zelari-code 2.53.0 → 2.54.0

package/dist/cli/main.bundled.js

@@ -6537,9 +6537,9 @@ var init_schemas = __esm({
         };
         doc.write(`const input = payload.value;`);
         const ids = /* @__PURE__ */ Object.create(null);
-        let counter = 0;
+        let counter2 = 0;
         for (const key of normalized.keys) {
-          ids[key] = `key_${counter++}`;
+          ids[key] = `key_${counter2++}`;
         }
         doc.write(`const newResult = {};`);
         for (const key of normalized.keys) {
@@ -21738,9 +21738,51 @@ function buildSystemPrompt(agent, options) {
 
 ${volatile}`;
 }
+function resolvePromptLayout(env = process.env) {
+  if (promptLayoutCache === void 0) {
+    promptLayoutCache = env[PROMPT_LAYOUT_ENV]?.trim().toLowerCase() === "legacy" ? "legacy" : "trailing";
+  }
+  return promptLayoutCache;
+}
+function resetPromptLayoutCache() {
+  promptLayoutCache = void 0;
+}
+function wrapTrailingContext(volatile) {
+  const body = volatile.trim();
+  if (lastTrailingRender?.volatile === body)
+    return lastTrailingRender.body;
+  const wrapped = body ? `${TRAILING_CONTEXT_OPEN_TAG}
+${body}
+${TRAILING_CONTEXT_CLOSE_TAG}` : "";
+  lastTrailingRender = { volatile: body, body: wrapped };
+  return wrapped;
+}
+function trailingContextFromSplit(split) {
+  return wrapTrailingContext(split.volatile);
+}
+function isTrailingContextContent(content) {
+  const trimmed = content.trim();
+  return trimmed.startsWith(TRAILING_CONTEXT_OPEN_TAG) && trimmed.endsWith(TRAILING_CONTEXT_CLOSE_TAG);
+}
+function trailingContextMessagesFromSplit(split) {
+  const body = trailingContextFromSplit(split);
+  return body ? [{ role: "user", content: body }] : [];
+}
+function assembleRequestMessages(input) {
+  const layout = input.layout ?? resolvePromptLayout();
+  const systemMessages = systemMessagesFromSplit(input.split, {
+    includeVolatile: layout === "legacy"
+  });
+  const trailingMessages = layout === "legacy" ? [] : trailingContextMessagesFromSplit(input.split);
+  return {
+    messages: [...systemMessages, ...input.history, ...trailingMessages, ...input.turn],
+    systemCount: systemMessages.length,
+    trailingCount: trailingMessages.length
+  };
+}
 function systemMessagesFromSplit(split, opts) {
   const stable = split.stable.trim();
-  const volatile = split.volatile.trim();
+  const volatile = opts?.includeVolatile ? split.volatile.trim() : "";
   if (!stable && !volatile)
     return [];
   if (opts?.singleSystem) {
@@ -21754,12 +21796,16 @@ function systemMessagesFromSplit(split, opts) {
     msgs.push({ role: "system", content: volatile });
   return msgs;
 }
+var PROMPT_LAYOUT_ENV, TRAILING_CONTEXT_OPEN_TAG, TRAILING_CONTEXT_CLOSE_TAG, promptLayoutCache, lastTrailingRender;
 var init_systemPromptBuilder = __esm({
   "packages/core/dist/agents/systemPromptBuilder.js"() {
     "use strict";
     init_promptModules();
     init_secrecyPolicy();
     init_skills();
+    PROMPT_LAYOUT_ENV = "ZELARI_PROMPT_LAYOUT";
+    TRAILING_CONTEXT_OPEN_TAG = "<context-update>";
+    TRAILING_CONTEXT_CLOSE_TAG = "</context-update>";
   }
 });
 
@@ -22137,10 +22183,14 @@ __export(skills_exports, {
   KRAKEN_SELECTION_PLAYBOOK_MODULE: () => KRAKEN_SELECTION_PLAYBOOK_MODULE,
   LANGUAGE_POLICY_MODULE_TYPE: () => LANGUAGE_POLICY_MODULE_TYPE,
   NATIVE_TOOL_PROTOCOL_MODULE: () => NATIVE_TOOL_PROTOCOL_MODULE,
+  PROMPT_LAYOUT_ENV: () => PROMPT_LAYOUT_ENV,
   SINGLE_AGENT_IDENTITY_MODULE: () => SINGLE_AGENT_IDENTITY_MODULE,
   SKILL_CATALOG: () => SKILL_CATALOG,
   TOOL_DEFINITIONS: () => TOOL_DEFINITIONS,
+  TRAILING_CONTEXT_CLOSE_TAG: () => TRAILING_CONTEXT_CLOSE_TAG,
+  TRAILING_CONTEXT_OPEN_TAG: () => TRAILING_CONTEXT_OPEN_TAG,
   VAULT_TOOL_DEFINITIONS: () => VAULT_TOOL_DEFINITIONS,
+  assembleRequestMessages: () => assembleRequestMessages,
   buildCustomParameters: () => buildCustomParameters,
   buildLanguageDirective: () => buildLanguageDirective,
   buildLanguagePolicyModule: () => buildLanguagePolicyModule,
@@ -22164,20 +22214,26 @@ __export(skills_exports, {
   getSkillById: () => getSkillById,
   getSkillMetadata: () => getSkillMetadata,
   getSkillsByCategory: () => getSkillsByCategory,
+  isTrailingContextContent: () => isTrailingContextContent,
   isValidTool: () => isValidTool,
   listCodingSkills: () => listCodingSkills,
   listSkills: () => listSkills,
   registerCodingSkill: () => registerCodingSkill,
   registerCustomTool: () => registerCustomTool,
   registerSkill: () => registerSkill,
+  resetPromptLayoutCache: () => resetPromptLayoutCache,
   resolveAgentSkills: () => resolveAgentSkills,
+  resolvePromptLayout: () => resolvePromptLayout,
   resolveResponseLanguage: () => resolveResponseLanguage,
   resolveSkillDependencies: () => resolveSkillDependencies,
   setWorkspaceStubs: () => setWorkspaceStubs,
   systemMessagesFromSplit: () => systemMessagesFromSplit,
+  trailingContextFromSplit: () => trailingContextFromSplit,
+  trailingContextMessagesFromSplit: () => trailingContextMessagesFromSplit,
   unregisterCustomTool: () => unregisterCustomTool,
   unregisterSkill: () => unregisterSkill,
-  validateCodingSkillRequires: () => validateCodingSkillRequires
+  validateCodingSkillRequires: () => validateCodingSkillRequires,
+  wrapTrailingContext: () => wrapTrailingContext
 });
 var init_skills2 = __esm({
   "packages/core/dist/skills/index.js"() {
@@ -36997,6 +37053,7 @@ __export(council_exports, {
   NFR_KEYWORDS: () => NFR_KEYWORDS,
   NON_RETRY_AGENTS: () => NON_RETRY_AGENTS,
