@mmerterden/multi-agent-pipeline 8.6.1 → 10.0.6

package/pipeline/scripts/check-md-links.mjs

@@ -0,0 +1,59 @@
+#!/usr/bin/env node
+// Internal markdown link checker for the user-facing docs surface.
+//
+// Scope: root *.md + docs/**/*.md + examples/**/*.md. Verifies that every
+// relative link target ([text](./path) or [text](path#anchor)) resolves to
+// an existing file or directory. External (http/https/mailto) links and
+// pure #anchors are skipped - this is a drift gate for file moves/renames,
+// not a network crawler.
+
+import { readFileSync, existsSync, readdirSync, statSync } from "node:fs";
+import { dirname, join, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..", "..");
+
+function mdFiles(dir, acc = []) {
+  for (const entry of readdirSync(dir)) {
+    if (entry === "node_modules" || entry.startsWith(".")) continue;
+    const full = join(dir, entry);
+    if (statSync(full).isDirectory()) mdFiles(full, acc);
+    else if (entry.endsWith(".md")) acc.push(full);
+  }
+  return acc;
+}
+
+const targets = [
+  ...readdirSync(repoRoot).filter((f) => f.endsWith(".md")).map((f) => join(repoRoot, f)),
+  ...["docs", "examples"].filter((d) => existsSync(join(repoRoot, d))).flatMap((d) => mdFiles(join(repoRoot, d))),
+  // CHANGELOG-archive.md is frozen history - it legitimately references docs
+  // that were deleted in later releases.
+].filter((f) => !f.endsWith("CHANGELOG-archive.md"));
+
+const LINK_RE = /\[[^\]]*\]\(([^)\s]+)(?:\s+"[^"]*")?\)/g;
+const errors = [];
+
+for (const file of targets) {
+  // Drop fenced blocks and inline code spans - links inside code are
+  // illustrative placeholders, not navigation.
+  const text = readFileSync(file, "utf-8")
+    .replace(/```[\s\S]*?```/g, "")
+    .replace(/`[^`\n]*`/g, "");
+  for (const match of text.matchAll(LINK_RE)) {
+    const raw = match[1];
+    if (/^(https?:|mailto:|#)/.test(raw)) continue;
+    const target = raw.split("#")[0];
+    if (target === "") continue;
+    const resolved = resolve(dirname(file), decodeURIComponent(target));
+    if (!existsSync(resolved)) {
+      errors.push(`${file.slice(repoRoot.length + 1)}: broken link -> ${raw}`);
+    }
+  }
+}
+
+if (errors.length > 0) {
+  console.error(`check-md-links: ${errors.length} broken internal link(s)`);
+  for (const e of errors) console.error("  " + e);
+  process.exit(1);
+}
+console.log(`check-md-links: ${targets.length} files OK`);

