@rpamis/comet 0.4.0-beta.3 → 0.4.0-beta.5

package/eval/scaffold/python/validation/authoring_rubric.py

@@ -0,0 +1,326 @@
+"""Rubric for generated `/comet-any` authoring Skill packages."""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+from typing import Any
+
+from scaffold.python.validation.generic_rubric import generic_rubric_validator
+
+AUTHORING_RUBRIC_DIMENSIONS = (
+    "completion",
+    "skill_invocation",
+    "artifact_presence",
+    "generated_package",
+    "resolved_skill_evidence",
+    "engine_contract",
+    "workflow_route_conformance",
+    "authoring_lanes",
+    "review_gate",
+    "review_readiness",
+    "safety_boundary",
+)
+
+
+def _fmt(dim: str, score: float, reason: str) -> str:
+    return f"[RUBRIC] {dim}: {score:.2f} - {reason}"
+
+
+def _fmt_na(dim: str, reason: str) -> str:
+    return f"[RUBRIC] {dim}: N/A - {reason}"
+
+
+def _package_path(test_dir: Path, outputs: dict[str, Any]) -> Path | None:
+    raw = outputs.get("skill_package_path")
+    if not raw:
+        return None
+    path = Path(raw)
+    return path if path.is_absolute() else test_dir / path
+
+
+def _read_json(path: Path) -> dict[str, Any] | None:
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None
+
+
+def _binary_score(checks: list[bool]) -> float:
+    return sum(1 for item in checks if item) / len(checks) if checks else 0.0
+
+
+def _workflow_route(protocol: dict[str, Any]) -> list[str]:
+    return [
+        str(node.get("id"))
+        for node in protocol.get("nodes") or []
+        if isinstance(node, dict) and node.get("id") and not node.get("disabled")
+    ]
+
+
+def _slug(value: str) -> str:
+    slug = re.sub(r"[^a-z0-9._-]+", "-", value.lower()).strip("-")
+    return slug or "node"
+
+
+def _generated_node_skill_name(protocol: dict[str, Any], node_id: str) -> str:
+    workflow_name = _slug(str(protocol.get("name") or "workflow"))
+    return f"{workflow_name}-{_slug(node_id)}"
+
+
+def _check_package(package: Path | None) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    skill = package / "SKILL.md"
+    if not skill.exists():
+        return 0.0, "SKILL.md missing", ["SKILL.md missing"]
+    text = skill.read_text(encoding="utf-8")
+    checks = [
+        (
+            bool(re.search(r"Workflow Nodes|工作流节点|workflow route|route", text, re.I)),
+            "workflow route guidance missing",
+        ),
+        (
+            bool(re.search(r"用户停顿点|decision point|stop", text, re.I)),
+            "decision-point guidance missing",
+        ),
+        (
+            "reference/workflow-protocol.json" in text and "reference/resolved-skills.json" in text,
+            "workflow and resolved-skill references missing",
+        ),
+        (
+            "workflow-guard" in text or "workflow-state" in text or "恢复" in text.lower(),
+            "workflow guard or recovery guidance missing",
+        ),
+    ]
+    failures = [reason for passed, reason in checks if not passed]
+    score = (len(checks) - len(failures)) / len(checks)
+    return score, f"{len(checks) - len(failures)}/{len(checks)} package checks passed", failures
+
+
+def _check_resolved_evidence(package: Path | None) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    evidence = package / "reference" / "resolved-skills.json"
+    if not evidence.exists():
+        return 0.0, "resolved-skills.json missing", ["resolved-skills.json missing"]
+    data = _read_json(evidence)
+    if data is None:
+        return 0.0, "resolved-skills.json invalid", ["resolved-skills.json invalid"]
+    summaries = data.get("sourceSummaries") or []
+    if not summaries:
+        return 0.0, "sourceSummaries empty", ["sourceSummaries empty"]
+    return 1.0, f"{len(summaries)} source summaries", []
+
+
+def _check_engine_contract(package: Path | None) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    comet = package / "comet"
+    expected = ["skill.yaml", "guardrails.yaml", "checks.yaml", "eval.yaml"]
+    present = [name for name in expected if (comet / name).exists()]
+    missing = [f"{name} missing" for name in expected if name not in present]
+    if not present and not comet.exists():
+        return 1.0, "Engine disabled for lightweight package", []
+    legacy = (comet / "evals.yaml").exists()
+    checks = [name in present for name in expected] + [not legacy]
+    failures = missing + (["legacy evals.yaml present"] if legacy else [])
+    return (