   OUTPUT_QUALITY_DIRECTIVE: () => OUTPUT_QUALITY_DIRECTIVE,
+  PROMPT_LAYOUT_ENV: () => PROMPT_LAYOUT_ENV,
   PROMPT_MODULES: () => PROMPT_MODULES,
   PROPRIETARY_REFUSAL_TEXT: () => PROPRIETARY_REFUSAL_TEXT,
   PROPRIETARY_SECRECY_MARKER: () => PROPRIETARY_SECRECY_MARKER,
@@ -37005,6 +37062,8 @@ __export(council_exports, {
   STRUCTURED_REASONING_DIRECTIVE: () => STRUCTURED_REASONING_DIRECTIVE,
   TIER_RANK: () => TIER_RANK,
   TOOL_USE_PROTOCOL_DIRECTIVE: () => TOOL_USE_PROTOCOL_DIRECTIVE,
+  TRAILING_CONTEXT_CLOSE_TAG: () => TRAILING_CONTEXT_CLOSE_TAG,
+  TRAILING_CONTEXT_OPEN_TAG: () => TRAILING_CONTEXT_OPEN_TAG,
   TURN_COMPLETION_MODULE: () => TURN_COMPLETION_MODULE,
   UnknownMemberError: () => UnknownMemberError,
   applyCompletionRetry: () => applyCompletionRetry,
@@ -37013,6 +37072,7 @@ __export(council_exports, {
   applyInlineJsAutofix: () => applyInlineJsAutofix,
   applyMotionAutofix: () => applyMotionAutofix,
   applyRetryIfMissing: () => applyRetryIfMissing,
+  assembleRequestMessages: () => assembleRequestMessages,
   auditDegradedBanner: () => auditDegradedBanner,
   auditSynthesisTiers: () => auditSynthesisTiers,
   buildCouncilCompletion: () => buildCouncilCompletion,
@@ -37054,6 +37114,7 @@ __export(council_exports, {
   getToolDescriptions: () => getToolDescriptions,
   hasInteractiveClarification: () => hasInteractiveClarification,
   isAnswerLeak: () => isAnswerLeak,
+  isTrailingContextContent: () => isTrailingContextContent,
   isVerifyToolCheckSkipped: () => isVerifyToolCheckSkipped,
   jaccardSimilarity: () => jaccardSimilarity,
   lintSynthesisHonesty: () => lintSynthesisHonesty,
@@ -37071,7 +37132,9 @@ __export(council_exports, {
   recallLessons: () => recallLessons,
   renderSkillMarkdown: () => renderSkillMarkdown,
   replayChairmanTextTools: () => replayChairmanTextTools,
+  resetPromptLayoutCache: () => resetPromptLayoutCache,
   resolveCouncilRunMode: () => resolveCouncilRunMode,
+  resolvePromptLayout: () => resolvePromptLayout,
   resolveRoleSystemPrompt: () => resolveRoleSystemPrompt,
   resolveVerifyRetryTool: () => resolveVerifyRetryTool,
   restrictImplementationWrites: () => restrictImplementationWrites,
@@ -37094,8 +37157,11 @@ __export(council_exports, {
   taskMatchesNfrKeywords: () => taskMatchesNfrKeywords,
   tierAtLeast: () => tierAtLeast,
   tokenizeForSignature: () => tokenizeForSignature,
+  trailingContextFromSplit: () => trailingContextFromSplit,
+  trailingContextMessagesFromSplit: () => trailingContextMessagesFromSplit,
   verifyCitations: () => verifyCitations,
   warnIfNfrSpecMissing: () => warnIfNfrSpecMissing,
+  wrapTrailingContext: () => wrapTrailingContext,
   writeCouncilCompletion: () => writeCouncilCompletion,
   writeVerificationReport: () => writeVerificationReport
 });
@@ -40038,7 +40104,7 @@ var CORE_VERSION;
 var init_version = __esm({
   "packages/core/dist/version.js"() {
     "use strict";
-    CORE_VERSION = "2.53.0";
+    CORE_VERSION = "2.54.0";
   }
 });
 
@@ -40196,6 +40262,7 @@ __export(dist_exports, {
   OUTPUT_QUALITY_DIRECTIVE: () => OUTPUT_QUALITY_DIRECTIVE,
   ObserverBus: () => ObserverBus,
   PROFILE_RESOURCE_POLICIES: () => PROFILE_RESOURCE_POLICIES,
+  PROMPT_LAYOUT_ENV: () => PROMPT_LAYOUT_ENV,
   PROMPT_MODULES: () => PROMPT_MODULES,
   PROPRIETARY_REFUSAL_TEXT: () => PROPRIETARY_REFUSAL_TEXT,
   PROPRIETARY_SECRECY_MARKER: () => PROPRIETARY_SECRECY_MARKER,
@@ -40243,6 +40310,8 @@ __export(dist_exports, {
   TOOL_CALL_TRUNCATED_RECOVERY_USER: () => TOOL_CALL_TRUNCATED_RECOVERY_USER,
   TOOL_DEFINITIONS: () => TOOL_DEFINITIONS,
   TOOL_USE_PROTOCOL_DIRECTIVE: () => TOOL_USE_PROTOCOL_DIRECTIVE,
+  TRAILING_CONTEXT_CLOSE_TAG: () => TRAILING_CONTEXT_CLOSE_TAG,
+  TRAILING_CONTEXT_OPEN_TAG: () => TRAILING_CONTEXT_OPEN_TAG,
   TURN_COMPLETION_MODULE: () => TURN_COMPLETION_MODULE,
   TaskConstraintSchema: () => TaskConstraintSchema,
   TaskContractConflictError: () => TaskContractConflictError,
@@ -40273,6 +40342,7 @@ __export(dist_exports, {
   applyMotionAutofix: () => applyMotionAutofix,
   applyRetryIfMissing: () => applyRetryIfMissing,
   applyTaskContractUpdate: () => applyTaskContractUpdate,
+  assembleRequestMessages: () => assembleRequestMessages,
   auditDegradedBanner: () => auditDegradedBanner,
   auditSynthesisTiers: () => auditSynthesisTiers,
   budgetPressure: () => budgetPressure,
@@ -40434,6 +40504,7 @@ __export(dist_exports, {
   isSettled: () => isSettled,
   isStatusTheaterUnit: () => isStatusTheaterUnit,
   isSteerControlEvent: () => isSteerControlEvent,
+  isTrailingContextContent: () => isTrailingContextContent,
   isValidTool: () => isValidTool,
   isVerificationReserveProtected: () => isVerificationReserveProtected,
   isVerifyParallelEnabled: () => isVerifyParallelEnabled,
@@ -40510,6 +40581,7 @@ __export(dist_exports, {
   renderSkillMarkdown: () => renderSkillMarkdown,
   renderSteers: () => renderSteers,
   replayChairmanTextTools: () => replayChairmanTextTools,
+  resetPromptLayoutCache: () => resetPromptLayoutCache,
   resolveAgentSkills: () => resolveAgentSkills,
   resolveCommandConcurrency: () => resolveCommandConcurrency,
   resolveCouncilRunMode: () => resolveCouncilRunMode,
@@ -40519,6 +40591,7 @@ __export(dist_exports, {
   resolveMaxTentacles: () => resolveMaxTentacles,
   resolvePlanTimeoutMs: () => resolvePlanTimeoutMs,
   resolveProfile: () => resolveProfile,
+  resolvePromptLayout: () => resolvePromptLayout,
   resolveRequestSnapshotMode: () => resolveRequestSnapshotMode,
