chainlesschain 0.166.101 → 0.166.103

package/src/runtime/agent-core.js

@@ -1,3 +1,8 @@
+import {
+  providerFetch,
+  requireUnscopedToolExecution,
+} from "../lib/formal-budget-provider-scope.js";
+import { isFormalBudgetError } from "../lib/formal-budget-errors.js";
 /**
  * Agent Core — transport-independent agentic logic
  *
@@ -61,7 +66,11 @@ import {
 } from "../lib/task-progress-tracker.js";
 import { TOOL_RECOVERY_PAUSED_CODE } from "../lib/tool-recovery-result.js";
 import { haiku55ThinkingParams } from "../lib/provider-options.js";
-import { projectUsagePricingContext } from "../lib/usage-pricing-context.js";
+import {
+  projectUsagePricingContext,
+  projectEnvelopePricingContext,
+  preserveRequestedServiceTierEvidence,
+} from "../lib/usage-pricing-context.js";
 import broker from "../lib/process-execution-broker/index.js";
 import os from "os";
 import { createHash, randomUUID } from "node:crypto";
@@ -88,6 +97,7 @@ import sharedHookEvents from "../lib/settings-hook-events.js";
 import {
   resolveAgentOutputBudget,
   resolveModelCapabilityProfile,
+  resolveModelServiceTier,
 } from "../lib/model-capabilities.js";
 import {
   createOpenAIResponsesBody,
@@ -1539,18 +1549,19 @@ function filesystemIdentity(stats) {
 function sameFilesystemObjectIdentity(left, right) {
   return Boolean(
     left &&
-    right &&
-    left.dev === right.dev &&
-    left.ino === right.ino &&
-    left.type === right.type,
+      right &&
+      left.dev === right.dev &&
+      left.ino === right.ino &&
+      left.type === right.type,
   );
 }
 
 function sameFilesystemIdentity(left, right, { content = false } = {}) {
   return Boolean(
     sameFilesystemObjectIdentity(left, right) &&
-    left.nlink === right.nlink &&
-    (!content || (left.size === right.size && left.mtimeNs === right.mtimeNs)),
+      left.nlink === right.nlink &&
+      (!content ||
+        (left.size === right.size && left.mtimeNs === right.mtimeNs)),
   );
 }
 
