@kmugalkhod/lightcode 0.18.4 → 0.19.0

package/server.js

@@ -635,7 +635,7 @@ var init_web_auth = __esm(() => {
 });
 
 // packages/ai/src/constants.ts
-var MAX_TOOL_TEXT_OUTPUT_CHARS = 12000, DEFAULT_TOOL_TEXT_OUTPUT_CHARS = 6000, MAX_TOOL_LIST_ENTRIES = 200, MAX_TOOL_SEARCH_RESULTS = 200, ANTHROPIC_TOOL_OPTIONAL_PARAMETER_BUDGET = 24, TODO_CONTENT_MAX_CHARS = 200, TODO_ACTIVE_FORM_MAX_CHARS = 80, PROGRESS_NOTE_PRUNE_MIN_CHARS = 200;
+var MAX_TOOL_TEXT_OUTPUT_CHARS = 12000, DEFAULT_TOOL_TEXT_OUTPUT_CHARS = 6000, READ_FILE_OUTPUT_CHARS = 24000, MAX_TOOL_LIST_ENTRIES = 200, MAX_TOOL_SEARCH_RESULTS = 200, ANTHROPIC_TOOL_OPTIONAL_PARAMETER_BUDGET = 24, TODO_CONTENT_MAX_CHARS = 200, TODO_ACTIVE_FORM_MAX_CHARS = 80, PROGRESS_NOTE_PRUNE_MIN_CHARS = 200;
 
 // packages/ai/src/coding-agent-modes.ts
 import { z as z2 } from "zod";
@@ -2002,14 +2002,15 @@ var init_schema8 = __esm(() => {
 
 // packages/ai/src/read-file/schema.ts
 import { z as z19 } from "zod";
-var readFileDescription = "Read file content from the working directory. Prefer a startLine/endLine range over whole files \u2014 large files are truncated head+tail.", readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
+var readFileDescription, readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
 var init_schema9 = __esm(() => {
   init_base_schemas();
+  readFileDescription = "Read file content from the working directory. Returns whole lines, up to about 24,000 characters (roughly 600 lines) per call, so read a file in one call instead of in slices. " + "If the file is longer, the result stops at endLine and nextStartLine says where to continue. " + "To inspect one spot in a large file, grep for it first, then read a generous startLine/endLine range around it.";
   readFileInputSchema = z19.object({
     path: boundedPathSchema,
     startLine: positiveLineSchema.optional(),
     endLine: positiveLineSchema.optional(),
-    maxChars: integerRangeSchema(1, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional()
+    maxChars: integerRangeSchema(1, READ_FILE_OUTPUT_CHARS, "maxChars").optional()
   }).superRefine((value, ctx) => {
     if (value.startLine !== undefined && value.endLine !== undefined && value.endLine < value.startLine) {
       ctx.addIssue({
@@ -2030,7 +2031,8 @@ var init_schema9 = __esm(() => {
     endLine: z19.number().int().min(1),
     content: z19.string(),
     totalLines: z19.number().int().nonnegative(),
-    truncated: z19.boolean()
+    truncated: z19.boolean(),
+    nextStartLine: z19.number().int().min(1).optional()
   });
 });
 
@@ -2350,7 +2352,7 @@ var init_tool_registry = __esm(() => {
       permissionMode: "read-only",
       execution: "server",
       availability: "git-workspace",
-      activation: "specialized",
+      activation: "core",
       outputPolicy: "artifact-if-large"
     },
     git_diff: {
@@ -2363,7 +2365,7 @@ var init_tool_registry = __esm(() => {
       permissionMode: "read-only",
       execution: "server",
       availability: "git-workspace",
-      activation: "specialized",
+      activation: "core",
       outputPolicy: "artifact-if-large"
     },
     git_log: {
@@ -2376,7 +2378,7 @@ var init_tool_registry = __esm(() => {
       permissionMode: "read-only",
       execution: "server",
       availability: "git-workspace",
-      activation: "specialized",
+      activation: "core",
       outputPolicy: "artifact-if-large"
     },
     git_show: {
@@ -2389,7 +2391,7 @@ var init_tool_registry = __esm(() => {
       permissionMode: "read-only",
       execution: "server",
       availability: "git-workspace",
-      activation: "specialized",
+      activation: "core",
       outputPolicy: "artifact-if-large"
     },
     tool_search: {
@@ -2790,6 +2792,7 @@ var init_agent_tools = __esm(() => {
     contextWindow: z27.number().int().positive().optional(),
     preserveRecentTokens: z27.number().int().nonnegative().optional(),
     originalInputTokens: z27.number().int().nonnegative().optional(),
+    promptCacheActive: z27.boolean().optional(),
     assemblyAbortSignal: z27.custom().optional(),
     turnKey: z27.string().min(1).max(200).optional()
   });
@@ -3068,10 +3071,10 @@ ${environmentContext.trim()}` : "");
 var cwdPromptPlaceholder = "{cwd}", defaultCodingAgentSystemPrompt, toolCallingDisciplineAddendum;
 var init_coding_agent_prompt = __esm(() => {
   init_coding_agent_modes();
-  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. " + "Check the available skill descriptions for every task. When a skill is explicitly named or clearly matches the task, load it with the skill tool before acting and follow its instructions within the user's scope and tool permissions. Skill files may be installed outside the workspace; use the skill tool to read their relative supporting resources. Never claim to have used a skill you have not loaded. " + `Inspect context before risky operations. Progress notes: before a major tool action you may emit ONE short note (one sentence, ~15 words) saying what you are about to do. After a tool returns, mention a finding only if it changes your plan, and state it as fact grounded in what the tool actually returned \u2014 cite the file/line/symbol. NEVER claim you found, changed, or fixed something before a tool result confirms it. Skip notes for routine reads, and do not restate the todo list in prose.
+  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. " + "Gather context efficiently \u2014 every tool round trip costs the user seconds. The environment section maps the repository's files: use it to go straight to the right paths instead of listing directories. Locate before you read: use grep or glob_search to find the relevant files and line numbers, then read whole files (read_file returns up to about 600 lines per call) or, in larger files, a range of a few hundred lines around the spot \u2014 never many small slices, and never re-read content you already have. When several reads or searches do not depend on each other, request them together in one step. Skip generated output (dist/, build/, minified bundles, node_modules) unless the task is about it. Stop exploring once you have enough evidence to act or answer. " + "Check the available skill descriptions for every task. When a skill is explicitly named or clearly matches the task, load it with the skill tool before acting and follow its instructions within the user's scope and tool permissions. Skill files may be installed outside the workspace; use the skill tool to read their relative supporting resources. Never claim to have used a skill you have not loaded. " + `Inspect context before risky operations. Progress notes: before a major tool action you may emit ONE short note (one sentence, ~15 words) saying what you are about to do. After a tool returns, mention a finding only if it changes your plan, and state it as fact grounded in what the tool actually returned \u2014 cite the file/line/symbol. NEVER claim you found, changed, or fixed something before a tool result confirms it. Skip notes for routine reads, and do not restate the todo list in prose.
