@rudderhq/agent-runtime-gemini-local 0.2.1 → 0.2.2-canary.0

package/skills/conversation-to-skill/scripts/run_loop.py

@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+"""Run the eval + improve loop until all pass or max iterations reached."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import random
+import sys
+import tempfile
+import time
+import webbrowser
+from pathlib import Path
+
+if __package__ in (None, ""):
+    sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from scripts.generate_report import generate_html
+from scripts.improve_description import improve_description
+from scripts.model_backends import detect_backend
+from scripts.run_eval import find_project_root, run_eval
+from scripts.utils import parse_skill_md
+
+
+def split_eval_set(eval_set: list[dict], holdout: float, seed: int = 42) -> tuple[list[dict], list[dict]]:
+    random.seed(seed)
+    trigger = [item for item in eval_set if item["should_trigger"]]
+    no_trigger = [item for item in eval_set if not item["should_trigger"]]
+    random.shuffle(trigger)
+    random.shuffle(no_trigger)
+    n_trigger_test = max(1, int(len(trigger) * holdout))
+    n_no_trigger_test = max(1, int(len(no_trigger) * holdout))
+    test_set = trigger[:n_trigger_test] + no_trigger[:n_no_trigger_test]
+    train_set = trigger[n_trigger_test:] + no_trigger[n_no_trigger_test:]
+    return train_set, test_set
+
+
+def run_loop(
+    *,
+    eval_set: list[dict],
+    skill_path: Path,
+    description_override: str | None,
+    num_workers: int,
+    timeout: int,
+    max_iterations: int,
+    runs_per_query: int,
+    trigger_threshold: float,
+    holdout: float,
+    model: str | None,
+    backend: str,
+    verbose: bool,
+    live_report_path: Path | None = None,
+    log_dir: Path | None = None,
+) -> dict:
+    project_root = find_project_root()
+    name, original_description, content = parse_skill_md(skill_path)
+    current_description = description_override or original_description
+    backend = detect_backend(backend)
+
+    if holdout > 0:
+        train_set, test_set = split_eval_set(eval_set, holdout)
+        if verbose:
+            print(f"Split: {len(train_set)} train, {len(test_set)} test (holdout={holdout})", file=sys.stderr)
+    else:
+        train_set = eval_set
+        test_set = []
+
+    history = []
+    exit_reason = "unknown"
+
+    for iteration in range(1, max_iterations + 1):
+        if verbose:
+            print(f"\n{'=' * 60}", file=sys.stderr)
+            print(f"Iteration {iteration}/{max_iterations}", file=sys.stderr)
+            print(f"Backend: {backend}", file=sys.stderr)
+            print(f"Description: {current_description}", file=sys.stderr)
+            print(f"{'=' * 60}", file=sys.stderr)
+
+        all_queries = train_set + test_set
+        t0 = time.time()
+        all_results = run_eval(
+            eval_set=all_queries,
+            skill_name=name,
+            description=current_description,
+            num_workers=num_workers,
+            timeout=timeout,
+            project_root=project_root,
+            runs_per_query=runs_per_query,
+            trigger_threshold=trigger_threshold,
+            model=model,
+            backend=backend,
+        )
+        eval_elapsed = time.time() - t0
+
+        train_queries = {item["query"] for item in train_set}
+        train_result_list = [item for item in all_results["results"] if item["query"] in train_queries]
+        test_result_list = [item for item in all_results["results"] if item["query"] not in train_queries]
+
+        train_passed = sum(1 for item in train_result_list if item["pass"])
+        train_total = len(train_result_list)
+        train_summary = {"passed": train_passed, "failed": train_total - train_passed, "total": train_total}
+        train_results = {"results": train_result_list, "summary": train_summary}
+
+        if test_set:
+            test_passed = sum(1 for item in test_result_list if item["pass"])
+            test_total = len(test_result_list)
+            test_summary = {"passed": test_passed, "failed": test_total - test_passed, "total": test_total}
+            test_results = {"results": test_result_list, "summary": test_summary}
+        else:
+            test_summary = None
+            test_results = None
+
+        history.append({
+            "iteration": iteration,
+            "description": current_description,
+            "train_passed": train_summary["passed"],
+            "train_failed": train_summary["failed"],