@@ -2607,6 +2618,7 @@ function snapshotApprovalPolicyAuthority(approvalGate, sessionId) {
  * @returns {Promise<object>} tool result
  */
 export async function executeTool(name, args, context = {}) {
+  requireUnscopedToolExecution();
   const liveExecutionContext = context;
   try {
     const executionPolicy = captureAgentExecutionPolicy(context);
@@ -3441,8 +3453,9 @@ export async function executeTool(name, args, context = {}) {
     GUARDED_FILE_MUTATION_TOOLS.has(name) &&
     settingsVerdict.decision !== "allow"
   ) {
-    const { sensitiveFileReason, autoExecConfigReason } =
-      await import("../lib/sensitive-file-guard.js");
+    const { sensitiveFileReason, autoExecConfigReason } = await import(
+      "../lib/sensitive-file-guard.js"
+    );
     const guarded = fileMutationPaths(name, args)
       .map((candidatePath) => ({
         path: candidatePath,
@@ -3477,8 +3490,9 @@ export async function executeTool(name, args, context = {}) {
     GUARDED_FILE_MUTATION_TOOLS.has(name) &&
     settingsVerdict.decision !== "allow"
   ) {
-    const { sessionStorePathReason } =
-      await import("../lib/session-store-guard.js");
+    const { sessionStorePathReason } = await import(
+      "../lib/session-store-guard.js"
+    );
     const guarded = fileMutationPaths(name, args)
       .map((candidatePath) => ({
         path: candidatePath,
@@ -4223,8 +4237,9 @@ export async function executeTool(name, args, context = {}) {
     context.externalToolExecutors
   ) {
     try {
-      const { collectIdeDiagnostics, formatIdeDiagnostics } =
-        await import("../lib/ide-context.js");
+      const { collectIdeDiagnostics, formatIdeDiagnostics } = await import(
+        "../lib/ide-context.js"
+      );
       const diags = await collectIdeDiagnostics(
         {
           mcpClient: context.mcpHostClient || context.mcpClient,
@@ -4766,8 +4781,9 @@ export async function _getSharedCodeIntel(cwd) {
   const root = path.resolve(cwd || process.cwd());
   let entry = _codeIntelPool.get(root);
   if (!entry) {
-    const { CodeIntelligence } =
-      await import("../lib/lsp/code-intelligence.js");
+    const { CodeIntelligence } = await import(
+      "../lib/lsp/code-intelligence.js"
+    );
     // Re-check AFTER the await: two concurrent callers (the parallel read-only
     // batch runs several `code_intelligence` calls at once) both observed an
     // empty pool before this `await import`, so without this guard both would
@@ -4856,8 +4872,9 @@ async function _postEditDiagnostics(filePath, cwd, additionalDirectories) {
   }
   let hasServer = false;
   try {
-    const { languageIdForFile, resolveServer } =
-      await import("../lib/lsp/lsp-server-registry.js");
+    const { languageIdForFile, resolveServer } = await import(
+      "../lib/lsp/lsp-server-registry.js"
+    );
     const languageId = languageIdForFile(filePath);
     if (!languageId) return null; // unsupported file type — no LSP, no cost
     hasServer = Boolean(resolveServer(languageId, root));
@@ -4880,8 +4897,9 @@ async function _postEditDiagnostics(filePath, cwd, additionalDirectories) {
       // the model sees the most severe diagnostics first and can't be flooded by
       // a pathological diagnostics dump. Budget is overridable via
       // CC_EDIT_DIAGNOSTICS_TOKENS (default 2000).
-      const { capDiagnostics } =
-        await import("../lib/lsp/diagnostics-scheduler.js");
+      const { capDiagnostics } = await import(
+        "../lib/lsp/diagnostics-scheduler.js"
+      );
       const budget = Number(process.env.CC_EDIT_DIAGNOSTICS_TOKENS) || 2000;
       const { kept } = capDiagnostics(diags, { maxTokens: budget });
       return kept.length ? kept : null;
@@ -4945,12 +4963,14 @@ async function _getSubtreeLoader(cwd, sessionScope = "__legacy__") {
   }
   let loader = scopedLoaders.get(sessionScope);
   if (!loader) {
-    const { SubtreeInstructionLoader } =
-      await import("../lib/project-instructions.js");
+    const { SubtreeInstructionLoader } = await import(
+      "../lib/project-instructions.js"
+    );
     let instructionExcludes;
     try {
-      const { readStringArraySetting } =
-        await import("../lib/settings-loader.cjs");
+      const { readStringArraySetting } = await import(
+        "../lib/settings-loader.cjs"
+      );
       instructionExcludes = readStringArraySetting("instructionExcludes", {
         cwd: root,
       });
@@ -4995,8 +5015,9 @@ async function _prepareSubtreeInstructions(
     }
     const fresh = [...discovered.values()];
     if (!fresh || !fresh.length) return null;
-    const { DEFAULT_MAX_FILE_BYTES } =
-      await import("../lib/project-instructions.js");
+    const { DEFAULT_MAX_FILE_BYTES } = await import(
+      "../lib/project-instructions.js"
+    );
     const out = [];
     const errors = [];
     for (const f of fresh) {
@@ -5103,8 +5124,9 @@ async function _withSubtreeInstructions(
 async function filterSkillsByCwd(skills, cwd) {
   try {
     if (!Array.isArray(skills) || skills.every((s) => !s?.paths)) return skills;
-    const { filterSkillsByRelCwd, relCwdForCwd } =
-      await import("../lib/skill-path-scope.js");
+    const { filterSkillsByRelCwd, relCwdForCwd } = await import(
+      "../lib/skill-path-scope.js"
+    );
     const { findProjectRoot } = await import("../lib/project-detector.js");
     const root = (cwd && findProjectRoot(cwd)) || cwd || "";
     return filterSkillsByRelCwd(skills, relCwdForCwd(cwd, root));
@@ -5119,8 +5141,9 @@ async function promoteBrowserGeneratedArtifact(
 ) {
   if (!filePath) return { artifact: null, error: null };
   try {
-    const { ArtifactStore, publicArtifactMetadata } =
-      await import("../lib/artifact-store.js");
+    const { ArtifactStore, publicArtifactMetadata } = await import(
+      "../lib/artifact-store.js"
+    );
     const entry = new ArtifactStore().publish({
       filePath,
       title,
@@ -5173,8 +5196,9 @@ async function promoteBrowserDownload(
 async function publishBrowserEvidence(envelope, sessionId) {
   if (!envelope?.envelopeDigest) return { artifact: null, error: null };
   try {
-    const { ArtifactStore, publicArtifactMetadata } =
-      await import("../lib/artifact-store.js");
+    const { ArtifactStore, publicArtifactMetadata } = await import(
+      "../lib/artifact-store.js"
+    );
     const {
       browserEvidenceDigest,
       canonicalBrowserEvidenceJson,
@@ -5237,8 +5261,9 @@ async function publishBrowserEvidence(envelope, sessionId) {
 }
 
