edsger 0.55.4 → 0.56.1

package/dist/phases/quality-benchmark/prompts.d.ts

@@ -0,0 +1,31 @@
+/**
+ * Prompt construction for the quality-benchmark SDK session.
+ *
+ * System prompt = the full rubric document (rubric.md) appended to
+ * Claude Code's default preset. The rubric is the single source of
+ * truth for: the 6-phase pipeline, the 8 dimensions, scoring anchors,
+ * N/A and Unmeasured rules, evidence/recommendation formats, the tool
+ * catalog (which the LLM cross-references against the MCP tools we
+ * expose), and the mandatory JSON output schema.
+ *
+ * User prompt = run-specific task context (product, repo path, options).
+ */
+export declare function createQualityBenchmarkSystemPrompt(): string;
+export interface QualityBenchmarkUserPromptParams {
+    productName: string;
+    productId: string;
+    repoRoot: string;
+    branch?: string;
+    commitSha: string;
+    runId: string;
+    /** Did the user grant consent for on-demand tool installation? */
+    installEnabled: boolean;
+    /** Optional caller hint (e.g. "first run", "scheduled rerun"). */
+    reason?: string;
+}
+export declare function createQualityBenchmarkUserPrompt(params: QualityBenchmarkUserPromptParams): string;
+/**
+ * Extract the final JSON report from the LLM's response. The rubric
+ * requires exactly one ```json block at the end of the response.
+ */
+export declare function extractReportJson(text: string): unknown | null;

