pantheon-opencode 1.5.2 → 1.6.0-beta.2

package/src/skills/README.md

@@ -4,7 +4,7 @@
 
 Skills are reference modules loaded on demand by agents. Each skill provides domain-specific knowledge — architecture patterns, security checklists, optimization strategies, and implementation guides. Agents load skills when their task matches the skill's description, keeping context focused and token-efficient.
 
-There are **21 skills** divided into **5 domains**.
+There are **20 skills** divided into **5 domains**.
 
 ## Skills by Domain
 
@@ -30,75 +30,71 @@ There are **21 skills** divided into **5 domains**.
    - Memory bank rules, structure, and optimization — complete guide to Pantheon memory management.
    - Used by: Mnemosyne, Athena
 
-6. **orchestration-workflow** — `orchestration-workflow/SKILL.md`
-   - Practical step-by-step walkthrough for orchestrating features end-to-end using the multi-agent system, from planning through deployment.
-   - Used by: Zeus
-
-7. **session-goal** — `session-goal/SKILL.md`
+6. **session-goal** — `session-goal/SKILL.md`
    - Pin session objectives to prevent scope creep across long multi-agent sessions.
    - Used by: Zeus, Mnemosyne
 
 ### Domain 2: Development Tools
 
-8. **clonedeps** — `clonedeps/SKILL.md`
+7. **clonedeps** — `clonedeps/SKILL.md`
    - Clones dependency source locally so agents can inspect library internals, understand behavior, and debug integration issues.
    - Used by: Hermes, Hephaestus
 
-9. **git-workflow-and-versioning** — `git-workflow-and-versioning/SKILL.md`
+8. **git-workflow-and-versioning** — `git-workflow-and-versioning/SKILL.md`
    - Atomic commits, conventional commits, trunk-based development workflow.
    - Used by: Hermes, Prometheus, Iris
 
-10. **incremental-implementation** — `incremental-implementation/SKILL.md`
+9. **incremental-implementation** — `incremental-implementation/SKILL.md`
     - Implement in thin vertical slices — one commit per task, testable, rollback-safe.
     - Used by: Zeus, Hermes, Aphrodite, Demeter, Prometheus, Talos
 
-11. **loop-engineering** — `loop-engineering/SKILL.md`
+10. **loop-engineering** — `loop-engineering/SKILL.md`
     - Iterative refinement pattern for complex problems — solve, review, improve, repeat in controlled cycles.
     - Used by: All agents
 
-12. **reflect** — `reflect/SKILL.md`
+11. **reflect** — `reflect/SKILL.md`
     - Turns repeated workflow friction into reusable skills, agents, or configuration improvements.
     - Used by: All agents
 
-13. **simplify** — `simplify/SKILL.md`
+12. **simplify** — `simplify/SKILL.md`
     - Behavior-preserving code simplification for readability, maintainability, and reduced complexity.
     - Used by: All agents
 
-14. **worktrees** — `worktrees/SKILL.md`
+13. **worktrees** — `worktrees/SKILL.md`
     - Git worktrees as safe, isolated coding lanes for risky or parallel work without affecting the main workspace.
     - Used by: All agents
 
 ### Domain 3: Quality & Security
 
-15. **code-review-checklist** — `code-review-checklist/SKILL.md`
+14. **code-review-checklist** — `code-review-checklist/SKILL.md`
     - Systematic code review with quality gates, security audit, and parallel checks for structured feedback.
     - Used by: Themis
 
-16. **security-hardening** — `security-hardening/SKILL.md`
+15. **security-hardening** — `security-hardening/SKILL.md`
     - Security gate — OWASP, SAST, secrets detection, MCP hardening, dependency audit.
     - Used by: Themis, Nyx
 
-17. **tdd-with-agents** — `tdd-with-agents/SKILL.md`
+16. **tdd-with-agents** — `tdd-with-agents/SKILL.md`
     - TDD enforcement with RED→GREEN→REFACTOR cycle and advanced testing patterns across all layers.
     - Used by: Hermes, Aphrodite, Demeter, Themis, Hephaestus
 
 ### Domain 4: Planning
 
-18. **codemap** — `codemap/SKILL.md`
+17. **codemap** — `codemap/SKILL.md`
     - Generates hierarchical repository maps so agents understand codebase structure without re-reading every file.
     - Used by: Apollo, Zeus
 
