cleveragents-core/tools/_pr_classification_cache.py

"""Delta-cached PR enumeration for the reviewer dispatcher.

Replaces the 5 per-cycle ``list_prs_*.ts`` subprocess calls (the
flaky path that timed out at 120 s in runs 14 / 15 / 16) with native
Python that:

- Fetches the open-PR list ONCE per cycle, sorted by ``updated_at``
  descending.
- For each PR, checks the on-disk cache (``ForgejoCache.pr_classifications``):
  if ``head_sha`` matches, ``updated_at`` hasn't advanced, schema
  version matches, and last-checked is within TTL → reuse the cached
  classification with ZERO per-PR API calls.
- Re-classifies (fetching CI status + reviews + commits) only for
  PRs that actually changed since the last check.
- Applies the 5 reviewer-filter predicates to all classifications.

See ``.drew/planning/fix list_prs_by_filter.md`` for the full plan,
including phased migration (Phase 1 ships this module + the MCP tool
with the dispatcher unchanged; Phase 2 cuts the dispatcher over
behind a feature flag).

Cache hit predicate (all four must hold):

  cached.head_sha               == pr.head.sha
  cached.updated_at             >= pr.updated_at
  cached.classification_schema  == current
  now - cached.last_checked_at  <  ttl_seconds

The schema_version axis is the lever for invalidating the entire
cache en masse when a new classification axis is added (e.g. adding
``has_unmerged_dependency`` would bump from v1 → v2 and all v1 rows
would cache-miss-and-refetch on first read).

This module is pure logic over the cache + the existing
``_review_fetch`` helpers — no agent-side I/O, no LLM calls.
"""
from __future__ import annotations

import datetime as _dt
import json
import logging
import os
import sys
from datetime import datetime, timedelta, timezone
from pathlib import Path
from types import SimpleNamespace
from typing import Any

sys.path.insert(0, str(Path(__file__).resolve().parent))
from _loader import load_sibling  # noqa: E402

_claim_runtime = load_sibling("_claim_runtime", "_claim_runtime.py")
_review_fetch = load_sibling("_review_fetch", "_review_fetch.py")
_forgejo_cache = load_sibling("_forgejo_cache", "_forgejo_cache.py")
_backoff = load_sibling("_backoff", "_backoff.py")
_pr_state_cache = load_sibling("_pr_state_cache", "_pr_state_cache.py")

_logger = logging.getLogger(__name__)


# The 5 reviewer-side filter names. Lifted verbatim from
# ``dispatch_review._make_work_group`` calls so a typo here surfaces
# as a clean ValueError, not a silent no-match.
FILTER_NAMES = (
    "addressed_changes_ci_passing",
    "addressed_changes_ci_failing",
    "no_active_review_ci_passing",
    "no_active_review_ci_failing",
    "missing_ci_checks",
)

DEFAULT_TTL_SECONDS = 300

# Forgejo /pulls flakiness resilience (R1, 2026-05-16). Live tests
# saw the ``/repos/{owner}/{repo}/pulls?state=open&...`` endpoint
# time out at the default 30 s request_timeout_s on roughly 1 in 5
# cycles (4 timeouts observed in run-18 alone). Each timeout
# triggered a 30 s wait followed by a fallback to the legacy
# ``npx --yes tsx list_prs_*.ts`` script that ALSO times out at
# 120 s in npx-cold-start conditions, costing 30-150 s per affected
# cycle. Two cheap fixes:
#
# - **Shorter per-call timeout** (default 8 s) for the listing
#   fetch. Forgejo serves a clean list in <500 ms on the happy
#   path; 8 s is plenty of headroom while bounding the worst case
#   so we fall back fast.
# - **Last-good listing cache + exponential backoff** so the next
#   N cycles inside the backoff window short-circuit to the cached
#   listing instead of paying the timeout again.
# 30s default chosen after run-21 probe (2026-05-16): Forgejo's
# ``/pulls`` endpoint takes ~24s on a cold cache because the response
# inlines full ``head``/``base`` repo+owner metadata (90 KB of 127 KB
# total for 8 PRs). The original 8s default fail-fast caused the
# dispatcher to time out before the cold-cache rebuild completed,
# triggering backoff while Forgejo's already-in-flight rebuild ran
# to completion server-side. Bumping to 30s lets a single attempt
# succeed; subsequent attempts within Forgejo's hot-cache window
# return in <1s.
_LIST_TIMEOUT_S_DEFAULT = 30
_LIST_TIMEOUT_ENV = "REVIEW_DISPATCHER_PR_LIST_TIMEOUT_S"
_LIST_CACHE_DIR_DEFAULT = "/tmp/cleveragents-pr-list-cache"
_LIST_CACHE_DIR_ENV = "REVIEW_DISPATCHER_PR_LIST_CACHE_DIR"
_LIST_BACKOFF = _backoff.Backoff(
    base_env="REVIEW_DISPATCHER_PR_LIST_BACKOFF_BASE_S",
    max_env="REVIEW_DISPATCHER_PR_LIST_BACKOFF_MAX_S",
    base_default=60,
    max_default=1800,
)
_LIST_DISABLE_ENV = "REVIEW_DISPATCHER_PR_LIST_CACHE_DISABLE"
_LIST_CACHE_SCHEMA_VERSION = 1

