@gajae-code/coding-agent 0.7.1 → 0.7.3

package/src/web/search/providers/text-citations.ts

@@ -0,0 +1,111 @@
+/**
+ * Inline citation extraction shared by native web-search providers.
+ *
+ * Web-search-capable models sometimes return a genuinely grounded answer whose
+ * sources are written inline (markdown links or bare URLs) instead of as
+ * structured citation annotations. When a provider has independent proof that a
+ * web search actually ran, these helpers recover sources from the answer text so
+ * the result is not discarded.
+ */
+import type { SearchSource } from "../types";
+
+/** Append a source, de-duplicating by URL. */
+export function addSource(sources: SearchSource[], source: SearchSource): void {
+	if (!sources.some(existing => existing.url === source.url)) {
+		sources.push(source);
+	}
+}
+
+function countCharacter(text: string, target: string): number {
+	let count = 0;
+	for (const char of text) {
+		if (char === target) count += 1;
+	}
+	return count;
+}
+
+/**
+ * Strips prose punctuation and unmatched closing delimiters from extracted URLs.
+ * Models often return links embedded in markdown or sentence text.
+ */
+export function normalizeExtractedUrl(candidate: string): string | null {
+	let url = candidate.trim();
+
+	while (url.length > 0) {
+		const lastCharacter = url.at(-1);
+		if (!lastCharacter) break;
+		if (/[.,!?;:'"]/u.test(lastCharacter)) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === ")" && countCharacter(url, ")") > countCharacter(url, "(")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === "]" && countCharacter(url, "]") > countCharacter(url, "[")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === "}" && countCharacter(url, "}") > countCharacter(url, "{")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		break;
+	}
+
+	if (!/^https?:\/\//.test(url)) return null;
+
+	try {
+		return new URL(url).toString();
+	} catch {
+		return null;
+	}
+}
+
+function findMarkdownLinkUrlEnd(text: string, openParenIndex: number): number | null {
+	let depth = 0;
+
+	for (let index = openParenIndex; index < text.length; index += 1) {
+		const character = text[index];
+		if (!character || character === "\n") return null;
+		if (character === "(") {
+			depth += 1;
+			continue;
+		}
+		if (character !== ")") continue;
+		depth -= 1;
+		if (depth === 0) return index;
+		if (depth < 0) return null;
+	}
+
+	return null;
+}
+
+/**
+ * Extracts citation sources from markdown links and bare URLs in answer text.
+ * Used only as a fallback when a provider confirms a search ran but omits
+ * structured citation annotations.
+ */
+export function extractTextSources(text: string): SearchSource[] {
+	const sources: SearchSource[] = [];
+
+	for (let index = 0; index < text.length; index += 1) {
+		if (text[index] !== "[") continue;
+		const titleEnd = text.indexOf("]", index + 1);
+		if (titleEnd === -1 || text[titleEnd + 1] !== "(") continue;
+		const urlEnd = findMarkdownLinkUrlEnd(text, titleEnd + 1);
+		if (urlEnd === null) continue;
+		const title = text.slice(index + 1, titleEnd).trim();
+		const url = normalizeExtractedUrl(text.slice(titleEnd + 2, urlEnd));
+		if (url) addSource(sources, { title: title || url, url });
+		index = urlEnd;
+	}
+
+	for (const match of text.matchAll(/https?:\/\/\S+/g)) {
+		const url = normalizeExtractedUrl(match[0] ?? "");
+		if (!url) continue;
+		addSource(sources, { title: url, url });
+	}
+
+	return sources;
+}

package/src/web/search/types.ts

@@ -7,6 +7,7 @@
 /** Supported web search providers */
 export type SearchProviderId =
 	| "duckduckgo"
+	| "insane"
 	| "exa"
 	| "brave"
 	| "jina"
@@ -38,6 +39,7 @@ export interface ActiveSearchModelContext {
 
 export const CONFIGURABLE_SEARCH_PROVIDER_IDS = [
 	"duckduckgo",
+	"insane",
 	"exa",
 	"brave",
 	"jina",
@@ -162,6 +164,15 @@ export interface AnthropicCitation {
 	encrypted_index: string;
 }
 