 
 ` + `Task list discipline \u2014 use the todo_write tool to keep a visible, evolving checklist:
-` + `- Create the list FIRST. For any task with roughly three or more distinct steps, or any multi-file or multi-stage work, call todo_write as your very first action \u2014 before reading or editing \u2014 with every step listed and all of them set to pending.
+` + `- Create the list in your first step. For any task with roughly three or more distinct steps, or any multi-file or multi-stage work, call todo_write in your first response with every step listed and all of them set to pending \u2014 request it together with your first reads or searches, not as a separate step, and always before any edit.
 ` + `- Keep exactly ONE item in_progress at a time. Mark an item in_progress immediately before you start it; never have two items in_progress at once.
 ` + `- Mark an item completed only after a tool result confirms it is actually done (test passed, file written, build clean) \u2014 not on intent. Do this the MOMENT it is confirmed; do not batch completions or wait until the end of the turn.
 ` + `- Completions are monotonic: once a task is completed, do not silently move it back to pending or in_progress. Reopen it only if it genuinely regressed.
@@ -3083,7 +3086,7 @@ var init_coding_agent_prompt = __esm(() => {
 ` + `When the task is complete, end with a short summary (a few lines) grounded in the todos you actually completed and the real file diffs \u2014 name the files you edited, do not list steps you did not verify, and do not repeat the todo card verbatim. If anything is incomplete or you are blocked, say so plainly.
 
 ` + "Long-running processes (dev servers, watchers): never run them as plain foreground bash \u2014 the bash tool enforces a timeout and will kill them. Start them detached with output redirected to a log " + '(Windows: start /B cmd /c "npm run dev > dev.log 2>&1"; Unix: nohup npm run dev > dev.log 2>&1 &), then read the log to confirm startup and the actual port (do not assume the default). ' + "Never kill processes or free ports you did not open yourself.";
-  toolCallingDisciplineAddendum = "Invoke tools only through the function-calling interface. " + "Never write tool calls as XML or plain text such as <tool_call>, <function_call>, <invoke>, or bare JSON in your message. " + "Tool arguments must be valid JSON matching the tool's schema. " + "Call one tool at a time and wait for its result before deciding the next step.";
+  toolCallingDisciplineAddendum = "Invoke tools only through the function-calling interface. " + "Never write tool calls as XML or plain text such as <tool_call>, <function_call>, <invoke>, or bare JSON in your message. " + "Tool arguments must be valid JSON matching the tool's schema. " + "You may request several independent tool calls in one response; wait for their results before making calls that depend on them.";
 });
 
 // packages/ai/src/common/coerce-tool-input.ts
@@ -3481,20 +3484,26 @@ function collectTextFromUnknown(value) {
   }
   return "";
 }
-function getLastUserText(messages, prompt) {
+function getUserTexts(messages, prompt) {
+  const texts = [];
   if (Array.isArray(messages)) {
-    for (let index = messages.length - 1;index >= 0; index -= 1) {
-      const message = messages[index];
+    for (const message of messages) {
       if (!isRecord4(message) || message.role !== "user") {
         continue;
       }
       const text = collectTextFromUnknown(message).trim();
       if (text) {
-        return text;
+        texts.push(text);
       }
     }
   }
-  return collectTextFromUnknown(prompt).trim();
+  if (texts.length === 0) {
+    const promptText = collectTextFromUnknown(prompt).trim();
+    if (promptText) {
+      texts.push(promptText);
+    }
+  }
+  return texts;
 }
 function escapeRegExp(value) {
   return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
@@ -3520,23 +3529,20 @@ function selectCodingAgentIntentTools({
   messages,
   availableSkillNames = []
 }) {
-  const userText = getLastUserText(messages, prompt);
-  const normalizedText = userText.toLowerCase();
-  if (!normalizedText || casualPromptPattern.test(normalizedText)) {
+  const userTexts = getUserTexts(messages, prompt);
+  if (userTexts.every((text) => casualPromptPattern.test(text))) {
     return [];
   }
+  const latestText = userTexts.at(-1) ?? "";
   const modeTools = getRegistryToolsForMode(mode);
   const selected = new Set(modeTools.filter((toolName) => codingToolRegistry[toolName].activation === "core"));
-  if (gitIntentPattern.test(userText)) {
-    addTools(selected, gitToolNames);
-  }
-  if (webIntentPattern.test(userText)) {
+  if (webIntentPattern.test(latestText)) {
     addTools(selected, webToolNames);
   }
-  if (mcpIntentPattern.test(userText)) {
+  if (userTexts.some((text) => mcpIntentPattern.test(text))) {
     addTools(selected, mcpToolNames);
   }
-  if (skillIntentPattern.test(userText) || requestsAvailableSkill(userText, availableSkillNames)) {
+  if (userTexts.some((text) => skillIntentPattern.test(text) || requestsAvailableSkill(text, availableSkillNames))) {
     selected.add("skill");
   }
   return modeTools.filter((toolName) => selected.has(toolName));
@@ -3616,21 +3622,14 @@ function limitProviderActiveTools(tools, maxTools = maxProviderActiveTools) {
   const toolSet = new Set(tools);
   return providerToolPriority.filter((toolName) => toolSet.has(toolName)).slice(0, maxTools);
 }
-var casualPromptPattern, gitIntentPattern, webIntentPattern, mcpIntentPattern, skillIntentPattern, gitToolNames, webToolNames, mcpToolNames, maxProviderActiveTools = 24, providerToolPriority;
+var casualPromptPattern, webIntentPattern, mcpIntentPattern, skillIntentPattern, webToolNames, mcpToolNames, maxProviderActiveTools = 24, providerToolPriority;
 var init_intent_tool_selection = __esm(() => {
   init_coding_agent_modes();
   init_tool_registry();
   casualPromptPattern = /^(hi|hello|hey|how are you|how r you|what'?s up|whats up|sup|thanks|thank you|ok|okay|yes|no|nice|cool|great)[\s?!.,]*$/i;