# Label names that mark a PR as claimed by another agent. Mirrors
# ``list_prs.ts``'s CLAIM_LABELS set. Coordinated with
# ``tools/setup_auto_labels.py`` and the worker-side ``claim_pr.ts``;
# mutating this set is a breaking change against the 4 agents that
# read these labels.
CLAIM_LABELS = frozenset(
    (
        "auto/claimed-merge",
        "auto/claimed-implementer",
        "auto/claimed-reviewer",
    )
)


def refresh_then_filter(
    cfg: Any,
    filter_name: str,
    ttl_seconds: int = DEFAULT_TTL_SECONDS,
    cache: Any = None,
) -> list[dict[str, Any]]:
    """Return open PRs matching ``filter_name``, populating the cache
    as needed.

    Order is the order Forgejo returned (sorted by ``updated_at`` desc
    — most-recently-touched PRs first), so the most-actionable PRs
    surface first in the dispatcher's iteration.

    Pass ``cache`` to share a connection (the dispatcher should open
    one per cycle to avoid SQLite open/close overhead); omit it for
    standalone calls (e.g. from the MCP tool) and the function will
    open/close its own.
    """
    if filter_name not in FILTER_NAMES:
        raise ValueError(
            f"unknown filter {filter_name!r}; "
            f"valid: {list(FILTER_NAMES)}"
        )
    own_cache = cache is None
    if own_cache:
        cache = _forgejo_cache.ForgejoCache.open()
    try:
        prs = _list_open_prs_sorted_by_updated(cfg)
        results: list[dict[str, Any]] = []
        for pr in prs:
            try:
                classification = _ensure_classified(
                    cache, cfg, pr, ttl_seconds,
                )
            except Exception as exc:  # noqa: BLE001
                # Per-PR classification failure (transient API blip
                # on this specific PR's reviews/CI) does NOT abort the
                # whole enumeration — log + skip the PR. Better to
                # return a slightly-incomplete list than to fail the
                # whole reviewer cycle on one bad PR.
                _logger.warning(
                    "classify PR #%s failed; skipping for this cycle: %s",
                    pr.get("number"), exc,
                )
                continue
            if _evaluate_filter(filter_name, classification):
                results.append(_project_pr(pr, classification))
        return results
    finally:
        if own_cache:
            cache.close()


def _list_timeout_s() -> int:
    raw = os.environ.get(_LIST_TIMEOUT_ENV)
    if raw:
        try:
            return max(1, int(raw))
        except ValueError:
            pass
    return _LIST_TIMEOUT_S_DEFAULT


def _list_cache_dir() -> Path:
    return Path(os.environ.get(_LIST_CACHE_DIR_ENV) or _LIST_CACHE_DIR_DEFAULT)


def _list_cache_path(cfg: Any) -> Path:
    """Per-(owner, repo) cache file. We sanitize defensively (the
    same path-traversal check the ci-logs cache uses) so a malformed
    owner/repo couldn't be persuaded to write outside the cache dir."""
    import re as _re
    owner = _re.sub(r"[^a-zA-Z0-9_.-]", "", str(getattr(cfg, "owner", "")))[:64]
    repo = _re.sub(r"[^a-zA-Z0-9_.-]", "", str(getattr(cfg, "repo", "")))[:64]
    return _list_cache_dir() / f"{owner or 'INVALID'}.{repo or 'INVALID'}.json"


def _list_cache_disabled() -> bool:
    raw = os.environ.get(_LIST_DISABLE_ENV, "").strip().lower()
    return raw in {"1", "true", "yes", "on"}