package/pipeline/scripts/classify-plan-safety.mjs

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-// classify-plan-safety.mjs — v7.0.G
+// classify-plan-safety.mjs  -  v7.0.G
 //
 // Heuristic safety classifier for a Phase 2 plan. Autopilot's "zero user
 // interaction" contract is fine for small, predictable tasks but dangerous
 // when a plan touches the security path, deletes files without paired
 // tests, or sprawls across many files. This script inspects the plan and
-// returns a JSON verdict — the orchestrator reads it and decides whether
+// returns a JSON verdict  -  the orchestrator reads it and decides whether
 // to inject a one-time pause even in autopilot mode.
 //
 // Usage:
@@ -17,21 +17,21 @@
 //     "score": 0-100,
 //     "recommendPause": true|false,
 //     "reasons": [ {rule, weight, detail} ],
-//     "summary": "…"
+//     "summary": "..."
 //   }
 //
 // Rules + weights (v1, tune via telemetry later):
-//   - file-count-high       30   — plan.tasks[].files[] total > 15
-//   - file-count-medium     10   — total 8..15
-//   - destructive-verb      25   — task title or approach mentions
+//   - file-count-high       30    -  plan.tasks[].files[] total > 15
+//   - file-count-medium     10    -  total 8..15
+//   - destructive-verb      25    -  task title or approach mentions
 //                                    delete|drop|remove|migrate|rename|purge|truncate
-//   - security-path         35   — any planned file matches
+//   - security-path         35    -  any planned file matches
 //                                    auth, keychain, credentials, secret, security, token,
 //                                    cert, encrypt, signing, privacy
-//   - delete-without-test   30   — refactor/code task that deletes files
+//   - delete-without-test   30    -  refactor/code task that deletes files
 //                                    AND no sibling `type: test` task touches the same area
-//   - schema-migration      25   — plan touches schemas/ OR migrations/ OR *.sql
-//   - infrastructure        20   — touches Dockerfile, CI yaml, Gradle wrapper,
+//   - schema-migration      25    -  plan touches schemas/ OR migrations/ OR *.sql
+//   - infrastructure        20    -  touches Dockerfile, CI yaml, Gradle wrapper,
 //                                    package.json resolutions / engines, install.js
 //
 // recommendPause = true when score >= 50 (tuned: any two medium signals or
@@ -94,20 +94,26 @@ if (destructiveHits.length > 0) {
 }
 
 // --- security path ---
-const securityRe = /(\/|^)(auth|keychain|credentials?|secret|security|tokens?|cert|encrypt|signing|privacy)(\/|$|\.)/i;
+// Match when a path segment STARTS WITH a security keyword (prefix match,
+// mirroring diff-risk SECURITY_GLOBS). The leading (\/|^) anchor keeps it at a
+// segment boundary  -  so "AuthManager.swift" / "authentication/" match, but
+// mid-word hits like "Reauthorize" do not. A trailing anchor here would miss
+// the common "AuthManager"/"TokenStore" suffix-naming and let sensitive files
+// slip past the autopilot pause heuristic.
+const securityRe = /(\/|^)(auth|keychain|credentials?|secret|security|tokens?|cert|encrypt|signing|privacy)/i;
 const securityHits = [];
 for (const f of allFiles) if (securityRe.test(f)) securityHits.push(f);
 if (securityHits.length > 0) {
   reasons.push({
     rule: "security-path",
     weight: 35,
-    detail: `${securityHits.length} security-path file(s): ${securityHits.slice(0, 3).join(", ")}${securityHits.length > 3 ? "…" : ""}`,
+    detail: `${securityHits.length} security-path file(s): ${securityHits.slice(0, 3).join(", ")}${securityHits.length > 3 ? "..." : ""}`,
   });
 }
 
 // --- delete-without-test ---
 // Heuristic: a task whose type is "code" or "refactor" and title includes a
-// destructive verb — but no sibling task (any type=test) touches the same dir.
+// destructive verb  -  but no sibling task (any type=test) touches the same dir.
 const testFileDirs = new Set();
 for (const t of tasks) {
   if (t.type === "test") {
@@ -144,7 +150,7 @@ if (schemaHits.length > 0) {
   reasons.push({
     rule: "schema-migration",
     weight: 25,
-    detail: `${schemaHits.length} schema/migration file(s): ${schemaHits.slice(0, 3).join(", ")}${schemaHits.length > 3 ? "…" : ""}`,
+    detail: `${schemaHits.length} schema/migration file(s): ${schemaHits.slice(0, 3).join(", ")}${schemaHits.length > 3 ? "..." : ""}`,
   });
 }
 
@@ -156,7 +162,7 @@ if (infraHits.length > 0) {
   reasons.push({
     rule: "infrastructure",
     weight: 20,
-    detail: `${infraHits.length} infra file(s): ${infraHits.slice(0, 3).join(", ")}${infraHits.length > 3 ? "…" : ""}`,
+    detail: `${infraHits.length} infra file(s): ${infraHits.slice(0, 3).join(", ")}${infraHits.length > 3 ? "..." : ""}`,
   });
 }
 