+        _binary_score(checks),
+        f"{len(present)}/{len(expected)} engine files present; legacy_evals={legacy}",
+        failures,
+    )
+
+
+def _check_workflow_route_conformance(
+    package: Path | None,
+    outputs: dict[str, Any],
+) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    protocol = _read_json(package / "reference" / "workflow-protocol.json")
+    if protocol is None:
+        return 0.0, "workflow-protocol.json missing or invalid", [
+            "workflow-protocol.json missing or invalid"
+        ]
+    protocol_route = _workflow_route(protocol)
+    expected_nodes = outputs.get("route_conformance_expected_node_order") or protocol_route
+    generated_node_skills = outputs.get("generated_node_skills") or [
+        _generated_node_skill_name(protocol, node_id) for node_id in protocol_route
+    ]
+    if not expected_nodes:
+        return 0.0, "workflow route missing", ["workflow route missing"]
+    route_matches = protocol_route == expected_nodes
+    node_files = [
+        (package.parent / node_skill / "SKILL.md").exists()
+        for node_skill in generated_node_skills
+    ]
+    eval_manifest = package / "comet" / "eval.yaml"
+    eval_text = eval_manifest.read_text(encoding="utf-8") if eval_manifest.exists() else ""
+    eval_lists_route = (
+        True
+        if not eval_manifest.exists()
+        else all(node_id in eval_text for node_id in expected_nodes)
+        and all(node_skill in eval_text for node_skill in generated_node_skills)
+        and ("workflow-route-conformance" in eval_text)
+    )
+    checks = [route_matches, all(node_files), eval_lists_route]
+    failures = []
+    if not route_matches:
+        failures.append(
+            f"workflow route mismatch: expected {expected_nodes}, got {protocol_route}"
+        )
+    if not all(node_files):
+        missing = [
+            node_skill
+            for node_skill, exists in zip(generated_node_skills, node_files)
+            if not exists
+        ]
+        failures.append(f"internal Node Skill missing: {', '.join(missing)}")
+    if not eval_lists_route:
+        failures.append("comet/eval.yaml missing workflow-route-conformance route entries")
+    score = _binary_score(checks)
+    return score, f"route={len(expected_nodes)} node(s) checks={sum(checks)}/{len(checks)}", failures
+
+
+def _check_authoring_lanes(package: Path | None) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    lanes_path = package / "reference" / "authoring-lanes.json"
+    data = _read_json(lanes_path)
+    if data is None:
+        return 0.0, "authoring-lanes.json missing or invalid", [
+            "authoring-lanes.json missing or invalid"
+        ]
+    present = {
+        lane.get("lane")
+        for lane in data.get("lanes") or []
+        if isinstance(lane, dict)
+    }
+    required = {"skill-core", "script-contract", "reference", "pause-points", "eval", "skill-review"}
+    missing = sorted(required - present)
+    review_passed = bool((data.get("review") or {}).get("passed"))
+    checks = [not missing, review_passed]
+    failures = []
+    if missing:
+        failures.append(f"authoring lanes missing: {', '.join(missing)}")
+    if not review_passed:
+        failures.append("authoring-lanes review did not pass")
+    return _binary_score(checks), f"lanes={len(present)} review_passed={review_passed}", failures
+
+
+def _check_review_gate(package: Path | None) -> tuple[float, str, list[str]]:
+    if not package:
+        return 0.0, "skill_package_path missing", ["skill_package_path missing"]
+    review = package / "reference" / "skill-review.md"
+    lanes = _read_json(package / "reference" / "authoring-lanes.json") or {}
+    if not review.exists():
+        return 0.0, "skill-review.md missing", ["skill-review.md missing"]
+    text = review.read_text(encoding="utf-8")
+    markdown_passed = bool(re.search(r"Review passed|审查通过|通过", text, re.I))
+    lanes_passed = bool((lanes.get("review") or {}).get("passed"))
+    blockers = (lanes.get("review") or {}).get("blockingFindings") or []
+    checks = [markdown_passed, lanes_passed, len(blockers) == 0]
+    failures = []
+    if not markdown_passed:
+        failures.append("skill-review.md does not report a passed review")
+    if not lanes_passed:
+        failures.append("authoring-lanes review did not pass")
+    if blockers:
+        failures.append(f"blocking findings present: {len(blockers)}")