 async function resolveBrowserUploadArtifact(artifactId, sessionId) {
-  const { ArtifactStore, publicArtifactMetadata } =
-    await import("../lib/artifact-store.js");
+  const { ArtifactStore, publicArtifactMetadata } = await import(
+    "../lib/artifact-store.js"
+  );
   const store = new ArtifactStore();
   return store.withIndexSnapshot((entries) => {
     const matches = entries.filter((entry) => entry.id === artifactId);
@@ -6219,8 +6244,9 @@ async function executeToolInner(
       // through PowerShell via explicit argv (`powershell.exe -NoProfile
       // [-ExecutionPolicy <p>] -Command <cmd>`). The unconfigured path keeps
       // the historical default shell byte-identical (useDefaultShell=true).
-      const { resolveShellInvocation } =
-        await import("../lib/shell-selector.cjs");
+      const { resolveShellInvocation } = await import(
+        "../lib/shell-selector.cjs"
+      );
       let shellInv = resolveShellInvocation({
         command: args.command,
         requested: args.shell,
@@ -6993,8 +7019,9 @@ async function executeToolInner(
         let removeSettingsAuthorityListener = null;
         if (needsEgress) {
           try {
-            const { startEgressProxyWorker } =
-              await import("../lib/sandbox-egress-worker.js");
+            const { startEgressProxyWorker } = await import(
+              "../lib/sandbox-egress-worker.js"
+            );
             if (dockerEgress) {
               privateSocketDir = fs.mkdtempSync(
                 path.join(os.tmpdir(), "cc-egress-"),
@@ -7035,8 +7062,9 @@ async function executeToolInner(
           // Proxy startup and dynamic imports are asynchronous. Re-read the
           // live authority after they settle and before sandbox dispatch.
           if (dockerEgress && proxyHandle?.socketPath) {
-            const { createDockerEgressAuthorityMonitor } =
-              await import("../lib/sandbox-egress-authority-monitor.js");
+            const { createDockerEgressAuthorityMonitor } = await import(
+              "../lib/sandbox-egress-authority-monitor.js"
+            );
             authorityMonitor = createDockerEgressAuthorityMonitor({
               sessionId: sessionId || null,
               policyVersion:
@@ -7656,8 +7684,9 @@ async function executeToolInner(
         !ruleAllowed &&
         typeof approvalGate.decide === "function"
       ) {
-        const { APPROVAL_RISK, APPROVAL_DECISION } =
-          await import("@chainlesschain/session-core");
+        const { APPROVAL_RISK, APPROVAL_DECISION } = await import(
+          "@chainlesschain/session-core"
+        );
         const gate = await approvalGate.decide({
           sessionId,
           riskLevel: APPROVAL_RISK.HIGH,
@@ -7828,6 +7857,7 @@ async function executeToolInner(
         });
         return attachDescriptor(result);
       } catch (err) {
+        if (isFormalBudgetError(err)) throw err;
         return attachDescriptor({ error: `web_search failed: ${err.message}` });
       }
     }
@@ -7868,8 +7898,9 @@ async function executeToolInner(
 
     case "slash_command": {
       try {
-        const { getCommand, expandCommand, discoverCommands } =
-          await import("../lib/slash-commands.js");
+        const { getCommand, expandCommand, discoverCommands } = await import(
+          "../lib/slash-commands.js"
+        );
         const raw = String(args.command || "").trim();
         if (!raw) {
           return attachDescriptor({
@@ -8080,8 +8111,9 @@ async function executeToolInner(
 
