@kmugalkhod/lightcode 0.3.0 → 0.4.0

package/cli.js

@@ -124,7 +124,7 @@ var init_logger = __esm(() => {
 });
 
 // packages/shared/src/version.ts
-var lightcodeVersion = "0.3.0";
+var lightcodeVersion = "0.4.0";
 
 // packages/shared/src/index.ts
 var productName = "Lightcode";
@@ -998,7 +998,7 @@ var init_slash_page_menu = __esm(() => {
 });
 
 // packages/ai/src/constants.ts
-var MAX_TOOL_TEXT_OUTPUT_CHARS = 12000, MAX_TOOL_LIST_ENTRIES = 200, MAX_TOOL_SEARCH_RESULTS = 200, ANTHROPIC_TOOL_OPTIONAL_PARAMETER_BUDGET = 24;
+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;
 
 // packages/ai/src/coding-agent-modes.ts
 import { z } from "zod";
@@ -1594,13 +1594,13 @@ var init_base_schemas = __esm(() => {
 
 // packages/ai/src/bash/schema.ts
 import { z as z5 } from "zod";
-var bashDescription = "Run a shell command from workspace root with a timeout. Commands are classified before execution and may require approval.", bashInputSchema, bashProviderInputSchema, bashOutputSchema;
+var bashDescription = "Run a shell command from workspace root with a timeout. Bound noisy output at the source (head/tail/grep/wc, or write to a file then read a range) rather than dumping everything \u2014 output is truncated head+tail and wastes tokens. Commands are classified before execution and may require approval.", bashInputSchema, bashProviderInputSchema, bashOutputSchema;
 var init_schema = __esm(() => {
   init_base_schemas();
   bashInputSchema = z5.object({
     command: z5.string().min(1).max(4000),
     timeoutMs: integerRangeSchema(1000, 120000, "timeoutMs").optional().default(30000),
-    maxOutputChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxOutputChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxOutputChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxOutputChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   bashProviderInputSchema = z5.object({
     command: z5.string().min(1).max(4000)
@@ -1710,7 +1710,7 @@ var init_schema5 = __esm(() => {
     path: boundedPathSchema.optional(),
     staged: z9.boolean().optional().default(false),
     statOnly: z9.boolean().optional().default(false),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   gitDiffProviderInputSchema = z9.object({
     path: boundedPathSchema.optional(),
@@ -1725,7 +1725,7 @@ var init_schema5 = __esm(() => {
   gitShowInputSchema = z9.object({
     revision: z9.string().min(1).max(240),
     path: boundedPathSchema.optional(),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   gitShowProviderInputSchema = z9.object({
     revision: z9.string().min(1).max(240),
@@ -1822,7 +1822,7 @@ var init_schema6 = __esm(() => {
 
 // packages/ai/src/read-file/schema.ts
 import { z as z11 } from "zod";
-var readFileDescription = "Read file content from the working directory with optional line range limits.", readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
+var readFileDescription = "Read file content from the working directory. Prefer a startLine/endLine range to read only the region you need instead of the whole file \u2014 large files are truncated (head+tail) and waste tokens. Raise maxChars only when you truly need more.", readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
 var init_schema7 = __esm(() => {
   init_base_schemas();
   readFileInputSchema = z11.object({
@@ -2076,12 +2076,12 @@ var init_schema13 = __esm(() => {
 
 // packages/ai/src/web-fetch/schema.ts
 import { z as z19 } from "zod";
-var webFetchDescription = "Fetch a URL and return normalized, bounded text with source metadata. Requires network permission.", webFetchInputSchema, webFetchProviderInputSchema, webFetchErrorSchema, webFetchOutputSchema;
+var webFetchDescription = "Fetch a URL and return normalized, bounded text with source metadata. Keep maxChars modest and request more only if the answer wasn't in the returned text. Requires network permission.", webFetchInputSchema, webFetchProviderInputSchema, webFetchErrorSchema, webFetchOutputSchema;
 var init_schema14 = __esm(() => {
   init_base_schemas();
   webFetchInputSchema = z19.object({
     url: z19.string().url().max(4096),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS),
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS),
     timeoutMs: integerRangeSchema(1000, 30000, "timeoutMs").optional().default(1e4)
   });
   webFetchProviderInputSchema = z19.object({
@@ -2857,11 +2857,11 @@ ${toolCallingDisciplineAddendum}` : "");
 var cwdPromptPlaceholder = "{cwd}", defaultCodingAgentSystemPrompt, toolCallingDisciplineAddendum;
 var init_coding_agent_prompt = __esm(() => {
   init_coding_agent_modes();
-  defaultCodingAgentSystemPrompt = "You are a basic coding agent. Use tools for filesystem and codebase tasks instead of guessing. " + "Respect the user's intent, explain changes clearly, and prefer incremental, auditable actions. " + "You can only interact with files under this working directory: {cwd}. " + "Use glob_search for path discovery, grep for text search, structured git tools for repository inspection, and bash only when dedicated tools are not enough. " + "Use todo_write to keep multi-step implementation work visible and current. " + "For risky or uncertain operations, inspect context first and be explicit about assumptions. " + "While working, emit brief progress notes in natural language before major tool actions and after important findings. " + `Keep them short, human, and concrete. Vary wording naturally and avoid repetitive templates or rigid labels.
+  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.
 
