@jsonstudio/llms 0.6.633 → 0.6.743

package/dist/filters/special/response-tool-arguments-toon-decode.js

@@ -1,5 +1,6 @@
 import { isShellToolName, normalizeToolName } from '../../tools/tool-description-utils.js';
 import { repairFindMeta } from '../../conversion/shared/tooling.js';
+import { decodeToonToKeyValue, coerceToonValue } from '../../utils/toon.js';
 function envEnabled() {
     // Default ON. Allow disabling via env RCC_TOON_ENABLE/ROUTECODEX_TOON_ENABLE = 0|false|off
     const v = String(process?.env?.RCC_TOON_ENABLE || process?.env?.ROUTECODEX_TOON_ENABLE || '').toLowerCase();
@@ -8,71 +9,6 @@ function envEnabled() {
     return !(v === '0' || v === 'false' || v === 'off');
 }
 function isObject(v) { return !!v && typeof v === 'object' && !Array.isArray(v); }
-function decodeToonPairs(toon) {
-    try {
-        const out = {};
-        const lines = String(toon).split(/\r?\n/);
-        let currentKey = null;
-        let currentVal = '';
-        const flush = () => {
-            if (currentKey) {
-                out[currentKey] = currentVal;
-            }
-            currentKey = null;
-            currentVal = '';
-        };
-        for (const raw of lines) {
-            const line = raw.trim();
-            if (!line)
-                continue;
-            const m = line.match(/^([A-Za-z0-9_\-]+)\s*:\s*(.*)$/);
-            if (m) {
-                // 新的 key: value 行，先提交上一段，再开始累积新 key 的值
-                flush();
-                currentKey = m[1];
-                currentVal = m[2] ?? '';
-            }
-            else {
-                // 非 key: value 行视为上一 key 的续行（例如多行脚本）
-                if (!currentKey) {
-                    // 如果一开始就遇到无法识别的行，认为整个 TOON 不是我们支持的形态
-                    return null;
-                }
-                currentVal += (currentVal ? '\n' : '') + raw;
-            }
-        }
-        flush();
-        return Object.keys(out).length ? out : null;
-    }
-    catch {
-        return null;
-    }
-}
-function coerceToPrimitive(value) {
-    const trimmed = value.trim();
-    if (!trimmed)
-        return '';
-    const lower = trimmed.toLowerCase();
-    if (lower === 'true')
-        return true;
-    if (lower === 'false')
-        return false;
-    if (/^[+-]?\d+(\.\d+)?$/.test(trimmed)) {
-        const num = Number(trimmed);
-        if (Number.isFinite(num))
-            return num;
-    }
-    if ((trimmed.startsWith('{') && trimmed.endsWith('}')) ||
-        (trimmed.startsWith('[') && trimmed.endsWith(']'))) {
-        try {
-            return JSON.parse(trimmed);
-        }
-        catch {
-            // fall through
-        }
-    }
-    return value;
-}
 /**
  * Decode arguments.toon to standard JSON ({command, workdir?}) and map tool name 'shell_toon' → 'shell'.
  * Stage: response_pre (before arguments stringify and invariants).
@@ -99,7 +35,6 @@ export class ResponseToolArgumentsToonDecodeFilter {
                         const toolName = typeof rawName === 'string' ? rawName : '';
                         const normalizedName = normalizeToolName(toolName);
                         const isShellLike = isShellToolName(toolName);
-                        const isApplyPatch = normalizedName === 'apply_patch';
                         const argIn = fn.arguments;
                         let parsed = undefined;
                         if (typeof argIn === 'string') {
@@ -118,7 +53,7 @@ export class ResponseToolArgumentsToonDecodeFilter {
                         const toon = parsed.toon;
                         if (typeof toon !== 'string' || !toon.trim())
                             continue;
-                        const kv = decodeToonPairs(toon);
+                        const kv = decodeToonToKeyValue(toon);
                         if (!kv) {
                             const preview = toon.split(/\r?\n/).slice(0, 5).join('\n');
                             const warnMsg = `response_tool_arguments_toon_decode: failed to decode TOON arguments for tool "${fn.name ?? 'unknown'}"`;
@@ -137,10 +72,6 @@ export class ResponseToolArgumentsToonDecodeFilter {
                             }
                             continue; // keep original if decode fails
                         }
-                        // apply_patch 的 toon 由专门的 ResponseApplyPatchToonDecodeFilter 处理，这里跳过，避免覆盖。
-                        if (isApplyPatch) {
-                            continue;
-                        }
                         if (isShellLike) {
                             const commandRaw = (typeof kv['command'] === 'string' && kv['command'].trim()
                                 ? kv['command']
@@ -191,16 +122,16 @@ export class ResponseToolArgumentsToonDecodeFilter {
                                 catch {
                                     /* keep original */
                                 }
-                                if (typeof fn.name === 'string' && fn.name === 'shell_toon') {
-                                    fn.name = 'shell';
-                                }
+                            }
+                            if (typeof fn.name === 'string' && fn.name === 'shell_toon') {
+                                fn.name = 'shell';
                             }
                         }
                         else {
                             // 通用 TOON → JSON 解码：除 shell / apply_patch 以外的工具，将 key: value 对映射为普通 JSON 字段。
                             const merged = {};
                             for (const [key, value] of Object.entries(kv)) {
-                                merged[key] = coerceToPrimitive(value);
+                                merged[key] = coerceToonValue(value);
                             }
                             try {
                                 fn.arguments = JSON.stringify(merged);