-19. **spec-driven-development** — `spec-driven-development/SKILL.md`
+18. **spec-driven-development** — `spec-driven-development/SKILL.md`
     - Define requirements via spec-first before any code — PRD, edge cases, acceptance criteria.
     - Used by: Athena
 
-20. **verification-planning** — `verification-planning/SKILL.md`
+19. **verification-planning** — `verification-planning/SKILL.md`
     - Plans an evidence path before non-trivial changes, ensuring changes are testable and verifiable.
     - Used by: Athena, Zeus
 
 ### Domain 5: Frontend Development
 
-21. **visual-review-pipeline** — `visual-review-pipeline/SKILL.md`
+20. **visual-review-pipeline** — `visual-review-pipeline/SKILL.md`
     - Automated visual review pipeline — Playwright screenshots, self-analysis, fix loop, escalation.
     - Used by: Aphrodite
 
@@ -130,22 +126,21 @@ The `SKILL.md` file contains YAML frontmatter with `name`, `description`, and pl
 | 3 | auto-continue | Orchestration & Workflow | Zeus, Apollo, Hephaestus, Nyx, Gaia |
 | 4 | context-compression | Orchestration & Workflow | Zeus, Mnemosyne, Themis |
 | 5 | memory-bank | Orchestration & Workflow | Mnemosyne, Athena |
-| 6 | orchestration-workflow | Orchestration & Workflow | Zeus |
-| 7 | session-goal | Orchestration & Workflow | Zeus, Mnemosyne |
-| 8 | clonedeps | Development Tools | Hermes, Hephaestus |
-| 9 | git-workflow-and-versioning | Development Tools | Hermes, Prometheus, Iris |
-| 10 | incremental-implementation | Development Tools | Zeus, Hermes, Aphrodite, Demeter, Prometheus, Talos |
-| 11 | loop-engineering | Development Tools | All agents |
-| 12 | reflect | Development Tools | All agents |
-| 13 | simplify | Development Tools | All agents |
-| 14 | worktrees | Development Tools | All agents |
-| 15 | code-review-checklist | Quality & Security | Themis |
-| 16 | security-hardening | Quality & Security | Themis, Nyx |
-| 17 | tdd-with-agents | Quality & Security | Hermes, Aphrodite, Demeter, Themis, Hephaestus |
-| 18 | codemap | Planning | Apollo, Zeus |
-| 19 | spec-driven-development | Planning | Athena |
-| 20 | verification-planning | Planning | Athena, Zeus |
-| 21 | visual-review-pipeline | Frontend Development | Aphrodite |
+| 6 | session-goal | Orchestration & Workflow | Zeus, Mnemosyne |
+| 7 | clonedeps | Development Tools | Hermes, Hephaestus |
+| 8 | git-workflow-and-versioning | Development Tools | Hermes, Prometheus, Iris |
+| 9 | incremental-implementation | Development Tools | Zeus, Hermes, Aphrodite, Demeter, Prometheus, Talos |
+| 10 | loop-engineering | Development Tools | All agents |
+| 11 | reflect | Development Tools | All agents |
+| 12 | simplify | Development Tools | All agents |
+| 13 | worktrees | Development Tools | All agents |
+| 14 | code-review-checklist | Quality & Security | Themis |
+| 15 | security-hardening | Quality & Security | Themis, Nyx |
+| 16 | tdd-with-agents | Quality & Security | Hermes, Aphrodite, Demeter, Themis, Hephaestus |
+| 17 | codemap | Planning | Apollo, Zeus |
+| 18 | spec-driven-development | Planning | Athena |
+| 19 | verification-planning | Planning | Athena, Zeus |
+| 20 | visual-review-pipeline | Frontend Development | Aphrodite |
 
 ---
 

package/src/skills/auto-continue/SKILL.md

@@ -43,7 +43,7 @@ Disciplined automatic continuation through multi-step tasks. Eliminates unnecess
 - A todo requires a decision not covered by the plan
 - An unexpected error changes the approach
 - A dependency is missing or broken
-- A task would exceed remaining `steps` budget
+- Remaining context is too thin to carry the task through — write an explicit handoff instead
 
 ---
 
@@ -70,7 +70,7 @@ Disciplined automatic continuation through multi-step tasks. Eliminates unnecess
 - [ ] Previous step completed successfully (tests pass, no errors)
 - [ ] Next step is within approved plan scope
 - [ ] No new blocking issues emerged
-- [ ] `steps` counter has sufficient budget remaining
+- [ ] Remaining context can carry the next step, or an explicit handoff will be written
 
 ---
 
@@ -104,27 +104,18 @@ Only expand to full cooldown when hitting a mandatory gate.
 