-  gitIntentPattern = /(?:\bgit\b|\bcommits?\b|\bstaged?\b|\bunstaged\b|\buntracked\b|\bworking tree\b|\brepositor(?:y|ies)\s+(?:status|diff|history|changes))/i;
   webIntentPattern = /(?:\bhttps?:\/\/|\bwww\.|\bweb(?:site|page)?\b|\binternet\b|\bonline\b|\burl\b|\bbrowser\b|\bsearch\s+(?:the\s+)?web\b|\bbrowse\s+(?:the\s+web|online|for)\b|\blook\s+up\b|\b(?:latest|current|recent|today'?s)\s+(?:news|price|weather|docs?|documentation|release|version)\b)/i;
   mcpIntentPattern = /(?:\bmcp\b|\bmodel context protocol\b)/i;
   skillIntentPattern = /\bskills?\b/i;
-  gitToolNames = [
-    "git_status",
-    "git_diff",
-    "git_log",
-    "git_show"
-  ];
   webToolNames = [
     "web_fetch",
     "web_search"
@@ -5160,19 +5159,44 @@ async function executeReadFile(input, workspaceContext = getDefaultWorkspaceCont
   const fileContent = await fs4.readFile(resolvedPath, "utf8");
   const lines = fileContent.split(/\r?\n/);
   const { start, end } = clampLineRange(lines.length, parsedInput.startLine, parsedInput.endLine);
-  const selectedContent = lines.slice(start - 1, end).join(`
-`);
-  const maxChars = parsedInput.maxChars ?? DEFAULT_TOOL_TEXT_OUTPUT_CHARS;
-  const truncated = truncateText(selectedContent, maxChars);
+  const maxChars = parsedInput.maxChars ?? READ_FILE_OUTPUT_CHARS;
+  const taken = takeWholeLines(lines.slice(start - 1, end), maxChars);
+  const lastLine = start + taken.lineCount - 1;
   return readFileOutputSchema.parse({
     path: relativePath,
     startLine: start,
-    endLine: end,
-    content: truncated.text,
+    endLine: lastLine,
+    content: taken.text,
     totalLines: lines.length,
-    truncated: truncated.truncated
+    truncated: taken.truncated,
+    ...lastLine < end ? { nextStartLine: lastLine + 1 } : {}
   });
 }
+function takeWholeLines(lines, maxChars) {
+  let used = 0;
+  let lineCount = 0;
+  for (const line of lines) {
+    const cost = line.length + (lineCount > 0 ? 1 : 0);
+    if (used + cost > maxChars) {
+      break;
+    }
+    used += cost;
+    lineCount += 1;
+  }
+  if (lineCount === 0 && lines.length > 0) {
+    return {
+      text: truncateText(lines[0] ?? "", maxChars).text,
+      lineCount: 1,
+      truncated: true
+    };
+  }
+  return {
+    text: lines.slice(0, lineCount).join(`
+`),
+    lineCount,
+    truncated: lineCount < lines.length
+  };
+}
 var init_runtime8 = __esm(() => {
   init_resolve_within_workspace();
   init_output_utils();
@@ -6444,10 +6468,12 @@ function buildProviderView(options) {
   }
   const tokenBoundedCutoffIndex = computeTokenBoundedCompactionCutoffIndex(messages, preserveRecentTokens, afterAnchorIndex);
   const userIndicesAfterAnchor = messages.map((message, index) => ({ message, index })).filter(({ message, index }) => index >= afterAnchorIndex && message.role === "user").map(({ index }) => index);
-  const rollingTurnCutoffIndex = userIndicesAfterAnchor.length > config.uncachedRollingCompactionUserTurns ? userIndicesAfterAnchor[userIndicesAfterAnchor.length - config.uncachedRollingCompactionUserTurns] ?? afterAnchorIndex : afterAnchorIndex;
+  const rollingKeepUserTurns = Math.max(1, Math.ceil(config.uncachedRollingCompactionUserTurns / 2));
+  const rollingTurnCutoffIndex = userIndicesAfterAnchor.length > rollingKeepUserTurns ? userIndicesAfterAnchor[userIndicesAfterAnchor.length - rollingKeepUserTurns] ?? afterAnchorIndex : afterAnchorIndex;
   const overCompactThreshold = triggerTokens(estimate.tokens) > config.compactAtFraction * inputBudgetTokens;
+  const rollingActive = (options.rollingCompaction ?? options.cacheActive === false) && config.aggressivePruneWhenUncached;
   const userTurnsAfterAnchor = messages.slice(afterAnchorIndex).filter((message) => message.role === "user").length;
-  const rollingThresholdReached = uncachedActive && userTurnsAfterAnchor > config.uncachedRollingCompactionUserTurns;
+  const rollingThresholdReached = rollingActive && userTurnsAfterAnchor > (options.continueRolling ? rollingKeepUserTurns : config.uncachedRollingCompactionUserTurns);
   const compactionCutoffIndex = rollingThresholdReached ? Math.max(tokenBoundedCutoffIndex, rollingTurnCutoffIndex) : tokenBoundedCutoffIndex;
   const coveredMessages = capCoverageToTokenBudget(messages.slice(afterAnchorIndex, compactionCutoffIndex), config.maxCoverageTokensPerCompaction);
   let compactionBlockedReason = null;
@@ -6901,6 +6927,14 @@ import path12 from "path";
 import {
   toolModelMessageSchema
 } from "ai";
+function resolveToolOutputArtifactThresholds({
+  promptCacheActive
+}) {
+  return {
+    thresholdCharacters: promptCacheActive ? DEFAULT_CURRENT_TURN_TOOL_OUTPUT_THRESHOLD_CHARACTERS : DEFAULT_TOOL_OUTPUT_ARTIFACT_THRESHOLD_CHARACTERS,
+    currentTurnThresholdCharacters: DEFAULT_CURRENT_TURN_TOOL_OUTPUT_THRESHOLD_CHARACTERS
+  };
+}
 function isRecord10(value) {
   return typeof value === "object" && value !== null;
 }
@@ -7080,12 +7114,19 @@ async function artifactizeLargeToolOutputs(messages, options = {}) {
   if (previewCharacters < MIN_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS) {
     throw new RangeError(`Tool-output artifact previews must allow at least ${MIN_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS} characters.`);
   }
+  const currentTurnThreshold = Math.max(0, Math.floor(options.currentTurnThresholdCharacters ?? thresholdCharacters));
+  let currentTurnStart = 0;
+  messages.forEach((message, index) => {
+    if (message.role === "user")
+      currentTurnStart = index + 1;
+  });
   const uniqueArtifacts = new Map;
   let artifactizedOutputs = 0;
   let savedCharacters = 0;
   const providerMessages = [];
-  for (const message of messages) {
+  for (const [index, message] of messages.entries()) {
     options.signal?.throwIfAborted();
+    const limit = index >= currentTurnStart ? currentTurnThreshold : thresholdCharacters;
     if (!Array.isArray(message.content)) {
       providerMessages.push(message);
       continue;
@@ -7099,7 +7140,7 @@ async function artifactizeLargeToolOutputs(messages, options = {}) {
       }