+            "train_total": train_summary["total"],
+            "train_results": train_results["results"],
+            "test_passed": test_summary["passed"] if test_summary else None,
+            "test_failed": test_summary["failed"] if test_summary else None,
+            "test_total": test_summary["total"] if test_summary else None,
+            "test_results": test_results["results"] if test_results else None,
+            "passed": train_summary["passed"],
+            "failed": train_summary["failed"],
+            "total": train_summary["total"],
+            "results": train_results["results"],
+        })
+
+        if live_report_path:
+            partial_output = {
+                "backend": backend,
+                "original_description": original_description,
+                "best_description": current_description,
+                "best_score": "in progress",
+                "iterations_run": len(history),
+                "holdout": holdout,
+                "train_size": len(train_set),
+                "test_size": len(test_set),
+                "history": history,
+            }
+            live_report_path.write_text(generate_html(partial_output, auto_refresh=True, skill_name=name))
+
+        if verbose:
+            def print_eval_stats(label: str, results: list[dict], elapsed: float):
+                pos = [item for item in results if item["should_trigger"]]
+                neg = [item for item in results if not item["should_trigger"]]
+                tp = sum(item["triggers"] for item in pos)
+                pos_runs = sum(item["runs"] for item in pos)
+                fn = pos_runs - tp
+                fp = sum(item["triggers"] for item in neg)
+                neg_runs = sum(item["runs"] for item in neg)
+                tn = neg_runs - fp
+                total = tp + tn + fp + fn
+                precision = tp / (tp + fp) if (tp + fp) > 0 else 1.0
+                recall = tp / (tp + fn) if (tp + fn) > 0 else 1.0
+                accuracy = (tp + tn) / total if total > 0 else 0.0
+                print(
+                    f"{label}: {tp + tn}/{total} correct, precision={precision:.0%} "
+                    f"recall={recall:.0%} accuracy={accuracy:.0%} ({elapsed:.1f}s)",
+                    file=sys.stderr,
+                )
+                for item in results:
+                    status = "PASS" if item["pass"] else "FAIL"
+                    rate_str = f"{item['triggers']}/{item['runs']}"
+                    print(f"  [{status}] rate={rate_str} expected={item['should_trigger']}: {item['query'][:60]}", file=sys.stderr)
+
+            print_eval_stats("Train", train_results["results"], eval_elapsed)
+            if test_summary:
+                print_eval_stats("Test ", test_results["results"], 0)
+
+        if train_summary["failed"] == 0:
+            exit_reason = f"all_passed (iteration {iteration})"
+            if verbose:
+                print(f"\nAll train queries passed on iteration {iteration}.", file=sys.stderr)
+            break
+
+        if iteration == max_iterations:
+            exit_reason = f"max_iterations ({max_iterations})"
+            if verbose:
+                print(f"\nMax iterations reached ({max_iterations}).", file=sys.stderr)
+            break
+
+        if verbose:
+            print("\nImproving description...", file=sys.stderr)
+
+        t0 = time.time()
+        blinded_history = [
+            {k: v for k, v in item.items() if not k.startswith("test_")}
+            for item in history
+        ]
+        new_description = improve_description(
+            backend=backend,
+            skill_name=name,
+            skill_content=content,
+            current_description=current_description,
+            eval_results=train_results,
+            history=blinded_history,
+            model=model,
+            log_dir=log_dir,
+            iteration=iteration,
+        )
+        improve_elapsed = time.time() - t0
+
+        if verbose:
+            print(f"Proposed ({improve_elapsed:.1f}s): {new_description}", file=sys.stderr)
+
+        current_description = new_description
+
+    if test_set:
+        best = max(history, key=lambda item: item["test_passed"] or 0)
+        best_score = f"{best['test_passed']}/{best['test_total']}"
+    else:
+        best = max(history, key=lambda item: item["train_passed"])
+        best_score = f"{best['train_passed']}/{best['train_total']}"
+
+    if verbose:
+        print(f"\nExit reason: {exit_reason}", file=sys.stderr)