@@ -166,11 +172,11 @@ const recommendPause = score >= 50;
 
 let summary;
 if (recommendPause) {
-  summary = `High-risk plan (score ${score}) — recommend manual approval even in autopilot`;
+  summary = `High-risk plan (score ${score})  -  recommend manual approval even in autopilot`;
 } else if (score > 0) {
-  summary = `Moderate plan (score ${score}) — autopilot safe`;
+  summary = `Moderate plan (score ${score})  -  autopilot safe`;
 } else {
-  summary = `Low-risk plan — autopilot safe`;
+  summary = `Low-risk plan  -  autopilot safe`;
 }
 
 process.stdout.write(JSON.stringify({ score, recommendPause, reasons, summary }, null, 2) + "\n");

package/pipeline/scripts/cost-budget-check.mjs

@@ -0,0 +1,160 @@
+#!/usr/bin/env node
+
+/**
+ * @file cost-budget-check.mjs  -  v9.2.0
+ *
+ * Proactive per-task cost ceiling. Reads the phase-tracker token accumulators
+ * for a run, prices them via cost-table.json at a (conservative by default)
+ * model rate, and compares the estimate against the user's budget. Meant to be
+ * called after each phase so a runaway run is caught live rather than only in
+ * the end-of-run cost summary.
+ *
+ * Opaque / spiralling cost is the single most-cited reason users abandon an AI
+ * coding tool; this turns "will this burn my tokens?" into a visible number.
+ *
+ * Inputs (CLI overrides prefs):
+ *   --task-id <id>        Locate .worktrees/<id>/phase-tracker.json
+ *   --tracker  <path>     Explicit tracker JSON (skips lookup)
+ *   --prefs    <path>     Preferences JSON; reads .global.costBudget
+ *   --max-usd  <n>        Hard ceiling (USD)
+ *   --warn-pct <n>        Warn threshold as percent of max-usd (1..99)
+ *   --pricing-model <m>   opus | sonnet | haiku  (rate to price tokens at)
+ *   --enabled             Force-enable even if prefs say otherwise
+ *
+ * Output: one human-readable status line on stdout.
+ * Exit codes:
+ *   0   ok (under warn threshold), disabled, or no tracker data yet
+ *   10  warn  (>= warnPct% of maxUsd, still under the ceiling)
+ *   11  exceeded (> maxUsd)
+ *   2   usage / environment error
+ *
+ * @module pipeline/scripts/cost-budget-check
+ */
+
+import { readFileSync, existsSync } from "fs";
+import { dirname, join } from "path";
+import { fileURLToPath } from "url";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+
+function die(msg) {
+  process.stderr.write(`cost-budget-check: ${msg}\n`);
+  process.exit(2);
+}
+
+const argv = process.argv.slice(2);
+const flags = {};
+for (let i = 0; i < argv.length; i++) {
+  const a = argv[i];
+  if (!a.startsWith("--")) continue;
+  const key = a.slice(2);
+  const next = argv[i + 1];
+  if (next !== undefined && !next.startsWith("--")) {
+    flags[key] = next;
+    i++;
+  } else {
+    flags[key] = true;
+  }
+}
+
+if (flags.help || flags.h) {
+  process.stdout.write(
+    "Usage: cost-budget-check.mjs [--task-id <id> | --tracker <path>] [--prefs <path>]\n" +
+      "       [--max-usd <n>] [--warn-pct <n>] [--pricing-model opus|sonnet|haiku] [--enabled]\n",
+  );
+  process.exit(0);
+}
+
+// --- resolve config: prefs first, CLI overrides -----------------------------
+const cfg = { enabled: false, maxUsd: 5.0, warnPct: 80, onExceed: "warn", pricingModel: "opus" };
+
+if (flags.prefs) {
