger7.warn("chat_finish_empty_assistant_skipped", { sessionId });
+      logger8.warn("chat_finish_empty_assistant_skipped", { sessionId });
       return;
     }
     const validatedFinishResult = await safeValidateUIMessages2({
       messages: normalizedMessages
     });
     if (!validatedFinishResult.success) {
-      logger7.warn("chat_invalid_finish_payload", {
+      logger8.warn("chat_invalid_finish_payload", {
         sessionId,
         message: validatedFinishResult.error.message
       });
@@ -7420,7 +7852,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         messages: validatedFinishResult.data
       });
     } catch (error) {
-      logger7.error("chat_finish_persist_failed", {
+      logger8.error("chat_finish_persist_failed", {
         sessionId,
         error: getErrorMessage(error)
       });
@@ -7443,22 +7875,30 @@ 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 });
   try {
     if (shouldUseFastChatPath({
       messages: providerMessages,
       mode,
       allowedTools
     })) {
-      logger7.info("chat_fast_path", {
+      logger8.info("chat_fast_path", {
         sessionId,
         model: chatModelId
       });
       const result = streamText({
         model: resolvedProviderModel.model,
-        system: fastChatSystemPrompt,
+        system: `${fastChatSystemPrompt}
+
+${environmentContext}`,
         messages: buildFastChatModelMessages(providerMessages),
         maxOutputTokens: Math.min(lightcodeConfigResult.config.maxOutputTokens, fastChatMaxOutputTokens),
         maxRetries: lightcodeConfigResult.config.maxRetries,
@@ -7488,6 +7928,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         onError: (error) => {
           const classified = classifyChatError(error);
           if (classified.kind === "context_overflow") {
+            noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
             noteOverflowForSession();
           }
           incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), { sessionId, statusCode: classified.statusCode });
@@ -7505,6 +7946,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     const mapStreamError = (error) => {
       const classified = classifyChatError(error);
       if (classified.kind === "context_overflow") {
+        noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
         noteOverflowForSession();
       }
       incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), {
@@ -7528,7 +7970,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         permissionMode,
         allowedTools,
         permissionRules,
-        sandbox
+        sandbox,
+        environmentContext
       },
       abortSignal: c.req.raw.signal,
       generateMessageId: generateId3,
@@ -7558,6 +8001,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
   } catch (error) {
     const classified = classifyChatError(error);
     if (classified.kind === "context_overflow") {
+      noteLearnedContextLimit(chatModelId, classified.contextLimitTokens);
       noteContextOverflow(sessionId);
     }
     incrementChatFailureCounter(chatFailureClassForErrorKind(classified.kind), {
@@ -7588,7 +8032,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     }, 500);
   }
 }
-var logger7, 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();
@@ -7597,13 +8041,15 @@ 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();
+  init_workspace_context();
   init_runtime_config();
-  logger7 = createLogger("chat-stream");
+  logger8 = createLogger("chat-stream");
   providerBillingOrQuotaMessage = "The configured model provider rejected this request due to billing or quota limits. " + "Update provider credits/quota and retry.";
-  fastChatSystemPrompt = "You are Lightcode's friendly coding assistant. For casual conversation, reply briefly and naturally. " + "If the user asks for coding work, say you can help and ask them what they want to change.";
+  fastChatSystemPrompt = "You are Lightcode's coding assistant, running inside the user's project. " + "For casual conversation, reply briefly and naturally. " + "If the user asks about or wants work on their code, do not ask them to paste it \u2014 you can read the project files directly; offer to look and proceed.";
 });
 
 // apps/server/src/routes/route-helpers.ts
@@ -7612,16 +8058,16 @@ function internalErrorResponse(c, {
   message,
   error
 }) {
-  logger8.error(event, { error: getErrorMessage(error) });
+  logger9.error(event, { error: getErrorMessage(error) });
   return c.json({
     error: message,
     details: Bun.env.NODE_ENV === "production" ? undefined : getErrorMessage(error)
   }, 500);
 }
-var logger8;
+var logger9;
 var init_route_helpers = __esm(() => {
   init_src();
-  logger8 = createLogger("routes");
+  logger9 = createLogger("routes");
 });
 
 // apps/server/src/routes/chat-routes.ts
@@ -7916,7 +8362,7 @@ var init_config_routes = __esm(() => {
     provider: lightcodeProviderSchema,
     model: z35.string().min(1).max(240)
   });
-  configRoutes = new Hono2().get("/status", (c) => c.json(configStatus)).get("/models", async (c) => {
+  configRoutes = new Hono2().get("/status", (c) => c.json(configStatus)).post("/reload", (c) => c.json({ ok: true, configStatus: reloadRuntimeConfig() })).get("/models", async (c) => {
     const provider = c.req.query("provider") ?? "openrouter";
     if (provider !== "openrouter") {
       return c.json({ error: `Model listing is not supported for provider "${provider}".` }, 400);
@@ -7987,7 +8433,7 @@ var init_extension_runtime = __esm(() => {
 
