feat(ci): reshape ClawReview 🦀 into an informative mascot (+ SDK fix)

.github/workflows/issue-triage.yml

@@ -21,6 +21,10 @@ jobs:
       contents: read
     # Skip issues opened by bots to avoid loops.
     if: ${{ !endsWith(github.actor, '[bot]') }}
+    # Job-level so a step `if:` can read it — a step cannot see env it sets on
+    # itself. Gates the optional App-token step below.
+    env:
+      HAS_APP: ${{ secrets.CLAWREVIEW_APP_ID }}
     steps:
       - name: Checkout
         # SHA-pinned: this workflow runs on every opened issue with
@@ -63,12 +67,12 @@ jobs:
         # is used. The bot works either way.
         id: app-token
         if: env.HAS_APP != ''
-        env:
-          HAS_APP: ${{ secrets.CLAWREVIEW_APP_ID }}
         uses: actions/create-github-app-token@fee1f7d63c2ff003460e3d139729b119787bc349 # v2
         with:
           app-id: ${{ secrets.CLAWREVIEW_APP_ID }}
           private-key: ${{ secrets.CLAWREVIEW_APP_PRIVATE_KEY }}
+          # Triage only labels + comments on issues — scope the token to that.
+          permission-issues: write
 
       - name: Triage with Claude
         env:

.github/workflows/pr-review.yml

@@ -31,6 +31,10 @@ jobs:
     # Drafts are skipped too (reviewed at ready_for_review) — no API spend
     # while the author is still iterating.
     if: ${{ !endsWith(github.event.pull_request.user.login, '[bot]') && !github.event.pull_request.draft }}
+    # Job-level so a step `if:` can read it — a step cannot see env it sets on
+    # itself. Gates the optional App-token step below.
+    env:
+      HAS_APP: ${{ secrets.CLAWREVIEW_APP_ID }}
     steps:
       - name: Checkout (base repo only — never the PR head)
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
@@ -66,12 +70,14 @@ jobs:
         # github-actions[bot] — the bot works either way.
         id: app-token
         if: env.HAS_APP != ''
-        env:
-          HAS_APP: ${{ secrets.CLAWREVIEW_APP_ID }}
         uses: actions/create-github-app-token@fee1f7d63c2ff003460e3d139729b119787bc349 # v2
         with:
           app-id: ${{ secrets.CLAWREVIEW_APP_ID }}
           private-key: ${{ secrets.CLAWREVIEW_APP_PRIVATE_KEY }}
+          # Scope the token to exactly what this bot needs (comments + labels),
+          # not the App's full installation grant.
+          permission-pull-requests: write
+          permission-issues: write
 
       - name: Review with Claude
         env:

scripts/issue-triage.mjs

@@ -4,7 +4,7 @@
 // Needs: ANTHROPIC_API_KEY (repo secret) and GH_TOKEN (the workflow's GITHUB_TOKEN).
 import fs from "node:fs";
 import { execFileSync } from "node:child_process";
-import Anthropic from "@anthropic-ai/sdk";
+import { callClaude } from "./lib/ai-backend.mjs";
 
 // Haiku 4.5 — fast and cheap, ideal for a high-volume issue classifier.
 // Switch to "claude-opus-4-8" for maximum classification accuracy.
@@ -42,61 +42,11 @@ function gh(args) {
   return execFileSync("gh", args, { encoding: "utf8", stdio: ["ignore", "pipe", "inherit"] });
 }
 