       const output = part.output;
       const serialized = serializeToolResultOutput(output);
-      if (serialized.length <= thresholdCharacters) {
+      if (serialized.length <= limit) {
         providerContent.push(part);
         continue;
       }
@@ -7131,7 +7172,7 @@ async function artifactizeLargeToolOutputs(messages, options = {}) {
     savedCharacters
   };
 }
-var TOOL_OUTPUT_ARTIFACT_HANDLE_PREFIX = "lightcode-artifact://tool-output/sha256/", DEFAULT_TOOL_OUTPUT_ARTIFACT_THRESHOLD_CHARACTERS, DEFAULT_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS = 2400, MIN_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS = 512, SHA256_PATTERN, ARTIFACT_PREVIEW_MARKER = "[Lightcode tool-output artifact]";
+var TOOL_OUTPUT_ARTIFACT_HANDLE_PREFIX = "lightcode-artifact://tool-output/sha256/", DEFAULT_TOOL_OUTPUT_ARTIFACT_THRESHOLD_CHARACTERS, DEFAULT_CURRENT_TURN_TOOL_OUTPUT_THRESHOLD_CHARACTERS = 32000, DEFAULT_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS = 2400, MIN_TOOL_OUTPUT_ARTIFACT_PREVIEW_CHARACTERS = 512, SHA256_PATTERN, ARTIFACT_PREVIEW_MARKER = "[Lightcode tool-output artifact]";
 var init_tool_output_artifacts = __esm(() => {
   init_src();
   DEFAULT_TOOL_OUTPUT_ARTIFACT_THRESHOLD_CHARACTERS = DEFAULT_TOOL_TEXT_OUTPUT_CHARS;
@@ -8617,7 +8658,10 @@ function createCodingAgent({
         originalInputTokens: context.originalInputTokens
       });
       const artifactized = await artifactizeLargeToolOutputs(messages, {
-        signal: context.assemblyAbortSignal
+        signal: context.assemblyAbortSignal,
+        ...resolveToolOutputArtifactThresholds({
+          promptCacheActive: context.promptCacheActive ?? false
+        })
       });
       const assembled = assembler.assembleModelMessages(artifactized.messages, {
         preserveRecentTokens: context.preserveRecentTokens
@@ -8683,6 +8727,7 @@ function createCodingAgent({
           reservedOutputTokens: options.maxOutputTokens,
           preserveRecentTokens: options.preserveRecentTokens,
           originalInputTokens: options.originalInputTokens,
+          promptCacheActive: options.promptCacheActive,
           assemblyAbortSignal: options.assemblyAbortSignal
         },
         instructions
@@ -13342,7 +13387,8 @@ function summarizeRawModel(raw3) {
   const pricingRaw = Reflect.get(raw3, "pricing");
   const pricing = typeof pricingRaw === "object" && pricingRaw !== null ? {
     prompt: typeof Reflect.get(pricingRaw, "prompt") === "string" ? Reflect.get(pricingRaw, "prompt") : null,
-    completion: typeof Reflect.get(pricingRaw, "completion") === "string" ? Reflect.get(pricingRaw, "completion") : null
+    completion: typeof Reflect.get(pricingRaw, "completion") === "string" ? Reflect.get(pricingRaw, "completion") : null,
+    inputCacheRead: typeof Reflect.get(pricingRaw, "input_cache_read") === "string" ? Reflect.get(pricingRaw, "input_cache_read") : null
   } : null;
   return {
     id,
@@ -13372,12 +13418,12 @@ async function listOpenRouterModels({
   forceRefresh = false
 } = {}) {
   const cache = readCache();
-  if (!forceRefresh && cache && Date.now() - cache.fetchedAt < cacheTtlMs) {
+  if (!forceRefresh && cache && cache.version === modelsCacheVersion && Date.now() - cache.fetchedAt < cacheTtlMs) {
     return { models: cache.models, fromCache: true };
   }
   try {
     const models = await fetchModelsFromApi();
-    writeCache({ fetchedAt: Date.now(), models });
+    writeCache({ version: modelsCacheVersion, fetchedAt: Date.now(), models });
     return { models, fromCache: false };
   } catch (error) {
     logger5.warn("openrouter_models_fetch_failed", {
@@ -13415,12 +13461,14 @@ function toCapabilityPricing(pricing) {
   if (!Number.isFinite(promptPerToken) || !Number.isFinite(completionPerToken)) {
     return null;
   }
+  const cacheReadPerToken = pricing.inputCacheRead ? Number.parseFloat(pricing.inputCacheRead) : Number.NaN;
   return {
     inputPerMTok: promptPerToken * 1e6,
-    outputPerMTok: completionPerToken * 1e6
+    outputPerMTok: completionPerToken * 1e6,
+    cachedInputPerMTok: Number.isFinite(cacheReadPerToken) && cacheReadPerToken > 0 ? cacheReadPerToken * 1e6 : null
   };
 }
-var logger5, openRouterModelsUrl = "https://openrouter.ai/api/v1/models", cacheTtlMs, fetchTimeoutMs = 15000, openRouterModelSummarySchema, modelsCacheSchema;
+var logger5, openRouterModelsUrl = "https://openrouter.ai/api/v1/models", cacheTtlMs, fetchTimeoutMs = 15000, modelsCacheVersion = 2, openRouterModelSummarySchema, modelsCacheSchema;
 var init_openrouter_models = __esm(() => {
   init_src();
   init_network_fetch();
@@ -13435,10 +13483,12 @@ var init_openrouter_models = __esm(() => {
     supportsReasoning: z40.boolean(),
     pricing: z40.object({
       prompt: z40.string().nullable(),
-      completion: z40.string().nullable()
+      completion: z40.string().nullable(),
+      inputCacheRead: z40.string().nullable().default(null)
     }).nullable()
   });
   modelsCacheSchema = z40.object({
+    version: z40.number().int().optional(),
     fetchedAt: z40.number(),
     models: z40.array(openRouterModelSummarySchema)
   });
@@ -13722,6 +13772,7 @@ function resolveAnthropicModel({
     contextWindow: getModelContextWindow("anthropic", modelId),
     missingCredentialHints,
     supportsPromptCaching: true,
+    rollingCompaction: false,
     pricing: getAnthropicModelPricing(modelId),
     webSearchCapability,
     providerTools: webSearchCapability.execution === "provider" ? {
@@ -13737,7 +13788,7 @@ function resolveCapabilityPricing(capabilities, cachedInputFactor) {
   return {
     inputPerMTok: pricing.inputPerMTok,
     outputPerMTok: pricing.outputPerMTok,
-    cachedInputPerMTok: cachedInputFactor === null ? null : pricing.inputPerMTok * cachedInputFactor
+    cachedInputPerMTok: pricing.cachedInputPerMTok ?? (cachedInputFactor === null ? null : pricing.inputPerMTok * cachedInputFactor)
   };
 }
 function resolveOpenAITransportModel({
@@ -13786,6 +13837,8 @@ function resolveOpenAITransportModel({
     env,
     storedCredentials
