awoooi/apps/api/src/services/agent_claude_remediator_adapter.py

"""
Claude Agent SDK Remediator Replay Adapter
=========================================

Deterministic offline adapter for the `claude_agent_sdk_remediator` market
candidate. The Claude Agent SDK is not installed in this repo environment, so
this module models the remediation boundary without adding dependencies or
calling Anthropic/Claude APIs.

It never edits files, executes tools, writes production systems, sends
messages, or reads fixture labels.
"""

from __future__ import annotations

import json
import time
from dataclasses import dataclass
from typing import Any

from src.services.agent_market_candidate_adapter import get_market_candidate_spec
from src.services.agent_replay_input import assert_no_evaluation_label_leak

CLAUDE_REMEDIATOR_CANDIDATE_ID = "claude_agent_sdk_remediator"


@dataclass(frozen=True)
class ClaudeRemediatorDecision:
    """Candidate replay result produced by the Claude-shaped remediator."""

    payload: dict[str, Any]

    def to_dict(self) -> dict[str, Any]:
        return dict(self.payload)


def build_claude_remediator_candidate_result(
    candidate_input: dict[str, Any],
) -> ClaudeRemediatorDecision:
    """Build one offline Claude remediator replay result."""
    started = time.perf_counter()
    assert_no_evaluation_label_leak(candidate_input)
    spec = get_market_candidate_spec(CLAUDE_REMEDIATOR_CANDIDATE_ID)
    incident_id = str(candidate_input.get("incident_id", "")).strip()
    run_id = str(candidate_input.get("run_id", "")).strip()
    if not incident_id or not run_id:
        raise ValueError("candidate input must include incident_id and run_id")

    context = dict(candidate_input.get("incident_context") or {})
    state = _build_state(context)
    route = _remediation_route(state)
    plan = _plan_for_route(state, route)
    risk_level = _risk_level(state, plan)
    requires_human_approval = _requires_human_approval(risk_level, plan)
    trace_events = _trace_events(state, route, plan, risk_level, requires_human_approval)
    latency_ms = (time.perf_counter() - started) * 1000

    return ClaudeRemediatorDecision(
        payload={
            "schema_version": "agent_candidate_replay_result_v1",
            "run_id": run_id,
            "incident_id": incident_id,
            "candidate_id": spec.candidate_id,
            "candidate_role": spec.candidate_role,
            "proposed_action": plan["proposed_action"],
            "action_plan": plan["action_plan"],
            "risk_level": risk_level,
            "requires_human_approval": requires_human_approval,
            "blocked_by_policy": plan["blocked_by_policy"],
            "fallback_used": False,
            "trace_complete": True,
            "trace_events": trace_events,
            "rca_correct": None,
            "tool_dry_run_pass": None,
            "repair_success": None,
            "false_repair": False,
            "latency_ms": latency_ms,
            "cost_usd": 0,
            "error": None,
            "metadata": {
                "adapter_mode": "deterministic_offline_remediation_boundary",
                "candidate_framework": "claude_agent_sdk",
                "sdk_dependency": "claude_agent_sdk_package_not_installed",
                "anthropic_api_calls": False,
                "new_dependency_added": False,
                "tools_executed": False,
                "files_edited": False,
                "remediation_route": route,
                "guardrail_checks": [
                    "answer_key_leak_check",
                    "no_file_edit_without_approval",
                    "no_tool_execution_without_approval",
                    "controlled_apply_for_low_medium_high_patch_or_runtime_change",
                    "trace_required",
                ],
                "source": "claude_agent_sdk_remediator_offline_adapter",
            },
        }
    )


def build_claude_remediator_candidate_results(
    candidate_inputs: list[dict[str, Any]],
) -> list[ClaudeRemediatorDecision]:
    """Build many Claude remediator replay results."""
    return [
        build_claude_remediator_candidate_result(candidate_input)
        for candidate_input in candidate_inputs
    ]


def _build_state(context: dict[str, Any]) -> dict[str, Any]:
    haystack = json.dumps(context, ensure_ascii=False, sort_keys=True).lower()
    severity = str(context.get("severity") or "P3").strip().upper()
    status = str(context.get("status") or "").strip().lower()
    category = str(context.get("alert_category") or "general").strip().lower()
    alertname = str(context.get("alertname") or "").strip()
    service = _primary_service(context)
    namespace = _namespace(context)
    return {
        "alertname": alertname,
        "category": category,
        "severity": severity,
        "status": status,
        "service": service,
        "namespace": namespace,
        "haystack": haystack,
        "is_resolved": status == "resolved",
        "is_code": any(
            marker in haystack
            for marker in (
                "traceback",
                "exception",
                "build",
                "lint",
                "type error",
                "builderror",
                "importerror",
                "syntax",
                "module",
            )
        ),
        "is_config": any(
            marker in haystack
            for marker in ("config", "env", "secret", "token", "certificate", "tls", "ingress")
        ),
        "is_kubernetes": any(
            marker in haystack
            for marker in ("kubernetes", "k8s", "pod", "deployment", "namespace", "container")
        ),
        "is_database": any(marker in haystack for marker in ("postgres", "deadlock", "migration", "schema")),
        "is_backup": "backup" in haystack,
        "is_aiops": any(marker in haystack for marker in ("openclaw", "awooop", "agent", "flywheel")),
    }