-` + "You are an agent \u2014 keep working until the user's request is fully resolved before ending your turn. " + "Never stop with a partial answer. If you say you will do something, do it in the same turn by calling a tool. " + "Use todo_write to track multi-step work and do not finish while any todo is pending or in_progress. " + `Only end your turn when the task is complete or you are blocked on input that only the user can provide.
+` + `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.
 
-` + "Long-running processes: never run dev servers, watchers, or other non-exiting commands as plain foreground bash \u2014 " + "the bash tool enforces a timeout and the command will be killed. Start them detached with output redirected to a " + 'log file (e.g. on Windows: start /B cmd /c "npm run dev > dev.log 2>&1"; on Unix: nohup npm run dev > dev.log 2>&1 &), ' + "wait briefly, then verify by reading the log file or with a bounded probe like curl --max-time 5. " + "Dev servers do not always get their default port \u2014 read the actual port from the startup log instead of assuming it. " + "Never kill processes or free ports you did not open yourself; other tools (including this agent's own backend) may be using them.";
+` + "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.";
 });
 
@@ -3646,6 +3646,10 @@ var init_config3 = __esm(() => {
     preserveRecentMessages: z26.number().int().min(2).max(100).optional(),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).optional(),
     summaryMaxChars: z26.number().int().min(200).max(50000).optional(),
+    aggressivePruneWhenUncached: z26.boolean().optional(),
+    uncachedPruneAtFraction: z26.number().min(0.2).max(0.9).optional(),
+    uncachedPruneMinOutputChars: z26.number().int().min(200).max(12000).optional(),
+    uncachedQuantizeUserTurns: z26.number().int().min(1).max(20).optional(),
     maxInputTokens: z26.number().int().min(1).max(1e7).optional()
   }).strict();
   resolvedContextOptimizerConfigSchema = z26.object({
@@ -3654,15 +3658,23 @@ var init_config3 = __esm(() => {
     pruneAtFraction: z26.number().min(0.2).max(0.9),
     preserveRecentMessages: z26.number().int().min(2).max(100),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).nullable(),
-    summaryMaxChars: z26.number().int().min(200).max(50000)
+    summaryMaxChars: z26.number().int().min(200).max(50000),
+    aggressivePruneWhenUncached: z26.boolean(),
+    uncachedPruneAtFraction: z26.number().min(0.2).max(0.9),
+    uncachedPruneMinOutputChars: z26.number().int().min(200).max(12000),
+    uncachedQuantizeUserTurns: z26.number().int().min(1).max(20)
   });
   defaultContextOptimizerConfig = {
     autoCompact: true,
-    compactAtFraction: 0.8,
+    compactAtFraction: 0.7,
     pruneAtFraction: 0.6,
     preserveRecentMessages: 6,
     contextWindowOverride: null,
-    summaryMaxChars: 4000
+    summaryMaxChars: 4000,
+    aggressivePruneWhenUncached: true,
+    uncachedPruneAtFraction: 0.45,
+    uncachedPruneMinOutputChars: 600,
+    uncachedQuantizeUserTurns: 1
   };
 });
 
