feat(governance): 新增 owner approved result capture readback

apps/api/src/api/v1/agents.py

@@ -88,6 +88,9 @@ from src.services.ai_agent_owner_approved_learning_dry_run import (
 from src.services.ai_agent_owner_approved_result_capture_dry_run import (
     load_latest_ai_agent_owner_approved_result_capture_dry_run,
 )
+from src.services.ai_agent_owner_approved_result_capture_readback import (
+    load_latest_ai_agent_owner_approved_result_capture_readback,
+)
 from src.services.ai_agent_operation_permission_model import (
     load_latest_ai_agent_operation_permission_model,
 )
@@ -1227,6 +1230,36 @@ async def get_agent_owner_approved_result_capture_dry_run() -> dict[str, Any]:
         ) from exc
 
 
+@router.get(
+    "/agent-owner-approved-result-capture-readback",
+    response_model=dict[str, Any],
+    summary="取得 AI Agent owner-approved result capture readback",
+    description=(
+        "讀取最新已提交的 P2-107 owner-approved result capture readback；"
+        "此端點只回傳 result capture readback digest、promotion readiness review、failure lane、"
+        "reviewer queue preview 與 operator action，"
+        "不讀 canonical runtime target、不寫 score、不寫 result capture、不寫 learning、不更新 PlayBook trust、"
+        "不寫 reviewer queue、不寫 Gateway queue、不送 Telegram、不呼叫 Bot API、不寫 production target、不讀 secret。"
+    ),
+)
+async def get_agent_owner_approved_result_capture_readback() -> dict[str, Any]:
+    """Return the latest read-only owner-approved result capture readback contract."""
+    try:
+        payload = await asyncio.to_thread(load_latest_ai_agent_owner_approved_result_capture_readback)
+        return redact_public_lan_topology(payload)
+    except FileNotFoundError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=str(exc),
+        ) from exc
+    except (json.JSONDecodeError, ValueError) as exc:
+        logger.error("ai_agent_owner_approved_result_capture_readback_invalid", error=str(exc))
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="AI Agent owner-approved result capture readback 無效",
+        ) from exc
+
+
 @router.get(
     "/agent-owner-approved-fixture-dry-run",
     response_model=dict[str, Any],

apps/api/src/services/ai_agent_owner_approved_result_capture_readback.py

@@ -0,0 +1,391 @@
+"""
+AI Agent owner-approved result capture readback snapshot.
+
+Loads the latest committed P2-107 owner-approved result capture readback
+contract. This module validates repo-committed evidence only; it never writes
+scores, result capture rows, learning state, PlayBook trust, reviewer queues,
+Gateway queues, Telegram messages, production targets, or secrets.
+"""
+
+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_owner_approved_result_capture_readback_*.json"
+_SCHEMA_VERSION = "ai_agent_owner_approved_result_capture_readback_v1"
+_RUNTIME_AUTHORITY = "owner_approved_result_capture_readback_only_no_live_write"
+
+
+def load_latest_ai_agent_owner_approved_result_capture_readback(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed owner-approved result capture readback contract."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(
+            f"no AI Agent owner-approved result capture readback 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_prior_dry_run(payload, str(latest))
+    _require_readback_truth(payload, str(latest))
+    _require_readback_digests(payload, str(latest))
+    _require_promotion_reviews(payload, str(latest))
+    _require_failure_lanes(payload, str(latest))
+    _require_reviewer_queue_preview(payload, str(latest))
+    _require_operator_actions(payload, str(latest))
+    _require_display_redaction(payload, str(latest))
+    _require_no_forbidden_display_terms(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") != "P2-107":
+        raise ValueError(f"{label}: current_task_id must be P2-107")
+    if status.get("next_task_id") != "P2-108":
+        raise ValueError(f"{label}: next_task_id must be P2-108")
+
+
+def _require_prior_dry_run(payload: dict[str, Any], label: str) -> None:
+    prior = payload.get("prior_dry_run_readback") or {}
+    if prior.get("source_schema_version") != "ai_agent_owner_approved_result_capture_dry_run_v1":
+        raise ValueError(f"{label}: prior_dry_run_readback must chain from P2-106")
+    required_counts = {
+        "result_capture_template_count": 5,
+        "score_fixture_count": 5,
+        "dry_run_gate_count": 7,
+        "operator_action_count": 5,
+        "approval_required_gate_count": 2,
+        "blocked_gate_count": 1,
+        "approved_without_execution_meta_24h": 63,
+        "execution_failed_with_matched_24h": 1,
+        "owner_approval_received_count": 0,
+        "dry_run_preview_generated_count": 0,
+        "result_capture_write_count": 0,
+        "score_write_count": 0,
+        "learning_write_count": 0,