   });
+  const openRouterAnthropicFamily = isOpenRouter && isOpenRouterAnthropicFamilyModel(configuredModel);
+  const openRouterCatalogCaching = isOpenRouter && (capabilities?.pricing?.cachedInputPerMTok ?? 0) > 0;
   return {
     provider: config6.provider,
     configuredModel,
@@ -13799,8 +13852,9 @@ function resolveOpenAITransportModel({
     contextWindow: capabilities?.contextLength ?? getModelContextWindow(config6.provider, configuredModel),
     missingCredentialHints: apiKey ? [] : isOpenCodeZen ? ["Set OPENCODE_API_KEY to call OpenCode Zen models."] : isOpenRouter ? ["Set OPENROUTER_API_KEY to call OpenRouter models."] : [],
     maxCompletionTokens: capabilities?.maxCompletionTokens ?? null,
-    supportsPromptCaching: isOpenRouter && isOpenRouterAnthropicFamilyModel(configuredModel),
-    pricing: resolveCapabilityPricing(capabilities, isOpenRouter && isOpenRouterAnthropicFamilyModel(configuredModel) ? 0.1 : null),
+    supportsPromptCaching: openRouterAnthropicFamily || openRouterCatalogCaching,
+    rollingCompaction: !openRouterAnthropicFamily,
+    pricing: resolveCapabilityPricing(capabilities, openRouterAnthropicFamily ? 0.1 : null),
     webSearchCapability,
     providerTools: openRouterProvider && webSearchCapability.execution === "provider" ? {
       web_search: createOpenRouterWebSearchTool(openRouterProvider, webSearchCapability)
@@ -14058,6 +14112,7 @@ var init_subagent_task_store = __esm(() => {
 });
 
 // apps/server/src/lib/workspace-context.ts
+import { execFile as execFile3 } from "child_process";
 import { createHash as createHash5, randomUUID as randomUUID3 } from "crypto";
 import {
   chmod as chmod2,
@@ -14072,6 +14127,7 @@ import {
 } from "fs/promises";
 import { platform as platform3 } from "os";
 import path21 from "path";
+import { promisify as promisify4 } from "util";
 import { z as z41 } from "zod";
 function isContainedPath(root, candidate) {
   const relative = path21.relative(root, candidate);
@@ -14080,12 +14136,16 @@ function isContainedPath(root, candidate) {
 function escapeAttribute(value) {
   return value.replaceAll("&", "&amp;").replaceAll('"', "&quot;").replaceAll("<", "&lt;").replaceAll(">", "&gt;");
 }
-function instructionEpochStatePath({
+function sessionStatePath({
+  directory,
   sessionId,
   dataDir
 }) {
   const sessionKey = createHash5("sha256").update(sessionId).digest("hex");
-  return path21.join(dataDir ?? getLightcodeDataDir(), "instruction-epochs", `${sessionKey}.json`);
+  return path21.join(dataDir ?? getLightcodeDataDir(), directory, `${sessionKey}.json`);
+}
+function instructionEpochStatePath(options) {
+  return sessionStatePath({ directory: "instruction-epochs", ...options });
 }
 async function loadInstructionEpochs(statePath) {
   try {
@@ -14106,6 +14166,9 @@ async function loadInstructionEpochs(statePath) {
   return [];
 }
 async function persistInstructionEpochs(statePath, epochs) {
+  await writePrivateJsonFile(statePath, { epochs });
+}
+async function writePrivateJsonFile(statePath, value) {
   const directory = path21.dirname(statePath);
   await mkdir3(directory, { recursive: true, mode: 448 });
   await chmod2(directory, 448).catch(() => {
@@ -14113,7 +14176,7 @@ async function persistInstructionEpochs(statePath, epochs) {
   });
   const temporaryPath = `${statePath}.${process.pid}.${randomUUID3()}.tmp`;
   try {
-    await writeFile2(temporaryPath, `${JSON.stringify({ epochs }, null, 2)}
+    await writeFile2(temporaryPath, `${JSON.stringify(value, null, 2)}
 `, { encoding: "utf8", flag: "wx", mode: 384 });
     await rename2(temporaryPath, statePath);
     await chmod2(statePath, 384).catch(() => {
@@ -14362,6 +14425,138 @@ async function buildDirectoryListing(cwd) {
     return "(unable to list directory)";
   }
 }
+async function listRepositoryFiles(cwd) {
+  try {
+    const { stdout } = await execFileAsync3("git", ["ls-files", "--cached", "--others", "--exclude-standard", "-z"], {
+      cwd,
+      encoding: "utf8",
+      maxBuffer: 64 * 1024 * 1024,
+      timeout: repositoryFileListTimeoutMs,
+      windowsHide: true
+    });
+    return [...new Set(stdout.split("\x00").filter(Boolean))];
+  } catch {
+    return null;
+  }
+}
+function createRepositoryTreeNode() {
+  return { directories: new Map, files: [], fileCount: 0 };
+}
+function compareNames(left, right) {
+  return left < right ? -1 : left > right ? 1 : 0;
+}
+function visibleSubdirectories(node) {
+  const all = [...node.directories.entries()].sort(([left], [right]) => compareNames(left, right)).map(([name, child]) => ({ name, node: child }));
+  return {
+    shown: all.slice(0, maxRepositoryMapChildren),
+    hiddenCount: Math.max(0, all.length - maxRepositoryMapChildren)
+  };
+}
+function subdirectoryLines(node, depth) {
+  const indent = "  ".repeat(depth);
+  const { shown, hiddenCount } = visibleSubdirectories(node);
+  const lines = shown.map(({ name, node: child }) => `${indent}${name}/ (${child.fileCount})`);
+  if (hiddenCount > 0) {
+    lines.push(`${indent}\u2026 ${hiddenCount} more folders`);
+  }
+  return lines;
+}
+function fileNameLines(node, depth) {
+  const indent = "  ".repeat(depth);
+  const files = [...node.files].sort(compareNames);
+  const lines = files.slice(0, maxRepositoryMapChildren).map((name) => `${indent}${name}`);
+  if (files.length > lines.length) {
+    lines.push(`${indent}\u2026 ${files.length - lines.length} more files`);
+  }
+  return lines;
+}
+function filePlaceholderLines(node, depth) {
+  return node.files.length > 0 ? [`${"  ".repeat(depth)}\u2026 ${node.files.length} files`] : [];
+}
+function linesCost(lines) {
+  return lines.reduce((sum, line) => sum + line.length + 1, 0);
+}
+function formatRepositoryMap(paths, maxChars = maxRepositoryMapChars) {
+  const root = createRepositoryTreeNode();
+  for (const filePath of paths) {
+    const segments = filePath.split("/").filter(Boolean);
+    const fileName = segments.pop();
+    if (!fileName) {
+      continue;
+    }
+    let current = root;
+    current.fileCount += 1;