 ## Timeout & Retry Enforcement
 
-When a delegated agent does not respond in time, enforce the timeout policy from routing.yml.
-
-### Behavior by Agent Role
-
-| Agent Role | Timeout | Retry Policy | Fallback | Timeout Parcial? |
-|------------|---------|-------------|----------|------------------|
-| Explorer (apollo) | 60s | 2 retries, exponential backoff | athena | ✅ Yes — partial results OK |
-| Implementer (hermes, aphrodite) | 180s | 3 retries, exponential backoff | talos | ❌ No — must produce artifact |
-| Reviewer (themis) | 120s | 2 retries, exponential backoff | zeus | ❌ No — must produce verdict |
-| Infrastructure (prometheus) | 300s | 2 retries, exponential backoff | hermes | ❌ No |
-| Hotfix (talos) | 30s | 1 retry, no backoff | hermes | ✅ Yes — one-liner fix OK |
+When a delegated agent does not respond in time, enforce the timeout policy defined in
+`src/instructions/zeus-timeout-retry.instructions.md`.
 
+**That file is the single authority** for per-role timeouts, retry counts, fallback chains, and
+whether partial results are acceptable. It is force-fed into every session via `AGENTS.md`; this
+skill is opt-in. Do not restate those values here — a copy in this skill is precisely what drifted
+out of sync with them. If a dispatch times out, read the role's row from the instruction file.
 
 ### Timeout Parcial (Partial Results)
 
-Timeout parcial is **only** allowed for read-only, independent agents:
-- ✅ @apollo (codebase search) — can return partial file list
-- ✅ @gaia (literature review) — can return partial findings
-- ✅ @talos (hotfix confirmation) — can return without fix if timeout
-- ❌ Never for implementers (@hermes, @aphrodite, @demeter) — must complete or fail
-- ❌ Never for reviewers (@themis) — must produce a verdict
+Whether an agent may return partial results is the **Partial Results OK?** column in that table.
+Only agents marked ✅ there may return partial — never implementers, never reviewers.
 