     case "notify": {
       try {
-        const { sendAgentNotification } =
-          await import("../lib/agent-notify.js");
+        const { sendAgentNotification } = await import(
+          "../lib/agent-notify.js"
+        );
         const outcome = await sendAgentNotification({
           title: args.title,
           body: args.body,
@@ -8095,8 +8127,9 @@ async function executeToolInner(
 
     case "schedule": {
       try {
-        const { AgentScheduleStore } =
-          await import("../lib/agent-schedule-store.js");
+        const { AgentScheduleStore } = await import(
+          "../lib/agent-schedule-store.js"
+        );
         const { parseDuration } = await import("../lib/loop.js");
         const store = new AgentScheduleStore();
         const action = String(args.action || "").toLowerCase();
@@ -8219,8 +8252,9 @@ async function executeToolInner(
       // Only METADATA returns to the conversation (the transcript never
       // carries the file body); `cc artifacts` lists/inspects/cleans.
       try {
-        const { ArtifactStore, publicArtifactMetadata } =
-          await import("../lib/artifact-store.js");
+        const { ArtifactStore, publicArtifactMetadata } = await import(
+          "../lib/artifact-store.js"
+        );
         const store = new ArtifactStore();
         const entry = store.publish({
           filePath: path.resolve(cwd, String(args.path || "")),
@@ -8310,8 +8344,9 @@ async function executeToolInner(
           !ruleAllowed &&
           typeof approvalGate.decide === "function"
         ) {
-          const { APPROVAL_RISK, APPROVAL_DECISION } =
-            await import("@chainlesschain/session-core");
+          const { APPROVAL_RISK, APPROVAL_DECISION } = await import(
+            "@chainlesschain/session-core"
+          );
           const summary = Array.isArray(args.actions)
             ? args.actions
                 .map((a) => a?.type)
@@ -8436,8 +8471,9 @@ async function executeToolInner(
     case "search_files": {
       if (typeof args.path === "string" && args.path.length > 0) {
         const file = path.resolve(cwd, args.path);
-        const { credentialFileReason } =
-          await import("../lib/credential-guard.js");
+        const { credentialFileReason } = await import(
+          "../lib/credential-guard.js"
+        );
         const reason =
           credentialFileReason(file) ||
           (fs.existsSync(file)
@@ -9143,6 +9179,7 @@ async function executeToolInner(
         } catch (err) {
           if (
             isEvolutionIngressFailure(err) ||
+            isFormalBudgetError(err) ||
             err?.runtimeLedgerPersistence === true ||
             err?.workflowEffectOutcomeUnknown === true
           ) {
@@ -9182,6 +9219,22 @@ async function executeToolInner(
     }
 