@@ -4128,9 +4140,13 @@ function buildProviderView(options) {
   }
   let estimate = estimateContextTokens(providerMessages);
   let tier1 = null;
-  if (estimate.tokens > config.pruneAtFraction * contextWindow) {
+  const uncachedActive = options.cacheActive === false && config.aggressivePruneWhenUncached;
+  const pruneFraction = uncachedActive ? config.uncachedPruneAtFraction : config.pruneAtFraction;
+  if (estimate.tokens > pruneFraction * contextWindow) {
     tier1 = pruneToolOutputs(providerMessages, {
-      preserveRecentMessages: config.preserveRecentMessages
+      preserveRecentMessages: config.preserveRecentMessages,
+      minOutputChars: uncachedActive ? config.uncachedPruneMinOutputChars : undefined,
+      quantizeUserTurns: uncachedActive ? config.uncachedQuantizeUserTurns : undefined
     });
     if (tier1.savedChars > 0) {
       providerMessages = tier1.messages;
@@ -4314,7 +4330,11 @@ function readEnvConfig(env) {
     contextWindowOverride: parseNumberEnv(env.LIGHTCODE_CONTEXT_WINDOW, "LIGHTCODE_CONTEXT_WINDOW"),
     maxInputTokens: parseNumberEnv(env.LIGHTCODE_CONTEXT_MAX_INPUT_TOKENS, "LIGHTCODE_CONTEXT_MAX_INPUT_TOKENS"),
     preserveRecentMessages: parseNumberEnv(env.LIGHTCODE_CONTEXT_PRESERVE_RECENT_MESSAGES, "LIGHTCODE_CONTEXT_PRESERVE_RECENT_MESSAGES"),
-    summaryMaxChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS, "LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS")
+    summaryMaxChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS, "LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS"),
+    aggressivePruneWhenUncached: parseBooleanEnv(env.LIGHTCODE_CONTEXT_AGGRESSIVE_PRUNE_UNCACHED, "LIGHTCODE_CONTEXT_AGGRESSIVE_PRUNE_UNCACHED"),
+    uncachedPruneAtFraction: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_PRUNE_AT_FRACTION, "LIGHTCODE_CONTEXT_UNCACHED_PRUNE_AT_FRACTION"),
+    uncachedPruneMinOutputChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_PRUNE_MIN_OUTPUT_CHARS, "LIGHTCODE_CONTEXT_UNCACHED_PRUNE_MIN_OUTPUT_CHARS"),
+    uncachedQuantizeUserTurns: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_QUANTIZE_USER_TURNS, "LIGHTCODE_CONTEXT_UNCACHED_QUANTIZE_USER_TURNS")
   };
   const compactContextConfig = Object.fromEntries(Object.entries(contextConfig).filter(([, value]) => value !== undefined));
   const headroomConfig = {
@@ -5949,19 +5969,29 @@ var init_resolve_within_workspace = __esm(() => {
 });
 
 // packages/ai/src/common/output-utils.ts
-function truncateText(text, maxChars = MAX_TOOL_TEXT_OUTPUT_CHARS) {
+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 suffix = `
-
-... output truncated to ${maxChars} characters ...`;
-  const allowedLength = Math.max(0, maxChars - suffix.length);
+  const markerReserve = omissionMarker(text.length).length;
+  const usable = Math.max(0, maxChars - markerReserve);
+  const headChars = Math.floor(usable * 0.7);
+  const tailChars = usable - headChars;
+  const head = text.slice(0, headChars);
+  const tail = tailChars > 0 ? text.slice(text.length - tailChars) : "";
+  const omitted = text.length - head.length - tail.length;
   return {
-    text: `${text.slice(0, allowedLength)}${suffix}`,
+    text: `${head}${omissionMarker(omitted)}${tail}`,
     truncated: true
   };
 }
@@ -6971,7 +7001,7 @@ async function executeReadFile(input, workspaceContext = getDefaultWorkspaceCont
   const { start, end } = clampLineRange(lines.length, parsedInput.startLine, parsedInput.endLine);
   const selectedContent = lines.slice(start - 1, end).join(`
 `);
-  const maxChars = parsedInput.maxChars ?? MAX_TOOL_TEXT_OUTPUT_CHARS;
+  const maxChars = parsedInput.maxChars ?? DEFAULT_TOOL_TEXT_OUTPUT_CHARS;
   const truncated = truncateText(selectedContent, maxChars);
   return readFileOutputSchema.parse({
     path: relativePath,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kmugalkhod/lightcode",
-  "version": "0.3.0",
+  "version": "0.4.0",
   "description": "Lightcode — a personal AI coding agent with a terminal UI and local session storage.",
   "type": "module",
   "license": "MIT",

package/server.js

@@ -147,7 +147,7 @@ var init_src = __esm(() => {
 });
 