+/**
+ * Error payload returned in `web_search_tool_result.content` when a server-side
+ * web search fails. Unlike the success case, this is an object, not an array.
+ */
+export interface AnthropicWebSearchToolResultError {
+	type: "web_search_tool_result_error";
+	error_code?: string;
+}
+
 export interface AnthropicContentBlock {
 	type: string;
 	/** Text content (for type="text") */
@@ -172,8 +183,8 @@ export interface AnthropicContentBlock {
 	name?: string;
 	/** Tool input (for type="server_tool_use") */
 	input?: { query: string };
-	/** Search results (for type="web_search_tool_result") */
-	content?: AnthropicSearchResult[];
+	/** Search results array on success, or an error object on failure (type="web_search_tool_result") */
+	content?: AnthropicSearchResult[] | AnthropicWebSearchToolResultError;
 }
 
 export interface AnthropicApiResponse {

package/vendor/insane-search/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 fivetaku
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/vendor/insane-search/MANIFEST.json

@@ -0,0 +1,24 @@
+{
+  "upstream": {
+    "repo": "https://github.com/fivetaku/insane-search",
+    "commit": "49306346b59aa89b5e96d98e1104da0890deed72",
+    "license": "MIT"
+  },
+  "vendoredAt": "2026-06-23",
+  "includedPaths": [
+    "skills/insane-search/engine/**",
+    "skills/insane-search/engine/templates/**",
+    "LICENSE"
+  ],
+  "excludedPaths": [
+    "setup/setup.sh",
+    "setup/gptaku-update-check.cjs",
+    ".claude-plugin/**",
+    "skills/insane-search/SKILL.md",
+    "skills/insane-search/references/**",
+    "skills/insane-search/tests/**"
+  ],
+  "exclusionRationale": "Excludes upstream install hooks (SessionStart settings.json mutation), GitHub star-baiting (gh api user/starred), the update-notifier, and the past-session transcript-language scanner. Only the runtime Phase 0-3 engine and its Playwright/stealth templates are vendored.",
+  "localPatches": [],
+  "notes": "Runtime engine is invoked via `python3 -m engine \"<url>\" --json` with cwd=this directory and PYTHONPATH pointed at this directory. Phase 0-2 require python3 + curl_cffi; Phase 3 requires node + playwright/playwright-extra/puppeteer-extra-plugin-stealth installed under engine/templates. GJC never auto-installs these dependencies."
+}

package/vendor/insane-search/engine/__init__.py

@@ -0,0 +1,23 @@
+"""insane-search engine — generic WAF-profile-based fetch chain.
+
+No site-specific logic lives here. Site specifics belong to runtime hints or
+observations, never to code. See `../SKILL.md` for the No-Site-Name Rule.
+"""
+
+from .validators import Verdict, ValidationResult, validate, CHALLENGE_MARKERS
+from .waf_detector import detect
+from .url_transforms import TRANSFORMS, apply_transform
+from .fetch_chain import fetch, FetchResult, Attempt
+
+__all__ = [
+    "Verdict",
+    "ValidationResult",
+    "validate",
+    "CHALLENGE_MARKERS",
+    "detect",
+    "TRANSFORMS",
+    "apply_transform",
+    "fetch",
+    "FetchResult",
+    "Attempt",
+]

