@c4a/server-cli 0.4.15-alpha.5 → 0.4.15-alpha.7

package/index.js

@@ -220991,6 +220991,12 @@ class LlmServiceImpl {
       if (options?.systemPrompt) {
         callSettings.system = options.systemPrompt;
       }
+      if (options?.jsonMode) {
+        callSettings.providerOptions = {
+          openai: { responseFormat: { type: "json_object" } },
+          anthropic: { responseFormat: { type: "json_object" } }
+        };
+      }
       const result = await retry(() => generateText(callSettings), {
         shouldRetry: (error40) => isRetryableStatus(getHttpStatus(error40)) || isTimeoutError2(error40)
       });
@@ -221038,6 +221044,12 @@ class LlmServiceImpl {
       if (options?.systemPrompt) {
         callSettings.system = options.systemPrompt;
       }
+      if (options?.jsonMode) {
+        callSettings.providerOptions = {
+          openai: { responseFormat: { type: "json_object" } },
+          anthropic: { responseFormat: { type: "json_object" } }
+        };
+      }
       const result = await retry(() => Promise.resolve(streamText(callSettings)), {
         shouldRetry: (error40) => isRetryableStatus(getHttpStatus(error40)) || isTimeoutError2(error40)
       });
@@ -221856,6 +221868,7 @@ function parseExtractionOutput(raw, schema) {
     }
     parsed = normalizeFlatOutput(parsed);
     parsed = stripNulls(parsed);
+    parsed = patchAttributeDefaults(parsed);
     const result = schema.safeParse(parsed);
     if (!result.success) {
       return { success: false, error: result.error };
@@ -221923,7 +221936,37 @@ function tryParseJson(raw) {
   return repairAndParse(raw);
 }
 function repairAndParse(raw) {
-  return JSON.parse(jsonrepair(raw));
+  try {
+    return JSON.parse(jsonrepair(raw));
+  } catch {
+    const truncated = truncateToLastCompleteEntry(raw);
+    if (truncated) {
+      return JSON.parse(jsonrepair(truncated));
+    }
+    throw new Error(`JSON repair failed for output of length ${raw.length}`);
+  }
+}
+function truncateToLastCompleteEntry(raw) {
+  let text2 = raw.trim();
+  if (text2.startsWith("```")) {
+    const firstNewline = text2.indexOf(`
+`);
+    text2 = text2.slice(firstNewline + 1);
+  }
+  if (text2.endsWith("```")) {
+    text2 = text2.slice(0, -3);
+  }
+  const pattern = /\}\s*,\s*"P\d+"/g;
+  let lastMatch = null;
+  let m;
+  while ((m = pattern.exec(text2)) !== null) {
+    lastMatch = m;
+  }
+  if (!lastMatch)
+    return null;
+  const cutPos = lastMatch.index + 1;
+  const truncated = text2.slice(0, cutPos) + "}";
+  return truncated;
 }
 var PARAGRAPH_TAG_RE = /^P\d+$/;
 function normalizeFlatOutput(parsed) {
@@ -221956,6 +221999,43 @@ function stripNulls(value) {
   }
   return value;
 }
+function patchAttributeDefaults(value) {
+  if (!value || typeof value !== "object" || Array.isArray(value))
+    return value;
+  const obj = value;
+  if ("paragraphs" in obj && Array.isArray(obj.paragraphs)) {
+    return {
+      ...obj,
+      paragraphs: obj.paragraphs.map((p) => {
+        if (!p || typeof p !== "object")
+          return p;
+        const para = p;
+        const atoms2 = para.atoms;
+        if (!atoms2 || typeof atoms2 !== "object")
+          return p;
+        return { ...para, atoms: patchAttrsInAtoms(atoms2) };
+      })
+    };
+  }
+  return value;
+}
+function patchAttrsInAtoms(atoms2) {
+  const attrs = atoms2.attributes;
+  if (!Array.isArray(attrs))
+    return atoms2;
+  return {
+    ...atoms2,
+    attributes: attrs.map((attr) => {
+      if (!attr || typeof attr !== "object" || Array.isArray(attr))
+        return attr;
+      const a = attr;
+      if (typeof a.type !== "string" || a.type === "") {
+        return { ...a, type: "other" };
+      }
+      return attr;
+    })
+  };
+}
 function isRecord(value) {
   return !!value && typeof value === "object" && "key" in value && "value" in value && typeof value.key === "string";
 }
@@ -222212,20 +222292,20 @@ class GleaningExtractor {
 // ../llm/src/prompts/docAtomAnnotation.ts
 init_src();
 var DOC_ATOM_DEFS = [
-  ["entities", "Named things: systems, services, modules, documents, APIs (NOT people/teams — use roles for those)", entityAtomSchema],
+  ["entities", "Named things with independent identity — something you can ask questions about ('What does X do?', 'Who owns X?'). Examples: systems, services, modules, APIs, products. If it is a value, address, path, or configuration detail, it is an attribute of an entity, not an entity itself. (NOT people/teams — use roles for those). kind: implementation=internal systems/services, external=third-party dependencies, concept=abstract/not-yet-implemented", entityAtomSchema],
   ["relations", "Connections between entities", relationAtomSchema],
   ["behaviors", "Actions/operations: functions, API calls, user actions, workflows", behaviorAtomSchema],
   ["attributes", "Properties of entities", attributeAtomSchema],
   ["states", "Possible states of entities", stateAtomSchema],
-  ["rules", "Business/domain rules: validations, constraints expressed as logic", ruleAtomSchema],
+  ["rules", "Conditional business/domain logic: IF condition THEN consequence (e.g., 'IF user not authenticated THEN reject request')", ruleAtomSchema],
   ["transitions", "State changes: from→to triggered by events or guards", transitionAtomSchema],
   ["events", "Occurrences that trigger behaviors", eventAtomSchema],
   ["decisions", "Architectural or business decisions", decisionAtomSchema],
   ["metrics", "Measurable targets: SLA, throughput, error_rate, with thresholds", metricAtomSchema],
-  ["roles", "Actors: human roles, teams, personas that perform behaviors", roleAtomSchema],
-  ["constraints", "Hard/soft requirements", constraintAtomSchema],
+  ["roles", "Actors: human roles, teams, personas that perform behaviors. kind: human=individual role, team=group/department, persona=user archetype. System-triggered actions use entity relations, NOT roles", roleAtomSchema],
+  ["constraints", "Declarative requirements: 'X must/should/must-not Y' (e.g., 'passwords must be >= 8 chars'). Unlike rules, constraints have no IF-THEN condition — they are unconditional mandates or restrictions", constraintAtomSchema],
   ["comparisons", "Side-by-side evaluations", comparisonAtomSchema],
-  ["boundaries", "System/domain boundaries: what is included vs excluded", boundaryAtomSchema]
+  ["boundaries", "Explicit scope declarations: what is included vs excluded. Only extract when the text explicitly declares scope (e.g., 'this product covers X but NOT Y'). Implicit containment (A runs inside B) is expressed via entity relations, not boundaries", boundaryAtomSchema]
 ];
 function buildAtomTypesBlock() {
   return DOC_ATOM_DEFS.map(([name21, desc, schema], i) => {
@@ -222255,7 +222335,8 @@ Return a single JSON object keyed by paragraph tags. Only include paragraphs tha
     "relations": [{ "from": "UserService", "to": "Database", "type": "DEPENDS_ON", "confidence": 0.9 }]
   },
   "P3": {
-    "rules": [{ "description": "User must be authenticated before access", "expression": "user.isAuthenticated == true", "confidence": 0.85 }]
+    "constraints": [{ "description": "User must be authenticated before access", "severity": "must", "confidence": 0.9 }],
+    "rules": [{ "description": "Reject request if user is not authenticated", "expression": "IF !user.isAuthenticated THEN reject", "confidence": 0.85 }]
   }
 }
 