+        "playbook_trust_write_count": 0,
+        "gateway_queue_write_count": 0,
+        "telegram_send_count": 0,
+        "production_write_count": 0,
+    }
+    mismatches = {
+        key: {"expected": expected, "actual": prior.get(key)}
+        for key, expected in required_counts.items()
+        if prior.get(key) != expected
+    }
+    if mismatches:
+        raise ValueError(f"{label}: P2-106 prior dry-run counts mismatch: {mismatches}")
+
+
+def _require_readback_truth(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("readback_truth") or {}
+    required_true = {
+        "p2_106_dry_run_loaded",
+        "fixture_readback_allowed",
+        "readback_digest_ready",
+        "promotion_readiness_review_ready",
+        "owner_review_required_before_promotion",
+    }
+    missing = sorted(field for field in required_true if truth.get(field) is not True)
+    if missing:
+        raise ValueError(f"{label}: readback readiness flags must remain true: {missing}")
+
+    required_false = {
+        "canonical_runtime_readback_enabled",
+        "runtime_result_capture_write_enabled",
+        "runtime_score_write_enabled",
+        "runtime_learning_write_enabled",
+        "playbook_trust_write_enabled",
+        "reviewer_queue_write_enabled",
+        "gateway_queue_write_enabled",
+        "telegram_send_enabled",
+        "production_write_enabled",
+        "secret_read_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}: runtime read/write/send flags must remain false: {unsafe}")
+
+    zero_counts = {
+        "owner_approval_received_count",
+        "readback_digest_generated_count",
+        "promotion_approved_count",
+        "reviewer_queue_write_count",
+        "result_capture_write_count_24h",
+        "score_write_count_24h",
+        "learning_write_count_24h",
+        "playbook_trust_write_count_24h",
+        "gateway_queue_write_count_24h",
+        "telegram_send_count_24h",
+        "production_write_count_24h",
+        "secret_read_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}: readback live counters must remain zero: {non_zero}")
+
+
+def _require_readback_digests(payload: dict[str, Any], label: str) -> None:
+    digests = payload.get("result_capture_readback_digests") or []
+    digest_ids = {digest.get("digest_id") for digest in digests}
+    required = {
+        "readback_digest_approved_execution_result",
+        "readback_digest_execution_failed_candidate",
+        "readback_digest_pending_human_gate",
+        "readback_digest_manual_or_noop",
+        "readback_digest_post_write_verifier_receipt",
+    }
+    if digest_ids != required:
+        raise ValueError(f"{label}: readback digests must match {sorted(required)}")
+
+    valid_statuses = {"ready_for_owner_review", "approval_required", "blocked_by_policy"}
+    for digest in digests:
+        digest_id = digest.get("digest_id")
+        if digest.get("status") not in valid_statuses:
+            raise ValueError(f"{label}: digest {digest_id} status is invalid")
+        if digest.get("fixture_only") is not True:
+            raise ValueError(f"{label}: digest {digest_id} fixture_only must remain true")
+        if digest.get("runtime_read_enabled") is not False:
+            raise ValueError(f"{label}: digest {digest_id} runtime_read_enabled must remain false")
+        if digest.get("write_enabled") is not False:
+            raise ValueError(f"{label}: digest {digest_id} write_enabled must remain false")
+        if not digest.get("readback_fields") or not digest.get("verifier_checks"):
+            raise ValueError(f"{label}: digest {digest_id} must list fields and verifier checks")
+        if not digest.get("promotion_blocker"):
+            raise ValueError(f"{label}: digest {digest_id} must include promotion_blocker")
+        if not _is_redacted_sha256(digest.get("evidence_hash")):
+            raise ValueError(f"{label}: digest {digest_id} must expose evidence_hash")
+
+
+def _require_promotion_reviews(payload: dict[str, Any], label: str) -> None:
+    reviews = payload.get("promotion_readiness_reviews") or []
+    review_ids = {review.get("review_id") for review in reviews}
+    required = {
+        "promotion_review_owner_scope",
+        "promotion_review_score_fixture_parity",
+        "promotion_review_redaction_digest",
+        "promotion_review_failure_candidate",
+        "promotion_review_live_writer_gate",
+    }
+    if review_ids != required:
+        raise ValueError(f"{label}: promotion reviews must match {sorted(required)}")
+
+    valid_states = {"ready_for_owner_review", "needs_owner_review", "blocked_by_policy"}
+    valid_tiers = {"low", "medium", "high", "critical"}
+    for review in reviews:
+        review_id = review.get("review_id")
+        if review.get("readiness_state") not in valid_states:
