import json, os, sys, urllib.request, urllib.parse, urllib.error

# The Alert Triage API lives under /idr/at, a base of its own — not /idr/v1 or
# /idr/v2 like the investigation endpoints. Same host, same API key.
AT = "idr/at/"


def _cfg():
    s = json.loads(os.environ.get("INTEGRATION_SECRETS", "{}"))
    region = str(s.get("region") or "us").strip().lower()
    base = "https://" + region + ".api.insight.rapid7.com/"
    headers = {"X-Api-Key": s.get("api_key", ""), "Content-Type": "application/json", "Accept": "application/json"}
    multi = str(s.get("is_multi_customer") or "").strip().lower() in ("1", "true", "yes")
    return base, headers, multi


def _cfg_region():
    return str(json.loads(os.environ.get("INTEGRATION_SECRETS", "{}")).get("region") or "us").strip().lower()


def _inputs():
    return json.loads(os.environ.get("INTEGRATION_INPUTS", "{}"))


def _list(v):
    if v in (None, ""):
        return []
    if isinstance(v, list):
        return [str(x).strip() for x in v if str(x).strip()]
    return [p.strip() for p in str(v).split(",") if p.strip()]


def request(method, path, params=None, body=None):
    base, headers, _ = _cfg()
    url = base + path.lstrip("/")
    if params:
        clean = {k: v for k, v in params.items() if v not in (None, "")}
        if clean:
            url += ("&" if "?" in url else "?") + urllib.parse.urlencode(clean, doseq=True)
    data = json.dumps(body).encode("utf-8") if body is not None else None
    req = urllib.request.Request(url, data=data, headers=headers, method=method)
    with urllib.request.urlopen(req, timeout=90) as r:
        raw = r.read()
        return json.loads(raw) if raw else {}

import re
import time
from datetime import datetime, timedelta, timezone

ISO = "%Y-%m-%dT%H:%M:%SZ"
RRN_BATCH = 100
DEFAULT_BUDGET_S = 60

# idr/v2/investigations answers with 13 fields, three of them null: a title, a
# priority, a status, an assignee and timestamps. No host, no account, no
# process, no MITRE code — and not one of the alerts the investigation
# aggregates. An incident built from it indexes zero entities, so it cannot be
# pivoted on, cannot be correlated, and gives an analyst nothing to read.
#
# The detection content is two hops away: the investigation's alert list gives
# RRNs, and the Alert Triage API turns those into alerts carrying
# rule_keys_of_interest (the host, the process, its command line, its parent,
# the hashes) and rule.mitre_tcodes. This command walks both hops and returns
# the investigation with its alerts inlined, so the INVESTIGATION stays the unit
# that becomes an incident — results_path = data, dedup on rrn, exactly as the
# plain investigation feed. Only the command changes.


def _alert_rrn(a):
    """The alert's RRN, or None when this alert has none.

    The investigation alert list is not consistent about where the RRN sits.
    Some entries carry it in `id` (rrn:alerts:eu:<org>:alert:1:<hex>); others put
    a plain UUID there and the RRN, if any, in `rrn`. `rrn` is therefore read
    first, and anything that is not RRN-shaped is dropped rather than sent on.

    That last part is not defensive habit. /alerts/ops/rrns validates the whole
    list and rejects the entire request when one entry is malformed, so a single
    UUID slipping through costs every alert in the batch — which is exactly how
    this failed: HTTP 400, eight invalid RRNs, zero alerts fetched, the whole run
    lost. Dropping them here means the investigation still arrives, with its
    unretrievable alerts counted rather than silently missing.
    """
    for v in (a.get("rrn"), a.get("id")):
        if isinstance(v, str) and v.startswith("rrn:"):
            return v
    return None


def _rrn_region(rrn):
    """The region segment of an RRN, or None.