     case "list_skills": {
+      const limit = args.limit ?? 20;
+      const offset = args.offset ?? 0;
+      if (
+        !Number.isSafeInteger(limit) ||
+        limit < 1 ||
+        limit > 64 ||
+        !Number.isSafeInteger(offset) ||
+        offset < 0 ||
+        offset > 10_000
+      ) {
+        return attachDescriptor({
+          error:
+            "Skill directory limit must be 1–64 and offset must be 0–10000.",
+          code: "CC_SKILL_DIRECTORY_PAGE_INVALID",
+        });
+      }
       let skills = await filterSkillsByCwd(
         skillLoader.getResolvedSkills().filter(skillAllowed),
         cwd,
@@ -9271,6 +9324,19 @@ async function executeToolInner(
           }
         }
       }
+      if (!routing) {
+        const key = (skill) =>
+          `${skill.id}\u0000${skill.source || ""}\u0000${skill.executionIdentity?.contentDigest || ""}`;
+        skills = [...skills].sort((left, right) =>
+          key(left) < key(right) ? -1 : key(left) > key(right) ? 1 : 0,
+        );
+      }
+      const totalCount = routing?.matchedCount ?? skills.length;
+      const candidateCount = skills.length;
+      skills = skills.slice(offset, offset + limit);
+      const nextOffset =
+        offset + skills.length < candidateCount ? offset + skills.length : null;
+      const hasMore = offset + skills.length < totalCount;
       skillLoader.recordDescriptorUse?.(skills, {
         sessionId,
         turnId,
@@ -9278,6 +9344,19 @@ async function executeToolInner(
       });
       return attachDescriptor({
         count: skills.length,
+        totalCount,
+        offset,
+        limit,
+        hasMore,
+        nextOffset,
+        ...(hasMore
+          ? {
+              hint:
+                nextOffset == null
+                  ? "Refine query or category to search beyond the top 64 matches."
+                  : "Use nextOffset for the next page, or query/category to narrow the directory.",
+            }
+          : {}),
         ...(routing
           ? {
               routing: {
@@ -10572,6 +10651,7 @@ function _throwBackgroundSubAgentUsageFailure(entry) {
   const error = entry?.outcome?.fatalError;
   if (
     isEvolutionIngressFailure(error) ||
+    isFormalBudgetError(error) ||
     error?.runtimeLedgerPersistence === true ||
     error?.workflowEffectOutcomeUnknown === true
   ) {
@@ -10701,8 +10781,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   // the breadth counter increments so spawnedCount is this-spawn-exclusive.
   if (ctx.subAgentContract) {
     try {
-      const { enforceRecursionLimits } =
-        await import("../lib/subagent-contract.js");
+      const { enforceRecursionLimits } = await import(
+        "../lib/subagent-contract.js"
+      );
       const recur = enforceRecursionLimits({
         depth: currentDepth,
         spawnedCount: ctx.subAgentBudget?.spawned || 0,
@@ -10888,8 +10969,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   let inheritedMcp = null;
   let inheritedHooks = null;
   try {
-    const { filterInheritedMcp, filterInheritedHooks } =
-      await import("../lib/subagent-inheritance.js");
+    const { filterInheritedMcp, filterInheritedHooks } = await import(
+      "../lib/subagent-inheritance.js"
+    );
     inheritedMcp = filterInheritedMcp(
       {
         extraToolDefinitions: ctx.extraToolDefinitions,
@@ -10922,8 +11004,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   // repo, refuse the spawn instead of silently running in the parent checkout.
   if (subIsolation === "worktree") {
     try {
-      const { resolveIsolationFailClosed } =
-        await import("../lib/subagent-contract.js");
+      const { resolveIsolationFailClosed } = await import(
+        "../lib/subagent-contract.js"
+      );
       const { isGitRepo } = await import("../lib/git-integration.js");
       const iso = resolveIsolationFailClosed({
         requested: "worktree",
@@ -10942,8 +11025,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   let profile = null;
   if (profileName) {
     try {
-      const { getSubAgentProfile } =
-        await import("../lib/sub-agent-profiles.js");
+      const { getSubAgentProfile } = await import(
+        "../lib/sub-agent-profiles.js"
+      );
       profile = getSubAgentProfile(profileName);
       if (!profile) {
         return {
@@ -11011,8 +11095,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   let childApprovalGate = null;
   if (effectiveContract?.permissionMode) {
     try {
-      const { resolvePermissionMode, resolveEnabledTools } =
-        await import("./headless-runner.js");
+      const { resolvePermissionMode, resolveEnabledTools } = await import(
+        "./headless-runner.js"
+      );
       const perm = resolvePermissionMode(effectiveContract.permissionMode);
       if (perm.readOnly) {
         allowedTools = resolveEnabledTools({ allowedTools, readOnly: true });
@@ -11272,8 +11357,9 @@ async function _executeSpawnSubAgent(args, ctx) {
   let subWorktreeOptions = null;
   if (subIsolation === "worktree") {
     try {
-      const { normalizeSparsePaths } =
-        await import("../lib/worktree-sparse.js");
+      const { normalizeSparsePaths } = await import(
+        "../lib/worktree-sparse.js"
+      );
       const sparse = normalizeSparsePaths(
         args.sparsePaths ?? mdSparsePaths ?? null,
       );
@@ -11475,10 +11561,11 @@ async function _executeSpawnSubAgent(args, ctx) {
 