+    return _binary_score(checks), f"review_checks={sum(checks)}/{len(checks)}", failures
+
+
+def _weighted(scores: dict[str, float | None]) -> float:
+    weights = {
+        "completion": 1.5,
+        "skill_invocation": 1.0,
+        "artifact_presence": 1.0,
+        "generated_package": 1.4,
+        "resolved_skill_evidence": 1.2,
+        "engine_contract": 1.2,
+        "workflow_route_conformance": 1.5,
+        "authoring_lanes": 1.2,
+        "review_gate": 0.8,
+        "review_readiness": 0.8,
+        "safety_boundary": 1.2,
+    }
+    applicable = {k: v for k, v in scores.items() if v is not None and k in weights}
+    if not applicable:
+        return 0.0
+    total_weight = sum(weights[k] for k in applicable)
+    return sum(applicable[k] * weights[k] for k in applicable) / total_weight
+
+
+def _parse_generic_score(checks: list[str], name: str) -> float | None:
+    prefix = f"[RUBRIC] {name}:"
+    for item in checks:
+        if item.startswith(prefix):
+            remainder = item.removeprefix(prefix).strip()
+            if remainder.startswith("N/A"):
+                return None
+            try:
+                return float(remainder.split(" ", 1)[0])
+            except ValueError:
+                return 0.0
+    return 0.0
+
+
+def authoring_skill_rubric_validator(
+    test_dir: Path,
+    outputs: dict[str, Any],
+) -> tuple[list[str], list[str]]:
+    generic_passed, generic_failed = generic_rubric_validator(test_dir, outputs)
+    package = _package_path(test_dir, outputs)
+    package_score, package_reason, package_failures = _check_package(package)
+    evidence_score, evidence_reason, evidence_failures = _check_resolved_evidence(package)
+    engine_score, engine_reason, engine_failures = _check_engine_contract(package)
+    route_score, route_reason, route_failures = _check_workflow_route_conformance(package, outputs)
+    lanes_score, lanes_reason, lanes_failures = _check_authoring_lanes(package)
+    review_gate_score, review_gate_reason, review_gate_failures = _check_review_gate(package)
+    review_score = 1.0 if not generic_failed else 0.0
+    review_reason = "no hard validation failures" if review_score == 1.0 else "hard failures present"
+
+    scores = {
+        "completion": _parse_generic_score(generic_passed, "completion"),
+        "skill_invocation": _parse_generic_score(generic_passed, "skill_invocation"),
+        "artifact_presence": _parse_generic_score(generic_passed, "artifact_presence"),
+        "generated_package": package_score,
+        "resolved_skill_evidence": evidence_score,
+        "engine_contract": engine_score,
+        "workflow_route_conformance": route_score,
+        "authoring_lanes": lanes_score,
+        "review_gate": review_gate_score,
+        "review_readiness": review_score,
+        "safety_boundary": _parse_generic_score(generic_passed, "safety_boundary"),
+    }
+    def _fmt_dim(dim: str, score: float | None, fallback_reason: str) -> str:
+        if score is not None:
+            return _fmt(dim, score, fallback_reason)
+        return _fmt_na(dim, fallback_reason)
+
+    passed = [
+        _fmt_dim("completion", scores["completion"], "baseline completion score"),
+        _fmt_dim("skill_invocation", scores["skill_invocation"], "required Skill invocation score"),
+        _fmt_dim("artifact_presence", scores["artifact_presence"], "expected artifact score"),
+        _fmt("generated_package", package_score, package_reason),
+        _fmt("resolved_skill_evidence", evidence_score, evidence_reason),
+        _fmt("engine_contract", engine_score, engine_reason),
+        _fmt("workflow_route_conformance", route_score, route_reason),
+        _fmt("authoring_lanes", lanes_score, lanes_reason),
+        _fmt("review_gate", review_gate_score, review_gate_reason),
+        _fmt("review_readiness", review_score, review_reason),
+        _fmt_dim("safety_boundary", scores["safety_boundary"], "generic safety score"),
+        f"[RUBRIC] weighted_score: {_weighted(scores):.2f}",
+    ]
+    failed = (
+        generic_failed
+        + package_failures
+        + evidence_failures
+        + engine_failures
+        + route_failures
+        + lanes_failures
+        + review_gate_failures
+    )
+    return passed, failed

package/eval/scaffold/python/validation/comet_workflow.py

@@ -0,0 +1,187 @@
+"""Shared comet-workflow validation checks for eval tasks.
