@sanqianx/claude-ai-workbench 0.13.2 → 0.13.4

package/src/backend/lib/ai-adapter.js

@@ -1,144 +1,144 @@
-// AI analyzer adapter interface (TASK-005 + TASK-008).
-// Each adapter implements the same shape so the orchestrator can swap implementations.
-//
-//   analyzeCommitBatch({ project, commits, contextPack }) -> { changes: [...] }
-//   validateOutput(output) -> { valid: boolean, errors: string[] }
-//
-// All adapters MUST return objects that pass validateOutput. The orchestrator never
-// writes drafts to disk unless the output validates.
-
-const fs = require('fs');
-const path = require('path');
-
-function isString(v) { return typeof v === 'string'; }
-function isNonEmptyString(v) { return isString(v) && v.length > 0; }
-function isStringArray(v) { return Array.isArray(v) && v.every(isString); }
-function isObject(v) { return v && typeof v === 'object' && !Array.isArray(v); }
-
-function normalizeKnowledgeLanguage(value) {
-  return value === 'en-US' ? 'en-US' : 'zh-CN';
-}
-
-function languageName(value) {
-  return normalizeKnowledgeLanguage(value) === 'en-US' ? 'English' : 'Simplified Chinese';
-}
-
-const changeEntrySchema = {
-  commit: isNonEmptyString,
-  classification: (v) => ['new-feature', 'existing-feature-update', 'bug-fix', 'refactor', 'infrastructure', 'test-only', 'docs-only'].includes(v),
-  developmentIntent: isNonEmptyString,
-  goalImpact: isNonEmptyString,
-  evidence: isStringArray,
-  proposedOps: (v) => Array.isArray(v) && v.every(op => isObject(op) && isNonEmptyString(op.op) && isNonEmptyString(op.path)),
-};
-
-function validateCommitBatchOutput(output) {
-  const errors = [];
-  if (!isObject(output)) { errors.push('output is not an object'); return { valid: false, errors }; }
-  if (!Array.isArray(output.changes)) { errors.push('missing or invalid: changes'); return { valid: false, errors }; }
-  for (let i = 0; i < output.changes.length; i++) {
-    const c = output.changes[i];
-    if (!isObject(c)) { errors.push(`change[${i}] is not an object`); continue; }
-    for (const [key, check] of Object.entries(changeEntrySchema)) {
-      if (!(key in c)) { errors.push(`change[${i}] missing: ${key}`); continue; }
-      if (!check(c[key])) errors.push(`change[${i}].${key} is invalid`);
-    }
-  }
-  return { valid: errors.length === 0, errors };
-}
-
-// ---- claude-code-agent: real LLM via Anthropic-compatible API ----
-//
-// This adapter calls an Anthropic Messages API (real Anthropic or any
-// compatible proxy such as api.minimaxi.com/anthropic). The model returns
-// JSON that the KB orchestrator validates against the existing schemas.
-//
-// The adapter is intentionally minimal: it does ONE call per commit batch
-// and asks the model to return a strict JSON object matching the schema.
-// Validation happens in `validateOutput`; if the model misbehaves, the run
-// is marked failed and the user can re-trigger.
-
-const { completeJson } = require('./llm-client');
-
-const COMMIT_SYSTEM = `You are a senior software architect reviewing a batch of git commits. For each commit, decide what useful knowledge should be recorded in the project's knowledge base. Respond with a single JSON object matching the schema in the user message. Do not invent file paths. Cite the provided evidence (context pack entries) whenever possible. Summarize the user's development intent; do not preserve raw prompts. Do not propose operations that would write outside changes/, modules/, README.md, GOAL.md, or ARCHITECTURE.md.`;
-
-function buildCommitUser({ project, commits, contextPack }) {
-  const outputLanguage = languageName(project && project.knowledgeLanguage);
-  const ctx = (contextPack && contextPack.entries || []).slice(0, 30).map(e => `- ${e.path}: ${(e.summary || '').slice(0, 200)}`).join('\n');
-  const list = (commits || []).map(c => `- full=${c.hash} | short=${c.short || c.hash.slice(0,7)} | ${c.date || ''} | ${c.author || ''} | ${c.subject || ''}`).join('\n');
-  return `Project slug: ${project.slug}
-Knowledge output language: ${normalizeKnowledgeLanguage(project.knowledgeLanguage)} (${outputLanguage})
-
-Pending commits (chronological order, oldest first). For each line, the "full=" value is the COMPLETE 40-character commit hash; you MUST copy that exact value into the JSON "commit" field:
-${list || '(no commits)'}
-
-Context pack excerpts:
-${ctx || '(no context pack available)'}
-
-Return a JSON object of the form:
-{
-  "changes": [
-    {
-      "commit": string,                // FULL 40-char commit hash copied from the input
-      "classification": "new-feature" | "existing-feature-update" | "bug-fix" | "refactor" | "infrastructure" | "test-only" | "docs-only",
-      "developmentIntent": string,     // concise summary of the user's request or prompt intent; never raw prompt logs
-      "goalImpact": string,            // one short sentence
-      "evidence": [string],            // paths cited from the context pack
-      "proposedOps": [{ "op": "create-file" | "append-section" | "update-section", "path": string, "fromTemplate"?: string }]
-    }
-  ]
-}
-
-Constraints:
-- All human-readable natural language string values MUST be written in ${outputLanguage}. Keep JSON keys, enum values, file paths, and commit hashes exactly as specified.