+  if (!existsSync(flags.prefs)) die(`prefs file not found: ${flags.prefs}`);
+  let prefs;
+  try {
+    prefs = JSON.parse(readFileSync(flags.prefs, "utf-8"));
+  } catch (e) {
+    die(`prefs not valid JSON: ${e.message}`);
+  }
+  const b = prefs?.global?.costBudget;
+  if (b && typeof b === "object") Object.assign(cfg, b);
+}
+
+if (flags["max-usd"] !== undefined) cfg.maxUsd = Number(flags["max-usd"]);
+if (flags["warn-pct"] !== undefined) cfg.warnPct = Number(flags["warn-pct"]);
+if (flags["pricing-model"] !== undefined) cfg.pricingModel = String(flags["pricing-model"]);
+if (flags.enabled) cfg.enabled = true;
+
+if (!cfg.enabled) {
+  // Silent no-op so the pipeline behaves exactly as before when unconfigured.
+  process.exit(0);
+}
+
+if (!(cfg.maxUsd > 0)) die(`maxUsd must be a positive number (got ${cfg.maxUsd})`);
+if (!(cfg.warnPct >= 1 && cfg.warnPct <= 99)) die(`warnPct must be 1..99 (got ${cfg.warnPct})`);
+
+// --- locate tracker ---------------------------------------------------------
+function resolveTracker() {
+  if (flags.tracker) {
+    if (!existsSync(flags.tracker)) die(`tracker not found: ${flags.tracker}`);
+    return flags.tracker;
+  }
+  const id = flags["task-id"];
+  if (!id) die("either --tracker or --task-id is required");
+  const bare = String(id).replace(/^#/, "");
+  for (const c of [
+    join(process.cwd(), ".worktrees", String(id), "phase-tracker.json"),
+    join(process.cwd(), ".worktrees", bare, "phase-tracker.json"),
+    join(process.cwd(), ".worktrees", `task-${bare}`, "phase-tracker.json"),
+  ]) {
+    if (existsSync(c)) return c;
+  }
+  return null;
+}
+
+const trackerPath = resolveTracker();
+if (!trackerPath) {
+  // No run data yet (e.g. called before Phase 0 wrote anything). Nothing to gate.
+  process.stdout.write("cost-budget: no tracker data yet - skipping check\n");
+  process.exit(0);
+}
+
+let tracker;
+try {
+  tracker = JSON.parse(readFileSync(trackerPath, "utf-8"));
+} catch (e) {
+  die(`tracker not valid JSON: ${e.message}`);
+}
+
+// --- price tokens -----------------------------------------------------------
+const costTable = JSON.parse(readFileSync(join(__dirname, "cost-table.json"), "utf-8"));
+const rate = costTable.prices?.[cfg.pricingModel];
+if (!rate) die(`no price for model "${cfg.pricingModel}" in cost-table.json`);
+
+let tokensIn = 0;
+let tokensOut = 0;
+for (const p of Object.values(tracker.phases || {})) {
+  tokensIn += Number(p?.tokens_in || 0);
+  tokensOut += Number(p?.tokens_out || 0);
+}
+
+const usd = (tokensIn * rate.inPerMtok + tokensOut * rate.outPerMtok) / 1_000_000;
+const pctOfMax = (usd / cfg.maxUsd) * 100;
+const usdStr = `$${usd.toFixed(2)}`;
+const totalTok = tokensIn + tokensOut;
+
+const base =
+  `cost-budget: ${usdStr} of $${cfg.maxUsd.toFixed(2)} ` +
+  `(${pctOfMax.toFixed(0)}%, ${totalTok.toLocaleString("en-US")} tok @ ${cfg.pricingModel})`;
+
+if (usd > cfg.maxUsd) {
+  const action = cfg.onExceed === "halt" ? "HALT - run multi-agent:resume to continue" : "over budget";
+  process.stdout.write(`EXCEEDED: ${base} - ${action}\n`);
+  process.exit(11);
+}
+if (pctOfMax >= cfg.warnPct) {
+  process.stdout.write(`WARN: ${base} - approaching ceiling\n`);
+  process.exit(10);
+}
+process.stdout.write(`OK: ${base}\n`);
+process.exit(0);