+            raise ValueError(f"{label}: promotion review {review_id} readiness_state is invalid")
+        if review.get("risk_tier") not in valid_tiers:
+            raise ValueError(f"{label}: promotion review {review_id} risk_tier is invalid")
+        if review.get("promotion_allowed") is not False:
+            raise ValueError(f"{label}: promotion review {review_id} promotion_allowed must remain false")
+        if review.get("creates_runtime_write") is not False:
+            raise ValueError(f"{label}: promotion review {review_id} creates_runtime_write must remain false")
+        if not review.get("required_before_promotion") or not review.get("failure_if_missing"):
+            raise ValueError(f"{label}: promotion review {review_id} must list promotion requirements")
+
+
+def _require_failure_lanes(payload: dict[str, Any], label: str) -> None:
+    lanes = payload.get("failure_lanes") or []
+    lane_ids = {lane.get("lane_id") for lane in lanes}
+    required = {
+        "failure_lane_missing_owner_scope",
+        "failure_lane_redaction_mismatch",
+        "failure_lane_verifier_fixture_missing",
+        "failure_lane_agent_disagreement",
+    }
+    if lane_ids != required:
+        raise ValueError(f"{label}: failure lanes must match {sorted(required)}")
+
+    for lane in lanes:
+        lane_id = lane.get("lane_id")
+        if lane.get("status") not in {"ready_for_review", "blocked_by_policy"}:
+            raise ValueError(f"{label}: failure lane {lane_id} status is invalid")
+        if lane.get("aborts_promotion") is not True:
+            raise ValueError(f"{label}: failure lane {lane_id} aborts_promotion must remain true")
+        if lane.get("creates_runtime_write") is not False:
+            raise ValueError(f"{label}: failure lane {lane_id} creates_runtime_write must remain false")
+        if not lane.get("trigger_condition") or not lane.get("required_response"):
+            raise ValueError(f"{label}: failure lane {lane_id} must include trigger and response")
+
+
+def _require_reviewer_queue_preview(payload: dict[str, Any], label: str) -> None:
+    queue_items = payload.get("reviewer_queue_preview") or []
+    queue_ids = {item.get("queue_id") for item in queue_items}
+    required = {
+        "review_queue_owner_scope",
+        "review_queue_public_redaction",
+        "review_queue_failure_candidate",
+        "review_queue_live_writer_blocker",
+    }
+    if queue_ids != required:
+        raise ValueError(f"{label}: reviewer queue preview must match {sorted(required)}")
+
+    for item in queue_items:
+        queue_id = item.get("queue_id")
+        if item.get("status") not in {"queued_for_owner_review", "blocked_by_policy"}:
+            raise ValueError(f"{label}: reviewer queue {queue_id} status is invalid")
+        if item.get("queue_write_enabled") is not False:
+            raise ValueError(f"{label}: reviewer queue {queue_id} queue_write_enabled must remain false")
+        if item.get("telegram_send_enabled") is not False:
+            raise ValueError(f"{label}: reviewer queue {queue_id} telegram_send_enabled must remain false")
+        if not item.get("required_inputs"):
+            raise ValueError(f"{label}: reviewer queue {queue_id} must list required_inputs")
+        if not _is_redacted_sha256(item.get("evidence_hash")):
+            raise ValueError(f"{label}: reviewer queue {queue_id} must expose evidence_hash")
+
+
+def _require_operator_actions(payload: dict[str, Any], label: str) -> None:
+    actions = payload.get("operator_actions") or []
+    action_types = {action.get("action_type") for action in actions}
+    required = {
+        "review_readback",
+        "compare_digest",
+        "review_failure_lane",
+        "reject_or_rework",
+        "promote_to_next_gate",
+    }
+    if action_types != required:
+        raise ValueError(f"{label}: operator actions must match {sorted(required)}")
+    for action in actions:
+        if action.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: operator action {action.get('action_id')} must not allow runtime write")
+        if not action.get("operator_instruction"):
+            raise ValueError(f"{label}: operator action {action.get('action_id')} must include instruction")
+
+
+def _require_display_redaction(payload: dict[str, Any], label: str) -> None:
+    contract = payload.get("display_redaction_contract") or {}
+    if contract.get("redaction_required") is not True:
+        raise ValueError(f"{label}: display redaction must remain required")
+    required_false = {
+        "raw_prompt_display_allowed",
+        "private_reasoning_display_allowed",
+        "secret_value_display_allowed",
+        "raw_telegram_payload_display_allowed",
+        "work_window_transcript_display_allowed",
+    }