-- One entry per commit in the input list, in the same order.
-- "commit" MUST be the full 40-character hash from the input (the value after "full="), NOT the short hash.
-- proposedOps[].path MUST be a relative path under changes/ or modules/, or exactly README.md, GOAL.md, or ARCHITECTURE.md. Goal and architecture edits require human review.
-- Do not include any text outside the JSON object.`;
-}
-
-const claudeCodeAgent = {
-  id: 'claude-code-agent',
-  name: 'Claude Code Agent (Anthropic-compatible)',
-  description: 'Real LLM through an Anthropic-compatible Messages API. AI profiles can reuse this implementation with different model providers.',
-  async analyzeCommitBatch({ project, commits, contextPack }) {
-    const user = buildCommitUser({ project, commits, contextPack });
-    const r = await completeJson({ system: COMMIT_SYSTEM, user, maxTokens: 4096, profileId: project && project.aiProfileId });
-    if (!r.parsed) {
-      console.error(`[claude-code-agent] LLM did not return valid JSON. text head: ${r.text.slice(0, 500)} ... text tail: ${r.text.slice(-500)}`);
-      const e = new Error(`LLM returned no parseable JSON (${r.parseError && r.parseError.message || 'no output'})`);
-      e.llmText = r.text;
-      throw e;
-    }
-    // Sanity log: how many changes did the LLM return, and what do the commits look like?
-    const changes = (r.parsed && r.parsed.changes) || [];
-    console.log(`[claude-code-agent] parsed ${changes.length} changes; first change: ${JSON.stringify(changes[0]).slice(0, 300)}`);
-    return r.parsed;
-  },
-  validateOutput(output) {
-    return validateCommitBatchOutput(output);
-  },
-};
-
-const ADAPTERS = {
-  'claude-code-agent': claudeCodeAgent,
-};
-
-function getAdapter(id) {
-  return ADAPTERS[id] || null;
-}
-
-function listAdapters() {
-  return Object.values(ADAPTERS).map(a => ({
-    id: a.id,
-    name: a.name,
-    description: a.description || null,
-  }));
-}
-
-module.exports = {
-  ADAPTERS,
-  getAdapter,
-  listAdapters,
-  validateCommitBatchOutput,
-};
+// AI analyzer adapter interface (TASK-005 + TASK-008).
+// Each adapter implements the same shape so the orchestrator can swap implementations.
+//
+//   analyzeCommitBatch({ project, commits, contextPack }) -> { changes: [...] }
+//   validateOutput(output) -> { valid: boolean, errors: string[] }
+//
+// All adapters MUST return objects that pass validateOutput. The orchestrator never
+// writes drafts to disk unless the output validates.
+
+const fs = require('fs');
+const path = require('path');
+
+function isString(v) { return typeof v === 'string'; }
+function isNonEmptyString(v) { return isString(v) && v.length > 0; }
+function isStringArray(v) { return Array.isArray(v) && v.every(isString); }
+function isObject(v) { return v && typeof v === 'object' && !Array.isArray(v); }
+
+function normalizeKnowledgeLanguage(value) {
+  return value === 'en-US' ? 'en-US' : 'zh-CN';
+}
+
+function languageName(value) {
+  return normalizeKnowledgeLanguage(value) === 'en-US' ? 'English' : 'Simplified Chinese';
+}
+
+const changeEntrySchema = {
+  commit: isNonEmptyString,
+  classification: (v) => ['new-feature', 'existing-feature-update', 'bug-fix', 'refactor', 'infrastructure', 'test-only', 'docs-only'].includes(v),
+  developmentIntent: isNonEmptyString,
+  goalImpact: isNonEmptyString,
+  evidence: isStringArray,
+  proposedOps: (v) => Array.isArray(v) && v.every(op => isObject(op) && isNonEmptyString(op.op) && isNonEmptyString(op.path)),
+};
+
+function validateCommitBatchOutput(output) {
+  const errors = [];
+  if (!isObject(output)) { errors.push('output is not an object'); return { valid: false, errors }; }
+  if (!Array.isArray(output.changes)) { errors.push('missing or invalid: changes'); return { valid: false, errors }; }
+  for (let i = 0; i < output.changes.length; i++) {
+    const c = output.changes[i];
+    if (!isObject(c)) { errors.push(`change[${i}] is not an object`); continue; }
+    for (const [key, check] of Object.entries(changeEntrySchema)) {
+      if (!(key in c)) { errors.push(`change[${i}] missing: ${key}`); continue; }
+      if (!check(c[key])) errors.push(`change[${i}].${key} is invalid`);
+    }
+  }
+  return { valid: errors.length === 0, errors };
+}
+
+// ---- claude-code-agent: real LLM via Anthropic-compatible API ----
+//
+// This adapter calls an Anthropic Messages API (real Anthropic or any
+// compatible proxy such as api.minimaxi.com/anthropic). The model returns
+// JSON that the KB orchestrator validates against the existing schemas.