   resolveResponseLanguage: () => resolveResponseLanguage,
   resolveRoleSystemPrompt: () => resolveRoleSystemPrompt,
@@ -40574,6 +40647,8 @@ __export(dist_exports, {
   toolManifestHash: () => toolManifestHash,
   toolMatches: () => toolMatches,
   topoLevels: () => topoLevels,
+  trailingContextFromSplit: () => trailingContextFromSplit,
+  trailingContextMessagesFromSplit: () => trailingContextMessagesFromSplit,
   unregisterCustomTool: () => unregisterCustomTool,
   unregisterSkill: () => unregisterSkill,
   usageFromLedger: () => usageFromLedger,
@@ -40590,6 +40665,7 @@ __export(dist_exports, {
   weaknessFromVerdict: () => weaknessFromVerdict,
   weaknessScoreFromText: () => weaknessScoreFromText,
   wrapLegacyStream: () => wrapLegacyStream,
+  wrapTrailingContext: () => wrapTrailingContext,
   writeCouncilCompletion: () => writeCouncilCompletion,
   writeVerificationReport: () => writeVerificationReport
 });
@@ -43803,12 +43879,12 @@ function loadHandle(rootDir) {
   }
   const {
     tasks: rawTasks,
-    counter,
+    counter: counter2,
     schemaVersion: _schemaVersion,
     ...rootFields
   } = parsed;
   const tasks = (Array.isArray(rawTasks) ? rawTasks : []).map(normalizeTask).filter((t) => typeof t.id === "string" && t.id.length > 0);
-  const numericCounter = typeof counter === "number" && Number.isFinite(counter) && counter >= 0 ? Math.floor(counter) : 0;
+  const numericCounter = typeof counter2 === "number" && Number.isFinite(counter2) && counter2 >= 0 ? Math.floor(counter2) : 0;
   return { rootDir, tasks, counter: numericCounter, rootFields };
 }
 function saveHandle(rootDir, handle) {
@@ -47010,14 +47086,14 @@ function parseRecordLine(line) {
   }
   if (typeof raw !== "object" || raw === null) return null;
   const r = raw;
-  const num = (v) => typeof v === "number" && Number.isFinite(v);
+  const num2 = (v) => typeof v === "number" && Number.isFinite(v);
   const str2 = (v) => typeof v === "string";
   const strOrNull = (v) => v === null || typeof v === "string";
-  const numOrNull = (v) => v === null || num(v);
-  if (!num(r.ts) || !str2(r.repo) || r.repo === "" || !str2(r.role)) return null;
+  const numOrNull = (v) => v === null || num2(v);
+  if (!num2(r.ts) || !str2(r.repo) || r.repo === "" || !str2(r.role)) return null;
   if (!strOrNull(r.model) || !strOrNull(r.provider) || !strOrNull(r.language)) return null;
   if (!REPUTATION_OUTCOMES.includes(r.outcome)) return null;
-  if (typeof r.firstPass !== "boolean" || !num(r.repairCount) || r.repairCount < 0) return null;
+  if (typeof r.firstPass !== "boolean" || !num2(r.repairCount) || r.repairCount < 0) return null;
   if (!numOrNull(r.costUsd) || !numOrNull(r.latencyMs)) return null;
   return {
     ts: r.ts,
@@ -54905,6 +54981,7 @@ function parseCachedPromptTokens(usage) {
   if (!usage || typeof usage !== "object") return 0;
   const candidates = [
     usage.prompt_tokens_details?.cached_tokens,
+    usage.input_tokens_details?.cached_tokens,
     usage.prompt_cache_hit_tokens
   ];
   for (const c of candidates) {
@@ -55011,6 +55088,8 @@ function positiveEnvInt(name) {
   return Number.isFinite(n) && n > 0 ? n : void 0;
 }
 function openaiCompatibleProvider(config2) {
+  const sessionDeepSeekThinking = resolveDeepSeekThinking();
+  const sessionToolChoice = "auto";
   return async function* (params) {
     const capabilities = capabilitiesFor(params.model, config2.providerId);
     const streamTimeouts = resolveStreamTimeouts(capabilities);
@@ -55068,7 +55147,7 @@ function openaiCompatibleProvider(config2) {
     }
     const thinkingSpec = config2.thinking ?? "auto";
     if (config2.providerId === "deepseek" && thinkingSpec === "auto") {
-      const thinking = resolveDeepSeekThinking();
+      const thinking = sessionDeepSeekThinking;
       if (thinking.thinking) body.thinking = { type: thinking.thinking };
       if (thinking.reasoningEffort) body.reasoning_effort = thinking.reasoningEffort;
     } else if (thinkingSpec !== "auto") {
@@ -55093,7 +55172,7 @@ function openaiCompatibleProvider(config2) {
       }));
       const recoveryAttempt = generation?.recoveryAttempt ?? 1;
       const forceRecoveryTool = generation?.toolChoice === "required" && capabilities.buildRecovery.forceToolChoice && recoveryAttempt <= capabilities.buildRecovery.maxForcedTurns;
-      body.tool_choice = forceRecoveryTool ? "required" : "auto";
+      body.tool_choice = forceRecoveryTool ? "required" : sessionToolChoice;
       if (config2.providerId === "glm") body.tool_stream = true;
     }
     const headers3 = {
@@ -56665,15 +56744,30 @@ function withRollingCacheBreakpoint(msg, cacheControl) {
   }
   return msg;
 }
+function trailingContextIndex(messages) {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const content = messages[i].content;
+    if (typeof content === "string" && isTrailingContextContent(content)) return i;
+  }
+  return -1;
+}
 function anthropicMessagesProvider(config2) {
   return async function* (params) {
     const { systemParts, rest } = splitMessages(params.messages);
     const ttlPref = resolvePromptCacheTtl();
     const cacheControl = ttlPref === "1h" ? { type: "ephemeral", ttl: "1h" } : { type: "ephemeral" };
+    const trailingIdx = trailingContextIndex(rest);
+    const rollingIndices = /* @__PURE__ */ new Set();
+    if (rest.length > 0) {
+      if (trailingIdx >= 0) rollingIndices.add(trailingIdx);
+      rollingIndices.add(rest.length - 1);
+    }
     const body = {
       model: params.model,