 **How to signal timeout parcial:**
 When dispatching, set the expectation explicitly:
@@ -167,69 +158,28 @@ Ready for Phase 2? Waiting for go-ahead.
 Should I now run the migration tests? [waiting]
 ```
 
-## Session Heartbeat
-
-Auto-save checkpoint every N turns (configurable, default 5).
-
-### Heartbeat File
-
-`.pantheon/deepwork/<slug>/heartbeat.json`
+## Session Heartbeat & Checkpoints
 
-### Heartbeat JSON Format
+Session state is **not** file-based. It lives in **pantheon-persistence** under the
+`checkpoint:<slug>` namespace and is cleaned up automatically by a 4h TTL.
 
-```json
-{
-  "slug": "task-identifier",
-  "last_action": "2026-07-16T20:00:00Z",
-  "turn_count": 10,
-  "current_phase": 2,
-  "status": "alive | warning | stalled | paused | completed",
-  "next_action": "brief description of next planned action"
-}
-```
-
-### Heartbeat Rules
-1. Update every 5 turns minimum (by default)
-2. Contains NO context — only timestamps and counters
-3. Single file, always overwritten (not versioned)
-4. Used for quick "is this session alive?" checks
-
----
+Write state with `context_save(slug, key, value, session_id=...)` and read it back with
+`context_get(slug, "latest", session_id=...)`. Capture `session_id` from the first `context_save`
+of the session and reuse it for the rest of the session, so the `latest` pointer survives
+compaction.
 
-## Checkpoint Persistence
-
-### Checkpoint File Structure
-
-Each deepwork task maintains:
-
-```
-.pantheon/deepwork/<slug>/
-├── PLAN.md                 # Immutable plan (created at start)
-├── STATUS.md               # Human-readable current state (updated every phase)
-├── heartbeat.json          # Lightweight ping (updated every N turns)
-├── checkpoint-<N>.json     # Full state snapshot (created at phase boundaries)
-├── session.json            # Session metadata (created at start, updated on stop)
-└── REVIEW.md               # Final Themis review (created at end)
-```
-
-### Checkpoint JSON Schema
-
-```json
-{
-  "slug": "string — unique task identifier",
-  "phase": "integer — current phase number",
-  "turn_count": "integer — total turns elapsed",
-  "timestamp": "string — ISO 8601 timestamp",
-  "context_hash": "string — first 8 hex chars of SHA-256 of STATUS.md",
-  "version": "integer — schema version (current: 2)"
-}
-```
+### Heartbeat Check
+- If `context_get(slug, "heartbeat")` returns a checkin older than 300s, log a stall warning and resume
+- Write a heartbeat after every anti-stall recovery action
 
 ### Checkpoint Rules
-1. Save at every phase boundary AND before any delegate dispatch
-2. Numbered sequentially (`checkpoint-1.json`, `checkpoint-2.json`, …)
-3. Keep last 10 checkpoints; archive older ones
-4. On resume: read the highest-numbered checkpoint for full state
+1. Save a checkpoint before **every** delegate dispatch and after every agent returns
+2. Include current phase, turn count, and remaining tasks
+3. On resume: `context_get(slug, "latest")` restores the most recent checkpoint
+
+> Do not create `heartbeat.json`, `checkpoint-<N>.json`, or `session.json` under
+> `.pantheon/deepwork/`. That file-based mechanism is retired — TTL handles cleanup, and reading
+> a file that nothing writes back returns stale state.
 
 ---
 

package/.pantheon/code-mode/eval-llm-judge.py

@@ -1,255 +0,0 @@
-#!/usr/bin/env python3
-# ---
-# description: LLM judge for skill/agent quality — scores 4 dimensions (0-100) via an OpenAI-compatible endpoint
-# timeout: 120
-# ---
-"""LLM judge layer of the plugin-eval certification pipeline.
-
-Scores a skill/agent directory on 4 dimensions (correctness, maintainability,
-security, practicality), each 0-100, via ONE structured prompt to an
-OpenAI-compatible chat endpoint. Emits a single JSON document on stdout:
-
-    {name, dimensions: {correctness, maintainability, security, practicality},
-     overall, notes: {<dimension>: <justification>}}
-
-Without explicit opt-in it emits a structured skipped result; missing credentials
-or failed calls exit 2 with a clear message.
-
-Env vars:
-  OPENAI_API_KEY    required bearer token
-  OPENAI_BASE_URL   optional endpoint base (default https://api.openai.com/v1)
-  EVAL_JUDGE_MODEL  optional model id (default gpt-4o-mini)
-
-Stdlib only (urllib.request) — no third-party dependencies.
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import re
-import sys
-import urllib.error
-import urllib.request
-from pathlib import Path
-from typing import Any
-from urllib.parse import urlparse
-
-DEFAULT_MODEL = "gpt-4o-mini"
-DEFAULT_BASE_URL = "https://api.openai.com/v1"
-REQUEST_TIMEOUT = 90.0
-MAX_CONTENT_CHARS = 12_000
-MAX_SUPPORTING_CHARS = 8_000
-MAX_SUPPORTING_FILES = 6