+
+Each comet task validator imports :func:`run_comet_checks` to get the four
+generic checks (openspec artifacts, comet state, workflow phases, tests written)
+and appends its own task-specific checks. All checks run inside Docker and emit
+results via the standard ``_test_results.json`` protocol.
+
+A check is a dict ``{"check": <name>, "status": "passed"|"failed", "message": ...}``.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+
+WORKSPACE = Path("/workspace")
+RESULTS_FILE = "_test_results.json"
+
+
+def _passed(name: str, message: str = "") -> dict:
+    return {"check": name, "status": "passed", "message": message}
+
+
+def _failed(name: str, message: str) -> dict:
+    return {"check": name, "status": "failed", "message": message}
+
+
+def check_openspec_artifacts() -> dict:
+    """proposal.md + tasks.md exist in a change dir (active or archived)."""
+    changes_dir = WORKSPACE / "openspec" / "changes"
+    if not changes_dir.exists():
+        return _failed("openspec_artifacts", "openspec/changes/ directory not found")
+
+    candidates: list[Path] = []
+    for d in changes_dir.iterdir():
+        if not d.is_dir():
+            continue
+        if d.name == "archive":
+            candidates.extend(s for s in d.iterdir() if s.is_dir())
+        else:
+            candidates.append(d)
+
+    if not candidates:
+        return _failed("openspec_artifacts", "No change directories found")
+
+    for change_dir in candidates:
+        if (change_dir / "proposal.md").exists() and (change_dir / "tasks.md").exists():
+            return _passed("openspec_artifacts", f"found in {change_dir.name}")
+
+    return _failed("openspec_artifacts", f"proposal.md/tasks.md not found together (checked {len(candidates)} dirs)")
+
+
+def check_comet_state() -> dict:
+    """A .comet.yaml exists somewhere under openspec/changes with a sane phase."""
+    changes_dir = WORKSPACE / "openspec" / "changes"
+    if not changes_dir.exists():
+        return _failed("comet_state", "openspec/changes/ not found")
+
+    # Search active + archived change dirs for .comet.yaml.
+    state_files: list[Path] = []
+    for root, _dirs, files in os_walk(changes_dir):
+        for f in files:
+            if f == ".comet.yaml":
+                state_files.append(Path(root) / f)
+
+    if not state_files:
+        archive_dir = changes_dir / "archive"
+        if archive_dir.exists():
+            for change_dir in archive_dir.iterdir():
+                if not change_dir.is_dir():
+                    continue
+                if (change_dir / "proposal.md").exists() and (change_dir / "tasks.md").exists():
+                    return _passed("comet_state", "phase=archived")
+        return _failed("comet_state", "No .comet.yaml found under openspec/changes/")
+
+    # Read the first one and check it has a recognised phase.
+    try:
+        text = state_files[0].read_text(errors="ignore")
+    except OSError:
+        return _failed("comet_state", f"Cannot read {state_files[0]}")
+
+    m = re.search(r"^phase:\s*(\S+)", text, re.MULTILINE)
+    if not m:
+        return _failed("comet_state", "No 'phase:' field in .comet.yaml")
+    phase = m.group(1)
+    valid = {"open", "design", "build", "verify", "archive", "archived"}
+    if phase not in valid:
+        return _failed("comet_state", f"Unrecognised phase: {phase}")
+    return _passed("comet_state", f"phase={phase}")
+
+
+def check_workflow_phases() -> dict:
+    """Evidence of the 5 phases: proposal→design→plan/build→verify→archive."""