 // packages/ai/src/constants.ts
-var MAX_TOOL_TEXT_OUTPUT_CHARS = 12000, MAX_TOOL_LIST_ENTRIES = 200, MAX_TOOL_SEARCH_RESULTS = 200, ANTHROPIC_TOOL_OPTIONAL_PARAMETER_BUDGET = 24;
+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;
 
 // packages/ai/src/coding-agent-modes.ts
 import { z } from "zod";
@@ -724,13 +724,13 @@ var init_base_schemas = __esm(() => {
 
 // packages/ai/src/bash/schema.ts
 import { z as z5 } from "zod";
-var bashDescription = "Run a shell command from workspace root with a timeout. Commands are classified before execution and may require approval.", bashInputSchema, bashProviderInputSchema, bashOutputSchema;
+var bashDescription = "Run a shell command from workspace root with a timeout. Bound noisy output at the source (head/tail/grep/wc, or write to a file then read a range) rather than dumping everything \u2014 output is truncated head+tail and wastes tokens. Commands are classified before execution and may require approval.", bashInputSchema, bashProviderInputSchema, bashOutputSchema;
 var init_schema = __esm(() => {
   init_base_schemas();
   bashInputSchema = z5.object({
     command: z5.string().min(1).max(4000),
     timeoutMs: integerRangeSchema(1000, 120000, "timeoutMs").optional().default(30000),
-    maxOutputChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxOutputChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxOutputChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxOutputChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   bashProviderInputSchema = z5.object({
     command: z5.string().min(1).max(4000)
@@ -840,7 +840,7 @@ var init_schema5 = __esm(() => {
     path: boundedPathSchema.optional(),
     staged: z9.boolean().optional().default(false),
     statOnly: z9.boolean().optional().default(false),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   gitDiffProviderInputSchema = z9.object({
     path: boundedPathSchema.optional(),
@@ -855,7 +855,7 @@ var init_schema5 = __esm(() => {
   gitShowInputSchema = z9.object({
     revision: z9.string().min(1).max(240),
     path: boundedPathSchema.optional(),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS)
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS)
   });
   gitShowProviderInputSchema = z9.object({
     revision: z9.string().min(1).max(240),
@@ -952,7 +952,7 @@ var init_schema6 = __esm(() => {
 
 // packages/ai/src/read-file/schema.ts
 import { z as z11 } from "zod";
-var readFileDescription = "Read file content from the working directory with optional line range limits.", readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
+var readFileDescription = "Read file content from the working directory. Prefer a startLine/endLine range to read only the region you need instead of the whole file \u2014 large files are truncated (head+tail) and waste tokens. Raise maxChars only when you truly need more.", readFileInputSchema, readFileProviderInputSchema, readFileOutputSchema;
 var init_schema7 = __esm(() => {
   init_base_schemas();
   readFileInputSchema = z11.object({
@@ -1206,12 +1206,12 @@ var init_schema13 = __esm(() => {
 
 // packages/ai/src/web-fetch/schema.ts
 import { z as z19 } from "zod";
-var webFetchDescription = "Fetch a URL and return normalized, bounded text with source metadata. Requires network permission.", webFetchInputSchema, webFetchProviderInputSchema, webFetchErrorSchema, webFetchOutputSchema;
+var webFetchDescription = "Fetch a URL and return normalized, bounded text with source metadata. Keep maxChars modest and request more only if the answer wasn't in the returned text. Requires network permission.", webFetchInputSchema, webFetchProviderInputSchema, webFetchErrorSchema, webFetchOutputSchema;
 var init_schema14 = __esm(() => {
   init_base_schemas();
   webFetchInputSchema = z19.object({
     url: z19.string().url().max(4096),
-    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(MAX_TOOL_TEXT_OUTPUT_CHARS),
+    maxChars: integerRangeSchema(200, MAX_TOOL_TEXT_OUTPUT_CHARS, "maxChars").optional().default(DEFAULT_TOOL_TEXT_OUTPUT_CHARS),
     timeoutMs: integerRangeSchema(1000, 30000, "timeoutMs").optional().default(1e4)
   });
   webFetchProviderInputSchema = z19.object({
@@ -1899,11 +1899,11 @@ ${toolCallingDisciplineAddendum}` : "");
 var cwdPromptPlaceholder = "{cwd}", defaultCodingAgentSystemPrompt, toolCallingDisciplineAddendum;
 var init_coding_agent_prompt = __esm(() => {
   init_coding_agent_modes();
-  defaultCodingAgentSystemPrompt = "You are a basic coding agent. Use tools for filesystem and codebase tasks instead of guessing. " + "Respect the user's intent, explain changes clearly, and prefer incremental, auditable actions. " + "You can only interact with files under this working directory: {cwd}. " + "Use glob_search for path discovery, grep for text search, structured git tools for repository inspection, and bash only when dedicated tools are not enough. " + "Use todo_write to keep multi-step implementation work visible and current. " + "For risky or uncertain operations, inspect context first and be explicit about assumptions. " + "While working, emit brief progress notes in natural language before major tool actions and after important findings. " + `Keep them short, human, and concrete. Vary wording naturally and avoid repetitive templates or rigid labels.
+  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.
 