       max_tokens: 16384,
-      messages: rest.length > 0 ? rest.map((m, i) => i === rest.length - 1 ? withRollingCacheBreakpoint(m, cacheControl) : m) : rest,
+      messages: rest.length > 0 ? rest.map(
+        (m, i) => rollingIndices.has(i) ? withRollingCacheBreakpoint(m, cacheControl) : m
+      ) : rest,
       stream: true
     };
     if (systemParts.length > 0) {
@@ -56804,11 +56898,11 @@ function anthropicMessagesProvider(config2) {
           } else if (type === "message_start") {
             const usage = ev.message?.usage;
             if (usage) {
-              const num = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? v : 0;
+              const num2 = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? v : 0;
               startUsage = {
-                inputTokens: num(usage.input_tokens),
-                cacheReadTokens: num(usage.cache_read_input_tokens),
-                cacheCreationTokens: num(usage.cache_creation_input_tokens)
+                inputTokens: num2(usage.input_tokens),
+                cacheReadTokens: num2(usage.cache_read_input_tokens),
+                cacheCreationTokens: num2(usage.cache_creation_input_tokens)
               };
             }
           } else if (type === "message_delta") {
@@ -56854,6 +56948,7 @@ var ANTHROPIC_VERSION, ANTHROPIC_BETA, ANTHROPIC_BETA_EXTENDED_CACHE_TTL;
 var init_anthropic = __esm({
   "src/cli/provider/anthropic.ts"() {
     "use strict";
+    init_skills2();
     init_openai_compatible();
     init_chatStats();
     init_thinking();
@@ -57043,7 +57138,8 @@ function chatgptResponsesProvider(config2) {
                 usage: {
                   promptTokens: usage.input_tokens ?? usage.prompt_tokens ?? 0,
                   completionTokens: usage.output_tokens ?? usage.completion_tokens ?? 0,
-                  totalTokens: usage.total_tokens ?? (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0)
+                  totalTokens: usage.total_tokens ?? (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0),
+                  ...parseCachedPromptTokens(usage) > 0 ? { cachedPromptTokens: parseCachedPromptTokens(usage) } : {}
                 }
               };
             }
@@ -57068,6 +57164,7 @@ var init_chatgpt = __esm({
     "use strict";
     init_openai_compatible();
     init_thinking();
+    init_openai_compatible();
   }
 });
 
@@ -57270,7 +57367,8 @@ function responsesApiProvider(config2) {
                 usage: {
                   promptTokens: usage.input_tokens ?? usage.prompt_tokens ?? 0,
                   completionTokens: usage.output_tokens ?? usage.completion_tokens ?? 0,
-                  totalTokens: usage.total_tokens ?? (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0)
+                  totalTokens: usage.total_tokens ?? (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0),
+                  ...parseCachedPromptTokens(usage) > 0 ? { cachedPromptTokens: parseCachedPromptTokens(usage) } : {}
                 }
               };
             }
@@ -57298,6 +57396,7 @@ var init_responsesApi = __esm({
     init_chatgpt();
     init_thinking();
     init_capabilities();
+    init_openai_compatible();
     RETRYABLE_STATUSES2 = /* @__PURE__ */ new Set([429, 500, 502, 503, 504]);
     MAX_RETRIES2 = (() => {
       const raw = process.env.ZELARI_PROVIDER_MAX_RETRIES;
@@ -59931,6 +60030,45 @@ var init_metrics3 = __esm({
   }
 });
 
+// src/cli/budget/messageUsage.ts
+function counter(value) {
+  return typeof value === "number" && Number.isFinite(value) && value > 0 ? Math.round(value) : 0;
+}
+function recordMessageUsage(input) {
+  try {
+    const promptTokens = counter(input.promptTokens);
+    const completionTokens = counter(input.completionTokens);
+    const cachedPromptTokens = Math.min(counter(input.cachedPromptTokens), promptTokens);
+    const model = input.model ?? "";
+    const record2 = {
+      kind: "message",
+      ts: input.ts ?? Date.now(),
+      ...input.sessionId ? { sessionId: input.sessionId } : {},
+      ...input.provider ? { provider: input.provider } : {},
+      ...model ? { model } : {},
+      promptTokens,
+      completionTokens,
+      cachedPromptTokens,
+      costUsd: calculateCost(model, promptTokens, completionTokens, cachedPromptTokens)
+    };
+    getMetricsLogger().record(record2);
+  } catch {
+  }
+}
+async function flushMessageUsage() {
+  try {
+    await getMetricsLogger().flush();
+  } catch {
+  }
+}
+var init_messageUsage = __esm({
+  "src/cli/budget/messageUsage.ts"() {
+    "use strict";
+    init_metrics3();
+    init_modelPricing();
+  }
+});
+
 // src/cli/kraken/turnRuntime.ts
 var WRITE_TOOLS3, KrakenTurnRuntime;
 var init_turnRuntime = __esm({
@@ -69114,7 +69252,7 @@ async function runAgentMissionSlice(deps) {
     deps.ragContext
   );
   async function runPass(messages, passId) {
-    const counter = createWriteCounter();
+    const counter2 = createWriteCounter();
     const harness = new AgentHarness({
       model: deps.model,
       provider,
@@ -69135,7 +69273,7 @@ async function runAgentMissionSlice(deps) {
     let completionTokens = 0;
     try {
       for await (const event of harness.run()) {
-        counter.onEvent(event);
+        counter2.onEvent(event);
         if (deps.onEvent) await deps.onEvent(event);
         if (event.type === "message_delta" && typeof event.delta === "string") {
           text += event.delta;
@@ -69167,8 +69305,8 @@ async function runAgentMissionSlice(deps) {