+    for (const segment of segments) {
+      let child = current.directories.get(segment);
+      if (!child) {
+        child = createRepositoryTreeNode();
+        current.directories.set(segment, child);
+      }
+      child.fileCount += 1;
+      current = child;
+    }
+    current.files.push(fileName);
+  }
+  if (root.fileCount === 0) {
+    return "(no files)";
+  }
+  const order = [
+    { node: root, depth: 0 }
+  ];
+  const expanded = new Set;
+  let used = 0;
+  for (let index = 0;index < order.length; index += 1) {
+    const { node, depth } = order[index];
+    const cost = linesCost(subdirectoryLines(node, depth)) + linesCost(filePlaceholderLines(node, depth));
+    if (node !== root && used + cost > maxChars) {
+      continue;
+    }
+    used += cost;
+    expanded.add(node);
+    for (const { name, node: child } of visibleSubdirectories(node).shown) {
+      if (!name.startsWith(".")) {
+        order.push({ node: child, depth: depth + 1 });
+      }
+    }
+  }
+  const fileCandidates = order.filter(({ node }) => expanded.has(node) && node.files.length > 0).sort((left, right) => left.node === root ? -1 : right.node === root ? 1 : left.node.files.length - right.node.files.length);
+  const filesShown = new Set;
+  for (const { node, depth } of fileCandidates) {
+    const delta = linesCost(fileNameLines(node, depth)) - linesCost(filePlaceholderLines(node, depth));
+    if (used + delta > maxChars) {
+      continue;
+    }
+    used += delta;
+    filesShown.add(node);
+  }
+  const lines = [];
+  const render = (node, depth) => {
+    const { shown } = visibleSubdirectories(node);
+    subdirectoryLines(node, depth).forEach((line, index) => {
+      lines.push(line);
+      const child = shown[index]?.node;
+      if (child && expanded.has(child)) {
+        render(child, depth + 1);
+      }
+    });
+    lines.push(...filesShown.has(node) ? fileNameLines(node, depth) : filePlaceholderLines(node, depth));
+  };
+  render(root, 0);
+  return lines.join(`
+`);
+}
+async function buildRepositoryOverview(cwd) {
+  const files = await listRepositoryFiles(cwd);
+  if (!files) {
+    return ["Top-level entries:", await buildDirectoryListing(cwd)].join(`
+`);
+  }
+  return [
+    `Repository files (${files.length}; dir/ (N) = N files inside, "\u2026" = more not shown; use glob_search or list_files for the rest):`,
+    formatRepositoryMap(files)
+  ].join(`
+`);
+}
 function buildGitSummary(cwd) {
   const cached = gitSummaryCache.get(cwd);
   if (cached && cached.expiresAt > Date.now()) {
@@ -14441,13 +14636,98 @@ async function buildWorkspaceInstructionBaseline({
 `)
   };
 }
+function memoizeEnvironmentSnapshot(statePath, snapshot) {
+  environmentSnapshotMemo.delete(statePath);
+  environmentSnapshotMemo.set(statePath, snapshot);
+  if (environmentSnapshotMemo.size > maxMemoizedEnvironmentSnapshots) {
+    const oldest = environmentSnapshotMemo.keys().next().value;
+    if (oldest !== undefined) {
+      environmentSnapshotMemo.delete(oldest);
+    }
+  }
+  return snapshot;
+}
+async function captureEnvironmentSnapshot(cwd) {
+  const [overview, git] = await Promise.all([
+    buildRepositoryOverview(cwd),
+    buildGitSummary(cwd)
+  ]);
+  const capturedAt = new Date().toISOString();
+  return {
+    cwd,
+    capturedAt,
+    body: [
+      `Working directory: ${cwd}`,
+      `Platform: ${describePlatform()}`,
+      `Date: ${capturedAt.slice(0, 10)} (session start)`,
+      "",
+      "Git (snapshot from session start; call git_status for the current state):",
+      git,
+      "",
+      overview
+    ].join(`
+`)
+  };
+}
+async function loadEnvironmentSnapshot(statePath, cwd) {
+  const memoized = environmentSnapshotMemo.get(statePath);
+  if (memoized?.cwd === cwd) {
+    return memoized;
+  }
+  try {
+    const parsed = environmentSnapshotSchema.safeParse(JSON.parse(await readFile4(statePath, "utf8")));
+    if (parsed.success && parsed.data.cwd === cwd) {
+      return memoizeEnvironmentSnapshot(statePath, parsed.data);
+    }
+    if (!parsed.success) {
+      logger7.warn("workspace_environment_snapshot_invalid", { statePath });
+    }
+  } catch (error) {
+    const code = isRecord3(error) ? Reflect.get(error, "code") : undefined;
+    if (code !== "ENOENT") {
+      logger7.warn("workspace_environment_snapshot_load_failed", {
+        statePath,
+        error: getErrorMessage(error)
+      });
+    }
+  }
+  return null;
+}
+async function resolveEnvironmentSnapshot({
+  cwd,
+  sessionId,
+  dataDir
+}) {
+  if (!sessionId) {
+    return captureEnvironmentSnapshot(cwd);
+  }
+  const statePath = sessionStatePath({
+    directory: "session-environment",
+    sessionId,
+    dataDir
+  });
+  const existing = await loadEnvironmentSnapshot(statePath, cwd);
+  if (existing) {
+    return existing;
+  }
+  const snapshot = memoizeEnvironmentSnapshot(statePath, await captureEnvironmentSnapshot(cwd));
+  try {
+    await writePrivateJsonFile(statePath, snapshot);
+  } catch (error) {
+    logger7.warn("workspace_environment_snapshot_persist_failed", {
+      statePath,
+      error: getErrorMessage(error)
+    });
+  }
+  return snapshot;
+}
 async function buildWorkspaceContext({
   cwd,
   sessionId,
   relatedPaths,
   dataDir
 }) {
-  const [baseline, nestedInstructions, listing, git] = await Promise.all([
+  const [baseline, nestedInstructions, snapshot] = await Promise.all([
     buildWorkspaceInstructionBaseline({ cwd }),
     buildNestedWorkspaceInstructionBlock({
       cwd,
@@ -14455,26 +14735,16 @@ async function buildWorkspaceContext({
       relatedPaths,
       dataDir
     }),
-    buildDirectoryListing(cwd),
-    buildGitSummary(cwd)
+    resolveEnvironmentSnapshot({ cwd, sessionId, dataDir })
   ]);
   const skills = buildAvailableSkills(cwd);
-  const today = new Date().toISOString().slice(0, 10);
   const parts = [
     baseline.block,
     nestedInstructions ? `
 ${nestedInstructions}` : "",
     "",
     "<environment>",
-    `Working directory: ${cwd}`,
-    `Platform: ${describePlatform()}`,
-    `Date: ${today}`,