@@ -222267,7 +222348,13 @@ Return a single JSON object keyed by paragraph tags. Only include paragraphs tha
 - **Enum fields MUST use ONLY the listed values.** For example, entity.kind must be one of "implementation"|"external"|"concept" — do NOT use values from other atom types (e.g., do NOT put "team" or "human" in entity.kind; those belong to roles.kind).
 - Every atom MUST include a "confidence" field (0.0-1.0) indicating how confident you are in the extraction. Use higher values (0.85-1.0) for explicitly stated facts and lower values (0.5-0.7) for inferred or ambiguous information.
 - **Classify correctly:** People, teams, and personas → "roles" (not "entities"). Technical systems, services, modules → "entities".
+- **Entity naming — extract the subject, not the document:** Entity names should represent the actual system/product/framework being described, not the document itself. If the text says "Next.js is a React framework", the entity is "Next.js" — not "Next.js 官方文档" or "Next.js API 参考". Use names that work as standalone knowledge graph nodes, independent of any document context.
 - **Entity reference consistency (CRITICAL):** Every entity name referenced in relation.from, relation.to, behavior.subject, or any other cross-reference field MUST also appear in the "entities" array of the SAME paragraph (or a preceding paragraph in the same chunk). If an entity is mentioned for the first time in a relation, you MUST also extract it as an entity. This ensures no "dangling references" — every name used in relations has a corresponding entity declaration.
+- **Cross-atom reference consistency:** transitions[].from and transitions[].to values MUST exist in states[].values of the same entity. roles[].performs values MUST match names declared in behaviors[].name.
+- **Constraints vs rules distinction:** Use "constraints" for unconditional declarative mandates ('X must Y'). Use "rules" for conditional logic ('IF X THEN Y'). Do not mix them — a requirement with no condition is a constraint, a requirement triggered by a condition is a rule. Do NOT invent a rule for every constraint — only create a rule when the text explicitly states conditional logic.
+- **One statement, multiple atoms:** A single sentence can produce several atom types simultaneously. Do NOT force a choice — extract all that apply. Example: "system uptime must be ≥ 99.9%" → constraint (severity: must) + metric (threshold: "≥ 99.9%").
+- **Relation types:** Use standard relation types when possible: CONTAINS (parent→child composition), DEPENDS_ON (runtime dependency), IMPLEMENTS (code/component→spec realization), PRODUCES (process→output), TRIGGERS (event/process triggering), REFERENCES (weak cross-reference). Only invent a new type when none of these fit.
+- **Decisions:** Extract as "decisions" when the text records a deliberate choice between alternatives with rationale (e.g., "we chose X because Y", "after evaluating A/B/C, selected B"). Do not extract routine descriptions as decisions.
 - Respond in the same language as the input text (e.g., Chinese input → Chinese descriptions, English input → English descriptions).
 - JSON structure keys (tag, atom type names, field names) must always be in English.
 - Be thorough: extract ALL relevant atoms from each paragraph.
@@ -222321,14 +222408,15 @@ var ENTITY_RESOLUTION_SYSTEM_PROMPT = `You are an entity resolution assistant. Y
 - Prefer the LONGER, more descriptive name as the canonical name
 - Do NOT merge names that share a substring but refer to different things
 - When uncertain, do NOT merge — add to "ambiguous" instead
-- Chinese and English names for the same entity SHOULD be merged (e.g. "Vmok" → "Vmok 微模块框架")
-- Abbreviations should be merged with their full forms (e.g. "AGW" → "API Gateway")
+- Chinese and English names for the same entity SHOULD be merged (e.g. "Webpack" → "Webpack 打包工具")
+- Abbreviations should be merged with their full forms (e.g. "K8s" → "Kubernetes")
+- Document-descriptive names (containing "文档", "参考", "指南" etc.) should be merged to the actual product/system name. The entity is the product, not the document about it
 