     }
     return {
       text,
-      successfulWrites: counter.state.successfulWrites,
-      emittedWrites: counter.state.emittedWrites,
+      successfulWrites: counter2.state.successfulWrites,
+      emittedWrites: counter2.state.emittedWrites,
       errored: errored2,
       messages: harness.getMessages(),
       promptTokens,
@@ -73125,6 +73263,7 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
   }));
   const toolNames = tools.map((t) => t.name);
   let systemMessages;
+  let wireSplit;
   let languageDirectiveContent;
   try {
     languageDirectiveContent = buildLanguagePolicyModuleFor(opts.task).content;
@@ -73220,7 +73359,8 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
         }
       }
     );
-    systemMessages = systemMessagesFromSplit(split);
+    systemMessages = systemMessagesFromSplit(split, { includeVolatile: true });
+    wireSplit = split;
   } catch {
     systemMessages = [
       {
@@ -73235,6 +73375,7 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
         ].join("\n")
       }
     ];
+    wireSplit = { stable: systemMessages[0].content, volatile: "" };
   }
   await spine.beginResourceTurn();
   const modelContext = await buildModelContext({
@@ -73322,6 +73463,16 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
       for await (const event of harness.run()) {
         progressRuntime.observe(event);
         spine.observe(event);
+        if (event.type === "message_end" && event.usage) {
+          recordMessageUsage({
+            sessionId: passSessionId,
+            provider,
+            model,
+            promptTokens: event.usage.promptTokens,
+            completionTokens: event.usage.completionTokens,
+            cachedPromptTokens: event.usage.cachedPromptTokens ?? 0
+          });
+        }
         if (event.type === "message_start") {
           scrub.reset();
         }
@@ -73370,6 +73521,7 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
       };
     }
     const buildProgress = readBuildProgress();
+    await flushMessageUsage();
     return {
       finalReason,
       exitCode,
@@ -73389,15 +73541,18 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
     }
   });
   progressRuntime.beginTurn();
-  const initialMessages = [
-    ...systemMessages,
-    ...historySeed,
-    {
-      role: "user",
-      content: effectiveTask,
-      ...opts.images && opts.images.length > 0 ? { images: opts.images } : {}
-    }
-  ];
+  const initialMessages = assembleRequestMessages({
+    split: wireSplit,
+    history: historySeed,
+    turn: [
+      {
+        role: "user",
+        content: effectiveTask,
+        ...opts.images && opts.images.length > 0 ? { images: opts.images } : {}
+      }
+    ],
+    layout: resolvePromptLayout()
+  }).messages;
   let pass = await runSinglePass(initialMessages, sessionId2);
   let strictExit = 0;
   const verifierReviewDeps = {
@@ -73450,11 +73605,17 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
           "[zelari-code --headless] Kraken BUILD: required checks unresolved \u2014 forcing repair pass\n"
         );
       }
-      const withSystem = [
-        ...systemMessages,
-        ...pass.messages.filter((m) => m.role !== "system"),
-        { role: "user", content: repairPrompt }
-      ];
+      const withSystem = assembleRequestMessages({
+        split: wireSplit,
+        history: [],
+        turn: [
+          ...pass.messages.filter(
+            (m) => m.role !== "system" && !isTrailingContextContent(m.content)
+          ),
+          { role: "user", content: repairPrompt }
+        ],
+        layout: resolvePromptLayout()
+      }).messages;
       progressRuntime.beginPass(true);
       markRepairTriggered();
       const repair = await runSinglePass(withSystem, `${sessionId2}-check-repair`);
@@ -73599,6 +73760,7 @@ var init_runOneTurn = __esm({
     init_verifierLifecycle();
     init_modelContextBuilder();
     init_metrics3();
+    init_messageUsage();
     init_headlessSpine();
     init_harnessStateEmit();
     init_runtime();
@@ -79268,6 +79430,76 @@ var init_contextGrowthSummary = __esm({
   }
 });
 
+// src/cli/budget/cacheHitReport.ts
+function billable(record2) {
+  return typeof record2.promptTokens === "number" && record2.promptTokens > 0;
+}
+function num(value) {
+  return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
+}
+function summarizeCacheHits(records, window = 500) {
+  const usable = records.filter(billable);
+  if (usable.length === 0) return null;
+  const tail2 = usable.slice(-window);
+  const rows = /* @__PURE__ */ new Map();
+  let promptTokens = 0;
+  let cachedPromptTokens = 0;
+  for (const record2 of tail2) {
+    const provider = record2.provider ?? "unknown";
+    const model = record2.model ?? "unknown";
+    const prompt = num(record2.promptTokens);
+    const cached2 = Math.min(num(record2.cachedPromptTokens), prompt);
+    promptTokens += prompt;
+    cachedPromptTokens += cached2;
+    const key = `${provider}::${model}`;
+    const row = rows.get(key) ?? {
+      provider,
+      model,
+      messages: 0,
+      promptTokens: 0,
+      cachedPromptTokens: 0,
+      hitRate: 0
+    };
+    row.messages += 1;
+    row.promptTokens += prompt;
+    row.cachedPromptTokens += cached2;
+    row.hitRate = row.promptTokens > 0 ? row.cachedPromptTokens / row.promptTokens : 0;
+    rows.set(key, row);
+  }
+  return {
+    messages: tail2.length,
+    promptTokens,
+    cachedPromptTokens,
