feat(governance): 新增 matched PlayBook 學習缺口

2026-06-13 01:05:48 +08:00
parent f9619b137e
commit 3928e3ae67
14 changed files with 1548 additions and 9 deletions
--- a/apps/api/src/services/ai_agent_matched_playbook_learning_gap.py
+++ b/apps/api/src/services/ai_agent_matched_playbook_learning_gap.py
@@ -0,0 +1,273 @@
+"""
+AI Agent matched PlayBook learning gap snapshot.
+
+Loads the latest committed P2-104 matched_playbook_id learning gap contract.
+This module validates repo-committed evidence only; it never reads live DB rows,
+updates PlayBook trust, writes KM, writes timelines, writes Gateway queues,
+sends Telegram messages, reads secrets, or starts runtime work.
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from src.services.snapshot_paths import default_evaluations_dir
+
+_DEFAULT_EVALUATIONS_DIR = default_evaluations_dir(Path(__file__))
+_SNAPSHOT_PATTERN = "ai_agent_matched_playbook_learning_gap_*.json"
+_SCHEMA_VERSION = "ai_agent_matched_playbook_learning_gap_v1"
+_RUNTIME_AUTHORITY = "matched_playbook_learning_gap_contract_only_no_live_trust_write"
+
+
+def load_latest_ai_agent_matched_playbook_learning_gap(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed matched PlayBook learning gap contract."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(f"no AI Agent matched PlayBook learning gap snapshots found in {directory}")
+
+    latest = candidates[-1]
+    with latest.open(encoding="utf-8") as handle:
+        payload = json.load(handle)
+
+    if not isinstance(payload, dict):
+        raise ValueError(f"{latest}: expected JSON object")
+    _require_schema(payload, str(latest))
+    _require_no_live_boundaries(payload, str(latest))
+    _require_learning_checkpoints(payload, str(latest))
+    _require_gap_items(payload, str(latest))
+    _require_test_contracts(payload, str(latest))
+    _require_owner_gates(payload, str(latest))
+    _require_redaction_contract(payload, str(latest))
+    _require_rollup_consistency(payload, str(latest))
+    return payload
+
+
+def _require_schema(payload: dict[str, Any], label: str) -> None:
+    if payload.get("schema_version") != _SCHEMA_VERSION:
+        raise ValueError(f"{label}: expected schema_version={_SCHEMA_VERSION}")
+    status = payload.get("program_status") or {}
+    if status.get("read_only_mode") is not True:
+        raise ValueError(f"{label}: program_status.read_only_mode must be true")
+    if status.get("runtime_authority") != _RUNTIME_AUTHORITY:
+        raise ValueError(f"{label}: runtime_authority must remain {_RUNTIME_AUTHORITY}")
+    if status.get("current_task_id") != "P2-104":
+        raise ValueError(f"{label}: current_task_id must be P2-104")
+    if status.get("next_task_id") != "P2-105":
+        raise ValueError(f"{label}: next_task_id must be P2-105")
+
+
+def _require_no_live_boundaries(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("learning_gap_truth") or {}
+    required_true = {
+        "p2_103_result_audit_loaded",
+        "matched_playbook_source_contract_ready",
+        "approval_record_contract_ready",
+        "learning_service_contract_ready",
+        "execution_learning_await_contract_ready",
+        "e2e_test_contract_ready",
+        "owner_gate_contract_ready",
+        "all_checkpoints_have_evidence_ref",
+    }
+    missing = sorted(field for field in required_true if truth.get(field) is not True)
+    if missing:
+        raise ValueError(f"{label}: learning gap readiness flags must remain true: {missing}")
+
+    required_false = {
+        "runtime_execution_enabled",
+        "live_db_read_enabled",
+        "playbook_trust_write_enabled",
+        "km_write_enabled",
+        "timeline_write_enabled",
+        "gateway_queue_write_enabled",
+        "telegram_send_enabled",
+        "production_write_enabled",
+        "secret_value_read_enabled",
+        "host_or_cluster_command_enabled",
+        "destructive_operation_enabled",
+    }
+    unsafe = sorted(field for field in required_false if truth.get(field) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: live read/write/send/execution flags must remain false: {unsafe}")