package/vendor/insane-search/engine/__main__.py

@@ -0,0 +1,128 @@
+#!/usr/bin/env python3
+"""CLI entrypoint for the insane-search engine.
+
+Usage:
+    python3 -m engine URL [--selector CSS] [--device auto|desktop|mobile]
+                          [--timeout N] [--max-attempts N] [--json] [--trace]
+
+Examples:
+    python3 -m engine "https://example.com/" --selector "h1"
+    python3 -m engine "https://example.com/" --json
+    python3 -m engine "https://example.com/" --device mobile --trace
+
+Exit codes:
+    0   strong_ok or weak_ok
+    1   ok=False (all attempts failed)
+    2   CLI arg error
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+
+from . import fetch
+
+
+def build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(prog="python3 -m engine",
+                                description="Generic WAF-profile fetch chain.")
+    p.add_argument("url", help="URL to fetch.")
+    p.add_argument("--selector", "-s", action="append", default=None,
+                   dest="selectors", metavar="CSS",
+                   help="Positive-proof CSS selector. Repeatable.")
+    p.add_argument("--device", choices=("auto", "desktop", "mobile"), default="auto",
+                   help="Device class pin.")
+    p.add_argument("--timeout", type=int, default=25,
+                   help="Per-attempt timeout seconds (default 25).")
+    p.add_argument("--max-attempts", type=int, default=None,
+                   help="TOTAL curl-attempt budget. Default: None = exhaustive (honours R6).")
+    p.add_argument("--no-playwright", action="store_true",
+                   help="Skip Playwright fallback (curl-only).")
+    p.add_argument("--no-phase0", action="store_true",
+                   help="Skip the Phase 0 official-API router (generic grid only).")
+    p.add_argument("--json", action="store_true",
+                   help="Emit FetchResult as JSON to stdout with bounded content included.")
+    p.add_argument("--json-content-limit", type=int, default=4_000_000,
+                   help="Maximum content characters to include in --json output (default 4000000).")
+    p.add_argument("--trace", action="store_true",
+                   help="Print per-attempt trace to stderr.")
+    return p
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    try:
+        result = fetch(
+            args.url,
+            success_selectors=args.selectors,
+            device_class=args.device,
+            timeout=args.timeout,
+            max_attempts=args.max_attempts,
+            enable_playwright=not args.no_playwright,
+            enable_phase0=not args.no_phase0,
+        )
+    except Exception as e:
+        print(f"engine fatal: {type(e).__name__}: {e}", file=sys.stderr)
+        return 2
+
+    if args.trace:
+        print("=== trace ===", file=sys.stderr)
+        for att in result.trace:
+            d = att.to_dict()
+            imp = d.get("impersonate") or "-"
+            ref = d.get("referer") or "-"
+            print(
+                f"[{d['phase']:<8}] {d['executor']:<18} "
+                f"xform={d['url_transform']:<16} imp={imp:<14} ref={ref:<14} "
+                f"status={d['status']:>4} size={d['body_size']:>8} "
+                f"verdict={d['verdict']} {('err=' + d['error'][:60]) if d.get('error') else ''}",
+                file=sys.stderr,
+            )
+        print(f"=== summary: {result.summary} ===", file=sys.stderr)
+
+    # Surface R7 hint (API-first route) prominently when summary contains it,
+    # regardless of --trace flag — this is actionable guidance, not noise.
+    if "R7 API-first" in (result.summary or ""):
+        print(
+            "\n════════════════════════════════════════════════════════════════\n"
+            "⚠️  R7 triggered — consider API-first route instead of HTML grid.\n"
+            "   See summary below (or re-run with --trace for full attempt log).\n"
+            "════════════════════════════════════════════════════════════════",
+            file=sys.stderr,
+        )
+        # Also print the full summary (which includes the hint) so caller sees it.
+        print(result.summary, file=sys.stderr)
+
+    # Failure gate (R6): giving up is NOT permission to stop. Surface the routes
+    # the engine could not run by itself so the caller keeps escalating instead
+    # of reporting "blocked" prematurely (the exact bug this hardening fixes).
+    if not result.ok and (result.untried_routes or result.must_invoke_playwright_mcp):
+        print(
+            "\n════════════════════════════════════════════════════════════════\n"
+            "⛔ NOT EXHAUSTED — do not declare failure yet (R6).\n"
+            f"   grid_exhausted={result.grid_exhausted}  stop_reason={result.stop_reason}\n"
+            "   Routes the engine cannot run itself — try these before giving up:",
+            file=sys.stderr,
+        )
+        for r in result.untried_routes:
+            print(f"     • {r}", file=sys.stderr)
+        if result.must_invoke_playwright_mcp:
+            print("   ➜ must_invoke_playwright_mcp = TRUE — drive MCP Playwright from the agent session.", file=sys.stderr)
+        print("════════════════════════════════════════════════════════════════", file=sys.stderr)
+
+    if args.json:
+        payload = result.to_dict(include_content=True, content_limit=args.json_content_limit)
+        print(json.dumps(payload, ensure_ascii=False, indent=2))
+    else:
+        # Default: HTML to stdout, status to stderr.
+        print(result.content, end="")
+        print(f"\n[engine] ok={result.ok} verdict={result.verdict} "
+              f"profile={result.profile_used} attempts={len(result.trace)}",
+              file=sys.stderr)
+
+    return 0 if result.ok else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/vendor/insane-search/engine/bias_check.py