+        print(f"Best score: {best_score} (iteration {best['iteration']})", file=sys.stderr)
+
+    return {
+        "backend": backend,
+        "exit_reason": exit_reason,
+        "original_description": original_description,
+        "best_description": best["description"],
+        "best_score": best_score,
+        "best_train_score": f"{best['train_passed']}/{best['train_total']}",
+        "best_test_score": f"{best['test_passed']}/{best['test_total']}" if test_set else None,
+        "final_description": current_description,
+        "iterations_run": len(history),
+        "holdout": holdout,
+        "train_size": len(train_set),
+        "test_size": len(test_set),
+        "history": history,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Run eval + improve loop")
+    parser.add_argument("--eval-set", required=True, help="Path to eval set JSON file")
+    parser.add_argument("--skill-path", required=True, help="Path to skill directory")
+    parser.add_argument("--description", default=None, help="Override starting description")
+    parser.add_argument("--num-workers", type=int, default=10, help="Number of parallel workers")
+    parser.add_argument("--timeout", type=int, default=60, help="Timeout per query in seconds")
+    parser.add_argument("--max-iterations", type=int, default=5, help="Max improvement iterations")
+    parser.add_argument("--runs-per-query", type=int, default=3, help="Number of runs per query")
+    parser.add_argument("--trigger-threshold", type=float, default=0.5, help="Trigger rate threshold")
+    parser.add_argument("--holdout", type=float, default=0.4, help="Fraction of eval set to hold out for testing (0 to disable)")
+    parser.add_argument("--model", default=None, help="Optional backend model identifier")
+    parser.add_argument("--backend", default="auto", choices=["auto", "claude", "codex"], help="Optimization backend")
+    parser.add_argument("--verbose", action="store_true", help="Print progress to stderr")
+    parser.add_argument("--report", default="auto", help="Generate HTML report at this path (default: 'auto', 'none' to disable)")
+    parser.add_argument("--results-dir", default=None, help="Save outputs to a timestamped subdirectory here")
+    args = parser.parse_args()
+
+    eval_set = json.loads(Path(args.eval_set).read_text())
+    skill_path = Path(args.skill_path)
+
+    if not (skill_path / "SKILL.md").exists():
+        print(f"Error: No SKILL.md found at {skill_path}", file=sys.stderr)
+        sys.exit(1)
+
+    name, _, _ = parse_skill_md(skill_path)
+
+    if args.report != "none":
+        if args.report == "auto":
+            timestamp = time.strftime("%Y%m%d_%H%M%S")
+            live_report_path = Path(tempfile.gettempdir()) / f"skill_description_report_{skill_path.name}_{timestamp}.html"
+        else:
+            live_report_path = Path(args.report)
+        live_report_path.write_text("<html><body><h1>Starting optimization loop...</h1><meta http-equiv='refresh' content='5'></body></html>")
+        webbrowser.open(str(live_report_path))
+    else:
+        live_report_path = None
+
+    if args.results_dir:
+        timestamp = time.strftime("%Y-%m-%d_%H%M%S")
+        results_dir = Path(args.results_dir) / timestamp
+        results_dir.mkdir(parents=True, exist_ok=True)
+    else:
+        results_dir = None
+
+    log_dir = results_dir / "logs" if results_dir else None
+
+    output = run_loop(
+        eval_set=eval_set,
+        skill_path=skill_path,
+        description_override=args.description,
+        num_workers=args.num_workers,
+        timeout=args.timeout,
+        max_iterations=args.max_iterations,
+        runs_per_query=args.runs_per_query,
+        trigger_threshold=args.trigger_threshold,
+        holdout=args.holdout,
+        model=args.model,
+        backend=args.backend,
+        verbose=args.verbose,
+        live_report_path=live_report_path,
+        log_dir=log_dir,
+    )
+
+    json_output = json.dumps(output, indent=2)
+    print(json_output)
+    if results_dir:
+        (results_dir / "results.json").write_text(json_output)
+
+    if live_report_path:
+        live_report_path.write_text(generate_html(output, auto_refresh=False, skill_name=name))
+        print(f"\nReport: {live_report_path}", file=sys.stderr)
+
+    if results_dir and live_report_path:
+        (results_dir / "report.html").write_text(generate_html(output, auto_refresh=False, skill_name=name))
+
+    if results_dir:
+        print(f"Results saved to: {results_dir}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