 ## Task 2: Remove Noise
-- Remove names that are NOT meaningful named entities — they are generic words, actions, or descriptions
-- Examples of noise: common verbs/nouns (登录, 路由, 直连), generic technical terms (Env, query), action descriptions (Kill 3001 进程)
-- Examples of REAL entities to KEEP: product names (TTAstra, Gulux), tools (nvm, Rush), services (Op Main 服务), platforms (AGW 平台)
-- When uncertain, KEEP the name — only remove if clearly not a named entity
+- Apply the **identity test**: a real entity is something you can discuss independently ("What is X?", "How does X work?", "Who owns X?"). Names that fail this test — values, addresses, actions, generic descriptions — are noise.
+- Remove names that are NOT meaningful named entities: generic words, action descriptions, or things that are attributes/values rather than independent subjects
+- Examples of REAL entities to KEEP: product names (Next.js, Vite), tools (nvm, Turborepo), services (Nginx, Redis), platforms (Kubernetes, Docker) — these all pass the identity test
+- When uncertain, KEEP the name — only remove if it clearly fails the identity test
 
 ## Output
 Valid JSON only. No markdown fences, no explanation.`;
@@ -222571,7 +222659,7 @@ Each diagram paragraph is tagged with [P0], [P1], etc. You must classify the dia
    Relation schema: ${relationFields2}
 3. Extract action nodes as behaviors (what the process does at each step).
    Behavior schema: ${behaviorFields2}
-4. Extract diamond/decision nodes as decisions.
+4. Extract diamond/condition nodes: if it represents a deliberate choice with rationale → "decisions"; if it represents conditional branching logic (IF-THEN) → "rules".
    Decision schema: ${decisionFields}
 
 ### Sequence → entities + relations + behaviors + events
@@ -222604,9 +222692,9 @@ Each diagram paragraph is tagged with [P0], [P1], etc. You must classify the dia
    Relation schema: ${relationFields2}
 
 ### Architecture → entities + relations + constraints
-1. Extract each system/service/container/component as an entity.
+1. Extract each system/service/container/component as an entity. Use kind to indicate origin: "implementation" for internal systems/services, "external" for third-party dependencies (databases, cloud services, external APIs).
    Entity schema: ${entityFields2}
-2. Extract connections between components as relations.
+2. Extract connections between components as relations. Use standard types: CONTAINS (parent→child), DEPENDS_ON (runtime dependency), TRIGGERS (event/process triggering).
    Relation schema: ${relationFields2}
 3. Extract deployment constraints, technology choices.
    Constraint schema: ${constraintFields2}
@@ -222668,7 +222756,7 @@ Return ONLY a valid JSON object. No markdown fences, no explanation.`;
 }
 var DOC_DIAGRAM_ANNOTATION_SYSTEM_PROMPT = DIAGRAM_SYSTEM_PROMPT;
 // ../llm/src/chunking/markdownChunker.ts
-var DEFAULT_MAX_TOKENS2 = 4000;
+var DEFAULT_MAX_TOKENS2 = 3600;
 var DEFAULT_PARAGRAPH_MAX_TOKENS = 500;
 function estimateTokens(text2) {
   return Math.ceil(text2.length / 4);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@c4a/server-cli",
-  "version": "0.4.15-alpha.5",
+  "version": "0.4.15-alpha.7",
   "type": "module",
   "bin": {
     "c4a-server": "./index.js"