+//
+// The adapter is intentionally minimal: it does ONE call per commit batch
+// and asks the model to return a strict JSON object matching the schema.
+// Validation happens in `validateOutput`; if the model misbehaves, the run
+// is marked failed and the user can re-trigger.
+
+const { completeJson } = require('./llm-client');
+
+const COMMIT_SYSTEM = `You are a senior software architect reviewing a batch of git commits. For each commit, decide what useful knowledge should be recorded in the project's knowledge base. Respond with a single JSON object matching the schema in the user message. Do not invent file paths. Cite the provided evidence (context pack entries) whenever possible. Summarize the user's development intent; do not preserve raw prompts. Do not propose operations that would write outside changes/, modules/, README.md, GOAL.md, or ARCHITECTURE.md.`;
+
+function buildCommitUser({ project, commits, contextPack }) {
+  const outputLanguage = languageName(project && project.knowledgeLanguage);
+  const ctx = (contextPack && contextPack.entries || []).slice(0, 30).map(e => `- ${e.path}: ${(e.summary || '').slice(0, 200)}`).join('\n');
+  const list = (commits || []).map(c => `- full=${c.hash} | short=${c.short || c.hash.slice(0,7)} | ${c.date || ''} | ${c.author || ''} | ${c.subject || ''}`).join('\n');
+  return `Project slug: ${project.slug}
+Knowledge output language: ${normalizeKnowledgeLanguage(project.knowledgeLanguage)} (${outputLanguage})
+
+Pending commits (chronological order, oldest first). For each line, the "full=" value is the COMPLETE 40-character commit hash; you MUST copy that exact value into the JSON "commit" field:
+${list || '(no commits)'}
+
+Context pack excerpts:
+${ctx || '(no context pack available)'}
+
+Return a JSON object of the form:
+{
+  "changes": [
+    {
+      "commit": string,                // FULL 40-char commit hash copied from the input
+      "classification": "new-feature" | "existing-feature-update" | "bug-fix" | "refactor" | "infrastructure" | "test-only" | "docs-only",
+      "developmentIntent": string,     // concise summary of the user's request or prompt intent; never raw prompt logs
+      "goalImpact": string,            // one short sentence
+      "evidence": [string],            // paths cited from the context pack
+      "proposedOps": [{ "op": "create-file" | "append-section" | "update-section", "path": string, "fromTemplate"?: string }]
+    }
+  ]
+}
+
+Constraints:
+- All human-readable natural language string values MUST be written in ${outputLanguage}. Keep JSON keys, enum values, file paths, and commit hashes exactly as specified.
+- One entry per commit in the input list, in the same order.
+- "commit" MUST be the full 40-character hash from the input (the value after "full="), NOT the short hash.
+- proposedOps[].path MUST be a relative path under changes/ or modules/, or exactly README.md, GOAL.md, or ARCHITECTURE.md. Goal and architecture edits require human review.
+- Do not include any text outside the JSON object.`;
+}
+
+const claudeCodeAgent = {
+  id: 'claude-code-agent',
+  name: 'Claude Code Agent (Anthropic-compatible)',
+  description: 'Real LLM through an Anthropic-compatible Messages API. AI profiles can reuse this implementation with different model providers.',
+  async analyzeCommitBatch({ project, commits, contextPack }) {
+    const user = buildCommitUser({ project, commits, contextPack });
+    const r = await completeJson({ system: COMMIT_SYSTEM, user, maxTokens: 4096, profileId: project && project.aiProfileId });
+    if (!r.parsed) {
+      console.error(`[claude-code-agent] LLM did not return valid JSON. text head: ${r.text.slice(0, 500)} ... text tail: ${r.text.slice(-500)}`);
+      const e = new Error(`LLM returned no parseable JSON (${r.parseError && r.parseError.message || 'no output'})`);
+      e.llmText = r.text;
+      throw e;
+    }
+    // Sanity log: how many changes did the LLM return, and what do the commits look like?
+    const changes = (r.parsed && r.parsed.changes) || [];
+    console.log(`[claude-code-agent] parsed ${changes.length} changes; first change: ${JSON.stringify(changes[0]).slice(0, 300)}`);
+    return r.parsed;
+  },
+  validateOutput(output) {
+    return validateCommitBatchOutput(output);
+  },
+};
+
+const ADAPTERS = {
+  'claude-code-agent': claudeCodeAgent,
+};
+
+function getAdapter(id) {
+  return ADAPTERS[id] || null;
+}
+
+function listAdapters() {
+  return Object.values(ADAPTERS).map(a => ({
+    id: a.id,
+    name: a.name,
+    description: a.description || null,
+  }));
+}
+
+module.exports = {
+  ADAPTERS,
+  getAdapter,
+  listAdapters,
+  validateCommitBatchOutput,
+};