+    evidence = 0
+    found: list[str] = []
+
+    # Phase 1 (open): proposal/tasks
+    if _glob_exists("openspec/changes/**/proposal.md") or _glob_exists("openspec/changes/**/tasks.md"):
+        evidence += 1; found.append("open")
+    # Phase 2 (design): design.md or docs/superpowers/specs/
+    if _glob_exists("openspec/changes/**/design.md") or _glob_exists("docs/superpowers/specs/*.md"):
+        evidence += 1; found.append("design")
+    # Phase 3 (build): plan.md, docs/superpowers/plans/, or .comet/ handoff
+    if _glob_exists("openspec/changes/**/plan.md") or _glob_exists("docs/superpowers/plans/*.md") or _glob_exists("openspec/changes/**/.comet/"):
+        evidence += 1; found.append("build")
+    # Phase 4 (verify): verification report or docs/superpowers/reports/
+    if (
+        _glob_exists("openspec/changes/**/verification.md")
+        or _glob_exists("openspec/changes/**/verification-report.md")
+        or _glob_exists("docs/superpowers/reports/*.md")
+    ):
+        evidence += 1; found.append("verify")
+    # Phase 5 (archive): openspec/changes/archive/
+    if (WORKSPACE / "openspec" / "changes" / "archive").exists():
+        evidence += 1; found.append("archive")
+
+    if evidence >= 4:
+        return _passed("workflow_phases", f"{evidence}/5 phases ({','.join(found)})")
+    if evidence >= 2:
+        return _failed("workflow_phases", f"Only {evidence}/5 phases evidenced ({','.join(found)})")
+    return _failed("workflow_phases", f"Only {evidence}/5 phases ({','.join(found)})")
+
+
+def check_tests_written() -> dict:
+    """The agent wrote or extended tests (test_*.py with real assertions)."""
+    test_files = list(WORKSPACE.glob("test_*.py")) + list(WORKSPACE.glob("**/test_*.py"))
+    # Filter out the validator itself and scaffold copies.
+    test_files = [f for f in test_files if "scaffold" not in str(f) and "validation" not in str(f)]
+    has_assert = False
+    for tf in test_files:
+        try:
+            if "assert" in tf.read_text(errors="ignore").lower():
+                has_assert = True
+                break
+        except OSError:
+            continue
+    if has_assert:
+        return _passed("tests_written", f"{len(test_files)} test file(s) with assertions")
+    if test_files:
+        return _failed("tests_written", f"{len(test_files)} test file(s) but none with assertions")
+    return _failed("tests_written", "No test files written by the agent")
+
+
+def run_comet_checks() -> list[dict]:
+    """Run all four generic comet-workflow checks."""
+    return [
+        check_openspec_artifacts(),
+        check_comet_state(),
+        check_workflow_phases(),
+        check_tests_written(),
+    ]
+
+
+def write_results(results: list[dict]) -> None:
+    """Write the combined results to _test_results.json and print a summary."""
+    out = {"checks": results}
+    results_path = WORKSPACE / RESULTS_FILE
+    try:
+        results_path.write_text(json.dumps(out))
+    except OSError:
+        pass
+    failed = [r for r in results if r["status"] == "failed"]
+    for r in results:
+        status = "✓" if r["status"] == "passed" else "✗"
+        print(f"  {status} {r['check']}: {r.get('message', '')}")
+    print(f"\n{len(results) - len(failed)}/{len(results)} checks passed")
+    if failed:
+        import sys
+        sys.exit(1)
+
+
+# --- helpers ---------------------------------------------------------------
+
+def os_walk(root: Path):
+    """os.walk that tolerates permission errors."""
+    import os
+    for dirpath, dirnames, filenames in os.walk(root):
+        yield dirpath, dirnames, filenames
+
+
+def _glob_exists(pattern: str) -> bool:
+    """True if any file matches the glob pattern under WORKSPACE."""
+    try:
+        return any(WORKSPACE.glob(pattern))
+    except Exception:
+        return False