package/dist/phases/quality-benchmark/prompts.js

@@ -0,0 +1,154 @@
+/**
+ * Prompt construction for the quality-benchmark SDK session.
+ *
+ * System prompt = the full rubric document (rubric.md) appended to
+ * Claude Code's default preset. The rubric is the single source of
+ * truth for: the 6-phase pipeline, the 8 dimensions, scoring anchors,
+ * N/A and Unmeasured rules, evidence/recommendation formats, the tool
+ * catalog (which the LLM cross-references against the MCP tools we
+ * expose), and the mandatory JSON output schema.
+ *
+ * User prompt = run-specific task context (product, repo path, options).
+ */
+import { readFileSync } from 'node:fs';
+import { dirname, join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { RUBRIC_VERSION } from './types.js';
+const moduleDir = dirname(fileURLToPath(import.meta.url));
+// rubric.md is shipped alongside the compiled prompts.ts (copied by the
+// build step) and lives next to it in src/. Try both locations so dev
+// (tsx / vitest) and prod (dist) both work.
+function loadRubric() {
+    const candidates = [
+        join(moduleDir, 'rubric.md'),
+        join(moduleDir, '..', '..', '..', 'src', 'phases', 'quality-benchmark', 'rubric.md'),
+    ];
+    for (const path of candidates) {
+        try {
+            return readFileSync(path, 'utf8');
+        }
+        catch {
+            // try next
+        }
+    }
+    throw new Error('quality-benchmark: rubric.md not found next to prompts.ts. ' +
+        'Ensure the build step copies it into dist/.');
+}
+const RUBRIC = loadRubric();
+// ---------------------------------------------------------------------------
+// System prompt
+// ---------------------------------------------------------------------------
+export function createQualityBenchmarkSystemPrompt() {
+    return [
+        `You are running the Edsger Quality Benchmark, rubric version ${RUBRIC_VERSION}.`,
+        '',
+        'Your task is to produce one JSON report at the end of your response that',
+        'matches the schema in the rubric. Drive the 6-phase pipeline using the',
+        'MCP tools exposed under the "quality-benchmark" server, NOT the raw Bash',
+        'tool, for anything that runs a code-quality tool. Bash is fine for git',
+        'commands, reading manifests, and other repo inspection.',
+        '',
+        'MCP tools available to you (all under the `mcp__quality-benchmark__` prefix):',
+        '  list_applicable_tools — returns the catalog subset for the detected repo',
+        '  probe_tool            — check whether a tool is installed and reachable',
+        '  install_tool          — install a missing tool (user-space, no sudo)',
+        '  run_tool              — execute a tool and receive its parsed summary',
+        '  verify_finding        — confirm a file:line is real before citing it',
+        '  record_progress       — surface a status message to the UI',
+        '',
+        'IMPORTANT discipline rules (also in the rubric — repeated for emphasis):',
+        '  1. Never run a quality tool through Bash. Always use run_tool.',
+        '     The MCP server captures, parses, and saves outputs deterministically.',
+        '     Using Bash bypasses parsing and verification.',
+        '  2. Always probe_tool before run_tool. If unavailable, decide whether to',
+        '     install_tool (if install_command is non-null) or accept "unmeasured".',
+        '  3. Every evidence entry MUST cite a real file:line. Call verify_finding',
+        '     before including any finding sourced from llm_judgment. Drop ones',
+        '     that fail verification.',
+        '  4. End your response with exactly ONE ```json code block matching the',
+        "     rubric's output schema. No prose after it.",
+        '',
+        '--- BEGIN RUBRIC ---',
+        RUBRIC,
+        '--- END RUBRIC ---',
+    ].join('\n');
+}
+export function createQualityBenchmarkUserPrompt(params) {
+    const lines = [];
+    lines.push(`Quality benchmark run for product **${params.productName}** (id: ${params.productId}).`);
+    lines.push('');
+    lines.push(`Repo: \`${params.repoRoot}\``);
+    if (params.branch) {
+        lines.push(`Branch: \`${params.branch}\``);
+    }
+    lines.push(`Commit: \`${params.commitSha}\``);
+    lines.push(`Run id: \`${params.runId}\``);
+    lines.push(`Install consent: ${params.installEnabled ? 'GRANTED — you may call install_tool for missing tools' : 'NOT GRANTED — do not call install_tool; missing tools must be marked unmeasured'}`);
+    if (params.reason) {
+        lines.push(`Context: ${params.reason}`);
+    }
+    lines.push('');
+    lines.push('Run the 6-phase pipeline now. Use record_progress at each phase');
+    lines.push('boundary so the user sees what you are doing. End with the');
+    lines.push('mandatory JSON report.');
+    return lines.join('\n');
+}
+// ---------------------------------------------------------------------------
+// Output extraction
+// ---------------------------------------------------------------------------
+/**
+ * Extract the final JSON report from the LLM's response. The rubric
+ * requires exactly one ```json block at the end of the response.
+ */
+export function extractReportJson(text) {
+    if (!text) {
+        return null;
+    }
+    // Scan for fenced ```json blocks; take the last one (rubric says final).
+    const blocks = [];
+    const re = /```json\s*\n([\s\S]*?)```/g;
+    let m;
+    while ((m = re.exec(text)) !== null) {
+        blocks.push(m[1].trim());
+    }
+    if (blocks.length === 0) {
+        return tryParseLooseJson(text);
+    }
+    const last = blocks[blocks.length - 1];
+    try {
+        return JSON.parse(last);
+    }
+    catch {
+        return null;
+    }
+}
+/**
+ * Fallback: scan the end of the text for the last balanced { ... } block
+ * (handles the case where the LLM omitted the fence markers).
+ */
+function tryParseLooseJson(text) {
+    const trimmed = text.trim();
+    let depth = 0;
+    let end = -1;
+    for (let i = trimmed.length - 1; i >= 0; i--) {
+        const ch = trimmed[i];
+        if (ch === '}') {
+            if (end === -1) {
+                end = i;
+            }
+            depth++;
+        }
+        else if (ch === '{') {
+            depth--;
+            if (depth === 0) {
+                try {
+                    return JSON.parse(trimmed.slice(i, end + 1));
+                }
+                catch {
+                    return null;
+                }
+            }
+        }
+    }
+    return null;
+}