package/serve.js

@@ -194996,6 +194996,12 @@ class LlmServiceImpl {
       if (options?.systemPrompt) {
         callSettings.system = options.systemPrompt;
       }
+      if (options?.jsonMode) {
+        callSettings.providerOptions = {
+          openai: { responseFormat: { type: "json_object" } },
+          anthropic: { responseFormat: { type: "json_object" } }
+        };
+      }
       const result = await retry(() => generateText(callSettings), {
         shouldRetry: (error40) => isRetryableStatus(getHttpStatus(error40)) || isTimeoutError2(error40)
       });
@@ -195043,6 +195049,12 @@ class LlmServiceImpl {
       if (options?.systemPrompt) {
         callSettings.system = options.systemPrompt;
       }
+      if (options?.jsonMode) {
+        callSettings.providerOptions = {
+          openai: { responseFormat: { type: "json_object" } },
+          anthropic: { responseFormat: { type: "json_object" } }
+        };
+      }
       const result = await retry(() => Promise.resolve(streamText(callSettings)), {
         shouldRetry: (error40) => isRetryableStatus(getHttpStatus(error40)) || isTimeoutError2(error40)
       });
@@ -195861,6 +195873,7 @@ function parseExtractionOutput(raw5, schema2) {
     }
     parsed = normalizeFlatOutput(parsed);
     parsed = stripNulls(parsed);
+    parsed = patchAttributeDefaults(parsed);
     const result = schema2.safeParse(parsed);
     if (!result.success) {
       return { success: false, error: result.error };
@@ -195928,7 +195941,37 @@ function tryParseJson(raw5) {
   return repairAndParse(raw5);
 }
 function repairAndParse(raw5) {
-  return JSON.parse(jsonrepair(raw5));
+  try {
+    return JSON.parse(jsonrepair(raw5));
+  } catch {
+    const truncated = truncateToLastCompleteEntry(raw5);
+    if (truncated) {
+      return JSON.parse(jsonrepair(truncated));
+    }
+    throw new Error(`JSON repair failed for output of length ${raw5.length}`);
+  }
+}
+function truncateToLastCompleteEntry(raw5) {
+  let text2 = raw5.trim();
+  if (text2.startsWith("```")) {
+    const firstNewline = text2.indexOf(`
+`);
+    text2 = text2.slice(firstNewline + 1);
+  }
+  if (text2.endsWith("```")) {
+    text2 = text2.slice(0, -3);
+  }
+  const pattern = /\}\s*,\s*"P\d+"/g;
+  let lastMatch = null;
+  let m;
+  while ((m = pattern.exec(text2)) !== null) {
+    lastMatch = m;
+  }
+  if (!lastMatch)
+    return null;
+  const cutPos = lastMatch.index + 1;
+  const truncated = text2.slice(0, cutPos) + "}";
+  return truncated;
 }
 var PARAGRAPH_TAG_RE = /^P\d+$/;
 function normalizeFlatOutput(parsed) {
@@ -195961,6 +196004,43 @@ function stripNulls(value) {
   }
   return value;
 }
+function patchAttributeDefaults(value) {
+  if (!value || typeof value !== "object" || Array.isArray(value))
+    return value;
+  const obj = value;
+  if ("paragraphs" in obj && Array.isArray(obj.paragraphs)) {
+    return {
+      ...obj,
+      paragraphs: obj.paragraphs.map((p4) => {
+        if (!p4 || typeof p4 !== "object")
+          return p4;
+        const para = p4;
+        const atoms2 = para.atoms;
+        if (!atoms2 || typeof atoms2 !== "object")
+          return p4;
+        return { ...para, atoms: patchAttrsInAtoms(atoms2) };
+      })
+    };
+  }
+  return value;
+}
+function patchAttrsInAtoms(atoms2) {
+  const attrs = atoms2.attributes;
+  if (!Array.isArray(attrs))
+    return atoms2;
+  return {
+    ...atoms2,
+    attributes: attrs.map((attr) => {
+      if (!attr || typeof attr !== "object" || Array.isArray(attr))
+        return attr;
+      const a = attr;
+      if (typeof a.type !== "string" || a.type === "") {
+        return { ...a, type: "other" };
+      }
+      return attr;
+    })
+  };
+}
 function isRecord(value) {
   return !!value && typeof value === "object" && "key" in value && "value" in value && typeof value.key === "string";
 }
@@ -196217,20 +196297,20 @@ class GleaningExtractor {
 // ../llm/src/prompts/docAtomAnnotation.ts
 init_src();
 var DOC_ATOM_DEFS = [
-  ["entities", "Named things: systems, services, modules, documents, APIs (NOT people/teams — use roles for those)", entityAtomSchema],
+  ["entities", "Named things with independent identity — something you can ask questions about ('What does X do?', 'Who owns X?'). Examples: systems, services, modules, APIs, products. If it is a value, address, path, or configuration detail, it is an attribute of an entity, not an entity itself. (NOT people/teams — use roles for those). kind: implementation=internal systems/services, external=third-party dependencies, concept=abstract/not-yet-implemented", entityAtomSchema],
   ["relations", "Connections between entities", relationAtomSchema],
   ["behaviors", "Actions/operations: functions, API calls, user actions, workflows", behaviorAtomSchema],
   ["attributes", "Properties of entities", attributeAtomSchema],
   ["states", "Possible states of entities", stateAtomSchema],
-  ["rules", "Business/domain rules: validations, constraints expressed as logic", ruleAtomSchema],
+  ["rules", "Conditional business/domain logic: IF condition THEN consequence (e.g., 'IF user not authenticated THEN reject request')", ruleAtomSchema],
   ["transitions", "State changes: from→to triggered by events or guards", transitionAtomSchema],
   ["events", "Occurrences that trigger behaviors", eventAtomSchema],
   ["decisions", "Architectural or business decisions", decisionAtomSchema],
   ["metrics", "Measurable targets: SLA, throughput, error_rate, with thresholds", metricAtomSchema],
-  ["roles", "Actors: human roles, teams, personas that perform behaviors", roleAtomSchema],
-  ["constraints", "Hard/soft requirements", constraintAtomSchema],
+  ["roles", "Actors: human roles, teams, personas that perform behaviors. kind: human=individual role, team=group/department, persona=user archetype. System-triggered actions use entity relations, NOT roles", roleAtomSchema],
+  ["constraints", "Declarative requirements: 'X must/should/must-not Y' (e.g., 'passwords must be >= 8 chars'). Unlike rules, constraints have no IF-THEN condition — they are unconditional mandates or restrictions", constraintAtomSchema],
   ["comparisons", "Side-by-side evaluations", comparisonAtomSchema],
-  ["boundaries", "System/domain boundaries: what is included vs excluded", boundaryAtomSchema]
+  ["boundaries", "Explicit scope declarations: what is included vs excluded. Only extract when the text explicitly declares scope (e.g., 'this product covers X but NOT Y'). Implicit containment (A runs inside B) is expressed via entity relations, not boundaries", boundaryAtomSchema]
 ];
 function buildAtomTypesBlock() {
   return DOC_ATOM_DEFS.map(([name21, desc, schema2], i) => {
@@ -196260,7 +196340,8 @@ Return a single JSON object keyed by paragraph tags. Only include paragraphs tha
     "relations": [{ "from": "UserService", "to": "Database", "type": "DEPENDS_ON", "confidence": 0.9 }]
   },
   "P3": {
-    "rules": [{ "description": "User must be authenticated before access", "expression": "user.isAuthenticated == true", "confidence": 0.85 }]
+    "constraints": [{ "description": "User must be authenticated before access", "severity": "must", "confidence": 0.9 }],
+    "rules": [{ "description": "Reject request if user is not authenticated", "expression": "IF !user.isAuthenticated THEN reject", "confidence": 0.85 }]
   }
 }
 