@@ -0,0 +1,183 @@
+#!/usr/bin/env python3
+"""No-Site-Name Rule checker.
+
+Run in CI / pre-commit. Scans engine/** for hard-coded site names or
+domains that would bias the generic fetch chain toward one site.
+
+Exit code 0 if clean, 1 if violations found.
+
+    python3 engine/bias_check.py
+    python3 engine/bias_check.py --strict    # also check references/*.md (usually off)
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import re
+import sys
+from pathlib import Path
+
+
+# Known brand / domain substrings that should NOT appear in engine code.
+# This is a non-exhaustive deny list. CI should treat hits as warnings that
+# require human review; false positives (e.g. "github" in comments) can be
+# whitelisted via EXPLICIT_ALLOW.
+BRAND_SUBSTRINGS = [
+    "coupang", "11st", "11번가", "musinsa", "무신사",
+    "fmkorea", "에펨코리아", "dcinside", "디시인사이드",
+    "ohou", "오늘의집", "kurly", "마켓컬리",
+    "daangn", "당근",
+    # Naver is allowed in Phase 0 references (official APIs) but not in engine code.
+    "naver.com", "blog.naver", "shopping.naver",
+    # Korean portal brand names
+    "daum.net", "kakao.com",
+]
+
+# Regex for bare URLs / domains. Used as a secondary pass to flag hardcoded
+# site hosts that slipped past the brand denylist.
+URL_PATTERN = re.compile(
+    r"https?://[\w\.-]+|[\w-]+\.(?:com|net|org|co\.kr|kr|io)\b",
+    re.IGNORECASE,
+)
+
+# Generic / neutral hosts that are allowed anywhere (examples, specs, stdlib,
+# and domains that legitimately appear as non-site-specific referrers / test
+# fixtures — Google search as a generic Referer strategy, httpbin for transport
+# tests, etc.). Anything in this set must be provably unrelated to a specific
+# target-site preference.
+URL_ALLOWLIST = {
+    "example.com", "example.org", "example.net",
+    "localhost", "127.0.0.1",
+    # Official API / documentation sources cited in code comments.
+    "curl.se", "playwright.dev", "nodejs.org", "npmjs.com",
+    # Generic Referer strategy target (used as a neutral off-site referer).
+    "www.google.com", "google.com",
+    # Generic HTTP test endpoint for infrastructure / transport tests.
+    "httpbin.org",
+}
+
+# Files / dirs that must be clean.
+SCAN_ROOTS_STRICT_OFF = ["engine"]
+SCAN_ROOTS_STRICT_ON = ["engine", "references"]
+
+# Directory names skipped during scan (third-party code, build artefacts).
+# `tests` is excluded because test fixtures legitimately use concrete hosts and
+# IP literals (e.g. SSRF/redirect cases, per-host session keys) — same exemption
+# rationale as SKILL.md examples; tests are not the generic fetch path.
+EXCLUDED_DIR_NAMES = {
+    "node_modules", "__pycache__", ".git", ".venv", "dist", "build", "tests",
+}
+
+# Comment markers within which a brand mention is OK (explanation).
+# Keyed per-extension; any line containing these is skipped.
+COMMENT_OK_MARKERS = {
+    ".py": ("# NOTE-BIAS-OK", "# EXAMPLE-ONLY"),
+    ".js": ("// NOTE-BIAS-OK", "// EXAMPLE-ONLY"),
+    ".yaml": ("# NOTE-BIAS-OK", "# EXAMPLE-ONLY"),
+    ".yml": ("# NOTE-BIAS-OK", "# EXAMPLE-ONLY"),
+    ".md": ("<!-- NOTE-BIAS-OK -->", "<!-- EXAMPLE-ONLY -->"),
+}
+
+# File paths explicitly exempted (full match against relative path from scan root).
+EXPLICIT_ALLOW_FILES = {
+    # Phase 0 official-API router. Per SKILL.md R5, naming platform hosts here is
+    # the SANCTIONED exception — these are official no-auth public endpoints, not
+    # a bias toward one target. This is the ONLY engine/ file allowed to do so;
+    # keeping it isolated is precisely why the rest of engine/ stays site-agnostic.
+    # NOTE: rel paths are computed against skill_root.parent, so they include the
+    # skill dir name (e.g. "insane-search/engine/phase0.py").