-    "",
-    "Git:",
-    git,
-    "",
-    "Top-level entries:",
-    listing,
+    snapshot.body,
     skills ? `
 ${skills}` : "",
     "</environment>"
@@ -14485,48 +14755,12 @@ ${skills}` : "",
 \u2026 (environment truncated)
 </environment>` : block;
 }
-async function buildWorkspaceContextDelta({
-  cwd,
-  sessionId,
-  relatedPaths,
-  dataDir
-}) {
-  const [baseline, nestedInstructions, git] = await Promise.all([
-    buildWorkspaceInstructionBaseline({ cwd }),
-    buildNestedWorkspaceInstructionBlock({
-      cwd,
-      sessionId,
-      relatedPaths,
-      dataDir
-    }),
-    buildGitSummary(cwd)
-  ]);
-  const skills = buildAvailableSkills(cwd);
-  const today = new Date().toISOString().slice(0, 10);
-  return [
-    baseline.block,
-    nestedInstructions ? `
-${nestedInstructions}` : "",
-    "",
-    "<environment>",
-    `Working directory: ${cwd}`,
-    `Platform: ${describePlatform()}`,
-    `Date: ${today}`,
-    "(Top-level listing is available through list_files; current git state follows.)",
-    "",
-    "Git:",
-    git,
-    skills ? `
-${skills}` : "",
-    "</environment>"
-  ].join(`
-`);
-}
-var logger7, projectInstructionNames, maxListedEntries = 40, maxInstructionChars = 6000, maxNestedInstructionChars = 6000, maxNestedInstructionFileChars = 1600, maxBlockChars = 18000, instructionEpochSchema, instructionEpochStateSchema, gitSummaryCacheTtlMs = 5000, gitSummaryCache;
+var execFileAsync3, logger7, projectInstructionNames, maxListedEntries = 40, maxInstructionChars = 6000, maxNestedInstructionChars = 6000, maxNestedInstructionFileChars = 1600, maxRepositoryMapChars = 6000, maxRepositoryMapChildren = 32, maxBlockChars = 24000, repositoryFileListTimeoutMs = 5000, instructionEpochSchema, instructionEpochStateSchema, gitSummaryCacheTtlMs = 5000, gitSummaryCache, environmentSnapshotSchema, maxMemoizedEnvironmentSnapshots = 256, environmentSnapshotMemo;
 var init_workspace_context = __esm(() => {
   init_runtime17();
   init_src2();
   init_src();
+  execFileAsync3 = promisify4(execFile3);
   logger7 = createLogger("workspace-context");
   projectInstructionNames = ["AGENTS.md", "CLAUDE.md"];
   instructionEpochSchema = z41.object({
@@ -14540,6 +14774,12 @@ var init_workspace_context = __esm(() => {
     epochs: z41.array(instructionEpochSchema)
   });
   gitSummaryCache = new Map;
+  environmentSnapshotSchema = z41.object({
+    cwd: z41.string().min(1),
+    capturedAt: z41.string().min(1),
+    body: z41.string()
+  });
+  environmentSnapshotMemo = new Map;
 });
 
 // apps/server/src/lib/subagent-runner.ts
@@ -14621,7 +14861,8 @@ async function runSubagentToolTask(input, context2, deps) {
       abortSignal: context2.abortSignal,
       prepareStep: async ({ messages }) => {
         const artifactized = await artifactizeLargeToolOutputs(messages, {
-          signal: context2.abortSignal
+          signal: context2.abortSignal,
+          currentTurnThresholdCharacters: DEFAULT_CURRENT_TURN_TOOL_OUTPUT_THRESHOLD_CHARACTERS
         });
         const turn = assembler.assembleModelMessages(artifactized.messages, {
           preserveRecentTokens
@@ -15129,6 +15370,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
   });
   const contextConfig = lightcodeConfigResult.config.context;
   const cacheActive = resolvedProviderModel.supportsPromptCaching;
+  const rollingCompaction = resolvedProviderModel.rollingCompaction ?? !cacheActive;
   const endpointLimitTokens = getLearnedContextLimit(chatModelId);
   const effectiveContextWindow = resolveContextWindowTokens({
     config: contextConfig,
@@ -15136,9 +15378,12 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     endpointLimitTokens
   });
   const reservedOutputTokens = Math.min(resolveMaxOutputTokens(lightcodeConfigResult.config.maxOutputTokens, lightcodeConfigDefaults.maxOutputTokens, resolvedProviderModel.maxCompletionTokens), Math.max(1024, Math.floor(effectiveContextWindow / 2)));
-  const isFirstAssistantTurn = !validatedMessages.some((message) => message.role === "assistant");
   const relatedPaths = collectRelatedWorkspacePaths(validatedMessages);
-  const environmentContext = isFirstAssistantTurn ? await buildWorkspaceContext({ cwd, sessionId, relatedPaths }) : await buildWorkspaceContextDelta({ cwd, sessionId, relatedPaths });
+  const environmentContext = await buildWorkspaceContext({
+    cwd,
+    sessionId,
+    relatedPaths
+  });
   const availableSkillNames = listSkills({ cwd }).map((skill) => skill.name);
   const useFastChatPath = shouldUseFastChatPath({
     messages: providerInputMessages,
@@ -15208,6 +15453,7 @@ ${environmentContext}` : agentCallSettings.instructions;
     endpointLimitTokens,
     pendingInteractionCount,
     cacheActive,
+    rollingCompaction,
     fixedInputTokens: turnAssembler.fixedInputTokens
   });
   if (contextConfig.autoCompact) {
@@ -15238,6 +15484,7 @@ ${environmentContext}` : agentCallSettings.instructions;
           endpointLimitTokens,
           pendingInteractionCount,
           cacheActive,
+          rollingCompaction,
           fixedInputTokens: turnAssembler.fixedInputTokens
         });
         logger11.info("context_auto_compacted", {
@@ -15280,6 +15527,7 @@ ${environmentContext}` : agentCallSettings.instructions;
         endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive,
+        rollingCompaction,
         fixedInputTokens: turnAssembler.fixedInputTokens
       });
       if (overflowView.coveredMessages.length > 0) {
@@ -15306,6 +15554,7 @@ ${environmentContext}` : agentCallSettings.instructions;
         endpointLimitTokens,
         pendingInteractionCount: 0,
         cacheActive,
+        rollingCompaction,
         fixedInputTokens: turnAssembler.fixedInputTokens
       });
       logger11.info("context_overflow_recovery", {
@@ -15359,7 +15608,7 @@ ${environmentContext}` : agentCallSettings.instructions;
     noteContextOverflow(sessionId);
   };
   const maybeRunRollingCompaction = async (finishedMessages) => {
-    if (cacheActive || !contextConfig.autoCompact) {