package/pipeline/scripts/cost-table.json

@@ -1,27 +1,37 @@
 {
-  "_readme": "Per-model unit prices in USD per million tokens. Source: Anthropic public pricing (verified 2026-04-21). Update when Anthropic publishes new tiers. Unknown models render USD as '—' and emit a footnote — never block PR-body generation.",
-  "schemaVersion": "1.0.0",
+  "_readme": "Per-model unit prices in USD per million tokens. Source: Anthropic public pricing (verified 2026-04-21). Update when Anthropic publishes new tiers. Unknown models render USD as ' - ' and emit a footnote  -  never block PR-body generation. cacheReadPerMtok is the discounted rate for prompt-cache hits (~10% of inPerMtok); the renderer prices a phase's tokens_cached at this rate when the tracker records it, so resume/cache reuse is visible in the ledger.",
+  "schemaVersion": "1.1.0",
   "prices": {
+    "fable": {
+      "inPerMtok": 10.0,
+      "outPerMtok": 50.0,
+      "cacheReadPerMtok": 1.0,
+      "modelId": "claude-fable-5"
+    },
     "opus": {
-      "inPerMtok": 15.00,
-      "outPerMtok": 75.00,
-      "modelId": "claude-opus-4-7"
+      "inPerMtok": 5.0,
+      "outPerMtok": 25.0,
+      "cacheReadPerMtok": 0.5,
+      "modelId": "claude-opus-4-8"
     },
     "sonnet": {
-      "inPerMtok": 3.00,
-      "outPerMtok": 15.00,
+      "inPerMtok": 3.0,
+      "outPerMtok": 15.0,
+      "cacheReadPerMtok": 0.3,
       "modelId": "claude-sonnet-4-6"
     },
     "haiku": {
-      "inPerMtok": 1.00,
-      "outPerMtok": 5.00,
+      "inPerMtok": 1.0,
+      "outPerMtok": 5.0,
+      "cacheReadPerMtok": 0.1,
       "modelId": "claude-haiku-4-5-20251001"
     },
     "gpt-5.4": {
-      "inPerMtok": 10.00,
-      "outPerMtok": 30.00,
+      "inPerMtok": 10.0,
+      "outPerMtok": 30.0,
+      "cacheReadPerMtok": 1.0,
       "modelId": "gpt-5.4",
-      "note": "Copilot CLI Reviewer 2 — approximate; verify against OpenAI pricing page before relying on totals."
+      "note": "Copilot CLI Reviewer 2  -  approximate; verify against OpenAI pricing page before relying on totals."
     }
   }
-}
+}

package/pipeline/scripts/diff-explain.mjs

@@ -1,13 +1,13 @@
 #!/usr/bin/env node
 
 /**
- * @file diff-explain.mjs — v7.8.0 Paket B
+ * @file diff-explain.mjs  -  v7.8.0 Paket B
  *
  * Bridges Phase 4 triage output to the actual git diff. For each accepted /
  * deferred / rejected finding, locates the corresponding hunk in the branch's
  * diff against base and renders an annotated markdown report.
  *
- * Read-only — never modifies state. Safe to run after the fact.
+ * Read-only  -  never modifies state. Safe to run after the fact.
  *
  * Inputs (any one):
  *   --triage <path>     Path to triage-output.json
@@ -31,7 +31,7 @@
  */
 
 import { readFileSync, existsSync } from "fs";
-import { execSync } from "child_process";
+import { execFileSync } from "child_process";
 