-// Extract a JSON object from possibly-fenced/wrapped model text.
-function parseModelJson(text) {
-  const stripped = text.replace(/^```(?:json)?\s*/m, "").replace(/```\s*$/m, "").trim();
-  try { return JSON.parse(stripped); } catch { /* fall through */ }
-  const m = stripped.match(/\{[\s\S]*\}/);
-  if (!m) throw new Error("no JSON object in model response");
-  return JSON.parse(m[0]);
-}
-
-// Subscription transport: headless Claude Code CLI, authed by
-// CLAUDE_CODE_OAUTH_TOKEN (the official Pro/Max path). Keep in sync with the
-// same pattern in scripts/pr-review.mjs.
-function classifyViaClaudeCli(userContent) {
-  const prompt = [
-    SYSTEM,
-    "\nRespond with ONLY a JSON object matching this schema (no prose, no fences):",
-    JSON.stringify(SCHEMA),
-    "\n---\n",
-    userContent,
-  ].join("\n");
-  const out = execFileSync("claude", ["-p", "--model", MODEL, "--output-format", "json"], {
-    encoding: "utf8",
-    input: prompt,
-    stdio: ["pipe", "pipe", "inherit"],
-    timeout: 180_000,
-    maxBuffer: 8 * 1024 * 1024,
-  });
-  const wrapper = JSON.parse(out);
-  if (wrapper.is_error) throw new Error(`claude cli error: ${String(wrapper.result).slice(0, 200)}`);
-  return parseModelJson(String(wrapper.result));
-}
-
-async function classifyViaSdk(userContent) {
-  const client = new Anthropic(); // reads ANTHROPIC_API_KEY from env
-  const resp = await client.messages.create({
-    model: MODEL,
-    max_tokens: 1024,
-    system: SYSTEM,
-    output_config: { format: { type: "json_schema", schema: SCHEMA } },
-    messages: [{ role: "user", content: userContent }],
-  });
-  const text = resp.content.find((b) => b.type === "text")?.text;
-  // Throw rather than default to "{}" — an empty object here would create
-  // and apply labels literally named "undefined"; the outer catch logs and
-  // exits 0 (triage must never fail issue creation).
-  if (!text) throw new Error("no text block in model response");
-  return JSON.parse(text);
-}
-
 async function main() {
   const userContent = `Triage this issue. Respond ONLY with the JSON object.\n\n<title>${title}</title>\n\n<body>\n${body.slice(0, 8000)}\n</body>`;
-  // Subscription OAuth preferred (team choice); API key as fallback backend.
-  const t = process.env.CLAUDE_CODE_OAUTH_TOKEN
-    ? classifyViaClaudeCli(userContent)
-    : await classifyViaSdk(userContent);
+  // Transport (OAuth CLI / SDK) lives in the shared backend so both bots stay
+  // in sync. OAuth is preferred; the API-key SDK is the fallback.
+  const t = await callClaude({ system: SYSTEM, schema: SCHEMA, userContent, model: MODEL, maxTokens: 1024, timeoutMs: 180_000, maxBuffer: 8 * 1024 * 1024 });
 
   // Ensure the priority/area labels exist (idempotent), then apply.
   const ensure = (name, color, desc) => {
@@ -119,24 +69,37 @@ async function main() {
   ensure(t.category, catColor, "Auto-triage category");
 
   const labels = [t.category, `priority: ${t.priority}`, `area: ${t.area}`];
-  gh(["issue", "edit", String(number), "--repo", REPO, ...labels.flatMap((l) => ["--add-label", l])]);
+  if (!process.env.DRY_RUN) gh(["issue", "edit", String(number), "--repo", REPO, ...labels.flatMap((l) => ["--add-label", l])]);
 
+  // Same crab mascot as ClawReview (the PR bot) — one friendly character
+  // across issues and PRs. Greeting picked by issue number for stability.
+  const GREETINGS = [
+    "Scuttled over to help sort this one 🦀",
+    "Your friendly reef crab, here to get this filed.",
+    "Thanks for the report — let me get you oriented.",
+    "Claws on the case. Here's how I've tagged it:",
+  ];
+  const priIcon = { high: "🔴", medium: "🟡", low: "🟢" }[t.priority] ?? "⚪";
   const comment = [
-    "### 🤖 Auto-triage",
+    "## 🦀 ClawReview",
+    "",
+    `*${GREETINGS[number % GREETINGS.length]}*`,
     "",
-    "| | |",
-    "|---|---|",
-    `| **Category** | \`${t.category}\` |`,
-    `| **Priority** | \`${t.priority}\` |`,
-    `| **Area** | \`${t.area}\` |`,
+    t.summary,
     "",
-    `**Summary:** ${t.summary}`,
+    "**At a glance**",
+    `- Category: \`${t.category}\` · Area: \`${t.area}\``,
+    `- Priority: ${priIcon} \`${t.priority}\``,
     "",
     `**Suggested next step:** ${t.suggested_action}`,
     "",
-    "<sub>Auto-classified on open — labels are advisory; adjust as needed.</sub>",
+    "<sub>— ClawReview 🦀. Labels auto-applied on open — advisory, a maintainer will follow up. Conventions: <a href=\"https://docs.clawbox.tech/llms.txt\">docs</a>.</sub>",
   ].join("\n");
 