-` + "You are an agent \u2014 keep working until the user's request is fully resolved before ending your turn. " + "Never stop with a partial answer. If you say you will do something, do it in the same turn by calling a tool. " + "Use todo_write to track multi-step work and do not finish while any todo is pending or in_progress. " + `Only end your turn when the task is complete or you are blocked on input that only the user can provide.
+` + `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.
 
-` + "Long-running processes: never run dev servers, watchers, or other non-exiting commands as plain foreground bash \u2014 " + "the bash tool enforces a timeout and the command will be killed. Start them detached with output redirected to a " + 'log file (e.g. on Windows: start /B cmd /c "npm run dev > dev.log 2>&1"; on Unix: nohup npm run dev > dev.log 2>&1 &), ' + "wait briefly, then verify by reading the log file or with a bounded probe like curl --max-time 5. " + "Dev servers do not always get their default port \u2014 read the actual port from the startup log instead of assuming it. " + "Never kill processes or free ports you did not open yourself; other tools (including this agent's own backend) may be using them.";
+` + "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.";
 });
 
@@ -2674,6 +2674,10 @@ var init_config3 = __esm(() => {
     preserveRecentMessages: z26.number().int().min(2).max(100).optional(),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).optional(),
     summaryMaxChars: z26.number().int().min(200).max(50000).optional(),
+    aggressivePruneWhenUncached: z26.boolean().optional(),
+    uncachedPruneAtFraction: z26.number().min(0.2).max(0.9).optional(),
+    uncachedPruneMinOutputChars: z26.number().int().min(200).max(12000).optional(),
+    uncachedQuantizeUserTurns: z26.number().int().min(1).max(20).optional(),
     maxInputTokens: z26.number().int().min(1).max(1e7).optional()
   }).strict();
   resolvedContextOptimizerConfigSchema = z26.object({
@@ -2682,15 +2686,23 @@ var init_config3 = __esm(() => {
     pruneAtFraction: z26.number().min(0.2).max(0.9),
     preserveRecentMessages: z26.number().int().min(2).max(100),
     contextWindowOverride: z26.number().int().min(8000).max(1e7).nullable(),
-    summaryMaxChars: z26.number().int().min(200).max(50000)
+    summaryMaxChars: z26.number().int().min(200).max(50000),
+    aggressivePruneWhenUncached: z26.boolean(),
+    uncachedPruneAtFraction: z26.number().min(0.2).max(0.9),
+    uncachedPruneMinOutputChars: z26.number().int().min(200).max(12000),
+    uncachedQuantizeUserTurns: z26.number().int().min(1).max(20)
   });
   defaultContextOptimizerConfig = {
     autoCompact: true,
-    compactAtFraction: 0.8,
+    compactAtFraction: 0.7,
     pruneAtFraction: 0.6,
     preserveRecentMessages: 6,
     contextWindowOverride: null,
-    summaryMaxChars: 4000
+    summaryMaxChars: 4000,
+    aggressivePruneWhenUncached: true,
+    uncachedPruneAtFraction: 0.45,
+    uncachedPruneMinOutputChars: 600,
+    uncachedQuantizeUserTurns: 1
   };
 });
 