-DIMENSIONS = ("correctness", "maintainability", "security", "practicality")
-
-JUDGE_PROMPT = """You are a rigorous quality auditor for AI agent skills and agent definitions used in an agentic coding framework. You decide whether a skill or agent definition is production-quality or AI slop.
-
-Apply an explicit ANTI-SLOP bias: flag overengineering, unnecessary abstractions, persona proliferation (multiple fake personas/characters), unverifiable claims, hype language, and instructions that sound impressive but are not actionable.
-
-Score the content on 4 dimensions, each 0-100:
-1. correctness — instructions are accurate and internally consistent; no contradictions; no broken references (files, commands, or agents that do not exist or are misnamed).
-2. maintainability — clear structure, single responsibility, no overengineering, YAGNI-respecting, easy to update.
-3. security — no secrets or credentials; no prompt-injection vectors; no dangerous instructions (e.g. rm -rf, curl | bash, disabling safety checks).
-4. practicality — solves a real problem, actionable, not AI slop or hype; a human would actually use it.
-
-Return ONLY a JSON object with this exact shape:
-{"correctness": <int 0-100>, "maintainability": <int 0-100>, "security": <int 0-100>, "practicality": <int 0-100>, "notes": {"correctness": "<1-2 sentences>", "maintainability": "<1-2 sentences>", "security": "<1-2 sentences>", "practicality": "<1-2 sentences>"}}
-
-Content to evaluate:
---- BEGIN CONTENT ---
-{content}
---- END CONTENT ---
-"""
-
-_NAME_RE = re.compile(r"^name:\s*[\"']?([^\"'\n]+?)[\"']?\s*$", re.MULTILINE)
-
-
-class JudgeError(Exception):
-    """Expected, user-facing judge failure."""
-
-
-def _frontmatter_name(text: str) -> str | None:
-    """Extract the frontmatter `name` value, or None when absent."""
-    if not text.startswith("---"):
-        return None
-    end = text.find("\n---", 3)
-    if end == -1:
-        return None
-    match = _NAME_RE.search(text[3:end])
-    return match.group(1).strip() if match else None
-
-
-def _discover_doc(path: Path) -> Path:
-    """Return the primary doc (SKILL.md or agent .md) for a path."""
-    if path.is_file():
-        return path
-    skill = path / "SKILL.md"
-    if skill.is_file():
-        return skill
-    candidates = sorted(path.glob("*.md"))
-    for candidate in candidates:
-        try:
-            text = candidate.read_text(encoding="utf-8", errors="replace")
-        except OSError:
-            continue
-        if text.startswith("---") and text.find("\n---", 3) != -1:
-            return candidate
-    if candidates:
-        return candidates[0]
-    raise JudgeError(f"No SKILL.md or .md file found in {path}")
-
-
-def _supporting_files(doc: Path) -> list[Path]:
-    """Script files next to the doc or in its scripts/ subdir (capped)."""
-    roots = [doc.parent]
-    scripts_dir = doc.parent / "scripts"
-    if scripts_dir.is_dir():
-        roots.append(scripts_dir)
-    found: list[Path] = []
-    for root in roots:
-        for pattern in ("*.py", "*.sh", "*.js", "*.ts", "*.mjs"):
-            found.extend(sorted(root.glob(pattern)))
-    return [p for p in found if p.is_file() and p != doc][:MAX_SUPPORTING_FILES]
-
-
-def _load_content(doc: Path) -> tuple[str, str]:
-    """Return (name, content) with the doc plus small supporting files."""
-    try:
-        text = doc.read_text(encoding="utf-8", errors="replace")
-    except OSError as exc:
-        raise JudgeError(f"Cannot read {doc}: {exc}") from exc
-    name = _frontmatter_name(text) or doc.stem
-    parts = [f"# {name}\n", text[:MAX_CONTENT_CHARS]]
-    if len(text) > MAX_CONTENT_CHARS:
-        parts.append(f"\n[... truncated {len(text) - MAX_CONTENT_CHARS} chars ...]")
-    support: list[str] = []
-    total = 0
-    for sibling in _supporting_files(doc):
-        try:
-            data = sibling.read_text(encoding="utf-8", errors="replace")
-        except OSError:
-            continue
-        chunk = f"\n--- supporting file: {sibling.name} ---\n{data[:2000]}"
-        if total + len(chunk) > MAX_SUPPORTING_CHARS:
-            break
-        support.append(chunk)
-        total += len(chunk)
-    if support:
-        parts.append("\n".join(support))
-    return name, "".join(parts)
-
-
-def _call_llm(prompt: str, allow_external_llm: bool = False) -> str:
-    """Send one chat completion request via urllib; return the assistant text."""
-    if not allow_external_llm:
-        raise JudgeError("External LLM calls require --allow-external-llm or PANTHEON_ALLOW_EXTERNAL_LLM=1")
-    key = os.getenv("OPENAI_API_KEY")
-    if not key:
-        raise JudgeError(
-            "No LLM available: set OPENAI_API_KEY "