@@ -196272,7 +196353,13 @@ Return a single JSON object keyed by paragraph tags. Only include paragraphs tha
 - **Enum fields MUST use ONLY the listed values.** For example, entity.kind must be one of "implementation"|"external"|"concept" — do NOT use values from other atom types (e.g., do NOT put "team" or "human" in entity.kind; those belong to roles.kind).
 - Every atom MUST include a "confidence" field (0.0-1.0) indicating how confident you are in the extraction. Use higher values (0.85-1.0) for explicitly stated facts and lower values (0.5-0.7) for inferred or ambiguous information.
 - **Classify correctly:** People, teams, and personas → "roles" (not "entities"). Technical systems, services, modules → "entities".
+- **Entity naming — extract the subject, not the document:** Entity names should represent the actual system/product/framework being described, not the document itself. If the text says "Next.js is a React framework", the entity is "Next.js" — not "Next.js 官方文档" or "Next.js API 参考". Use names that work as standalone knowledge graph nodes, independent of any document context.
 - **Entity reference consistency (CRITICAL):** Every entity name referenced in relation.from, relation.to, behavior.subject, or any other cross-reference field MUST also appear in the "entities" array of the SAME paragraph (or a preceding paragraph in the same chunk). If an entity is mentioned for the first time in a relation, you MUST also extract it as an entity. This ensures no "dangling references" — every name used in relations has a corresponding entity declaration.
+- **Cross-atom reference consistency:** transitions[].from and transitions[].to values MUST exist in states[].values of the same entity. roles[].performs values MUST match names declared in behaviors[].name.
+- **Constraints vs rules distinction:** Use "constraints" for unconditional declarative mandates ('X must Y'). Use "rules" for conditional logic ('IF X THEN Y'). Do not mix them — a requirement with no condition is a constraint, a requirement triggered by a condition is a rule. Do NOT invent a rule for every constraint — only create a rule when the text explicitly states conditional logic.
+- **One statement, multiple atoms:** A single sentence can produce several atom types simultaneously. Do NOT force a choice — extract all that apply. Example: "system uptime must be ≥ 99.9%" → constraint (severity: must) + metric (threshold: "≥ 99.9%").
+- **Relation types:** Use standard relation types when possible: CONTAINS (parent→child composition), DEPENDS_ON (runtime dependency), IMPLEMENTS (code/component→spec realization), PRODUCES (process→output), TRIGGERS (event/process triggering), REFERENCES (weak cross-reference). Only invent a new type when none of these fit.
+- **Decisions:** Extract as "decisions" when the text records a deliberate choice between alternatives with rationale (e.g., "we chose X because Y", "after evaluating A/B/C, selected B"). Do not extract routine descriptions as decisions.
 - Respond in the same language as the input text (e.g., Chinese input → Chinese descriptions, English input → English descriptions).
 - JSON structure keys (tag, atom type names, field names) must always be in English.
 - Be thorough: extract ALL relevant atoms from each paragraph.
@@ -196326,14 +196413,15 @@ var ENTITY_RESOLUTION_SYSTEM_PROMPT = `You are an entity resolution assistant. Y
 - Prefer the LONGER, more descriptive name as the canonical name
 - Do NOT merge names that share a substring but refer to different things
 - When uncertain, do NOT merge — add to "ambiguous" instead
-- Chinese and English names for the same entity SHOULD be merged (e.g. "Vmok" → "Vmok 微模块框架")
-- Abbreviations should be merged with their full forms (e.g. "AGW" → "API Gateway")
+- Chinese and English names for the same entity SHOULD be merged (e.g. "Webpack" → "Webpack 打包工具")
+- Abbreviations should be merged with their full forms (e.g. "K8s" → "Kubernetes")
+- Document-descriptive names (containing "文档", "参考", "指南" etc.) should be merged to the actual product/system name. The entity is the product, not the document about it
 
 ## Task 2: Remove Noise
-- Remove names that are NOT meaningful named entities — they are generic words, actions, or descriptions
-- Examples of noise: common verbs/nouns (登录, 路由, 直连), generic technical terms (Env, query), action descriptions (Kill 3001 进程)
-- Examples of REAL entities to KEEP: product names (TTAstra, Gulux), tools (nvm, Rush), services (Op Main 服务), platforms (AGW 平台)
-- When uncertain, KEEP the name — only remove if clearly not a named entity
+- Apply the **identity test**: a real entity is something you can discuss independently ("What is X?", "How does X work?", "Who owns X?"). Names that fail this test — values, addresses, actions, generic descriptions — are noise.
+- Remove names that are NOT meaningful named entities: generic words, action descriptions, or things that are attributes/values rather than independent subjects
+- Examples of REAL entities to KEEP: product names (Next.js, Vite), tools (nvm, Turborepo), services (Nginx, Redis), platforms (Kubernetes, Docker) — these all pass the identity test
+- When uncertain, KEEP the name — only remove if it clearly fails the identity test
 