@@ -3156,9 +3168,13 @@ function buildProviderView(options) {
   }
   let estimate = estimateContextTokens(providerMessages);
   let tier1 = null;
-  if (estimate.tokens > config.pruneAtFraction * contextWindow) {
+  const uncachedActive = options.cacheActive === false && config.aggressivePruneWhenUncached;
+  const pruneFraction = uncachedActive ? config.uncachedPruneAtFraction : config.pruneAtFraction;
+  if (estimate.tokens > pruneFraction * contextWindow) {
     tier1 = pruneToolOutputs(providerMessages, {
-      preserveRecentMessages: config.preserveRecentMessages
+      preserveRecentMessages: config.preserveRecentMessages,
+      minOutputChars: uncachedActive ? config.uncachedPruneMinOutputChars : undefined,
+      quantizeUserTurns: uncachedActive ? config.uncachedQuantizeUserTurns : undefined
     });
     if (tier1.savedChars > 0) {
       providerMessages = tier1.messages;
@@ -3339,7 +3355,11 @@ function readEnvConfig(env) {
     contextWindowOverride: parseNumberEnv(env.LIGHTCODE_CONTEXT_WINDOW, "LIGHTCODE_CONTEXT_WINDOW"),
     maxInputTokens: parseNumberEnv(env.LIGHTCODE_CONTEXT_MAX_INPUT_TOKENS, "LIGHTCODE_CONTEXT_MAX_INPUT_TOKENS"),
     preserveRecentMessages: parseNumberEnv(env.LIGHTCODE_CONTEXT_PRESERVE_RECENT_MESSAGES, "LIGHTCODE_CONTEXT_PRESERVE_RECENT_MESSAGES"),
-    summaryMaxChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS, "LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS")
+    summaryMaxChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS, "LIGHTCODE_CONTEXT_SUMMARY_MAX_CHARS"),
+    aggressivePruneWhenUncached: parseBooleanEnv(env.LIGHTCODE_CONTEXT_AGGRESSIVE_PRUNE_UNCACHED, "LIGHTCODE_CONTEXT_AGGRESSIVE_PRUNE_UNCACHED"),
+    uncachedPruneAtFraction: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_PRUNE_AT_FRACTION, "LIGHTCODE_CONTEXT_UNCACHED_PRUNE_AT_FRACTION"),
+    uncachedPruneMinOutputChars: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_PRUNE_MIN_OUTPUT_CHARS, "LIGHTCODE_CONTEXT_UNCACHED_PRUNE_MIN_OUTPUT_CHARS"),
+    uncachedQuantizeUserTurns: parseNumberEnv(env.LIGHTCODE_CONTEXT_UNCACHED_QUANTIZE_USER_TURNS, "LIGHTCODE_CONTEXT_UNCACHED_QUANTIZE_USER_TURNS")
   };
   const compactContextConfig = Object.fromEntries(Object.entries(contextConfig).filter(([, value]) => value !== undefined));
   const headroomConfig = {
@@ -7230,6 +7250,7 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     }, 500);
   }
   const contextConfig = lightcodeConfigResult.config.context;
+  const cacheActive = resolvedProviderModel.provider === "anthropic";
   const pendingInteractionCount = await countPendingChatInteractions(sessionId);
   let contextState = await loadSessionContextStateSafe(sessionId);
   let view = buildProviderView({
@@ -7237,7 +7258,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
     contextState,
     config: contextConfig,
     modelContextWindow: resolvedProviderModel.contextWindow,
-    pendingInteractionCount
+    pendingInteractionCount,
+    cacheActive
   });
   if (view.needsCompaction && contextConfig.autoCompact) {
     try {
@@ -7258,7 +7280,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: contextConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
-        pendingInteractionCount
+        pendingInteractionCount,
+        cacheActive
       });
       logger7.info("context_auto_compacted", {
         sessionId,
@@ -7287,7 +7310,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
-        pendingInteractionCount: 0
+        pendingInteractionCount: 0,
+        cacheActive
       });
       if (overflowView.coveredMessages.length > 0) {
         const compaction = await compactSessionContext({
@@ -7307,7 +7331,8 @@ async function streamSessionChat(c, sessionIdentifier, messagesPayload, cwd, mod
         contextState,
         config: overflowConfig,
         modelContextWindow: resolvedProviderModel.contextWindow,
-        pendingInteractionCount: 0
+        pendingInteractionCount: 0,
+        cacheActive
       });
       logger7.info("context_overflow_recovery", {
         sessionId,