 const argv = process.argv.slice(2);
 const flags = {};
@@ -82,7 +82,7 @@ function loadTriage() {
   }
   if (flags["task-id"]) {
     die(
-      "--task-id resolution not implemented in this script — use --triage or --state directly. " +
+      "--task-id resolution not implemented in this script  -  use --triage or --state directly. " +
       "The slash command resolves task IDs via the orchestrator and passes --state to this script.",
     );
   }
@@ -97,9 +97,9 @@ function getDiff(filesFilter) {
   const base = flags.base || resolveBase();
   const branch = flags.branch || "HEAD";
   const filterArgs = filesFilter.length ? ["--", ...filesFilter] : [];
-  const cmd = ["git", "diff", "--unified=3", `${base}...${branch}`, ...filterArgs];
+  const args = ["diff", "--unified=3", `${base}...${branch}`, ...filterArgs];
   try {
-    return execSync(cmd.join(" "), { encoding: "utf-8", maxBuffer: 50 * 1024 * 1024 });
+    return execFileSync("git", args, { encoding: "utf-8", maxBuffer: 50 * 1024 * 1024 });
   } catch (e) {
     die(`git diff failed: ${e.message}`);
   }
@@ -108,13 +108,13 @@ function getDiff(filesFilter) {
 function resolveBase() {
   for (const ref of ["origin/main", "main", "origin/master", "master"]) {
     try {
-      execSync(`git rev-parse --verify ${ref}`, { stdio: "ignore" });
+      execFileSync("git", ["rev-parse", "--verify", ref], { stdio: "ignore" });
       return ref;
     } catch {
       /* try next */
     }
   }
-  die("no base ref found (tried origin/main, main, origin/master, master) — pass --base");
+  die("no base ref found (tried origin/main, main, origin/master, master)  -  pass --base");
 }
 
 /**
@@ -176,7 +176,7 @@ function severityLabel(s) {
 }
 
 function renderHunk(hunk) {
-  if (!hunk) return "_(line not in diff — finding may reference unchanged code)_";
+  if (!hunk) return "_(line not in diff  -  finding may reference unchanged code)_";
   const lines = ["```diff", hunk.header, ...hunk.lines, "```"];
   return lines.join("\n");
 }
@@ -210,7 +210,7 @@ const findings = flattenFindings(triage);
 const filesFromFindings = [...new Set(findings.map((f) => f.file).filter(Boolean))];
 
 if (findings.length === 0) {
-  console.log("# Phase 4 Triage Diff Bridge\n\n_(no findings to map — clean review.)_");
+  console.log("# Phase 4 Triage Diff Bridge\n\n_(no findings to map  -  clean review.)_");
   process.exit(0);
 }
 
@@ -252,7 +252,7 @@ for (const [file, items] of byFile) {
     const icon = severityIcon(item.severity);
     const bucketTag = bucketIcon(item._bucket);
     const reviewer = item.reviewer ? ` _(${item.reviewer})_` : "";
-    lines.push(`### ${bucketTag} ${icon} ${severityLabel(item.severity)} — line ${item.line ?? "?"}${reviewer}`);
+    lines.push(`### ${bucketTag} ${icon} ${severityLabel(item.severity)}  -  line ${item.line ?? "?"}${reviewer}`);
     lines.push("");
     if (item._reason) lines.push(`> _${item._bucket}: ${item._reason}_`);
     lines.push(`> ${item.issue}`);
@@ -266,7 +266,7 @@ for (const [file, items] of byFile) {
         const hunk = hunkAtLine(fileHunks, item.line);
         lines.push(renderHunk(hunk));
       } else {
-        lines.push("_(file not in diff — finding references unchanged or non-staged code)_");
+        lines.push("_(file not in diff  -  finding references unchanged or non-staged code)_");
       }
       lines.push("");
     }