    An RRN is rrn:<service>:<region>:<org>:<type>:… — so a resource states which
    region it lives in, while the base URL is built from the instance's
    configured region. Those two agree on a single-region tenant and need not on
    a multi-customer key, which spans organizations.
    """
    parts = str(rrn or "").split(":")
    return parts[2] if len(parts) > 4 and parts[0] == "rrn" and parts[2] else None


def _collect_region_failures(out, sink):
    """Carry a partial failure out of a response instead of dropping it.

    The Alert Triage API answers HTTP 200 with a region_failures list when it
    could not reach one of the regions behind the request — the spec says the
    presence of items there IS the partial failure. Reading only the payload
    turns that into missing alerts with no error anywhere, which for an
    ingestion feed is indistinguishable from a quiet period.
    """
    for f in (out.get("region_failures") or []):
        sink.append(f)
    return out


# key="value" / key=value, the shape syslog-derived log lines take. Non-matching
# sources simply yield nothing, which is the right answer for them.
_KV = re.compile(r'([A-Za-z_][\w.\-]*)=("(?:[^"\\]|\\.)*"|\S+)')


# Values a source writes when it has nothing to say. Mapped as-is they become an
# asset named "unknown" in the entity index, shared by every incident that had no
# asset — a pivot that groups things by the absence of a fact.
_PLACEHOLDERS = {"unknown", "n/a", "na", "none", "null", "-", ""}


def _clean(v):
    """Strip the quoting Rapid7's parser leaves behind.

    custom_data values arrive as '10.70.20.66"' — a trailing quote from the
    vendor parser. Left alone it produces an IP that matches no regex and an
    entity nobody can pivot on, so the cleanup is not cosmetic.
    """
    v = str(v).strip()
    if len(v) >= 2 and v[0] == '"' and v[-1] == '"':
        v = v[1:-1]
    v = v.replace('\\"', '"').strip().strip('"')
    return "" if v.strip().lower() in _PLACEHOLDERS else v


def _evidence_fields(ev):
    """One evidence's event as a flat, normalised dict.

    `data` is a JSON-encoded STRING. JSONata cannot deserialise, so a mapper can
    never reach into it; this is why the parsing happens here.

    It arrives in at least two shapes and the union of both is read, rather than
    one or the other. The first version of this handled only the wrapped shape
    and returned {} for everything else — silently, so an Azure AD sign-in
    carrying its account, its source address and its geolocation produced an
    incident with no observable at all.

      wrapped  {source_data: "<syslog line>", custom_data: {Forti_user: …}}
               An unparsed source. custom_data is the vendor's normalised view,
               its keys prefixed per source; source_data is the raw line.

      flat     {user: …, account: …, source_ip: …, geoip_country_code: …,
                source_json: {…}, r7_context: {…}, detection_context: {…}}
               A parsed source. The event IS the object.

    Scalars at the top level are always taken, so a third shape still yields its
    plain fields instead of nothing. source_json is deliberately skipped — it is
    the vendor's whole payload, several kilobytes of it, and it stays in the raw
    line. r7_context and detection_context are flattened one level because they
    are Rapid7's own annotations and are the same across parsed sources.
    """
    raw = ev.get("data")
    if not isinstance(raw, str) or not raw.strip():
        return {}, None
    try:
        blob = json.loads(raw)
    except ValueError:
        # Not JSON: keep the string itself as the raw line rather than lose it.
        return {}, raw
    if not isinstance(blob, dict):
        return {}, raw

    fields = {}

    # Flat shape: every scalar the event states about itself.
    for k, v in blob.items():
        if k in ("source_data", "custom_data", "source_json"):
            continue
        if isinstance(v, bool) or isinstance(v, (int, float)):
            fields[k.lower()] = v
        elif isinstance(v, str):
            fields[k.lower()] = _clean(v)
        elif isinstance(v, dict) and k in ("r7_context", "detection_context"):
            for sk, sv in v.items():
                if isinstance(sv, dict):
                    for ssk, ssv in sv.items():
                        if not isinstance(ssv, (str, int, float)) or isinstance(ssv, bool):
                            continue
                        # detection_context nests user.user_age_mins, so a blind
                        # join yields user_user_age_mins.
                        name = ssk if ssk.startswith(sk + "_") else sk + "_" + ssk
                        fields[name.lower()] = _clean(ssv) if isinstance(ssv, str) else ssv
                elif isinstance(sv, (str, int, float)):
                    fields[(k + "_" + sk).lower()] = _clean(sv) if isinstance(sv, str) else sv