+    hitRate: promptTokens > 0 ? cachedPromptTokens / promptTokens : 0,
+    byModel: [...rows.values()].sort((a, b) => b.promptTokens - a.promptTokens)
+  };
+}
+function fmtTokens(n) {
+  return n.toLocaleString("en-US");
+}
+function fmtRate(rate) {
+  return `${(rate * 100).toFixed(1)}%`;
+}
+function formatCacheHitSummary(summary, top = 3) {
+  const lines = [
+    `messages ${summary.messages} \xB7 prompt ${fmtTokens(summary.promptTokens)} tokens \xB7 cached ${fmtTokens(summary.cachedPromptTokens)} \xB7 hit ${fmtRate(summary.hitRate)}`
+  ];
+  const rows = summary.byModel.slice(0, top);
+  if (rows.length > 0) {
+    lines.push(
+      "top models: " + rows.map(
+        (r) => `${r.provider}/${r.model} ${fmtRate(r.hitRate)} (${r.messages} msg \xB7 ${fmtTokens(r.promptTokens)} prompt)`
+      ).join(" \xB7 ")
+    );
+  }
+  return lines;
+}
+var init_cacheHitReport = __esm({
+  "src/cli/budget/cacheHitReport.ts"() {
+    "use strict";
+  }
+});
+
 // src/cli/utils/doctor.ts
 var doctor_exports = {};
 __export(doctor_exports, {
@@ -79559,6 +79791,24 @@ async function checkContextGrowth() {
     return WARN(`metrics unreadable: ${err instanceof Error ? err.message : String(err)}`);
   }
 }
+async function checkPromptCache() {
+  try {
+    const logger = getMetricsLogger();
+    const records = await readMetrics(logger.filePath);
+    const summary = summarizeCacheHits(
+      records.filter((r) => r.kind === "message")
+    );
+    if (!summary) {
+      return WARN(
+        `no data: sessions running the new per-message telemetry are needed for the baseline (${logger.filePath})`
+      );
+    }
+    const lines = formatCacheHitSummary(summary);
+    return OK(lines.join("\n\n         "));
+  } catch (err) {
+    return WARN(`metrics unreadable: ${err instanceof Error ? err.message : String(err)}`);
+  }
+}
 function firstBlockingRed(entries) {
   return entries.find((e) => !e.ok && e.severity === "critical") ?? null;
 }
@@ -79633,6 +79883,10 @@ function buildDoctorChecks(pkg, pkgName) {
     // Informational aggregate over metrics.jsonl: tool round-trips,
     // intermediate tool bytes, history@request size, cache-hit tokens.
     { name: "context growth", run: () => checkContextGrowth() },
+    // --- prompt-cache metrics (cache-hit-rate plan M1.2) ---
+    // Provider-verified cache hit% over the `kind: 'message'` rows. WARN-only
+    // ("no data") until sessions run the M1.1 telemetry.
+    { name: "prompt cache", run: () => checkPromptCache() },
     // --- optional tool plugins (v1.5.0) ---
     // Playwright / eslint / ruff / LSP servers. WARN-only (never critical):
     // these power edge features that degrade silently when absent. Surfaced
@@ -79714,6 +79968,7 @@ var init_doctor = __esm({
     init_prereqChecks();
     init_metrics3();
     init_contextGrowthSummary();
+    init_cacheHitReport();
     init_updater();
     require3 = createRequire3(import.meta.url);
     __dirname3 = path112.dirname(fileURLToPath3(import.meta.url));
@@ -80776,12 +81031,111 @@ var init_turnAdapter = __esm({
 });
 
 // src/cli/acp/framing.ts
-function createFrameWriter(sink, onError) {
+function createDualWireDecoder(format) {
+  const fmt = format ?? { mode: "ndjson" };
+  let buf = Buffer.alloc(0);
+  let locked = false;
+  const lock = (mode) => {
+    if (locked) return;
+    locked = true;
+    fmt.mode = mode;
+  };
+  const step = (out) => {
+    while (buf.length > 0 && (buf[0] === 10 || buf[0] === 13)) buf = buf.subarray(1);
+    if (buf.length === 0) return false;
+    if (buf.subarray(0, HEADER_PREFIX.length).toString("latin1").toLowerCase() === HEADER_PREFIX) {
+      const sep5 = buf.indexOf("\r\n\r\n");
+      if (sep5 === -1) {
+        if (buf.length > MAX_PENDING_BYTES2) {
+          out.warnings.push("runaway header block without \\r\\n\\r\\n (dropped)");
+          buf = Buffer.alloc(0);
+          return true;
+        }
+        return false;
+      }
+      const header = buf.subarray(0, sep5).toString("latin1");
+      const m = /content-length:\s*(\d+)/i.exec(header);
+      if (!m) {
+        out.warnings.push(
+          `frame without a usable Content-Length: ${JSON.stringify(header.slice(0, 80))}`
+        );
+        buf = buf.subarray(sep5 + 4);
+        return true;
+      }
+      const bodyStart = sep5 + 4;
+      const bodyEnd = bodyStart + Number(m[1]);
+      if (buf.length < bodyEnd) return false;
+      const body = buf.subarray(bodyStart, bodyEnd).toString("utf8");
+      buf = buf.subarray(bodyEnd);
+      try {
+        out.messages.push(JSON.parse(body));
+        lock("lsp-frame");
+      } catch {
+        out.warnings.push(`unparseable frame body (dropped): ${JSON.stringify(body.slice(0, 80))}`);
+      }
+      return true;
+    }
+    const nl = buf.indexOf(10);
+    if (nl === -1) {
+      if (buf.length > MAX_PENDING_BYTES2) {
+        out.warnings.push("runaway line without a newline (dropped)");
+        buf = Buffer.alloc(0);
+        return true;
+      }
+      return false;
+    }
+    const line = buf.subarray(0, nl).toString("utf8").replace(/\r+$/, "");
+    buf = buf.subarray(nl + 1);
+    if (line.trim().length === 0) return true;
+    try {
+      out.messages.push(JSON.parse(line));
+      lock("ndjson");
+    } catch {
+      out.warnings.push(`unparseable JSON line (dropped): ${JSON.stringify(line.slice(0, 80))}`);
+    }
+    return true;
+  };
+  const drain = () => {
+    const out = { messages: [], warnings: [] };
+    while (step(out)) {
+    }
+    return out;