+
+    zero_counts = {
+        "runtime_execution_count_24h",
+        "live_db_read_count_24h",
+        "playbook_trust_write_count_24h",
+        "km_write_count_24h",
+        "timeline_write_count_24h",
+        "gateway_queue_write_count_24h",
+        "telegram_send_count_24h",
+        "production_write_count_24h",
+        "secret_value_read_count_24h",
+        "host_or_cluster_command_count_24h",
+        "destructive_operation_count_24h",
+    }
+    non_zero = sorted(field for field in zero_counts if truth.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: live read/write/send/execution counts must remain zero: {non_zero}")
+
+
+def _require_learning_checkpoints(payload: dict[str, Any], label: str) -> None:
+    checkpoints = payload.get("learning_checkpoints") or []
+    checkpoint_ids = {checkpoint.get("checkpoint_id") for checkpoint in checkpoints}
+    required = {
+        "checkpoint_proposal_playbook_match",
+        "checkpoint_approval_record_serialization",
+        "checkpoint_learning_service_update",
+        "checkpoint_execution_await_learning",
+        "checkpoint_repair_candidate_playbook_id",
+        "checkpoint_webhook_source_correlation",
+        "checkpoint_live_null_rate_gate",
+    }
+    if checkpoint_ids != required:
+        raise ValueError(f"{label}: learning checkpoints must match {sorted(required)}")
+
+    valid_status = {
+        "covered_by_test",
+        "covered_by_existing_tests",
+        "source_guarded",
+        "blocked_until_live_evidence",
+    }
+    for checkpoint in checkpoints:
+        checkpoint_id = checkpoint.get("checkpoint_id")
+        if checkpoint.get("current_status") not in valid_status:
+            raise ValueError(f"{label}: checkpoint {checkpoint_id} current_status is invalid")
+        if checkpoint.get("writes_live_state") is not False:
+            raise ValueError(f"{label}: checkpoint {checkpoint_id} writes_live_state must remain false")
+        if not isinstance(checkpoint.get("requires_owner_review"), bool):
+            raise ValueError(f"{label}: checkpoint {checkpoint_id} requires_owner_review must be boolean")
+        for field in {"source_component", "evidence_ref", "expected_signal", "gap_if_missing"}:
+            if not checkpoint.get(field):
+                raise ValueError(f"{label}: checkpoint {checkpoint_id} must list {field}")
+        if not _is_redacted_sha256(checkpoint.get("evidence_hash")):
+            raise ValueError(f"{label}: checkpoint {checkpoint_id} must expose evidence_hash")
+
+
+def _require_gap_items(payload: dict[str, Any], label: str) -> None:
+    gaps = payload.get("gap_items") or []
+    if not gaps:
+        raise ValueError(f"{label}: gap_items must not be empty")
+    for gap in gaps:
+        gap_id = gap.get("gap_id")
+        if gap.get("severity") not in {"critical", "high", "medium", "low"}:
+            raise ValueError(f"{label}: gap {gap_id} severity is invalid")
+        if not isinstance(gap.get("blocks_trust_write"), bool):
+            raise ValueError(f"{label}: gap {gap_id} blocks_trust_write must be boolean")
+        for field in {"display_name", "current_state", "required_evidence", "owner_next_action"}:
+            if not gap.get(field):
+                raise ValueError(f"{label}: gap {gap_id} must list {field}")
+
+
+def _require_test_contracts(payload: dict[str, Any], label: str) -> None:
+    tests = payload.get("test_contracts") or []
+    if len(tests) < 5:
+        raise ValueError(f"{label}: test_contracts must include at least five tests")
+    for test in tests:
+        test_id = test.get("test_id")
+        if test.get("live_execution_required") is not False:
+            raise ValueError(f"{label}: test {test_id} live_execution_required must remain false")
+        if not test.get("test_ref") or "::" not in test.get("test_ref", ""):
+            raise ValueError(f"{label}: test {test_id} must expose test_ref")
+        if not test.get("expected_result"):
+            raise ValueError(f"{label}: test {test_id} must list expected_result")
+
+