package/dist/filters/utils/snapshot-writer.js

@@ -13,22 +13,60 @@ function isSnapshotEnabled() {
     const v = String(process?.env?.RCC_FILTER_SNAPSHOT || process?.env?.RCC_HOOKS_VERBOSITY || '').toLowerCase();
     return v === '1' || v === 'true' || v === 'verbose';
 }
+function sanitizeToken(value, fallback) {
+    if (typeof value !== 'string') {
+        return fallback;
+    }
+    const trimmed = value.trim();
+    if (!trimmed) {
+        return fallback;
+    }
+    return trimmed.replace(/[^A-Za-z0-9_.-]/g, '_') || fallback;
+}
+function toErrorCode(error) {
+    if (!error || typeof error !== 'object') {
+        return undefined;
+    }
+    const code = error.code;
+    return typeof code === 'string' && code.trim() ? code : undefined;
+}
+async function writeUniqueFile(dir, baseName, contents) {
+    const parsed = path.parse(baseName);
+    const ext = parsed.ext || '.json';
+    const stem = parsed.name || 'snapshot';
+    for (let i = 0; i < 64; i += 1) {
+        const name = i === 0 ? `${stem}${ext}` : `${stem}_${i}${ext}`;
+        try {
+            await fsp.writeFile(path.join(dir, name), contents, { encoding: 'utf-8', flag: 'wx' });
+            return;
+        }
+        catch (error) {
+            if (toErrorCode(error) === 'EEXIST') {
+                continue;
+            }
+            throw error;
+        }
+    }
+    const fallback = `${stem}_${Date.now()}_${Math.random().toString(36).slice(2, 8)}${ext}`;
+    await fsp.writeFile(path.join(dir, fallback), contents, 'utf-8');
+}
 export async function writeFilterSnapshot(options) {
     try {
         if (!isSnapshotEnabled())
             return;
-        const rid = options.requestId || `req_${Date.now()}`;
+        const rid = sanitizeToken(options.requestId || '', `req_${Date.now()}`);
         const baseOverride = process?.env?.RCC_SNAPSHOT_DIR;
         const base = baseOverride && baseOverride.trim()
             ? baseOverride.trim()
             : path.join(os.homedir(), '.routecodex', 'codex-samples');
         const folder = mapEndpointToFolder(options.endpoint);
-        const dir = path.join(base, folder);
+        const provider = sanitizeToken(options.profile || '', '__pending__');
+        const dir = path.join(base, folder, provider, rid);
         await fsp.mkdir(dir, { recursive: true });
         const parts = ['filters', options.stage.replace(/\s+/g, ''), options.tag || (options.name ? `after_${options.name}` : 'after')]
             .filter(Boolean)
             .join('_');
-        const file = path.join(dir, `${rid}_${parts}.json`);
+        const file = `${sanitizeToken(parts, 'filters')}.json`;
         const payload = {
             meta: {
                 requestId: rid,
@@ -41,7 +79,7 @@ export async function writeFilterSnapshot(options) {
             },
             data: options.data
         };
-        await fsp.writeFile(file, JSON.stringify(payload, null, 2), 'utf-8');
+        await writeUniqueFile(dir, file, JSON.stringify(payload, null, 2));
     }
     catch { /* ignore snapshot errors */ }
 }