 ## Output
 Valid JSON only. No markdown fences, no explanation.`;
@@ -196576,7 +196664,7 @@ Each diagram paragraph is tagged with [P0], [P1], etc. You must classify the dia
    Relation schema: ${relationFields2}
 3. Extract action nodes as behaviors (what the process does at each step).
    Behavior schema: ${behaviorFields2}
-4. Extract diamond/decision nodes as decisions.
+4. Extract diamond/condition nodes: if it represents a deliberate choice with rationale → "decisions"; if it represents conditional branching logic (IF-THEN) → "rules".
    Decision schema: ${decisionFields}
 
 ### Sequence → entities + relations + behaviors + events
@@ -196609,9 +196697,9 @@ Each diagram paragraph is tagged with [P0], [P1], etc. You must classify the dia
    Relation schema: ${relationFields2}
 
 ### Architecture → entities + relations + constraints
-1. Extract each system/service/container/component as an entity.
+1. Extract each system/service/container/component as an entity. Use kind to indicate origin: "implementation" for internal systems/services, "external" for third-party dependencies (databases, cloud services, external APIs).
    Entity schema: ${entityFields2}
-2. Extract connections between components as relations.
+2. Extract connections between components as relations. Use standard types: CONTAINS (parent→child), DEPENDS_ON (runtime dependency), TRIGGERS (event/process triggering).
    Relation schema: ${relationFields2}
 3. Extract deployment constraints, technology choices.
    Constraint schema: ${constraintFields2}
@@ -196673,7 +196761,7 @@ Return ONLY a valid JSON object. No markdown fences, no explanation.`;
 }
 var DOC_DIAGRAM_ANNOTATION_SYSTEM_PROMPT = DIAGRAM_SYSTEM_PROMPT;
 // ../llm/src/chunking/markdownChunker.ts
-var DEFAULT_MAX_TOKENS2 = 4000;
+var DEFAULT_MAX_TOKENS2 = 3600;
 var DEFAULT_PARAGRAPH_MAX_TOKENS = 500;
 function estimateTokens(text2) {
   return Math.ceil(text2.length / 4);
@@ -197598,7 +197686,8 @@ async function extractTableAtoms(chunk, sections, result, llmService) {
   const prompt = buildDocTableAnnotationPrompt(tableText);
   try {
     const res = await llmService.generateText(prompt, {
-      systemPrompt: DOC_TABLE_ANNOTATION_SYSTEM_PROMPT
+      systemPrompt: DOC_TABLE_ANNOTATION_SYSTEM_PROMPT,
+      jsonMode: true
     });
     const parsed = parseExtractionOutput(res.text, docChunkResultSchema);
     if (!parsed.success) {
@@ -197690,7 +197779,8 @@ async function extractDiagramAtoms(chunk, sections, result, llmService) {
   const prompt = buildDocDiagramAnnotationPrompt(diagramText);
   try {
     const res = await llmService.generateText(prompt, {
-      systemPrompt: DOC_DIAGRAM_ANNOTATION_SYSTEM_PROMPT
+      systemPrompt: DOC_DIAGRAM_ANNOTATION_SYSTEM_PROMPT,
+      jsonMode: true
     });
     const parsed = parseExtractionOutput(res.text, docChunkResultSchema);
     if (!parsed.success) {
@@ -198208,7 +198298,9 @@ ${trimmed}
 Continue the JSON output from the exact point of truncation. Output ONLY the remaining JSON text.`;
   try {
     const result = await llmService.generateText(continuationPrompt, {
-      systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT
+      systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT,
+      jsonMode: true,
+      maxTokens: 16384
     });
     const combined = trimmed + result.text.trim();
     JSON.parse(jsonrepair(combined));
@@ -198250,9 +198342,12 @@ async function processChunk(chunk, chunkIndex, llmService, sections, onStep) {
   const chunkText = injectParagraphTags(chunk, sections);
   let llmCalls = 0;
   let totalTokens = 0;
+  const DOC_INDEX_MAX_OUTPUT_TOKENS = 16384;
   const t1Prompt = buildDocAtomAnnotationPrompt(chunkText);
   const t1Result = await llmService.generateText(t1Prompt, {
-    systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT
+    systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT,
+    jsonMode: true,
+    maxTokens: DOC_INDEX_MAX_OUTPUT_TOKENS
   });
   llmCalls++;
   totalTokens += t1Result.usage.totalTokens;
@@ -198262,8 +198357,17 @@ async function processChunk(chunk, chunkIndex, llmService, sections, onStep) {
   onStep?.("T1 done", llmCalls, totalTokens);
   let parseResult = parseExtractionOutput(continued.text, docChunkResultSchema);
   if (!parseResult.success) {
+    const rawLen = continued.text.length;
     const preview = continued.text.slice(0, 500).replace(/\n/g, "\\n");
-    console.warn(`[docIndexer] chunk ${chunkIndex} T1 strict parse failed, attempting lenient. ` + `Error: ${parseResult.error.message.slice(0, 200)}. ` + `LLM output preview: ${preview}`);
+    const errMsg = parseResult.error.message.slice(0, 300);
+    const posMatch = errMsg.match(/position\s+(\d+)/);
+    const errContext = posMatch ? continued.text.slice(Math.max(0, +posMatch[1] - 100), +posMatch[1] + 100).replace(/\n/g, "\\n") : "";
+    console.warn(`[docIndexer] chunk ${chunkIndex} T1 strict parse failed, attempting lenient.
+` + `  Error: ${errMsg}
+` + `  Output length: ${rawLen} chars | finishReason: ${t1Result.finishReason}
+` + `  Preview (first 500): ${preview}
+` + (errContext ? `  Context around error position: ...${errContext}...
+` : ""));
     const lenient = tryLenientParse(continued.text, chunkIndex);
     if (lenient) {
       parseResult = { success: true, data: lenient };
@@ -198282,7 +198386,8 @@ async function processChunk(chunk, chunkIndex, llmService, sections, onStep) {
     extract: async (input) => {
       const prompt = buildDocGleaningPrompt(input.chunkText, input.previousResult);
       const result = await llmService.generateText(prompt, {
-        systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT
+        systemPrompt: DOC_ANNOTATION_SYSTEM_PROMPT,
+        jsonMode: true
       });
       llmCalls++;
       totalTokens += result.usage.totalTokens;
@@ -198512,6 +198617,17 @@ function ensureAtomConfidence(atoms2) {
     }
   }
 }
+function sortAtomsByConfidence(sections) {
+  for (const section of sections) {
+    for (const para of section.paragraphs) {
+      for (const atomList of Object.values(para.atoms)) {
+        if (!Array.isArray(atomList) || atomList.length < 2)
+          continue;
+        atomList.sort((a, b) => (b.confidence ?? 0) - (a.confidence ?? 0));
+      }
+    }
+  }
+}
 function countAtoms(sections) {
   const counts = {};
   for (const section of sections.values()) {
@@ -198548,7 +198664,8 @@ async function runEntityResolution(sections, entityNames, llmService, onProgress
       ...noiseCandidates.length > 0 ? { noiseCandidates } : {}
     });
     const result = await llmService.generateText(prompt, {
-      systemPrompt: ENTITY_RESOLUTION_SYSTEM_PROMPT
+      systemPrompt: ENTITY_RESOLUTION_SYSTEM_PROMPT,
+      jsonMode: true
     });
     let resolution;
     try {
@@ -198673,7 +198790,13 @@ async function indexDocument(input) {
     const paragraphs = sortedParaKeys.map((pk) => {
       const p4 = section.paragraphs.get(pk);
       return { text: p4.text, atoms: p4.atoms };
+    }).filter((p4) => {
+      const hasAtoms = Object.values(p4.atoms).some((arr) => Array.isArray(arr) && arr.length > 0);
+      const hasText = p4.text.replace(/[\s\p{P}\p{S}]/gu, "").length > 0;
+      return hasAtoms || hasText;
     });
+    if (paragraphs.length === 0)
+      continue;
     digestSections.push({
       heading: section.heading,
       level: section.level,
@@ -198704,6 +198827,7 @@ async function indexDocument(input) {
     totalLlmCalls += resolutionResult.llmCalls;
     totalTokens += resolutionResult.totalTokens;
   }
+  sortAtomsByConfidence(digestSections);
   const atomCounts = countAtoms(sectionsMap);
   const paragraphCount = digestSections.reduce((sum, s) => sum + s.paragraphs.length, 0);
   if (paragraphCount === 0) {