   // Resolve the registry ONCE before the try so the failure path can also move
   // the sub-agent out of `_active` (the success path does this via complete()).
-  const { SubAgentRegistry } =
-    await import("../lib/sub-agent-registry.js").catch(() => ({
-      SubAgentRegistry: null,
-    }));
+  const { SubAgentRegistry } = await import(
+    "../lib/sub-agent-registry.js"
+  ).catch(() => ({
+    SubAgentRegistry: null,
+  }));
   let registry = null;
   if (SubAgentRegistry) {
     try {
@@ -11543,6 +11630,7 @@ async function _executeSpawnSubAgent(args, ctx) {
             result: subCtx.result,
             error: err.message,
             ...(isEvolutionIngressFailure(err) ||
+            isFormalBudgetError(err) ||
             err?.runtimeLedgerPersistence === true ||
             err?.workflowEffectOutcomeUnknown === true
               ? { fatalError: err }
@@ -11679,6 +11767,7 @@ async function _executeSpawnSubAgent(args, ctx) {
 
     if (
       isEvolutionIngressFailure(err) ||
+      isFormalBudgetError(err) ||
       err?.runtimeLedgerPersistence === true ||
       err?.workflowEffectOutcomeUnknown === true
     ) {
@@ -11798,6 +11887,12 @@ function _openAIProviderRequestReceipt(response, data, clientRequestId) {
  * @returns {Promise<object>} response with .message
  */
 export async function chatWithTools(rawMessages, options) {
+  if (options.serviceTier != null) {
+    resolveModelServiceTier(
+      resolveModelCapabilityProfile(options),
+      options.serviceTier,
+    );
+  }
   const evolutionIngress =
     options.evolutionIngress == null
       ? null
@@ -11827,8 +11922,9 @@ export async function chatWithTools(rawMessages, options) {
     options.contextMemoryPreplanned !== true &&
     options.contextMemorySkipPlanning !== true
   ) {
-    const { prepareCanonicalProviderContext } =
-      await import("../lib/context-memory-kernel/provider-context.js");
+    const { prepareCanonicalProviderContext } = await import(
+      "../lib/context-memory-kernel/provider-context.js"
+    );
     const prepared = await prepareCanonicalProviderContext(rawMessages, {
       ...options,
       contextMemoryToolDefinitions: tools,
@@ -11915,7 +12011,7 @@ export async function chatWithTools(rawMessages, options) {
         },
       );
     }
-    const response = await fetch(apiUrl, {
+    const response = await providerFetch(apiUrl, {
       method: "POST",
       headers: { "Content-Type": "application/json" },
       signal,
@@ -12071,7 +12167,7 @@ export async function chatWithTools(rawMessages, options) {
       );
     }
 
-    const response = await fetch(`${url}/messages`, {
+    const response = await providerFetch(`${url}/messages`, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
@@ -12190,6 +12286,7 @@ export async function chatWithTools(rawMessages, options) {
       reasoning: _openAIResponsesReasoningParams(options, modelProfile),
       stream: typeof options.onToken === "function",
       modelProfile,
+      serviceTier: options.serviceTier,
     });
     if (typeof options.onToken === "function") {
       return await _retryStreamingChat(
@@ -12226,7 +12323,7 @@ export async function chatWithTools(rawMessages, options) {
       );
     }
 
-    const response = await fetch(`${url}/responses`, {
+    const response = await providerFetch(`${url}/responses`, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
@@ -12243,6 +12340,11 @@ export async function chatWithTools(rawMessages, options) {
     }
     const data = await response.json();
     const out = normalizeOpenAIResponsesResponse(data);
+    if (out.usage)
+      out.usage = preserveRequestedServiceTierEvidence(
+        out.usage,
+        options.serviceTier,
+      );
     delete out.responseId;
     const providerReceipt = supportsOpenAIRequestIdentity
       ? _openAIProviderRequestReceipt(response, data, providerRequestId)
@@ -12299,7 +12401,7 @@ export async function chatWithTools(rawMessages, options) {
     );
   }
 