+  };
+  return {
+    push(chunk) {
+      const data = typeof chunk === "string" ? Buffer.from(chunk, "utf8") : Buffer.from(chunk);
+      buf = buf.length === 0 ? data : Buffer.concat([buf, data]);
+      return drain();
+    },
+    /** EOF: salvage a trailing NDJSON line that lacks the final newline. */
+    flush() {
+      const out = drain();
+      if (buf.length === 0) return out;
+      const line = buf.toString("utf8").trim();
+      buf = Buffer.alloc(0);
+      if (line.length > 0) {
+        try {
+          out.messages.push(JSON.parse(line));
+          lock("ndjson");
+        } catch {
+          out.warnings.push(
+            `unparseable JSON line at EOF (dropped): ${JSON.stringify(line.slice(0, 80))}`
+          );
+        }
+      }
+      return out;
+    }
+  };
+}
+function createFrameWriter(sink, onError, format) {
   const write = sink.write.bind(sink);
   return {
     write(message) {
       try {
-        write(encodeMessage(message));
+        write(
+          format && format.mode === "ndjson" ? `${JSON.stringify(message)}
+` : encodeMessage(message)
+        );
       } catch (err) {
         try {
           onError?.(err instanceof Error ? err.message : String(err));
@@ -80805,7 +81159,7 @@ function malformedFrameWarnings(chunk) {
   }
 }
 function attachFrameReader(stream, options) {
-  const parser = createMessageParser();
+  const parser = createDualWireDecoder(options.format);
   let closed = false;
   const report = (detail) => {
     try {
@@ -80813,36 +81167,43 @@ function attachFrameReader(stream, options) {
     } catch {
     }
   };
+  const deliver = (message) => {
+    try {
+      options.onMessage(message);
+    } catch (err) {
+      report(`consumer error: ${err instanceof Error ? err.message : String(err)}`);
+    }
+  };
   const finish2 = () => {
     if (closed) return;
     closed = true;
+    try {
+      const rest = parser.flush();
+      for (const warning of rest.warnings) report(warning);
+      for (const message of rest.messages) deliver(message);
+    } catch {
+    }
     try {
       options.onEof();
     } catch {
     }
   };
   const onData = (chunk) => {
-    for (const warning of malformedFrameWarnings(chunk)) report(warning);
-    let messages;
+    for (const warning of malformedFrameWarnings(chunk.toString("utf8"))) report(warning);
+    let decoded;
     try {
-      messages = parser.push(chunk);
+      decoded = parser.push(chunk);
     } catch (err) {
       report(`decoder error: ${err instanceof Error ? err.message : String(err)}`);
       return;
     }
-    for (const message of messages) {
-      try {
-        options.onMessage(message);
-      } catch (err) {
-        report(`consumer error: ${err instanceof Error ? err.message : String(err)}`);
-      }
-    }
+    for (const warning of decoded.warnings) report(warning);
+    for (const message of decoded.messages) deliver(message);
   };
   const onError = (err) => {
     report(`stdin error: ${err.message}`);
     finish2();
   };
-  stream.setEncoding("utf8");
   stream.on("data", onData);
   stream.on("end", finish2);
   stream.on("close", finish2);
@@ -80857,10 +81218,13 @@ function attachFrameReader(stream, options) {
     }
   };
 }
+var MAX_PENDING_BYTES2, HEADER_PREFIX;
 var init_framing = __esm({
   "src/cli/acp/framing.ts"() {
     "use strict";
     init_protocol();
+    MAX_PENDING_BYTES2 = 4 * 1024 * 1024;
+    HEADER_PREFIX = "content-length:";
   }
 });
 
@@ -80887,12 +81251,15 @@ function startAcpServer(deps) {
 `));
   const sessions = /* @__PURE__ */ new Map();
   let closed = false;
+  const format = { mode: "ndjson" };
   const writer = createFrameWriter(
     deps.output ?? process.stdout,
-    (message) => log(`[zelari-code acp] stdout write failed: ${message}`)
+    (message) => log(`[zelari-code acp] stdout write failed: ${message}`),
+    format
   );
   const shutdownWaiters = [];
   const reader = attachFrameReader(input, {
+    format,
     onMessage: (raw) => {
       void handleMessage(raw);
     },
@@ -81116,7 +81483,7 @@ function parseAcpFlags(argv) {
   return out;
 }
 function acpHelpText() {
-  return 'zelari-code acp \u2014 Agent Client Protocol server on stdio (for editors)\n\nSpeaks JSON-RPC 2.0 over stdin/stdout with LSP-style framing\n(`Content-Length: <bytes>\\r\\n\\r\\n<json>`). Point an ACP-capable editor\n(e.g. Zed) at this command as a custom agent server.\n\nMethods (implemented subset):\n  initialize                 protocolVersion + agent capabilities\n  session/new { cwd }        -> { sessionId }\n  session/prompt { sessionId, prompt: [{type:"text",text}] }\n                             -> { stopReason } when the turn ends\n  session/cancel { sessionId }\n\nNotifications sent to the client: session/update with\nagent_message_chunk, tool_call and tool_call_update (see\nsrc/cli/acp/protocol.ts for the exact subset and its non-goals).\n\nOptions:\n  --cwd <path>       Fallback workspace when a client omits cwd\n                     (default: current directory)\n  --provider <id>    Provider override (default: the active provider)\n  --model <id>       Model override (default: the provider default)\n  --help, -h         Print this help and exit\n\nTurns run through the same headless dispatch as `--headless`\n(kraken by default), one turn at a time; stdin EOF shuts down cleanly.\n';