package/web/assets/ContentDetail-C0zfArPg.js

@@ -0,0 +1 @@
+import{c as o,u as m,a as c,j as e,b as d,C as h}from"./index-BKETuM1m.js";const p=[["path",{d:"m12 19-7-7 7-7",key:"1l729n"}],["path",{d:"M19 12H5",key:"x3x0zl"}]],f=o("arrow-left",p);function j({icon:a,title:t,badges:s,meta:l}){const i=m(),{t:r}=c();return e.jsxs("div",{children:[e.jsxs("div",{className:"flex justify-between items-center",children:[e.jsxs("div",{className:"flex min-w-0 items-center gap-2",children:[e.jsx("span",{className:"shrink-0",children:a}),e.jsx("h1",{className:"min-w-0 break-all font-mono text-xl font-bold leading-none text-c4a-text-primary",children:t}),s&&s.length>0&&e.jsx("div",{className:"flex items-center gap-1.5 ml-2",children:s.map((n,x)=>e.jsx("span",{children:n},x))})]}),e.jsxs("button",{onClick:()=>i(-1),className:"flex items-center gap-1 text-c4a-text-muted hover:text-c4a-text-primary cursor-pointer transition-colors text-xs font-mono shrink-0 ml-4",children:[e.jsx(f,{size:14}),r("common.back")]})]}),l&&l.length>0&&e.jsx("div",{className:"mt-2 flex flex-wrap items-baseline gap-4 text-xs",children:l.map(n=>e.jsxs("span",{className:"text-c4a-text-muted",children:[n.label,"：",e.jsx("span",{className:"text-c4a-text-secondary",children:n.value})]},n.label))})]})}function N(){const{t:a}=c(),{hashId:t}=d();if(!t)return e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:a("common.loading")})});const s=t.length>16?t.slice(0,16)+"…":t;return e.jsxs("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:[e.jsx(j,{icon:e.jsx("span",{children:"📄"}),title:`Content: ${s}`}),e.jsx(h,{hashId:t})]})}export{N as ContentDetail};