def _remediation_route(state: dict[str, Any]) -> str:
    if state["is_resolved"]:
        return "observe_only"
    if state["is_code"]:
        return "code_patch_proposal"
    if state["is_config"]:
        return "config_patch_proposal"
    if state["is_database"]:
        return "migration_review"
    if state["is_backup"]:
        return "backup_runbook_patch"
    if state["is_aiops"]:
        return "agent_workflow_patch"
    if state["is_kubernetes"]:
        return "kubernetes_manifest_review"
    return "incident_runbook_patch"


def _plan_for_route(state: dict[str, Any], route: str) -> dict[str, Any]:
    if route == "observe_only":
        return _observe_plan(state)
    if route == "code_patch_proposal":
        return _code_patch_plan(state)
    if route == "config_patch_proposal":
        return _config_patch_plan(state)
    if route == "migration_review":
        return _migration_plan(state)
    if route == "backup_runbook_patch":
        return _backup_plan(state)
    if route == "agent_workflow_patch":
        return _agent_workflow_plan(state)
    if route == "kubernetes_manifest_review":
        return _kubernetes_manifest_plan(state)
    return _runbook_patch_plan(state)


def _observe_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            f"CLAUDE_OBSERVE_ONLY: incident is resolved; preserve evidence for "
            f"{state['alertname']} on {state['service']} and draft no patch"
        ),
        "blocked_by_policy": True,
        "action_plan": [
            _step("inspect-timeline", "awoooi-api", ["GET", "/api/v1/incidents/{incident_id}/timeline"]),
            _step("summarize-evidence", "remediator", ["no-patch-required"]),
            _step("handoff", "human", ["review-if-recurs"]),
        ],
    }


def _code_patch_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_PATCH_PROPOSAL: inspect traceback/build evidence, identify likely "
            "source file, draft a minimal patch, and require approval before editing"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-error", "logs", [state["alertname"], state["service"]]),
            _step("inspect-source", "repo", ["read-only", "related-files"]),
            _step("draft-patch", "remediator", ["minimal-diff", "no-write"]),
            _step("draft-tests", "remediator", ["targeted-tests", "no-execution"]),
            _step("approval-gate", "human", ["approve-before-apply-patch"]),
        ],
    }


def _config_patch_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_CONFIG_REVIEW: inspect env/config/TLS evidence, draft a redacted "
            "configuration change, and require approval before secret or deploy changes"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-config", "repo", ["read-only", "config-and-deploy-files"]),
            _step("inspect-runtime", "awoooi-api", ["read-only", state["service"]]),
            _step("draft-redacted-change", "remediator", ["no-secret-disclosure"]),
            _step("approval-gate", "human", ["approve-before-secret-or-config-change"]),
        ],
    }


def _migration_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_MIGRATION_REVIEW: inspect schema/migration evidence, draft an "
            "additive migration or rollback note, and require approval before DB writes"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-schema", "postgres", ["read-only", "information_schema"]),
            _step("inspect-migrations", "repo", ["read-only", "migrations"]),
            _step("draft-migration", "remediator", ["additive-only", "no-write"]),
            _step("approval-gate", "human", ["approve-before-db-write"]),
        ],
    }


def _backup_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_BACKUP_RUNBOOK_PATCH: inspect backup evidence and draft runbook or "
            "script patch; do not delete backups, rotate retention, or change secrets"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-backup-evidence", "logs", [state["service"], "backup"]),
            _step("inspect-scripts", "repo", ["read-only", "scripts/backup"]),
            _step("draft-runbook-patch", "remediator", ["no-write"]),
            _step("approval-gate", "human", ["approve-before-script-change"]),
        ],
    }