 // apps/server/src/routes/diagnostics-routes.ts
 import { existsSync as existsSync9, readdirSync as readdirSync4 } from "fs";
-import path11 from "path";
+import path12 from "path";
 import { Hono as Hono3 } from "hono";
 function combineStatuses(statuses) {
   if (statuses.includes("error")) {
@@ -8309,7 +8755,7 @@ async function buildLogsTailPayload(requestedLines) {
       lines: []
     };
   }
-  const filePath = path11.join(directory, latestFile);
+  const filePath = path12.join(directory, latestFile);
   const content = await Bun.file(filePath).text();
   const lines = content.split(`
 `).filter((line) => line.length > 0);
@@ -8432,36 +8878,36 @@ async function ensureHeadroomRouting() {
     health = await checkHeadroomHealth(headroom.proxyUrl);
   } catch (error) {
     health = "down";
-    logger9.warn("headroom_health_probe_failed", {
+    logger10.warn("headroom_health_probe_failed", {
       proxyUrl: headroom.proxyUrl,
       error: getErrorMessage(error)
     });
   }
   if (health === "healthy") {
-    logger9.info("headroom_routing_active", { proxyUrl: headroom.proxyUrl });
+    logger10.info("headroom_routing_active", { proxyUrl: headroom.proxyUrl });
     return;
   }
   if (headroom.failOpen) {
-    logger9.warn("headroom_unreachable_failing_open", {
+    logger10.warn("headroom_unreachable_failing_open", {
       proxyUrl: headroom.proxyUrl
     });
     disableHeadroomRouting();
     return;
   }
-  logger9.error("headroom_unreachable_fail_closed", {
+  logger10.error("headroom_unreachable_fail_closed", {
     proxyUrl: headroom.proxyUrl
   });
 }
-var logger9, healthCheckTimeoutMs = 1500;
+var logger10, healthCheckTimeoutMs = 1500;
 var init_headroom_proxy = __esm(() => {
   init_src();
   init_runtime_config();
-  logger9 = createLogger("headroom-proxy");
+  logger10 = createLogger("headroom-proxy");
 });
 
 // apps/server/src/index.ts
 init_src();
-var logger10 = createLogger("server");
+var logger11 = createLogger("server");
 var httpIdleTimeoutSeconds = 30;
 var shutdownForceExitMs = 5000;
 function isChatStreamingRoute(request) {
@@ -8474,7 +8920,7 @@ function isChatStreamingRoute(request) {
 async function startServer() {
   const logDirectory = enableFileLogSink(Bun.env);
   if (logDirectory) {
-    logger10.info("file_log_sink_enabled", { directory: logDirectory });
+    logger11.info("file_log_sink_enabled", { directory: logDirectory });
   }
   const port = Number(Bun.env.PORT ?? 4983);
   const hostname = Bun.env.LIGHTCODE_HOST ?? "127.0.0.1";
@@ -8495,12 +8941,12 @@ async function startServer() {
     process.exit(1);
   }
   Promise.resolve().then(() => (init_runtime_config(), exports_runtime_config)).then((runtime19) => runtime19.initializeModelCapabilities()).catch((error) => {
-    logger10.warn("model_capability_init_failed", {
+    logger11.warn("model_capability_init_failed", {
       error: getErrorMessage(error)
     });
   });
   Promise.resolve().then(() => (init_headroom_proxy(), exports_headroom_proxy)).then((headroom) => headroom.ensureHeadroomRouting()).catch((error) => {
-    logger10.warn("headroom_init_failed", {
+    logger11.warn("headroom_init_failed", {
       error: getErrorMessage(error)
     });
   });
@@ -8536,7 +8982,7 @@ async function startServer() {
       return;
     }
     shuttingDown = true;
-    logger10.info("server_shutdown", { signal });
+    logger11.info("server_shutdown", { signal });
     const forceExitTimer = setTimeout(() => {
       server.stop(true);
       process.exit(0);
@@ -8551,6 +8997,17 @@ async function startServer() {
   };
   process.on("SIGINT", () => void shutdown("SIGINT"));
   process.on("SIGTERM", () => void shutdown("SIGTERM"));
+  const parentPid = Number(Bun.env.LIGHTCODE_PARENT_PID);
+  if (Number.isInteger(parentPid) && parentPid > 0) {
+    const parentWatch = setInterval(() => {
+      try {
+        process.kill(parentPid, 0);
+      } catch {
+        shutdown("parent-exit");
+      }
+    }, 2000);
+    parentWatch.unref();
+  }
   console.log(`Server listening on http://${hostname}:${port}`);
 }
 if (import.meta.main) {