    # source_json is the vendor's own payload. Its NESTED part is skipped — the
    # Azure AD one runs to several kilobytes — but its top-level scalars are
    # small, bounded and often the only place something essential appears: a
    # Windows 4725 puts the domain controller in computerName and nowhere else,
    # and an Azure sign-in puts the caller address in callerIpAddress. Existing
    # keys win, so the event's own naming is never overwritten by the vendor's.
    sj = blob.get("source_json")
    if isinstance(sj, dict):
        for k, v in sj.items():
            key = k.lower()
            if key in fields:
                continue
            if isinstance(v, bool) or isinstance(v, (int, float)):
                fields[key] = v
            elif isinstance(v, str) and v.strip():
                fields[key] = _clean(v)

    # Wrapped shape, overlaid on top when present.
    custom = blob.get("custom_data")
    if isinstance(custom, dict):
        for k, v in custom.items():
            if not isinstance(v, (str, int, float)) or isinstance(v, bool):
                continue
            key = k.split("_", 1)[1] if "_" in k else k
            fields[key.lower()] = _clean(v) if isinstance(v, str) else v
    source = blob.get("source_data")
    if isinstance(source, str):
        for k, v in _KV.findall(source):
            fields[k.lower()] = _clean(v)

    # A key whose value cleaned away to nothing is dropped rather than kept
    # empty: an empty string is not nil, so it would still be written to its OCSF
    # field, and device.hostname = "" asserts an asset with no name.
    fields = {k: v for k, v in fields.items() if v != ""}
    return fields, (source if isinstance(source, str) else raw)


# Event types worth asking a process tree for. The endpoint sensor produces one;
# an authentication or firewall alert never will, and on a feed where those
# dominate, asking anyway is one wasted call per alert. include_process_tree=true
# overrides this and asks for every alert.
_PROCESS_EVENT_TYPES = {
    "process_start_event", "advanced_malware", "virus",
    "endpoint_activity", "file_modification", "local_service_installed",
}

# Guards on a tree of unknown shape: a wide one would otherwise expand into more
# chains than anybody reads, and a cycle would not terminate at all.
_MAX_CHAINS = 12
_MAX_DEPTH = 24


def _flatten_tree(root, out, path=None, depth=0):
    """Rapid7 nests a process tree through `children`; the layout block reads a
    flat chain, root first. So each root-to-leaf path becomes one chain.

    The block marks the LAST node of a chain as the target. Rapid7 flags no node
    as the one the alert fired on, so that mark means "end of this branch" here
    and nothing stronger — inventing a target would assert something the sensor
    never said.
    """
    if not isinstance(root, dict) or depth > _MAX_DEPTH or len(out) >= _MAX_CHAINS:
        return
    name = root.get("name")
    node = {}
    if name:
        node["name"] = str(name)
        if root.get("cmdline"):
            node["cmd_line"] = str(root["cmdline"])
        pid = root.get("pid")
        if pid not in (None, ""):
            try:
                node["pid"] = int(str(pid))
            except ValueError:
                pass
        h = root.get("hashes") or {}
        if isinstance(h, dict):
            for k in ("md5", "sha256"):
                if h.get(k):
                    node[k] = str(h[k])
        if root.get("start_time") not in (None, ""):
            node["started_at"] = root["start_time"]
        if root.get("username"):
            node["_username"] = str(root["username"])
    chain = (path or []) + ([node] if node else [])
    children = [c for c in (root.get("children") or []) if isinstance(c, dict)]
    if not children:
        if chain:
            out.append(chain)
        return
    for c in children:
        _flatten_tree(c, out, chain, depth + 1)