+  return 'zelari-code acp \u2014 Agent Client Protocol server on stdio (for editors)\n\nSpeaks JSON-RPC 2.0 over stdin/stdout as newline-delimited JSON\n(the ACP stdio transport, as spoken by Zed). LSP-style frames\n(`Content-Length: <bytes>\\r\\n\\r\\n<json>`) are also accepted and\nmirrored on output. Point an ACP-capable editor at this command as\na custom agent server.\n\nMethods (implemented subset):\n  initialize                 protocolVersion + agent capabilities\n  session/new { cwd }        -> { sessionId }\n  session/prompt { sessionId, prompt: [{type:"text",text}] }\n                             -> { stopReason } when the turn ends\n  session/cancel { sessionId }\n\nNotifications sent to the client: session/update with\nagent_message_chunk, tool_call and tool_call_update (see\nsrc/cli/acp/protocol.ts for the exact subset and its non-goals).\n\nOptions:\n  --cwd <path>       Fallback workspace when a client omits cwd\n                     (default: current directory)\n  --provider <id>    Provider override (default: the active provider)\n  --model <id>       Model override (default: the provider default)\n  --help, -h         Print this help and exit\n\nTurns run through the same headless dispatch as `--headless`\n(kraken by default), one turn at a time; stdin EOF shuts down cleanly.\n';
 }
 async function runAcpCommand(opts = {}) {
   if (opts.help === true) {
@@ -84729,6 +85096,7 @@ ${lines.join("\n")}`;
 init_harness();
 init_observationStore();
 init_metrics3();
+init_messageUsage();
 init_modelPricing();
 init_openai_compatible();
 init_resolveStream();
@@ -84937,6 +85305,8 @@ function useChatTurn(params) {
       let memoryAutoWrite = false;
       let historySeedLen = 0;
       let systemPrefixLen = 0;
+      let trailingSeedLen = 0;
+      let seedMessages;
       let turnSucceeded = false;
       try {
         const anchored = maybeAnchorShortAnswer(userText);
@@ -85268,6 +85638,7 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
           volatileSessionBlock,
           composedWorkspace
         ].filter(Boolean).join("\n\n");
+        const promptLayout = resolvePromptLayout();
         let systemMessages = [];
         let lastStableHash;
         try {
@@ -85295,7 +85666,16 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
             ragContext: void 0
           });
           lastStableHash = hashStablePrompt(split.stable);
-          systemMessages = systemMessagesFromSplit(split);
+          const assembled = assembleRequestMessages({
+            split,
+            history: historyForModel,
+            turn: [{ role: "user", content: effectiveUserText }],
+            layout: promptLayout
+          });
+          seedMessages = assembled.messages;
+          systemMessages = assembled.messages.slice(0, assembled.systemCount);
+          systemPrefixLen = assembled.systemCount;
+          trailingSeedLen = assembled.trailingCount;
         } catch {
           const languageModule = buildLanguagePolicyModuleFor(userText);
           const fallback = [
@@ -85309,8 +85689,10 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
           ].join("\n");
           lastStableHash = hashStablePrompt(fallback);
           systemMessages = [{ role: "system", content: fallback }];
+          seedMessages = void 0;
+          trailingSeedLen = 0;
         }
-        systemPrefixLen = systemMessages.length;
+        if (seedMessages === void 0) systemPrefixLen = systemMessages.length;
         const perTurnEnv = envNumber(process.env.ZELARI_MAX_TOOL_CALLS, {
           default: 25,
           min: 1
@@ -85329,7 +85711,11 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
           // hardcode "openai-compatible" (the transport family) so snapshots
           // and telemetry mislabeled deepseek/glm/minimax routing.
           provider: envConfig.providerId,
-          messages: [
+          // M2.1: the seed assembled by the layout helper when the split prompt
+          // built (trailing layout = [stable system][history][trailing
+          // context][user]); the degraded fallback keeps the plain
+          // [system][history][user] shape.
+          messages: seedMessages ?? [
             ...systemMessages,
             // v1.8.0: shared rolling history (agent/council/zelari) so short
             // answers bind to prior ---QUESTION--- blocks. Possibly empty
@@ -85466,6 +85852,14 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
                   totalTokens: event.usage.totalTokens,
                   cachedPromptTokens: event.usage.cachedPromptTokens
                 });
+                recordMessageUsage({
+                  sessionId: sessionId2,
+                  provider: envConfig.providerId,
+                  model: envConfig.model,
+                  promptTokens: event.usage.promptTokens,
+                  completionTokens: event.usage.completionTokens,
+                  cachedPromptTokens: event.usage.cachedPromptTokens ?? 0
+                });
               }
               if (streamContent) {
                 const sealed = streamScrub.finalize(streamContent);
@@ -85659,7 +86053,7 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il turno continua dopo
             const h = harnessRef.current;
             if (h && turnSucceeded) {
               const all = h.getMessages();
-              const seedLen = systemPrefixLen + historySeedLen + 1;
+              const seedLen = systemPrefixLen + historySeedLen + trailingSeedLen + 1;
               if (all.length > seedLen) {
                 appendMessages(
                   all.slice(seedLen).map((m) => {