-            "(optionally OPENAI_BASE_URL and EVAL_JUDGE_MODEL)."
-        )
-    base_url = os.getenv("OPENAI_BASE_URL", DEFAULT_BASE_URL).rstrip("/")
-    parsed_url = urlparse(base_url)
-    if parsed_url.scheme not in {"http", "https"} or not parsed_url.netloc:
-        raise JudgeError("OPENAI_BASE_URL must use an http or https URL")
-    endpoint = base_url + "/chat/completions"
-    model = os.getenv("EVAL_JUDGE_MODEL") or DEFAULT_MODEL
-    body = {
-        "model": model,
-        "messages": [{"role": "user", "content": prompt}],
-        "temperature": 0.2,
-    }
-    request = urllib.request.Request(
-        endpoint,
-        data=json.dumps(body).encode("utf-8"),
-        headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
-        method="POST",
-    )
-    try:
-        with urllib.request.urlopen(request, timeout=REQUEST_TIMEOUT) as response:
-            payload = json.loads(response.read().decode("utf-8"))
-    except urllib.error.HTTPError as exc:
-        detail = exc.read().decode("utf-8", errors="replace")[:200].replace(key, "[redacted]")
-        raise JudgeError(f"LLM endpoint returned HTTP {exc.code}: {detail}") from exc
-    except (urllib.error.URLError, TimeoutError, OSError) as exc:
-        reason = getattr(exc, "reason", exc)
-        raise JudgeError(f"LLM request failed: {reason}") from exc
-    except (json.JSONDecodeError, UnicodeDecodeError) as exc:
-        raise JudgeError(f"LLM returned invalid JSON payload: {exc}") from exc
-    choices = payload.get("choices") if isinstance(payload, dict) else None
-    message = choices[0].get("message") if choices else None
-    text = message.get("content") if isinstance(message, dict) else None
-    if not isinstance(text, str) or not text.strip():
-        raise JudgeError("LLM returned an empty response.")
-    return text.replace(key, "[redacted]")
-
-
-def _parse_json_object(value: str) -> dict[str, Any] | None:
-    """Parse strict JSON and the fenced JSON commonly returned by models."""
-    candidates = [value.strip()]
-    if candidates[0].startswith("```") and candidates[0].endswith("```"):
-        candidates.append(candidates[0].split("\n", 1)[-1][:-3].strip())
-    for candidate in candidates:
-        try:
-            parsed = json.loads(candidate)
-        except json.JSONDecodeError:
-            continue
-        if isinstance(parsed, dict):
-            return parsed
-    return None
-
-
-def _parse_scores(text: str) -> dict[str, Any]:
-    """Validate the LLM JSON into the report's dimensions/overall/notes."""
-    parsed = _parse_json_object(text)
-    if parsed is None:
-        raise JudgeError("LLM did not return valid JSON.")
-    dimensions: dict[str, int] = {}
-    for dim in DIMENSIONS:
-        raw = parsed.get(dim)
-        try:
-            value = int(raw)
-        except (TypeError, ValueError) as exc:
-            raise JudgeError(f"LLM returned non-integer score for {dim}: {raw!r}") from exc
-        if not 0 <= value <= 100:
-            raise JudgeError(f"LLM returned out-of-range score for {dim}: {value}")
-        dimensions[dim] = value
-    raw_notes = parsed.get("notes")
-    notes = (
-        {dim: str(raw_notes.get(dim, "")) for dim in DIMENSIONS}
-        if isinstance(raw_notes, dict)
-        else {dim: "" for dim in DIMENSIONS}
-    )
-    overall = round(sum(dimensions.values()) / len(dimensions), 1)
-    return {"dimensions": dimensions, "overall": overall, "notes": notes}
-
-
-def main(argv: list[str] | None = None) -> int:
-    """CLI entry point: `python eval-llm-judge.py <skill-or-agent-dir>`."""
-    parser = argparse.ArgumentParser(
-        description="LLM judge for skill/agent quality (plugin-eval layer 2)."
-    )
-    parser.add_argument("path", help="Path to a skill/agent directory or doc file")
-    parser.add_argument("--allow-external-llm", action="store_true", help="Permit sending content to the configured LLM endpoint")
-    args = parser.parse_args(argv)
-    allow_external_llm = args.allow_external_llm or os.getenv("PANTHEON_ALLOW_EXTERNAL_LLM") == "1"
-    try:
-        doc = _discover_doc(Path(args.path))
-        name, content = _load_content(doc)
-        if not allow_external_llm:
-            print(json.dumps({"name": name, "skipped": True, "reason": "external LLM calls not opted in"}, ensure_ascii=False, indent=2))
-            return 0
-        raw = _call_llm(JUDGE_PROMPT.replace("{content}", content), allow_external_llm=True)
-        result = _parse_scores(raw)
-    except JudgeError as exc:
-        print(f"eval-llm-judge: {exc}", file=sys.stderr)
-        return 2
-    print(json.dumps({"name": name, **result}, ensure_ascii=False, indent=2))
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())