def _tree_chains(payload, alert):
    """The process_tree block's own shape: N chains, each with its nodes."""
    chains = []
    for tree in (payload.get("process_trees") or []):
        if not isinstance(tree, dict):
            continue
        paths = []
        _flatten_tree(tree.get("process_tree"), paths)
        for nodes in paths:
            user = ""
            for n in nodes:
                user = n.pop("_username", user) or user
            if not nodes:
                continue
            chain = {"nodes": nodes, "alert_id": tree.get("alert_rrn") or alert.get("rrn")}
            if user:
                chain["user"] = user
            if alert.get("type"):
                chain["rule"] = alert["type"]
            host = (alert.get("evidence_fields") or {}).get("devname") or (alert.get("evidence_fields") or {}).get("hostname")
            if not host:
                for kv in (alert.get("rule_keys_of_interest") or []):
                    if kv.get("key") == "hostname" and kv.get("values"):
                        host = kv["values"][0]
            if host:
                chain["host"] = host
            chains.append(chain)
            if len(chains) >= _MAX_CHAINS:
                return chains
    return chains


# Keys already rendered elsewhere in the incident, or too noisy for a summary.
_MD_SKIP = {"entry_id", "timestamp"}
_MD_MAX_ROWS = 40
_MD_MAX_VALUE = 300


# The columns worth reading first, in this order. Everything else follows
# alphabetically. Column order is the ONLY ordering signal the table field has:
# it takes the union of keys in order of first appearance, so what the first row
# names first is what the analyst sees without scrolling.
_ROW_LEAD = [
    "alert", "time", "action", "result", "status",
    "source_account", "source_user", "target_account", "target_user",
    "hostname", "computername", "source_asset", "devname",
    "source_ip", "srcip", "destination_ip", "dstip",
    "user", "account", "process.name", "process.cmd_line",
]
_MAX_ROWS = 100
# A table wider than this is scrolled, not read. The cap keeps the useful
# columns — the lead list, then whichever others are populated on the most rows
# — rather than whichever happened to appear first.
_MAX_COLUMNS = 30


def _evidence_rows(alerts):
    """One row per evidence, across every alert of the investigation.

    The table custom field reads an array of objects directly, so this needs no
    reshaping on the mapper side. It exists beside summary_markdown rather than
    instead of it: the summary reads as a page, this sorts and filters.

    Each row carries its alert's name, because an investigation aggregates
    several and a row that does not say where it came from cannot be acted on.
    """
    rows = []
    for a in alerts:
        for ev in (a.get("evidences") or []):
            f = ev.get("fields") or {}
            if not f:
                continue
            row = {"alert": a.get("type") or a.get("title") or ""}
            t = ev.get("evented_at") or f.get("timestamp")
            if t:
                row["time"] = str(t)
            for k in _ROW_LEAD:
                if k in f and k not in row:
                    row[k] = f[k]
            for k in sorted(f):
                if k not in row:
                    row[k] = f[k]
            rows.append(row)
            if len(rows) >= _MAX_ROWS:
                break
        if len(rows) >= _MAX_ROWS:
            break
    if not rows:
        return [], 0

    # Trim the width. Counted over every row, so a column present once out of
    # eighty does not push out one present on all of them.
    counts = {}
    for r in rows:
        for k in r:
            counts[k] = counts.get(k, 0) + 1
    ordered = [k for k in _ROW_LEAD + ["alert", "time"] if k in counts]
    ordered += sorted((k for k in counts if k not in ordered),
                      key=lambda k: (-counts[k], k))
    # `alert` and `time` lead whatever else survives.
    keep = ["alert"] + [k for k in ordered if k != "alert"][:_MAX_COLUMNS - 1]
    dropped = len(counts) - len(keep)
    if dropped > 0:
        rows = [{k: r[k] for k in keep if k in r} for r in rows]
    return rows, max(dropped, 0)


def _markdown_summary(inv, alerts):
    """A readable rendering of what the detection actually says.