package/skills/conversation-to-skill/scripts/utils.py

@@ -0,0 +1,223 @@
+"""Shared utilities for skill-creator scripts."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+
+FRONTMATTER_PATTERN = re.compile(r"^---\s*\n(.*?)\n---\s*(?:\n|$)", re.DOTALL)
+
+
+def extract_frontmatter_text(content: str) -> str:
+    """Return the raw YAML frontmatter text from a SKILL.md file."""
+    match = FRONTMATTER_PATTERN.match(content)
+    if not match:
+        raise ValueError("SKILL.md missing frontmatter (expected opening and closing ---)")
+    return match.group(1)
+
+
+def _count_indent(line: str) -> int:
+    return len(line) - len(line.lstrip(" "))
+
+
+def _parse_scalar(value: str):
+    value = value.strip()
+    if not value:
+        return ""
+
+    if value[0] == value[-1] and value[0] in {'"', "'"}:
+        return value[1:-1]
+
+    lowered = value.lower()
+    if lowered in {"true", "false"}:
+        return lowered == "true"
+    if lowered in {"null", "~"}:
+        return None
+
+    if re.fullmatch(r"-?\d+", value):
+        return int(value)
+    if re.fullmatch(r"-?\d+\.\d+", value):
+        return float(value)
+
+    if value.startswith("[") and value.endswith("]"):
+        inner = value[1:-1].strip()
+        if not inner:
+            return []
+        return [_parse_scalar(part.strip()) for part in inner.split(",")]
+
+    return value
+
+
+def _fold_lines(lines: list[str]) -> str:
+    paragraphs: list[list[str]] = [[]]
+    for line in lines:
+        if line == "":
+            if paragraphs[-1]:
+                paragraphs.append([])
+            continue
+        paragraphs[-1].append(line)
+
+    folded = [" ".join(paragraph).strip() for paragraph in paragraphs if paragraph]
+    return "\n\n".join(part for part in folded if part)
+
+
+def _parse_block_scalar(lines: list[str], start: int, indent: int, style: str):
+    collected: list[str] = []
+    index = start
+    while index < len(lines):
+        raw = lines[index]
+        if not raw.strip():
+            collected.append("")
+            index += 1
+            continue
+
+        current_indent = _count_indent(raw)
+        if current_indent < indent:
+            break
+        collected.append(raw[indent:])
+        index += 1
+
+    if style.startswith("|"):
+        value = "\n".join(collected)
+    else:
+        value = _fold_lines(collected)
+    return value, index
+
+
+def _looks_like_mapping_entry(text: str) -> bool:
+    if text.startswith("- "):
+        return False
+    return bool(re.match(r"^[A-Za-z0-9_-]+:\s*.*$", text))
+
+
+def _next_nonempty_index(lines: list[str], start: int, min_indent: int):
+    index = start
+    while index < len(lines):
+        raw = lines[index]
+        if not raw.strip():
+            index += 1
+            continue
+        indent = _count_indent(raw)
+        if indent < min_indent:
+            return None
+        return index
+    return None
+
+
+def _parse_list(lines: list[str], start: int, indent: int):
+    items = []
+    index = start
+    while index < len(lines):
+        raw = lines[index]
+        if not raw.strip():
+            index += 1
+            continue
+
+        current_indent = _count_indent(raw)
+        if current_indent < indent:
+            break
+        if current_indent != indent or not raw[indent:].startswith("- "):
+            raise ValueError(f"Invalid list entry near line: {raw}")
+
+        remainder = raw[indent + 2 :].strip()
+        index += 1
+
+        if remainder:
+            items.append(_parse_scalar(remainder))
+            continue
+
+        next_index = _next_nonempty_index(lines, index, indent + 2)
+        if next_index is None:
+            items.append("")
+            continue
+
+        nested_indent = _count_indent(lines[next_index])
+        nested_text = lines[next_index][nested_indent:]
+        if nested_text.startswith("- "):
+            value, index = _parse_list(lines, next_index, nested_indent)
+        elif _looks_like_mapping_entry(nested_text):
+            value, index = _parse_mapping(lines, next_index, nested_indent)
+        else:
+            value, index = _parse_block_scalar(lines, next_index, nested_indent, ">")
+        items.append(value)
+
+    return items, index
+
+
+def _parse_mapping(lines: list[str], start: int, indent: int):
+    mapping = {}
+    index = start
+
+    while index < len(lines):
+        raw = lines[index]
+        if not raw.strip():
+            index += 1
+            continue
+
+        current_indent = _count_indent(raw)
+        if current_indent < indent:
+            break
+        if current_indent != indent:
+            raise ValueError(f"Unexpected indentation near line: {raw}")
+
+        text = raw[indent:]
+        if text.startswith("- "):
+            raise ValueError(f"Unexpected list item near line: {raw}")
+        if ":" not in text:
+            raise ValueError(f"Invalid mapping entry near line: {raw}")
+
+        key, remainder = text.split(":", 1)
+        key = key.strip()
+        remainder = remainder.strip()
+        index += 1
+
+        if remainder in {"|", ">", "|-", ">-"}:
+            value, index = _parse_block_scalar(lines, index, indent + 2, remainder)
+        elif remainder:
+            value = _parse_scalar(remainder)
+        else:
+            next_index = _next_nonempty_index(lines, index, indent + 2)
+            if next_index is None:
+                value = ""
+            else:
+                nested_indent = _count_indent(lines[next_index])
+                nested_text = lines[next_index][nested_indent:]
+                if nested_text.startswith("- "):
+                    value, index = _parse_list(lines, next_index, nested_indent)
+                elif _looks_like_mapping_entry(nested_text):
+                    value, index = _parse_mapping(lines, next_index, nested_indent)
+                else:
+                    value, index = _parse_block_scalar(lines, next_index, nested_indent, ">")
+        mapping[key] = value
+
+    return mapping, index
+
+
+def parse_frontmatter(frontmatter_text: str) -> dict:
+    """Parse a small YAML subset used by SKILL.md frontmatter without PyYAML."""
+    lines = frontmatter_text.splitlines()
+    mapping, index = _parse_mapping(lines, 0, 0)
+
+    trailing = [line for line in lines[index:] if line.strip()]
+    if trailing:
+        raise ValueError(f"Unexpected trailing content in frontmatter: {trailing[0]}")
+    return mapping
+
+
+def load_skill_frontmatter(skill_path: Path) -> tuple[dict, str]:
+    """Load and parse the frontmatter from a skill directory."""
+    content = (skill_path / "SKILL.md").read_text()
+    frontmatter_text = extract_frontmatter_text(content)
+    frontmatter = parse_frontmatter(frontmatter_text)
+    if not isinstance(frontmatter, dict):
+        raise ValueError("Frontmatter must be a mapping")
+    return frontmatter, content
+
+
+def parse_skill_md(skill_path: Path) -> tuple[str, str, str]:
+    """Parse a SKILL.md file, returning (name, description, full_content)."""
+    frontmatter, content = load_skill_frontmatter(skill_path)
+    name = frontmatter.get("name", "")
+    description = frontmatter.get("description", "")
+    return str(name or ""), str(description or ""), content