package/pipeline/scripts/diff-risk-score.mjs

@@ -1,20 +1,20 @@
 #!/usr/bin/env node
 
 /**
- * @file diff-risk-score.mjs — v8.3.0
+ * @file diff-risk-score.mjs  -  v8.3.0
  *
  * Heuristic risk score per file in a diff. Deterministic, zero LLM.
  * Output sorted by score descending so reviewer prompts can attach a
- * "priority files" list. Advisory only — never blocks the pipeline.
+ * "priority files" list. Advisory only  -  never blocks the pipeline.
  *
  * Signals (all sourced from `git diff --numstat` + targeted hunk grep):
- *   loc_changed      — base sensitivity to size of change         w=1.0
- *   security_path    — touches Auth/Keychain/Network/Crypto/...   w=3.0
- *   public_api       — adds/changes public func/class/Composable  w=2.0
- *   no_test_change   — source touched, test file untouched        w=2.5
- *   complexity_delta — added if/guard/case/switch/while count     w=1.5
- *   ui_critical      — *View.swift / *Screen.kt / Configuration   w=1.5
- *   migration        — DB schema / migration path                 w=4.0
+ *   loc_changed       -  base sensitivity to size of change         w=1.0
+ *   security_path     -  touches Auth/Keychain/Network/Crypto/...   w=3.0
+ *   public_api        -  adds/changes public func/class/Composable  w=2.0
+ *   no_test_change    -  source touched, test file untouched        w=2.5
+ *   complexity_delta  -  added if/guard/case/switch/while count     w=1.5
+ *   ui_critical       -  *View.swift / *Screen.kt / Configuration   w=1.5
+ *   migration         -  DB schema / migration path                 w=4.0
  *
  * Inputs:
  *   --base <ref>     Base ref. Default: origin/main, fallback: main
@@ -27,12 +27,12 @@
  *   --json           Pretty print (default: compact JSON)
  *
  * Output:
- *   stdout — JSON conforming to pipeline/schemas/diff-risk.schema.json
+ *   stdout  -  JSON conforming to pipeline/schemas/diff-risk.schema.json
  *
  * Exit codes:
- *   0 — success
- *   1 — git/diff invocation failed or unparseable
- *   2 — empty diff (no files changed) — caller can choose to skip
+ *   0  -  success
+ *   1  -  git/diff invocation failed or unparseable
+ *   2  -  empty diff (no files changed)  -  caller can choose to skip
  *
  * @module pipeline/scripts/diff-risk-score
  */
@@ -90,7 +90,7 @@ function loadDiff() {
 
 function loadNumstat() {
   if (DIFF_FILE) {
-    // Approximate from the diff text — count + and - lines per file.
+    // Approximate from the diff text  -  count + and - lines per file.
     const text = readFileSync(DIFF_FILE, "utf8");
     const map = new Map();
     let cur = null;
@@ -231,7 +231,7 @@ function buildRow(stat, addedLines, allChangedPaths) {
   if (matchesAny(path, SECURITY_GLOBS)) {
     const w = 3.0;
     signals.push({ name: "security_path", weight: w, value: 1 });
-    score += 30 * w; // boolean signal — fixed contribution
+    score += 30 * w; // boolean signal  -  fixed contribution
   }
 
   if (matchesAny(path, MIGRATION_GLOBS)) {
@@ -264,8 +264,9 @@ function buildRow(stat, addedLines, allChangedPaths) {
   if (!isTestPath(path)) {
     const candidates = pairedTestPaths(path).map((c) => c.toLowerCase());
     const anyTest = allChangedPaths.some((p) => {
-      const low = p.toLowerCase();
-      return isTestPath(p) && candidates.some((c) => low.includes(c));
+      if (!isTestPath(p)) return false;
+      const pBase = p.split("/").pop().toLowerCase();
+      return candidates.includes(pBase);
     });
     if (!anyTest) {
       const w = 2.5;