    The layout builder's field picker is a closed <select> over the OCSF
    catalogue, so a path that is not in it — `detection`, `evidence`, anything
    an unanticipated rule emits — cannot be placed on a layout at all. Mapping
    each key to a real OCSF field is the right answer where one exists, but
    rule_keys_of_interest is an open vocabulary and always will be: three
    vocabularies turned up in as many days.

    So the whole of it is also rendered as Markdown and mapped to a custom
    field, which a layout CAN show. Nothing is lost to the analyst even when
    nothing was mapped, and the mapping work stops being what stands between
    a new rule type and a readable incident.
    """
    lines = []
    for a in alerts:
        title = a.get("type") or a.get("title") or "Alert"
        lines.append("### " + str(title))
        rule = a.get("rule") or {}
        tcodes = [t for t in (rule.get("mitre_tcodes") or []) if t]
        meta = []
        if a.get("triggering_event_type"):
            meta.append("`" + str(a["triggering_event_type"]) + "`")
        if tcodes:
            meta.append("ATT&CK " + ", ".join(str(t) for t in tcodes))
        if a.get("status"):
            meta.append("status " + str(a["status"]))
        if a.get("disposition"):
            meta.append("disposition " + str(a["disposition"]))
        if meta:
            lines.append(" · ".join(meta))
        rows = []
        seen = set()
        # Keys of interest first: the rule itself says these are what matter.
        for kv in (a.get("rule_keys_of_interest") or []):
            k = kv.get("key")
            vals = [str(v) for v in (kv.get("values") or []) if v not in (None, "")]
            if k and vals and k not in seen:
                seen.add(k)
                rows.append((k, ", ".join(vals)))
        for k, v in sorted((a.get("evidence_fields") or {}).items()):
            if k in seen or k in _MD_SKIP or v in (None, ""):
                continue
            seen.add(k)
            rows.append((k, str(v)))
        if rows:
            lines.append("")
            lines.append("| Field | Value |")
            lines.append("| --- | --- |")
            for k, v in rows[:_MD_MAX_ROWS]:
                if len(v) > _MD_MAX_VALUE:
                    v = v[:_MD_MAX_VALUE] + "…"
                # A pipe inside a value would split the row into a new column.
                lines.append("| " + k.replace("|", "\\|") + " | " + v.replace("|", "\\|") + " |")
            if len(rows) > _MD_MAX_ROWS:
                lines.append("")
                lines.append("_%d more field(s) not shown._" % (len(rows) - _MD_MAX_ROWS))
        lines.append("")
    return "\n".join(lines).strip()


def run():
    _, _, multi = _cfg()
    inp = _inputs()
    limit = int(inp.get("limit") or 50)
    max_alerts = int(inp.get("max_alerts") or 500)
    budget = float(inp.get("time_budget_seconds") or DEFAULT_BUDGET_S)
    started = time.monotonic()

    start_time = inp.get("start_time")
    if not start_time:
        days = int(inp.get("default_days") or 1)
        start_time = (datetime.now(timezone.utc) - timedelta(days=days)).strftime(ISO)

    inv_page = request("GET", "idr/v2/investigations", params={
        "index": inp.get("index") or "0",
        "size": inp.get("page_size") or limit,
        # ",".join(_list(...)): the API rejects the whole call when a list value
        # carries whitespace, so "HIGH, CRITICAL" as typed must lose its space.
        "statuses": ",".join(_list(inp.get("statuses"))) or None,
        "priorities": ",".join(_list(inp.get("priorities"))) or None,
        "start_time": start_time,
        "end_time": inp.get("end_time"),
        "sort_field": "Created time",
        "sort_direction": "ASC",
        "multi-customer": "true" if multi else None,
    })
    investigations = (inv_page.get("data") or [])[:limit]