package/dist/guidance/index.js

@@ -66,7 +66,9 @@ function augmentApplyPatch(fn) {
         'Each change describes one operation, e.g. `{ "file": "src/foo.ts", "kind": "insert_after", "anchor": "const foo = 1;", "lines": ["const bar = 2;"] }`.',
         'Supported kinds: insert_after, insert_before, replace, delete, create_file, delete_file.',
         'Paths must stay relative to the workspace root (no leading "/" or drive letters).',
-        'Insert operations require `anchor` text; replace/delete require exact `target` snippets; `lines` omit "+/-" prefixes.'
+        'Insert operations require `anchor` text; replace/delete require exact `target` snippets; `lines` omit "+/-" prefixes.',
+        'If you must emit raw patch text, use "*** Begin Patch" / "*** End Patch" with "*** Update/Add/Delete File" headers (no "diff --git").',
+        'Do not output "*** Create File:"; use "*** Add File:".'
     ].join('\n');
     const params = ensureObjectSchema(fn.parameters);
     const props = params.properties;
@@ -230,7 +232,10 @@ export function augmentAnthropicTools(tools) {
                         'Before using apply_patch, always read the latest content of the target file (via shell or another tool) and base your changes on that content.',
                         'Provide structured changes (insert_after / insert_before / replace / delete / create_file / delete_file) instead of raw patch text.',
                         'Each change must include the target file (relative path) plus anchor/target snippets and the replacement lines.',
-                        '所有路径必须相对工作区根目录，禁止输出以 / 或盘符开头的绝对路径。'
+                        '所有路径必须相对工作区根目录，禁止输出以 / 或盘符开头的绝对路径。',
+                        'Example: {\"changes\":[{\"file\":\"src/app.ts\",\"kind\":\"replace\",\"target\":\"const answer = 41;\",\"lines\":[\"const answer = 42;\"]}]}（修改同一文件时尽量只修改一段连续区域，多处不相邻修改请拆成多次 apply_patch 调用）.',
+                        'Raw patch text must use "*** Begin Patch" / "*** End Patch" + "*** Update/Add/Delete File" headers（不要输出 "diff --git"）。',
+                        '不要输出 "*** Create File:"；请使用 "*** Add File:".'
                     ].join('\n');
                     copy.description = appendOnce(desc, guidance, marker);
                 }
@@ -267,6 +272,7 @@ export function buildSystemToolGuidance() {
     lines.push(bullet('File writes are FORBIDDEN via shell (no redirection, no here-doc, no sed -i, no ed -s, no tee). Use apply_patch ONLY. / 通过 shell 写文件一律禁止（不得使用重定向、heredoc、sed -i、ed -s、tee）；必须使用 apply_patch。'));
     lines.push(bullet('apply_patch: Before writing, always read the target file first and compute changes against the latest content using appropriate tools. / apply_patch 在写入前必须先通过合适的工具读取目标文件最新内容，并基于该内容生成变更。'));
     lines.push(bullet('apply_patch: Provide structured JSON arguments with a `changes` array (insert_after / insert_before / replace / delete / create_file / delete_file); omit "+/-" prefixes in `lines`; file paths必须是相对路径。 / apply_patch 仅接受结构化 JSON。'));
+    lines.push(bullet('apply_patch: For a given file, prefer one contiguous change block per call; if you need to touch non-adjacent regions, split them into multiple apply_patch calls. / apply_patch 修改同一文件时尽量只提交一段连续补丁，多个不相邻位置请拆成多次调用。'));
     lines.push(bullet('update_plan: Keep exactly one step in_progress; others pending/completed. / 仅一个 in_progress 步骤。'));
     lines.push(bullet('view_image: Path must be an image file (.png .jpg .jpeg .gif .webp .bmp .svg). / 仅图片路径。'));
     lines.push(bullet('Do NOT use view_image for text files (.md/.ts/.js/.json). Use shell: {"command":["cat","<path>"]}. / 文本文件请用 shell: cat。'));

package/dist/router/virtual-router/engine-health.js

@@ -275,10 +275,6 @@ export function applyQuotaRecoveryImpl(event, healthManager, clearProviderCooldo
         healthManager.recordSuccess(providerKey);
         resetRateLimitBackoffForProvider(providerKey);
         clearProviderCooldown(providerKey);
-        debug?.log?.('[virtual-router] quota recovery', {
-            providerKey,
-            reason: detail.reason
-        });
     }
     catch {
         // 恢复失败不得影响主路由流程

package/dist/router/virtual-router/engine-selection.d.ts

@@ -1,4 +1,4 @@
-import type { ClassificationResult, RoutePoolTier, RouterMetadataInput, RoutingFeatures } from './types.js';
+import type { ClassificationResult, RoutePoolTier, RouterMetadataInput, RoutingFeatures, ProviderQuotaView } from './types.js';
 import type { RoutingInstructionState } from './routing-instructions.js';
 import type { ContextAdvisor } from './context-advisor.js';
 import type { RouteLoadBalancer } from './load-balancer.js';
@@ -12,6 +12,7 @@ type SelectionDeps = {
     loadBalancer: RouteLoadBalancer;
     isProviderCoolingDown: (providerKey: string) => boolean;
     resolveStickyKey: (metadata: RouterMetadataInput) => string | undefined;
+    quotaView?: ProviderQuotaView;
 };
 export declare function selectProviderImpl(requestedRoute: string, metadata: RouterMetadataInput, classification: ClassificationResult, features: RoutingFeatures, activeState: RoutingInstructionState, deps: SelectionDeps, options?: {
     routingState?: RoutingInstructionState;