-  const response = await fetch(`${url}/chat/completions`, {
+  const response = await providerFetch(`${url}/chat/completions`, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -12379,6 +12481,7 @@ export async function chatWithTools(rawMessages, options) {
         input_tokens: promptTokens - cached,
         output_tokens: completionTokens,
         cache_read_input_tokens: cached,
+        ...projectEnvelopePricingContext(data.usage, data),
       };
     }
   }
@@ -12535,6 +12638,7 @@ export function _streamErrorDisposition(
   partialText,
   visible = false,
 ) {
+  if (isFormalBudgetError(err)) return "rethrow";
   if (err?.isOutputBackpressureFailure === true) return "rethrow";
   if (isAbortError(err)) return "rethrow";
   if (signal && signal.aborted) return "rethrow";
@@ -12779,7 +12883,7 @@ async function _chatOllamaStreaming(
   stallMs,
   stallTimeoutMs,
 ) {
-  const response = await fetch(apiUrl, {
+  const response = await providerFetch(apiUrl, {
     method: "POST",
     headers: { "Content-Type": "application/json" },
     signal,
@@ -12841,7 +12945,7 @@ async function _chatAnthropicStreaming(
   stallMs,
   stallTimeoutMs,
 ) {
-  const response = await fetch(apiUrl, {
+  const response = await providerFetch(apiUrl, {
     method: "POST",
     headers: { "Content-Type": "application/json", ...extraHeaders },
     signal,
@@ -12908,7 +13012,7 @@ async function _chatOpenAIStreaming(
   stallTimeoutMs,
   providerRequestId = null,
 ) {
-  const response = await fetch(apiUrl, {
+  const response = await providerFetch(apiUrl, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -12991,7 +13095,7 @@ async function _chatOpenAIResponsesStreaming(
   stallTimeoutMs,
   providerRequestId = null,
 ) {
-  const response = await fetch(apiUrl, {
+  const response = await providerFetch(apiUrl, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -13067,6 +13171,11 @@ async function _chatOpenAIResponsesStreaming(
     return out;
   }
   const out = finalizeOpenAIResponsesStream(state);
+  if (out.usage)
+    out.usage = preserveRequestedServiceTierEvidence(
+      out.usage,
+      body.service_tier,
+    );
   const responseId = out.responseId || state.responseId;
   delete out.responseId;
   const providerReceipt = _openAIProviderRequestReceipt(
@@ -13988,8 +14097,9 @@ async function _getAutoCompactor(options) {
   let canonicalRequired = false;
   try {
     const { feature } = await import("../lib/feature-flags.js");
-    const { resolveCliContextMemoryCutover } =
-      await import("../lib/context-memory-kernel/authority.js");
+    const { resolveCliContextMemoryCutover } = await import(
+      "../lib/context-memory-kernel/authority.js"
+    );
     const cutover = resolveCliContextMemoryCutover({
       env: options.contextMemoryEnv || process.env,
       scopeKey: options.sessionId
@@ -13998,8 +14108,9 @@ async function _getAutoCompactor(options) {
     });
     canonicalRequired = cutover.canonical;
     if (feature("PROMPT_COMPRESSOR") || cutover.canonical) {
-      const { PromptCompressor } =
-        await import("../harness/prompt-compressor.js");
+      const { PromptCompressor } = await import(
+        "../harness/prompt-compressor.js"
+      );
       const llmQuery =
         typeof options.compactionLlmQuery === "function"
           ? options.compactionLlmQuery
@@ -14049,8 +14160,9 @@ async function _getAutoCompactor(options) {
       compressor = _instrumentAutoCompactorUsage(compressor, options);
       if (cutover.canonical) {
         const compatibilitySummarizer = compressor;
-        const { compactLiveMessagesCanonical } =
-          await import("../lib/context-memory-kernel/live-compaction.js");
+        const { compactLiveMessagesCanonical } = await import(
+          "../lib/context-memory-kernel/live-compaction.js"
+        );
         compressor = {
           canonicalKernel: true,
           shouldAutoCompact: (messages) =>
@@ -14372,6 +14484,12 @@ function permissionDecision(callId, tool, result) {
 }
 