    # Alert RRNs per investigation. One call each — there is no bulk form — so
    # this is the step that grows with the window, and the one the time budget
    # guards. An investigation whose alerts cannot be read keeps its place in the
    # output with an alerts_error: losing the whole run over one of them, or
    # dropping it silently, would both be worse.
    rrns_by_inv, order, truncated, skipped = {}, [], False, 0
    region_failures, regions_seen = [], set()
    for inv in investigations:
        inv_rrn = inv.get("rrn") or ""
        if not inv_rrn:
            continue
        if time.monotonic() - started > budget or len(order) >= max_alerts:
            truncated = True
            break
        try:
            page = request("GET", "idr/v2/investigations/" + urllib.parse.quote(inv_rrn, safe="") + "/alerts",
                           params={"multi-customer": "true" if multi else None})
        except urllib.error.HTTPError as e:
            inv["alerts_error"] = "HTTP " + str(e.code)
            continue
        got = []
        for a in (page.get("data") or []):
            rrn = _alert_rrn(a)
            if rrn is None:
                skipped += 1
                continue
            got.append(rrn)
            r = _rrn_region(rrn)
            if r:
                regions_seen.add(r)
            if rrn not in order:
                order.append(rrn)
        rrns_by_inv[inv_rrn] = got

    # The alerts themselves, batched, then indexed so each investigation can be
    # handed its own.
    field_ids = _list(inp.get("field_ids"))
    by_rrn = {}
    for i in range(0, len(order), RRN_BATCH):
        body = {"rrns": order[i:i + RRN_BATCH]}
        if field_ids:
            body["field_ids"] = field_ids
        out = _collect_region_failures(request("POST", AT + "alerts/ops/rrns", body=body), region_failures)
        for a in (out.get("alerts") or []):
            if a.get("rrn"):
                by_rrn[a["rrn"]] = a

    # ④ The evidence behind each alert: the actual log line. An alert names what
    #    fired; the evidence is what happened. One call per alert — there is no
    #    bulk form — so it is optional, capped, and under the same time budget.
    evidence_calls, evidence_errors = 0, 0
    if str(inp.get("include_evidence") or "true").strip().lower() in ("1", "true", "yes"):
        per_alert = int(inp.get("max_evidence_per_alert") or 5)
        for rrn, a in by_rrn.items():
            if time.monotonic() - started > budget:
                truncated = True
                break
            try:
                page = request("GET", AT + "alerts/" + urllib.parse.quote(rrn, safe="") + "/evidences",
                               params={"index": "0", "size": per_alert})
                evidence_calls += 1
            except urllib.error.HTTPError as e:
                a["evidence_error"] = "HTTP " + str(e.code)
                evidence_errors += 1
                continue
            evs = (page.get("evidences") or [])[:per_alert]
            for ev in evs:
                ev["fields"], ev["raw_line"] = _evidence_fields(ev)
            a["evidences"] = evs
            if evs:
                # Lifted for the mapper: MapFields cannot index an array, and the
                # ocsf section reads alerts[0] only.
                a["evidence_fields"] = evs[0].get("fields") or {}
                a["evidence_raw_line"] = evs[0].get("raw_line")

    # ⑤ Process trees. A third call per alert, so it is gated: "auto" asks only
    #    for the event types an endpoint sensor produces one for, "true" asks
    #    always, "false" never. Chains are attached to the alert AND to the
    #    investigation, because the layout block reads ctx.process_tree at the
    #    top level.
    tree_mode = str(inp.get("include_process_tree") or "auto").strip().lower()
    tree_calls, tree_errors = 0, 0
    if tree_mode != "false":
        for rrn, a in by_rrn.items():
            if time.monotonic() - started > budget:
                truncated = True
                break
            if tree_mode == "auto" and str(a.get("triggering_event_type") or "") not in _PROCESS_EVENT_TYPES:
                continue
            try:
                page = request("POST", AT + "alerts/" + urllib.parse.quote(rrn, safe="") + "/process_trees/latest",
                               params={"index": "0", "size": "5"})
                tree_calls += 1
            except urllib.error.HTTPError as e:
                # 404 is the normal answer for an alert that has no tree, not a
                # fault: recorded only when it is something else.
                if e.code != 404:
                    a["process_tree_error"] = "HTTP " + str(e.code)
                    tree_errors += 1
                continue
            chains = _tree_chains(page, a)
            if chains:
                a["process_tree"] = chains