+  if (process.env.DRY_RUN) {
+    console.log(comment);
+    return;
+  }
   gh(["issue", "comment", String(number), "--repo", REPO, "--body", comment]);
   console.log(`Triaged #${number}: ${t.category} / ${t.priority} / ${t.area}`);
 }

scripts/lib/ai-backend.mjs

@@ -0,0 +1,61 @@
+// Shared Claude backend for the ClawBox CI bots (pr-review + issue-triage).
+// One place for the two transports so a change to either stays in sync:
+//   - CLAUDE_CODE_OAUTH_TOKEN set -> Claude Code CLI (`claude -p`), the
+//     official Pro/Max subscription path (same runtime as claude-code-action)
+//   - else -> the Anthropic SDK (API key)
+// The SDK is imported lazily so the OAuth-only install (CLI, no SDK) doesn't
+// crash at module load.
+import { execFileSync } from "node:child_process";
+
+// Extract a JSON object from possibly-fenced/wrapped model text.
+export function parseModelJson(text) {
+  const stripped = text.replace(/^```(?:json)?\s*/m, "").replace(/```\s*$/m, "").trim();
+  try { return JSON.parse(stripped); } catch { /* fall through */ }
+  const m = stripped.match(/\{[\s\S]*\}/);
+  if (!m) throw new Error("no JSON object in model response");
+  return JSON.parse(m[0]);
+}
+
+function viaClaudeCli({ system, schema, userContent, model, timeoutMs, maxBuffer }) {
+  const prompt = [
+    system,
+    "\nRespond with ONLY a JSON object matching this schema (no prose, no fences):",
+    JSON.stringify(schema),
+    "\n---\n",
+    userContent,
+  ].join("\n");
+  const out = execFileSync("claude", ["-p", "--model", model, "--output-format", "json"], {
+    encoding: "utf8",
+    input: prompt,
+    stdio: ["pipe", "pipe", "inherit"],
+    timeout: timeoutMs,
+    maxBuffer,
+  });
+  const wrapper = JSON.parse(out);
+  if (wrapper.is_error) throw new Error(`claude cli error: ${String(wrapper.result).slice(0, 200)}`);
+  return parseModelJson(String(wrapper.result));
+}
+
+async function viaSdk({ system, schema, userContent, model, maxTokens }) {
+  const { default: Anthropic } = await import("@anthropic-ai/sdk");
+  const client = new Anthropic(); // reads ANTHROPIC_API_KEY from env
+  const resp = await client.messages.create({
+    model,
+    max_tokens: maxTokens,
+    system,
+    output_config: { format: { type: "json_schema", schema } },
+    messages: [{ role: "user", content: userContent }],
+  });
+  const text = resp.content.find((b) => b.type === "text")?.text;
+  // Throw rather than default to "{}" — an empty object downstream would create
+  // and apply labels literally named "undefined". Callers' outer catch exits 0.
+  if (!text) throw new Error("no text block in model response");
+  return JSON.parse(text);
+}
+
+// Run one structured-output call and return the validated JSON object.
+// OAuth transport preferred; API-key SDK is the fallback.
+export function callClaude(opts) {
+  if (process.env.CLAUDE_CODE_OAUTH_TOKEN) return viaClaudeCli(opts);
+  return viaSdk(opts);
+}