# Thin wrappers preserved for the existing test suite + any direct
# imports; delegate to the shared :class:`_backoff.Backoff`.
def _list_compute_next_attempt_after(
    failures: int, now_dt: _dt.datetime,
) -> str | None:
    return _LIST_BACKOFF.next_attempt_after(failures, now_dt)


def _list_backoff_active(
    cached: dict[str, Any] | None, now_dt: _dt.datetime,
) -> bool:
    return _LIST_BACKOFF.is_active(cached, now_dt)


def _read_list_cache(cfg: Any) -> dict[str, Any] | None:
    path = _list_cache_path(cfg)
    if not path.exists():
        return None
    try:
        payload = json.loads(path.read_text(encoding="utf-8"))
    except (OSError, ValueError) as exc:
        _logger.warning(
            "PR-list cache read failed for %s: %s", path, exc,
        )
        return None
    if not isinstance(payload, dict):
        return None
    if payload.get("schema_version") != _LIST_CACHE_SCHEMA_VERSION:
        return None
    if not isinstance(payload.get("prs"), list):
        return None
    return payload


def _write_list_cache(cfg: Any, payload: dict[str, Any]) -> None:
    path = _list_cache_path(cfg)
    tmp = path.with_suffix(path.suffix + ".tmp")
    try:
        path.parent.mkdir(parents=True, exist_ok=True)
        tmp.write_text(
            json.dumps(payload, indent=2, default=str),
            encoding="utf-8",
        )
        tmp.replace(path)
    except (OSError, TypeError, ValueError) as exc:
        _logger.warning(
            "PR-list cache write failed for %s: %s", path, exc,
        )
        try:
            tmp.unlink(missing_ok=True)
        except OSError:
            pass


def _cfg_with_short_timeout(cfg: Any) -> Any:
    """Shallow shadow of cfg with a shorter ``request_timeout_s`` for
    the /pulls listing only. Keeps the rest of the dispatcher's call
    surface on the standard (longer) timeout. Returns the original
    cfg unchanged when no shortening is in effect.

    Uses ``copy.copy`` rather than a ``dir(cfg)`` walk: the latter
    triggered every descriptor / property getter for side effects we
    don't want, and was slower than the stdlib copy anyway."""
    import copy as _copy
    short = _list_timeout_s()
    current = int(getattr(cfg, "request_timeout_s", 30) or 30)
    if short >= current:
        return cfg
    shadow = _copy.copy(cfg)
    try:
        shadow.request_timeout_s = short
    except (AttributeError, TypeError):
        # AttributeError: slotted shape with no __dict__.
        # TypeError: frozen dataclass raises dataclasses.FrozenInstanceError
        # which subclasses AttributeError on 3.13 but TypeError on
        # older builds — catch the union for safety. Fall back to
        # the dir-walk reconstruction for both cases.
        attrs = {
            k: getattr(cfg, k) for k in dir(cfg)
            if not k.startswith("_")
            and not callable(getattr(cfg, k, None))
        }
        attrs["request_timeout_s"] = short
        shadow = SimpleNamespace(**attrs)
    return shadow


_WARMER_PREFER_ENV = "PR_STATE_WARMER_PREFER"
# How old the warmer's latest write can get before we treat the cache
# as stale and fall through to a live fetch. Default 5 min ≈ 10× the
# warmer's 30s interval — generous enough to absorb a slow warmer
# cycle without false-positive staleness, tight enough to catch a
# dead warmer process within a couple of dispatcher cycles.
_WARMER_STALE_AFTER_S_DEFAULT = 300
_WARMER_STALE_AFTER_S_ENV = "PR_STATE_WARMER_STALE_AFTER_S"
# Floor below which env-provided values are ignored as obvious config
# errors. Production should never let an operator set this below
# the warmer's interval (config drift would make every cycle fall
# through to a live fetch). Tests that need sub-floor values to
# drive the staleness gate end-to-end monkeypatch
# ``_WARMER_STALE_FLOOR_S`` to 1.
_WARMER_STALE_FLOOR_S = 30