package/skills/rudder/references/organization-skills.md

@@ -152,7 +152,7 @@ the full optional enabled-skill set intentionally.
 
 ## Notes
 
-- Built-in Rudder skills live in the organization library but are not auto-enabled.
+- Built-in Rudder skills live in the organization library and are always loaded for agent runs.
 - New organizations also seed optional community preset skills into the organization library. They stay organization-managed and default-off for agents.
 - If a skill reference is missing or ambiguous, Rudder returns `422`.
 - Prefer linking back to the relevant issue, approval, and agent when commenting about skill changes.

package/skills/skill-creator/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: skill-creator
+description: |
+  Create new skills, improve existing skills, and evaluate whether a skill definition is actually doing useful work.
+---
+
+# Skill Creator
+
+Use this skill when the task is to create a skill, refine a skill, or judge whether a skill definition should be changed.

package/skills/skill-optimizer/CHANGELOG.md

@@ -0,0 +1,29 @@
+# Changelog
+
+## v4.3 framing evidence hardening
+
+- Added sequencing guidance for requests that ask to optimize a skill after another live task: complete and verify the primary task first, then optimize from evidence.
+- Added explicit treatment of strong user corrections as high-signal evidence, especially framing corrections and wrong-abstraction-level failures.
+- Added framing checks for user outcome vs UI surface, scenario spine vs fixture rows, source of truth vs derivative signal, and product intent vs local convenience.
+- Relaxed the final response contract for larger workflows so the primary task result can be reported before concise skill changes.
+
+## v4.2 open-source package
+
+- Added package mode and open-source project structure guidance.
+- Added explicit adapter file lookup under `references/adapters/`.
+- Added packaging expectations for README, examples, evals, changelog, and distributable skill zip.
+- Preserved the generic analysis framework: core optimizer plus modular domain adapters.
+
+## v4.1 adapter hardening
+
+- Added explicit domain adapter use rule: source of truth, required inputs, review owner, authority gates, privacy, output template, validation cases, and must-not behaviors.
+- Added benchmark reporting split for trigger accuracy, patch-quality coverage, and downstream transfer.
+- Added warning that synthetic verifier scores are regression signals, not official leaderboard results.
+
+## v4.0 generic
+
+- Reframed Skill Optimizer from a software-focused hardening checklist into a domain-general analysis framework.
+- Added universal optimization lens covering purpose, triggers, inputs, workflow, tools, outputs, quality, safety, failure, and maintainability.
+- Moved domain-specific checks into modular adapter patterns.
+- Added trigger optimization guidance and benchmark mode.
+- Preserved strict patch safety around high-impact actions.