+def _require_owner_gates(payload: dict[str, Any], label: str) -> None:
+    gates = payload.get("owner_gates") or []
+    gate_ids = {gate.get("gate_id") for gate in gates}
+    required = {
+        "gate_read_only_live_null_rate",
+        "gate_playbook_trust_write",
+        "gate_km_timeline_write",
+    }
+    if gate_ids != required:
+        raise ValueError(f"{label}: owner gates must match {sorted(required)}")
+    for gate in gates:
+        gate_id = gate.get("gate_id")
+        if gate.get("approval_required") is not True:
+            raise ValueError(f"{label}: gate {gate_id} approval_required must remain true")
+        if gate.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: gate {gate_id} runtime_write_allowed must remain false")
+        if not gate.get("blocked_reason"):
+            raise ValueError(f"{label}: gate {gate_id} must list blocked_reason")
+
+
+def _require_redaction_contract(payload: dict[str, Any], label: str) -> None:
+    contract = payload.get("display_redaction_contract") or {}
+    required_false = {
+        "raw_prompt_display_allowed",
+        "private_reasoning_display_allowed",
+        "secret_value_display_allowed",
+        "raw_telegram_payload_display_allowed",
+        "work_window_transcript_display_allowed",
+    }
+    if contract.get("redaction_required") is not True:
+        raise ValueError(f"{label}: redaction_required must be true")
+    unsafe = sorted(field for field in required_false if contract.get(field) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: redaction flags must remain false: {unsafe}")
+    if not contract.get("allowed_display_fields"):
+        raise ValueError(f"{label}: allowed_display_fields must not be empty")
+    if not contract.get("blocked_display_fields"):
+        raise ValueError(f"{label}: blocked_display_fields must not be empty")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    checkpoints = payload.get("learning_checkpoints") or []
+    gaps = payload.get("gap_items") or []
+    tests = payload.get("test_contracts") or []
+    gates = payload.get("owner_gates") or []
+
+    expected = {
+        "learning_checkpoint_count": len(checkpoints),
+        "covered_by_test_count": sum(
+            1
+            for item in checkpoints
+            if item.get("current_status") in {"covered_by_test", "covered_by_existing_tests"}
+        ),
+        "source_guarded_count": sum(1 for item in checkpoints if item.get("current_status") == "source_guarded"),
+        "blocked_until_live_evidence_count": sum(1 for item in checkpoints if item.get("current_status") == "blocked_until_live_evidence"),
+        "gap_item_count": len(gaps),
+        "high_severity_gap_count": sum(1 for item in gaps if item.get("severity") == "high"),
+        "test_contract_count": len(tests),
+        "owner_gate_count": len(gates),
+        "approval_required_gate_count": sum(1 for item in gates if item.get("approval_required") is True),
+    }
+    mismatches = {
+        key: {"expected": value, "actual": rollups.get(key)}
+        for key, value in expected.items()
+        if rollups.get(key) != value
+    }
+    if mismatches:
+        raise ValueError(f"{label}: rollup counts mismatch: {mismatches}")
+
+    zero_rollups = {
+        "runtime_execution_count",
+        "live_db_read_count",
+        "playbook_trust_write_count",
+        "km_write_count",
+        "timeline_write_count",
+        "gateway_queue_write_count",
+        "telegram_send_count",
+        "production_write_count",
+        "secret_value_read_count",
+        "destructive_operation_count",
+    }
+    non_zero = sorted(field for field in zero_rollups if rollups.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: rollup live read/write/send/execution counts must remain zero: {non_zero}")
+
+
+def _is_redacted_sha256(value: Any) -> bool:
+    if not isinstance(value, str):
+        return False
+    if not value.startswith("sha256:"):
+        return False
+    digest = value.removeprefix("sha256:")
+    return len(digest) == 64 and all(char in "0123456789abcdef" for char in digest)