package/dist/router/virtual-router/engine-selection.js

@@ -1,11 +1,32 @@
 import { DEFAULT_ROUTE, ROUTE_PRIORITY, VirtualRouterError, VirtualRouterErrorCode } from './types.js';
 export function selectProviderImpl(requestedRoute, metadata, classification, features, activeState, deps, options = {}) {
     const state = options.routingState ?? activeState;
+    const quotaView = deps.quotaView;
+    const quotaNow = quotaView ? Date.now() : 0;
+    const isAllowedByQuota = (key) => {
+        if (!quotaView) {
+            return true;
+        }
+        const entry = quotaView(key);
+        if (!entry) {
+            return true;
+        }
+        if (!entry.inPool) {
+            return false;
+        }
+        if (entry.cooldownUntil && entry.cooldownUntil > quotaNow) {
+            return false;
+        }
+        if (entry.blacklistUntil && entry.blacklistUntil > quotaNow) {
+            return false;
+        }
+        return true;
+    };
     const excludedProviderKeys = extractExcludedProviderKeySet(features.metadata);
     const forcedResolution = state.forcedTarget ? resolveInstructionTarget(state.forcedTarget, deps.providerRegistry) : null;
     if (forcedResolution && forcedResolution.mode === 'exact') {
         const forcedKey = forcedResolution.keys[0];
-        if (!excludedProviderKeys.has(forcedKey) && !deps.isProviderCoolingDown(forcedKey)) {
+        if (!excludedProviderKeys.has(forcedKey) && !deps.isProviderCoolingDown(forcedKey) && isAllowedByQuota(forcedKey)) {
             return {
                 providerKey: forcedKey,
                 routeUsed: requestedRoute,
@@ -22,7 +43,8 @@ export function selectProviderImpl(requestedRoute, metadata, classification, fea
             const stickyKey = stickyResolution.keys[0];
             if (deps.healthManager.isAvailable(stickyKey) &&
                 !excludedProviderKeys.has(stickyKey) &&
-                !deps.isProviderCoolingDown(stickyKey)) {
+                !deps.isProviderCoolingDown(stickyKey) &&
+                isAllowedByQuota(stickyKey)) {
                 return {
                     providerKey: stickyKey,
                     routeUsed: requestedRoute,
@@ -34,7 +56,8 @@ export function selectProviderImpl(requestedRoute, metadata, classification, fea
         if (stickyResolution && stickyResolution.mode === 'filter' && stickyResolution.keys.length > 0) {
             const liveKeys = stickyResolution.keys.filter((key) => deps.healthManager.isAvailable(key) &&
                 !excludedProviderKeys.has(key) &&
-                !deps.isProviderCoolingDown(key));
+                !deps.isProviderCoolingDown(key) &&
+                isAllowedByQuota(key));
             if (liveKeys.length > 0) {
                 stickyKeySet = new Set(liveKeys);
             }
@@ -247,13 +270,62 @@ function trySelectFromTier(routeName, tier, stickyKey, estimatedTokens, features
     }
     const contextResult = deps.contextAdvisor.classify(targets, estimatedTokens, (key) => deps.providerRegistry.get(key));
     const prioritizedPools = buildContextCandidatePools(contextResult);
+    const quotaView = deps.quotaView;
+    const now = quotaView ? Date.now() : 0;
+    const selectWithQuota = (candidates) => {
+        if (!quotaView) {
+            return deps.loadBalancer.select({
+                routeName: `${routeName}:${tier.id}`,
+                candidates,
+                stickyKey: options.allowAliasRotation ? undefined : stickyKey,
+                availabilityCheck: (key) => deps.healthManager.isAvailable(key)
+            });
+        }
+        const buckets = new Map();
+        for (const key of candidates) {
+            const entry = quotaView(key);
+            if (!entry) {
+                const list = buckets.get(100) ?? [];
+                list.push(key);
+                buckets.set(100, list);
+                continue;
+            }
+            if (!entry.inPool) {
+                continue;
+            }
+            if (entry.cooldownUntil && entry.cooldownUntil > now) {
+                continue;
+            }
+            if (entry.blacklistUntil && entry.blacklistUntil > now) {
+                continue;
+            }
+            const tierPriority = typeof entry.priorityTier === 'number' && Number.isFinite(entry.priorityTier)
+                ? entry.priorityTier
+                : 100;
+            const list = buckets.get(tierPriority) ?? [];
+            list.push(key);
+            buckets.set(tierPriority, list);
+        }
+        const sortedPriorities = Array.from(buckets.keys()).sort((a, b) => a - b);
+        for (const priority of sortedPriorities) {
+            const bucketCandidates = buckets.get(priority) ?? [];
+            if (!bucketCandidates.length) {
+                continue;
+            }
+            const selected = deps.loadBalancer.select({
+                routeName: `${routeName}:${tier.id}`,
+                candidates: bucketCandidates,
+                stickyKey: options.allowAliasRotation ? undefined : stickyKey,
+                availabilityCheck: (key) => deps.healthManager.isAvailable(key)
+            });
+            if (selected) {
+                return selected;
+            }
+        }
+        return null;
+    };
     for (const candidatePool of prioritizedPools) {
-        const providerKey = deps.loadBalancer.select({
-            routeName: `${routeName}:${tier.id}`,
-            candidates: candidatePool,
-            stickyKey: options.allowAliasRotation ? undefined : stickyKey,
-            availabilityCheck: (key) => deps.healthManager.isAvailable(key)
-        });
+        const providerKey = selectWithQuota(candidatePool);
         if (providerKey) {
             return { providerKey, poolTargets: tier.targets, tierId: tier.id };
         }
@@ -277,6 +349,26 @@ export function selectFromStickyPool(stickyKeySet, metadata, features, state, de
     ]));
     const disabledModels = new Map(Array.from(state.disabledModels.entries()).map(([provider, models]) => [provider, new Set(models)]));
     let candidates = Array.from(stickyKeySet).filter((key) => !deps.isProviderCoolingDown(key));
+    const quotaView = deps.quotaView;
+    const now = quotaView ? Date.now() : 0;
+    if (quotaView) {
+        candidates = candidates.filter((key) => {
+            const entry = quotaView(key);
+            if (!entry) {
+                return true;
+            }
+            if (!entry.inPool) {
+                return false;
+            }
+            if (entry.cooldownUntil && entry.cooldownUntil > now) {
+                return false;
+            }
+            if (entry.blacklistUntil && entry.blacklistUntil > now) {
+                return false;
+            }
+            return true;
+        });
+    }
     if (allowedProviders.size > 0) {
         candidates = candidates.filter((key) => {
             const providerId = extractProviderId(key);