+    if (!rollingCompaction || !contextConfig.autoCompact) {
       return;
     }
     if (rollingCompactionInFlight.has(sessionId)) {
@@ -15367,38 +15616,45 @@ ${environmentContext}` : agentCallSettings.instructions;
     }
     rollingCompactionInFlight.add(sessionId);
     try {
-      const latestState = await loadSessionContextStateSafe(sessionId);
-      const rollingView = buildProviderView({
-        messages: finishedMessages,
-        contextState: latestState,
-        config: contextConfig,
-        modelContextWindow: resolvedProviderModel.contextWindow,
-        reservedOutputTokens: budgetReservedOutputTokens,
-        endpointLimitTokens,
-        pendingInteractionCount: await countPendingChatInteractions(sessionId),
-        cacheActive,
-        fixedInputTokens: turnAssembler.fixedInputTokens
-      });
-      if (!rollingView.needsRollingCompaction) {
-        return;
+      let latestState = await loadSessionContextStateSafe(sessionId);
+      const pendingInteractionCount2 = await countPendingChatInteractions(sessionId);
+      for (let round = 1;round <= MAX_ROLLING_COMPACTION_ROUNDS; round += 1) {
+        const rollingView = buildProviderView({
+          messages: finishedMessages,
+          contextState: latestState,
+          config: contextConfig,
+          modelContextWindow: resolvedProviderModel.contextWindow,
+          reservedOutputTokens: budgetReservedOutputTokens,
+          endpointLimitTokens,
+          pendingInteractionCount: pendingInteractionCount2,
+          cacheActive,
+          rollingCompaction,
+          continueRolling: round > 1,
+          fixedInputTokens: turnAssembler.fixedInputTokens
+        });
+        if (!rollingView.needsRollingCompaction) {
+          return;
+        }
+        const compaction = await compactSessionContext({
+          sessionId,
+          coveredMessages: rollingView.coveredMessages,
+          previousState: latestState,
+          model: resolvedProviderModel.model,
+          modelId: chatModelId,
+          cwd,
+          config: contextConfig,
+          estimatedTokens: rollingView.estimate.tokens,
+          contextWindow: rollingView.contextWindow
+        });
+        latestState = compaction.state;
+        logger11.info("context_rolling_compaction", {
+          sessionId,
+          round,
+          tier: compaction.state.tier,
+          usedFallback: compaction.usedFallback,
+          coveredMessageCount: compaction.state.coveredMessageCount
+        });
       }
-      const compaction = await compactSessionContext({
-        sessionId,
-        coveredMessages: rollingView.coveredMessages,
-        previousState: latestState,
-        model: resolvedProviderModel.model,
-        modelId: chatModelId,
-        cwd,
-        config: contextConfig,
-        estimatedTokens: rollingView.estimate.tokens,
-        contextWindow: rollingView.contextWindow
-      });
-      logger11.info("context_rolling_compaction", {
-        sessionId,
-        tier: compaction.state.tier,
-        usedFallback: compaction.usedFallback,
-        coveredMessageCount: compaction.state.coveredMessageCount
-      });
     } catch (error) {
       logger11.error("context_rolling_compaction_failed", {
         sessionId,
@@ -15615,6 +15871,7 @@ ${environmentContext}` : agentCallSettings.instructions;
         contextWindow: view.contextWindow,
         preserveRecentTokens: view.preserveRecentTokens,
         originalInputTokens: estimateStructuralTokens(providerInputMessages) + turnAssembler.fixedInputTokens,
+        promptCacheActive: cacheActive,
         assemblyAbortSignal: requestAbortSignal,
         turnKey: [...validatedMessages].reverse().find((message) => message.role === "user")?.id
       },
@@ -15680,7 +15937,7 @@ ${environmentContext}` : agentCallSettings.instructions;
     }, 500);
   }
 }
-var logger11, recoverableDisconnectMessage = "Connection interrupted. Please retry or regenerate your last message.", providerBillingOrQuotaMessage, fastChatSystemPrompt, fastChatMaxOutputTokens = 512, MAX_COMPACTION_ROUNDS = 1, rollingCompactionInFlight;
+var logger11, recoverableDisconnectMessage = "Connection interrupted. Please retry or regenerate your last message.", providerBillingOrQuotaMessage, fastChatSystemPrompt, fastChatMaxOutputTokens = 512, MAX_COMPACTION_ROUNDS = 1, MAX_ROLLING_COMPACTION_ROUNDS = 4, rollingCompactionInFlight;
 var init_chat_stream = __esm(() => {
   init_src2();
   init_src();
@@ -15742,11 +15999,7 @@ async function assembleStoredSessionProviderTurn({
     messages,
     cwd: session.cwd
   });
-  const environmentContext = messages.some((message) => message.role === "assistant") ? await buildWorkspaceContextDelta({
-    cwd: session.cwd,
-    sessionId: session.id,
-    relatedPaths: collectRelatedWorkspacePaths(messages)
-  }) : await buildWorkspaceContext({
+  const environmentContext = await buildWorkspaceContext({
     cwd: session.cwd,
     sessionId: session.id,
     relatedPaths: collectRelatedWorkspacePaths(messages)
@@ -15799,6 +16052,7 @@ async function assembleStoredSessionProviderTurn({
     endpointLimitTokens,
     reservedOutputTokens,
     cacheActive: resolvedProviderModel.supportsPromptCaching,
+    rollingCompaction: resolvedProviderModel.rollingCompaction ?? !resolvedProviderModel.supportsPromptCaching,
     fixedInputTokens: baseAssembler.fixedInputTokens
   });
   const assembler = new ProviderTurnAssembler({
@@ -15811,7 +16065,10 @@ async function assembleStoredSessionProviderTurn({
   });
   const modelMessages = await convertToModelMessages(normalizeProviderMessages(view.providerMessages), { tools });
   const artifactized = await artifactizeLargeToolOutputs(modelMessages, {
-    signal: abortSignal
+    signal: abortSignal,
+    ...resolveToolOutputArtifactThresholds({
+      promptCacheActive: resolvedProviderModel.supportsPromptCaching
+    })
   });
   const assembled = assembler.assembleModelMessages(artifactized.messages, { preserveRecentTokens: view.preserveRecentTokens });
   return { view, assembled };