 export async function* agentLoop(messages, options) {
+  if (options.serviceTier != null) {
+    resolveModelServiceTier(
+      resolveModelCapabilityProfile(options),
+      options.serviceTier,
+    );
+  }
   const executionPolicy = captureAgentExecutionPolicy(options, {
     hermetic: true,
   });
@@ -14386,8 +14504,9 @@ export async function* agentLoop(messages, options) {
       : captureAgentEvolutionIngress(options.evolutionIngress);
   let llmCall = configuredChatFn || chatWithTools;
   if (evolutionIngress !== null && configuredChatFn) {
-    const { captureCanonicalFallbackChatFn } =
-      await import("./fallback-model.js");
+    const { captureCanonicalFallbackChatFn } = await import(
+      "./fallback-model.js"
+    );
     llmCall = captureCanonicalFallbackChatFn(
       configuredChatFn,
       evolutionIngress,
@@ -14396,8 +14515,9 @@ export async function* agentLoop(messages, options) {
   // Shared iteration budget — replaces hardcoded MAX_ITERATIONS.
   // When options.iterationBudget is provided (e.g. from parent agent),
   // the same budget instance is shared, so parent+child consume from one pool.
-  const { IterationBudget, WarningLevel } =
-    await import("../lib/iteration-budget.js");
+  const { IterationBudget, WarningLevel } = await import(
+    "../lib/iteration-budget.js"
+  );
   const budget = options.iterationBudget || new IterationBudget();
   const { HostResourceBudget } = await import("../lib/host-resource-budget.js");
   const hostResourceBudget =
@@ -14747,8 +14867,9 @@ export async function* agentLoop(messages, options) {
     options.strictUsageTelemetry !== true &&
     workflowEffectId == null
   ) {
-    const { makeRunnableProviderFallback } =
-      await import("../lib/runnable-provider.js");
+    const { makeRunnableProviderFallback } = await import(
+      "../lib/runnable-provider.js"
+    );
     llmCall = makeRunnableProviderFallback(llmCall, {
       onFallback: ({ from, to, reason, fromModel, toModel }) => {
         // Telemetry: a provider/model fallback is a reliability signal — count
@@ -15370,6 +15491,7 @@ export async function* agentLoop(messages, options) {
         }
       } catch (_e) {
         if (isAbortError(_e) || signal?.aborted) throw _e;
+        if (isFormalBudgetError(_e)) throw _e;
         if (_e?.runtimeLedgerPersistence === true) throw _e;
         if (_e?.code === "CC_AGENT_EVOLUTION_INGRESS_FAILED") throw _e;
         if (
@@ -15388,8 +15510,9 @@ export async function* agentLoop(messages, options) {
         }
         compactionUsageCall =
           compactionUsageCall || compactionUsageState?.calls?.at(-1) || null;
-        const { getRejectedCompactionUsage } =
-          await import("../lib/context-memory-kernel/live-compaction.js");
+        const { getRejectedCompactionUsage } = await import(
+          "../lib/context-memory-kernel/live-compaction.js"
+        );
         const rejectedUsage = getRejectedCompactionUsage(_e);
         const returnedUsage = compactionUsageCall?.returnedUsage;
         const providerReceipt = _takeCompactionProviderReceipt(
@@ -15801,6 +15924,7 @@ export async function* agentLoop(messages, options) {
           ];
         }
       } catch (_e) {
+        if (isFormalBudgetError(_e)) throw _e;
         // prepareCall failures are non-critical — proceed with original messages
       }
     }
@@ -15830,8 +15954,9 @@ export async function* agentLoop(messages, options) {
     // the run-level ids (turn.id correlates with the agent.iteration counter).
     let canonicalProviderContext = null;
     if (options.contextMemorySkipPlanning !== true) {
-      const { prepareCanonicalProviderContext } =
-        await import("../lib/context-memory-kernel/provider-context.js");
+      const { prepareCanonicalProviderContext } = await import(
+        "../lib/context-memory-kernel/provider-context.js"
+      );
       canonicalProviderContext = await prepareCanonicalProviderContext(
         callMessages,
         {
@@ -16400,6 +16525,7 @@ export async function* agentLoop(messages, options) {
                 : {}),
             },
             error: err.message,
+            fatalError: isFormalBudgetError(err) ? err : null,
           }),
         );
         return { call, toolArgs, workflowBinding, promise };
@@ -16407,7 +16533,12 @@ export async function* agentLoop(messages, options) {
       for (const { call, toolArgs, workflowBinding, promise } of inflight) {
         throwIfAborted(signal);
         _throwBackgroundUsageFailureState(backgroundUsageFailureState);
-        const { result: toolResult, error: toolError } = await promise;
+        const {
+          result: toolResult,
+          error: toolError,
+          fatalError,
+        } = await promise;
+        if (fatalError) throw fatalError;
         throwIfAborted(signal);
         const taskAdvanced = taskProgressTracker.record(
           call.function.name,
@@ -16850,6 +16981,7 @@ export async function* agentLoop(messages, options) {
           } catch (err) {
             if (
               isEvolutionIngressFailure(err) ||
+              isFormalBudgetError(err) ||
               err?.runtimeLedgerPersistence === true
             )
               throw err;