+    "insane-search/engine/phase0.py",
+}
+
+
+def _line_is_exempt(line: str, ext: str) -> bool:
+    markers = COMMENT_OK_MARKERS.get(ext, ())
+    return any(m in line for m in markers)
+
+
+def _scan_file(path: Path, root: Path) -> list[str]:
+    """Return list of violation strings for this file."""
+    rel = path.relative_to(root.parent)
+    if str(rel) in EXPLICIT_ALLOW_FILES:
+        return []
+
+    ext = path.suffix.lower()
+    try:
+        text = path.read_text(encoding="utf-8", errors="ignore")
+    except Exception as e:
+        return [f"{rel}:0 — read error: {e}"]
+
+    violations: list[str] = []
+    for lineno, line in enumerate(text.splitlines(), start=1):
+        if _line_is_exempt(line, ext):
+            continue
+        lowered = line.lower()
+        # 1) Brand / domain denylist
+        hit_brand = None
+        for brand in BRAND_SUBSTRINGS:
+            if brand.lower() in lowered:
+                hit_brand = brand
+                break
+        if hit_brand:
+            violations.append(f"{rel}:{lineno} — brand `{hit_brand}` in: {line.strip()[:120]}")
+            continue  # one violation per line
+        # 2) URL/domain regex scan — catches hosts that aren't in the denylist.
+        for match in URL_PATTERN.finditer(line):
+            host = match.group(0).lower()
+            host = host.split("//", 1)[-1].split("/", 1)[0]
+            if host in URL_ALLOWLIST:
+                continue
+            if host.endswith(".example.com") or host.endswith(".example.org"):
+                continue
+            violations.append(f"{rel}:{lineno} — hardcoded host `{host}` in: {line.strip()[:120]}")
+            break
+    return violations
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Scan engine for site-name bias")
+    parser.add_argument("--strict", action="store_true",
+                        help="Also scan references/*.md (usually noisy — off by default)")
+    parser.add_argument("--root", default=None,
+                        help="Skill root directory. Defaults to parent of this file.")
+    args = parser.parse_args(argv)
+
+    skill_root = Path(args.root) if args.root else Path(__file__).parent.parent
+    scan_roots = SCAN_ROOTS_STRICT_ON if args.strict else SCAN_ROOTS_STRICT_OFF
+
+    total_violations: list[str] = []
+    scanned = 0
+    for name in scan_roots:
+        root = skill_root / name
+        if not root.exists():
+            continue
+        for dirpath, dirnames, filenames in os.walk(root):
+            # In-place filter so os.walk skips these subtrees.
+            dirnames[:] = [d for d in dirnames if d not in EXCLUDED_DIR_NAMES]
+            for fname in filenames:
+                p = Path(dirpath) / fname
+                if p.suffix.lower() not in (".py", ".js", ".yaml", ".yml", ".md", ".ts", ".mjs"):
+                    continue
+                if p.name == "bias_check.py":
+                    continue  # self-exempt (this file lists the brands)
+                scanned += 1
+                total_violations.extend(_scan_file(p, skill_root))
+
+    print(f"[bias-check] scanned {scanned} files under {skill_root}")
+    if total_violations:
+        print(f"[bias-check] ❌ {len(total_violations)} violation(s):")
+        for v in total_violations:
+            print(f"  - {v}")
+        print()
+        print("Fix options:")
+        print("  1) Remove the brand name (preferred)")
+        print("  2) If genuinely explanatory, add '# NOTE-BIAS-OK' on the same line")
+        print("  3) If this is a Phase 0 official API reference, move it to references/*.md and rerun without --strict")
+        return 1
+
+    print("[bias-check] ✅ clean")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())