    fetched = 0
    for inv in investigations:
        rrns = rrns_by_inv.get(inv.get("rrn") or "", [])
        alerts = [by_rrn[r] for r in rrns if r in by_rrn]
        inv["alerts"] = alerts
        inv["alert_count"] = len(rrns)
        # The block reads ctx.process_tree at the top level and draws one chain
        # per alert, so every alert's chains are gathered here rather than left
        # nested where no mapper expression could reach past alerts[0].
        if alerts:
            inv["summary_markdown"] = _markdown_summary(inv, alerts)
            ev_rows, ev_dropped = _evidence_rows(alerts)
            if ev_rows:
                inv["evidence_rows"] = ev_rows
                if ev_dropped:
                    inv["evidence_columns_dropped"] = ev_dropped
        rolled = []
        for a in alerts:
            rolled.extend(a.get("process_tree") or [])
        if rolled:
            inv["process_tree"] = rolled[:_MAX_CHAINS]
        # Say it when the alert list is a sample: alert_count is what the
        # investigation actually has, len(alerts) what this run retrieved.
        if len(alerts) < len(rrns):
            inv["alerts_truncated"] = True
        fetched += len(alerts)
        # Lift a few first-alert values to the top level. The incident-field
        # mapper resolves dotted paths only — it cannot index an array — so
        # `alerts[0].ai_suggested_disposition` in a field_mappings entry reads as
        # nothing, silently, and the custom field arrives empty with no error
        # anywhere. The OCSF entries are evaluated by full JSONata and do not
        # need this; only field_mappings does.
        if alerts:
            a = alerts[0]
            log = (a.get("log_details") or [{}])[0]
            inv.update({
                "first_alert_rrn": a.get("rrn"),
                "first_alert_type": a.get("type"),
                "first_alert_source": a.get("external_source"),
                "first_alert_status": a.get("status"),
                "first_alert_disposition": a.get("disposition"),
                "first_alert_ai_disposition": a.get("ai_suggested_disposition"),
                "first_alert_event_type": a.get("triggering_event_type"),
                "first_alert_rule_rrn": (a.get("rule") or {}).get("rrn"),
                "first_alert_log_id": log.get("log_id"),
                "first_alert_logset_id": log.get("logset_id"),
                "first_alert_log_entry_id": log.get("log_entry_id"),
            })
            ef = a.get("evidence_fields") or {}
            if ef:
                inv["first_evidence_message"] = ef.get("msg") or ef.get("logdesc") or ef.get("message")
                inv["first_evidence_raw_line"] = a.get("evidence_raw_line")

    print(json.dumps({
        "data": investigations,
        "metadata": {
            "investigations_returned": len(investigations),
            "alert_rrns_collected": len(order),
            "alerts_fetched": fetched,
            "evidence_calls": evidence_calls,
            "evidence_errors": evidence_errors,
            "process_tree_calls": tree_calls,
            "process_tree_errors": tree_errors,
            "alert_ids_without_rrn": skipped,
            "truncated": truncated,
            # HTTP 200 with a non-empty region_failures is a PARTIAL result:
            # alerts from that region are simply absent. Surfaced so a shrinking
            # feed has a cause attached to it.
            "region_failures": region_failures,
            "regions_seen": sorted(regions_seen),
            "configured_region": _cfg_region(),
            "region_mismatch": sorted(r for r in regions_seen if r != _cfg_region()),
            "elapsed_seconds": round(time.monotonic() - started, 1),
        },
    }))


try:
    run()
except urllib.error.HTTPError as e:
    detail = e.read().decode("utf-8", "replace")
    if e.code in (401, 403):
        detail += (" — the Alert Triage API (/idr/at) is only available to Managed Detection and"
                   " Response customers with access to the Alerts experience; a key that works on"
                   " the investigation endpoints can still be refused here")
    print(json.dumps({"error": "HTTP " + str(e.code), "detail": detail}))
    sys.exit(1)
except Exception as e:
    print(json.dumps({"error": str(e)}))
    sys.exit(1)