def _agent_workflow_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_AGENT_WORKFLOW_PATCH: inspect agent sessions, approval queue, and "
            "workflow code; draft a guardrail patch without changing production routing"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-agent-evidence", "database", ["read-only", "agent_sessions"]),
            _step("inspect-approval-chain", "database", ["read-only", "approval_records"]),
            _step("inspect-code", "repo", ["read-only", "agent-workflow-files"]),
            _step("draft-guardrail-patch", "remediator", ["no-write"]),
            _step("approval-gate", "human", ["approve-before-agent-routing-change"]),
        ],
    }


def _kubernetes_manifest_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            f"CLAUDE_K8S_MANIFEST_REVIEW: inspect workload manifests and runtime "
            f"events for {state['service']}; draft patch but do not rollout"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-manifest", "repo", ["read-only", "k8s", state["namespace"]]),
            _step("inspect-events", "kubectl", ["get", "events", "-n", state["namespace"]]),
            _step("draft-manifest-patch", "remediator", ["no-write"]),
            _step("approval-gate", "human", ["approve-before-rollout"]),
        ],
    }


def _runbook_patch_plan(state: dict[str, Any]) -> dict[str, Any]:
    return {
        "proposed_action": (
            "CLAUDE_RUNBOOK_PATCH: inspect incident evidence, draft runbook/playbook "
            "improvement, and require replay validation before production use"
        ),
        "blocked_by_policy": False,
        "action_plan": [
            _step("inspect-evidence", "awoooi-api", ["GET", "/api/v1/incidents/{incident_id}/evidence"]),
            _step("inspect-docs", "repo", ["read-only", "docs/runbooks"]),
            _step("draft-runbook-update", "remediator", ["no-write"]),
            _step("approval-gate", "human", ["approve-before-runbook-change"]),
        ],
    }


def _risk_level(state: dict[str, Any], plan: dict[str, Any]) -> str:
    if state["severity"] == "P0":
        return "critical"
    if state["severity"] == "P1" or state["is_config"]:
        return "high"
    action = json.dumps(plan, ensure_ascii=False).lower()
    if any(marker in action for marker in ("patch", "migration", "secret", "rollout", "db write")):
        return "medium"
    if state["severity"] == "P2":
        return "medium"
    return "low"


def _requires_human_approval(risk_level: str, plan: dict[str, Any]) -> bool:
    action = json.dumps(plan, ensure_ascii=False).lower()
    return risk_level == "critical" or any(
        marker in action
        for marker in (
            "break-glass",
            "migration",
            "secret",
            "credential",
            "authorization header",
            "private key",
            "drop database",
            "truncate",
            "delete pvc",
            "delete namespace",
            "force push",
            "ref deletion",
            "external attack",
            "paid provider",
        )
    )


def _trace_events(
    state: dict[str, Any],
    route: str,
    plan: dict[str, Any],
    risk_level: str,
    requires_human_approval: bool,
) -> list[dict[str, Any]]:
    return [
        {"type": "input_loaded", "alertname": state["alertname"], "service": state["service"]},
        {
            "type": "guardrails_checked",
            "answer_key_leak": False,
            "external_api_called": False,
            "files_edited": False,
            "tools_executed": False,
        },
        {"type": "remediation_route_selected", "route": route},
        {"type": "patch_boundary_set", "draft_only": True, "writes_allowed": False},
        {
            "type": "risk_reviewed",
            "risk_level": risk_level,
            "requires_human_approval": requires_human_approval,
        },
        {
            "type": "read_only_plan_built",
            "steps": len(plan["action_plan"]),
            "blocked_by_policy": plan["blocked_by_policy"],
        },
    ]


def _step(name: str, tool: str, args: list[str]) -> dict[str, Any]:
    return {
        "name": name,
        "tool": tool,
        "args": args,
        "mode": "read_only",
    }


def _primary_service(context: dict[str, Any]) -> str:
    affected = context.get("affected_services")
    if isinstance(affected, list) and affected:
        return str(affected[0]).strip() or "unknown-service"
    for signal in context.get("signals") or []:
        if not isinstance(signal, dict):
            continue
        labels = signal.get("labels") or {}
        if not isinstance(labels, dict):
            continue
        for key in ("deployment", "service", "container", "pod", "app", "instance"):
            if labels.get(key):
                return str(labels[key]).split(":")[0].strip() or "unknown-service"
    service = context.get("service") or context.get("target_service")
    return str(service or "unknown-service").strip()


def _namespace(context: dict[str, Any]) -> str:
    namespace = context.get("namespace") or context.get("kubernetes_namespace")
    if namespace:
        return str(namespace).strip()
    for signal in context.get("signals") or []:
        if not isinstance(signal, dict):
            continue
        labels = signal.get("labels") or {}
        if isinstance(labels, dict) and labels.get("namespace"):
            return str(labels["namespace"]).strip()
    return "awoooi-prod"