package/pipeline/scripts/eval-golden-tasks-live.mjs

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 /**
- * @file eval-golden-tasks-live.mjs — v7.8.0 Paket A
+ * @file eval-golden-tasks-live.mjs  -  v7.8.0 Paket A
  *
  * Opt-in live evaluation harness for golden-task fixtures. Unlike the
  * always-on contract harness (`eval-golden-tasks.mjs`) which only validates
@@ -9,13 +9,13 @@
  * `claude` CLI and compares the actual output against the fixture's expected
  * scope (stack, blocker count, deferral count).
  *
- * **Default behavior is dry-run** — no model is invoked unless `--live` is
+ * **Default behavior is dry-run**  -  no model is invoked unless `--live` is
  * passed AND `MULTI_AGENT_LIVE_EVAL=1` is set in the environment. This
  * double-gate prevents accidental cost.
  *
  * Cost guard:
- *   - Per-case budget cap (default $1 USD) — case skipped if exceeded
- *   - Run-wide max-cases cap (default 1) — even with --live, only N cases run
+ *   - Per-case budget cap (default $1 USD)  -  case skipped if exceeded
+ *   - Run-wide max-cases cap (default 1)  -  even with --live, only N cases run
  *   - Run-wide hard ceiling: budget * max_cases (default $1 total)
  *
  * Zero-dep: shells out to the user's installed `claude` CLI rather than the
@@ -113,11 +113,14 @@ function loadCase(slug) {
   const caseDir = join(evalDir, slug);
   const taskFile = join(caseDir, "task.json");
   if (!existsSync(taskFile)) return null;
-  return {
-    slug,
-    dir: caseDir,
-    task: JSON.parse(readFileSync(taskFile, "utf-8")),
-  };
+  let task;
+  try {
+    task = JSON.parse(readFileSync(taskFile, "utf-8"));
+  } catch (e) {
+    process.stderr.write(`[eval-live] malformed task.json for "${slug}": ${e.message}\n`);
+    return null;
+  }
+  return { slug, dir: caseDir, task };
 }
 
 function hasClaudeCLI() {
@@ -126,7 +129,7 @@ function hasClaudeCLI() {
 }
 
 /**
- * Live invocation — wraps the task as a prompt for `claude -p`. Returns the
+ * Live invocation  -  wraps the task as a prompt for `claude -p`. Returns the
  * raw stdout. The current iteration captures only the analysis-phase output
  * by asking for a JSON-only response per the analysis schema. Future work
  * extends this to full Phase 1/2/4 multi-call evaluation.

package/pipeline/scripts/eval-golden-tasks.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-// eval-golden-tasks.mjs — contract regression check for whole-pipeline fixtures.
+// eval-golden-tasks.mjs  -  contract regression check for whole-pipeline fixtures.
 //
 // Walks pipeline/eval/golden-tasks/<NN>-<name>/ subdirectories. For each case:
 //   1. Reads task.json (input) + expected/phase-{1,2,4-review,4-triage}.json.
@@ -72,17 +72,6 @@ function runValidator(scriptPath, payloadPath) {
   return { code: r.status, stdout: r.stdout || "", stderr: r.stderr || "" };
 }
 
-// eslint-disable-next-line no-unused-vars
-function runReviewerValidator(payloadPath) {
-  // reviewer validator reads stdin — pipe the JSON in.
-  const payload = readFileSync(payloadPath, "utf-8");
-  const r = spawnSync("node", [validators.reviewer, "-"], {
-    input: payload,
-    encoding: "utf-8",
-  });
-  return { code: r.status, stdout: r.stdout || "", stderr: r.stderr || "" };
-}
-
 function runCase(name) {
   const dir = join(evalDir, name);
   const errors = [];
@@ -124,7 +113,7 @@ function runCase(name) {
   const p = runValidator(validators.planning, planPath);
   if (p.code !== 0) errors.push(`phase-2-plan fails validate-planning (exit ${p.code}): ${p.stderr.trim()}`);
 
-  // phase-4-review is an array (one entry per reviewer) — validate each
+  // phase-4-review is an array (one entry per reviewer)  -  validate each
   if (!Array.isArray(review)) {
     errors.push("phase-4-review.json must be a JSON array (one entry per reviewer)");
   } else {
@@ -170,7 +159,7 @@ function runCase(name) {
     ...(triage.accepted || []).map((x) => x),
     ...(triage.deferred || []).map((x) => x.finding),
     ...(triage.rejected || []).map((x) => x.finding),
-  ];
+  ].filter((f) => f && typeof f === "object");
   const triageKeys = new Set(triageAll.map((f) => `${f.file}::${f.line}::${f.issue}`));
 
   for (const rev of (Array.isArray(review) ? review : [])) {