package/web/assets/ContentDetail-DZuam1p0.js

@@ -0,0 +1 @@
+import{c as o,u as m,a as c,j as e,b as d,C as h}from"./index-wBO4P6CB.js";const p=[["path",{d:"m12 19-7-7 7-7",key:"1l729n"}],["path",{d:"M19 12H5",key:"x3x0zl"}]],f=o("arrow-left",p);function j({icon:a,title:t,badges:s,meta:l}){const i=m(),{t:r}=c();return e.jsxs("div",{children:[e.jsxs("div",{className:"flex justify-between items-center",children:[e.jsxs("div",{className:"flex min-w-0 items-center gap-2",children:[e.jsx("span",{className:"shrink-0",children:a}),e.jsx("h1",{className:"min-w-0 break-all font-mono text-xl font-bold leading-none text-c4a-text-primary",children:t}),s&&s.length>0&&e.jsx("div",{className:"flex items-center gap-1.5 ml-2",children:s.map((n,x)=>e.jsx("span",{children:n},x))})]}),e.jsxs("button",{onClick:()=>i(-1),className:"flex items-center gap-1 text-c4a-text-muted hover:text-c4a-text-primary cursor-pointer transition-colors text-xs font-mono shrink-0 ml-4",children:[e.jsx(f,{size:14}),r("common.back")]})]}),l&&l.length>0&&e.jsx("div",{className:"mt-2 flex flex-wrap items-baseline gap-4 text-xs",children:l.map(n=>e.jsxs("span",{className:"text-c4a-text-muted",children:[n.label,"：",e.jsx("span",{className:"text-c4a-text-secondary",children:n.value})]},n.label))})]})}function N(){const{t:a}=c(),{hashId:t}=d();if(!t)return e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:a("common.loading")})});const s=t.length>16?t.slice(0,16)+"…":t;return e.jsxs("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:[e.jsx(j,{icon:e.jsx("span",{children:"📄"}),title:`Content: ${s}`}),e.jsx(h,{hashId:t})]})}export{N as ContentDetail};

package/web/assets/ContentDetail-qsl_01UW.js

@@ -0,0 +1 @@
+import{c as o,u as m,a as c,j as e,b as d,C as h}from"./index-CMVd9rlp.js";const p=[["path",{d:"m12 19-7-7 7-7",key:"1l729n"}],["path",{d:"M19 12H5",key:"x3x0zl"}]],f=o("arrow-left",p);function j({icon:a,title:t,badges:s,meta:l}){const i=m(),{t:r}=c();return e.jsxs("div",{children:[e.jsxs("div",{className:"flex justify-between items-center",children:[e.jsxs("div",{className:"flex min-w-0 items-center gap-2",children:[e.jsx("span",{className:"shrink-0",children:a}),e.jsx("h1",{className:"min-w-0 break-all font-mono text-xl font-bold leading-none text-c4a-text-primary",children:t}),s&&s.length>0&&e.jsx("div",{className:"flex items-center gap-1.5 ml-2",children:s.map((n,x)=>e.jsx("span",{children:n},x))})]}),e.jsxs("button",{onClick:()=>i(-1),className:"flex items-center gap-1 text-c4a-text-muted hover:text-c4a-text-primary cursor-pointer transition-colors text-xs font-mono shrink-0 ml-4",children:[e.jsx(f,{size:14}),r("common.back")]})]}),l&&l.length>0&&e.jsx("div",{className:"mt-2 flex flex-wrap items-baseline gap-4 text-xs",children:l.map(n=>e.jsxs("span",{className:"text-c4a-text-muted",children:[n.label,"：",e.jsx("span",{className:"text-c4a-text-secondary",children:n.value})]},n.label))})]})}function N(){const{t:a}=c(),{hashId:t}=d();if(!t)return e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:a("common.loading")})});const s=t.length>16?t.slice(0,16)+"…":t;return e.jsxs("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:[e.jsx(j,{icon:e.jsx("span",{children:"📄"}),title:`Content: ${s}`}),e.jsx(h,{hashId:t})]})}export{N as ContentDetail};

package/web/assets/EntityDetail-BafRMdUD.js

@@ -0,0 +1 @@
+import{a as n,b as a,j as e,E as i}from"./index-wBO4P6CB.js";function r(){const{t:s}=n(),{entityId:t}=a();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(i,{entityId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as EntityDetail};

package/web/assets/EntityDetail-C9k4cMVL.js

@@ -0,0 +1 @@
+import{a as n,b as a,j as e,E as i}from"./index-BKETuM1m.js";function r(){const{t:s}=n(),{entityId:t}=a();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(i,{entityId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as EntityDetail};

package/web/assets/EntityDetail-Dssp-tuW.js

@@ -0,0 +1 @@
+import{a as n,b as a,j as e,E as i}from"./index-CMVd9rlp.js";function r(){const{t:s}=n(),{entityId:t}=a();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(i,{entityId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as EntityDetail};

package/web/assets/RelationDetail-BK8C5waL.js

@@ -0,0 +1 @@
+import{a,b as n,j as e,R as l}from"./index-BKETuM1m.js";function r(){const{t:s}=a(),{relationId:t}=n();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(l,{relationId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as RelationDetail};

package/web/assets/RelationDetail-DH86ysel.js

@@ -0,0 +1 @@
+import{a,b as n,j as e,R as l}from"./index-CMVd9rlp.js";function r(){const{t:s}=a(),{relationId:t}=n();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(l,{relationId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as RelationDetail};

package/web/assets/RelationDetail-kDhcDsKg.js

@@ -0,0 +1 @@
+import{a,b as n,j as e,R as l}from"./index-wBO4P6CB.js";function r(){const{t:s}=a(),{relationId:t}=n();return t?e.jsx("div",{className:"flex flex-1 flex-col p-4 md:p-6",children:e.jsx(l,{relationId:t})}):e.jsx("div",{className:"flex flex-1 items-center justify-center p-8",children:e.jsx("p",{className:"text-sm text-c4a-text-muted",children:s("common.loading")})})}export{r as RelationDetail};