+    unsafe = sorted(field for field in required_false if contract.get(field) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: display redaction fields must remain false: {unsafe}")
+
+
+def _require_no_forbidden_display_terms(payload: dict[str, Any], label: str) -> None:
+    forbidden_terms = {
+        "工作視窗",
+        "對話內容",
+        "批准！繼續",
+        "In app browser",
+        "My request for Codex",
+        "browser_context",
+        "codex_user_message",
+        "prompt_text",
+        "raw prompt",
+        "private reasoning",
+        "chain of thought",
+        "private_reasoning",
+        "chain_of_thought",
+        "authorization_header",
+        "work window transcript",
+        "internal collaboration transcript",
+    }
+    hits: list[str] = []
+
+    def walk(value: Any, path: str) -> None:
+        if isinstance(value, dict):
+            for key, nested in value.items():
+                walk(nested, f"{path}.{key}" if path else str(key))
+            return
+        if isinstance(value, list):
+            for index, nested in enumerate(value):
+                walk(nested, f"{path}[{index}]")
+            return
+        if isinstance(value, str):
+            matched = sorted(term for term in forbidden_terms if term in value)
+            if matched:
+                hits.append(f"{path}: {', '.join(matched)}")
+
+    walk(payload, "")
+    if hits:
+        raise ValueError(f"{label}: forbidden display terms found: {hits}")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    truth = payload.get("readback_truth") or {}
+    prior = payload.get("prior_dry_run_readback") or {}
+    digests = payload.get("result_capture_readback_digests") or []
+    reviews = payload.get("promotion_readiness_reviews") or []
+    lanes = payload.get("failure_lanes") or []
+    queue_items = payload.get("reviewer_queue_preview") or []
+    actions = payload.get("operator_actions") or []
+    expected = {
+        "readback_digest_count": len(digests),
+        "promotion_review_count": len(reviews),
+        "failure_lane_count": len(lanes),
+        "reviewer_queue_preview_count": len(queue_items),
+        "operator_action_count": len(actions),
+        "approval_required_digest_count": sum(1 for item in digests if item.get("status") == "approval_required"),
+        "blocked_digest_count": sum(1 for item in digests if item.get("status") == "blocked_by_policy"),
+        "ready_review_count": sum(1 for item in reviews if item.get("readiness_state") == "ready_for_owner_review"),
+        "blocked_review_count": sum(1 for item in reviews if item.get("readiness_state") == "blocked_by_policy"),
+        "blocked_failure_lane_count": sum(1 for item in lanes if item.get("status") == "blocked_by_policy"),
+        "queued_reviewer_preview_count": sum(
+            1 for item in queue_items if item.get("status") == "queued_for_owner_review"
+        ),
+        "blocked_reviewer_preview_count": sum(1 for item in queue_items if item.get("status") == "blocked_by_policy"),
+        "approved_without_execution_meta_24h": prior.get("approved_without_execution_meta_24h"),
+        "execution_failed_with_matched_24h": prior.get("execution_failed_with_matched_24h"),
+        "owner_approval_received_count": truth.get("owner_approval_received_count"),
+        "readback_digest_generated_count": truth.get("readback_digest_generated_count"),
+        "promotion_approved_count": truth.get("promotion_approved_count"),
+        "reviewer_queue_write_count": truth.get("reviewer_queue_write_count"),
+        "result_capture_write_count": truth.get("result_capture_write_count_24h"),
+        "score_write_count": truth.get("score_write_count_24h"),
+        "learning_write_count": truth.get("learning_write_count_24h"),
+        "playbook_trust_write_count": truth.get("playbook_trust_write_count_24h"),
+        "gateway_queue_write_count": truth.get("gateway_queue_write_count_24h"),
+        "telegram_send_count": truth.get("telegram_send_count_24h"),
+        "production_write_count": truth.get("production_write_count_24h"),
+        "secret_read_count": truth.get("secret_read_count_24h"),
+        "destructive_operation_count": truth.get("destructive_operation_count_24h"),
+    }
+    mismatches = {
+        key: {"expected": expected_value, "actual": rollups.get(key)}
+        for key, expected_value in expected.items()
+        if rollups.get(key) != expected_value
+    }
+    if mismatches:
+        raise ValueError(f"{label}: rollup counts mismatch: {mismatches}")
+
+
+def _is_redacted_sha256(value: Any) -> bool:
+    if not isinstance(value, str):
+        return False
+    if not value.startswith("sha256:") or len(value) != 71:
+        return False
+    return all(char in "0123456789abcdef" for char in value.removeprefix("sha256:"))