package/dist/router/virtual-router/engine.d.ts

@@ -1,6 +1,7 @@
-import { type RoutingDecision, type RoutingDiagnostics, type RouterMetadataInput, type VirtualRouterConfig, type TargetMetadata, type ProviderFailureEvent, type ProviderErrorEvent, type VirtualRouterHealthStore } from './types.js';
+import { type RoutingDecision, type RoutingDiagnostics, type StopMessageStateSnapshot, type RouterMetadataInput, type VirtualRouterConfig, type TargetMetadata, type ProviderFailureEvent, type ProviderErrorEvent, type VirtualRouterHealthStore } from './types.js';
 import type { ProcessedRequest, StandardizedRequest } from '../../conversion/hub/types/standardized.js';
 import { type RoutingInstructionState } from './routing-instructions.js';
+import type { ProviderQuotaView } from './types.js';
 interface RoutingInstructionStateStore {
     loadSync(key: string): RoutingInstructionState | null;
     saveAsync(key: string, state: RoutingInstructionState | null): void;
@@ -22,9 +23,11 @@ export declare class VirtualRouterEngine {
     private healthStore?;
     private routingStateStore;
     private routingInstructionState;
+    private quotaView?;
     constructor(deps?: {
         healthStore?: VirtualRouterHealthStore;
         routingStateStore?: RoutingInstructionStateStore;
+        quotaView?: ProviderQuotaView;
     });
     initialize(config: VirtualRouterConfig): void;
     route(request: StandardizedRequest | ProcessedRequest, metadata: RouterMetadataInput): {
@@ -32,6 +35,7 @@ export declare class VirtualRouterEngine {
         decision: RoutingDecision;
         diagnostics: RoutingDiagnostics;
     };
+    getStopMessageState(metadata: RouterMetadataInput): StopMessageStateSnapshot | null;
     handleProviderFailure(event: ProviderFailureEvent): void;
     handleProviderError(event: ProviderErrorEvent): void;
     getStatus(): {