def _warmer_preferred() -> bool:
    """Whether to read from :mod:`_pr_state_cache` before falling back
    to a live Forgejo fetch. Default ON; ``=0`` rolls back to the
    pre-warmer code path."""
    raw = os.environ.get(_WARMER_PREFER_ENV)
    if raw is None:
        return True
    return raw.strip().lower() in {"1", "true", "yes", "on"}


def _warmer_stale_after_s() -> int:
    """Staleness threshold in seconds. Default 300s (5 min ≈ 10× the
    warmer's 30s interval).

    Floored at ``_WARMER_STALE_FLOOR_S`` (30s in prod) so operator
    config drift can't accidentally produce sub-interval thresholds
    that fall through to live fetch every cycle. Tests that need
    sub-floor values to drive the gate end-to-end monkeypatch the
    floor symbol directly."""
    raw = os.environ.get(_WARMER_STALE_AFTER_S_ENV)
    if raw:
        try:
            return max(_WARMER_STALE_FLOOR_S, int(raw))
        except ValueError:
            pass
    return _WARMER_STALE_AFTER_S_DEFAULT


def _warmer_cache_fresh(latest_at: str | None) -> bool:
    """True iff ``latest_at`` (ISO-8601) is recent enough to trust.
    None / unparseable → False (treat as stale, fall through).

    A naive ``latest_at`` (no tz suffix — e.g. a test that monkeypatched
    ``_now`` or a future schema-drift) is normalized to UTC before the
    aware-vs-aware comparison so this never raises ``TypeError`` into
    the dispatcher cycle.
    """
    if not latest_at:
        return False
    try:
        parsed = _dt.datetime.fromisoformat(str(latest_at))
    except (ValueError, TypeError):
        return False
    if parsed.tzinfo is None:
        parsed = parsed.replace(tzinfo=_dt.timezone.utc)
    deadline = parsed + _dt.timedelta(seconds=_warmer_stale_after_s())
    return _dt.datetime.now(_dt.timezone.utc) < deadline


def _list_open_prs_sorted_by_updated(cfg: Any) -> list[dict[str, Any]]:
    """List open PRs sorted by most-recently-updated, with three
    layered sources (preferred → fallback):

    1. **Warmer cache** (:mod:`_pr_state_cache`) — populated by the
       ``pr_state_warmer`` sidecar every ~30s. Local SQLite read,
       ~0ms, always paginated across all open PRs, never blocks on
       Forgejo flakes. Default path when warmer is running.
    2. **Last-good list cache** — populated by this module on a
       successful live fetch. Stored at ``/tmp/cleveragents-pr-list-cache``.
       Used when the warmer's cache is empty (warmer not running
       yet / disabled) AND when in active backoff window.
    3. **Live Forgejo fetch** — single-page ``/pulls?state=open&sort=newest&limit=50&page=1``
       call. Used when both caches are empty. Has the 50-PR hard cap
       and cold-cache flakiness this module was originally built to
       compensate for; the warmer is the real fix.

    ``PR_STATE_WARMER_PREFER=0`` skips layer 1 for rollback.
    """
    # Layer 1: warmer cache (preferred). If the warmer has populated
    # the local SQLite at any point, prefer it — it's always fresher
    # than the request-time fallback below (30s poll cadence beats
    # 120s dispatcher cycle) AND it doesn't block on Forgejo.
    #
    # Staleness check: if the latest warmer write is older than
    # ``PR_STATE_WARMER_STALE_AFTER_S`` (default 5 min ≈ 10× the
    # default 30s interval), the warmer process is likely dead and
    # the cache is stale — fall through to the live path so the
    # dispatcher doesn't blindly serve hour-old data.
    if _warmer_preferred() and not _pr_state_cache.is_disabled():
        try:
            warmer_prs = _pr_state_cache.list_open_prs(
                owner=cfg.owner, repo=cfg.repo,
            )
            latest_at = _pr_state_cache.latest_write_at(
                owner=cfg.owner, repo=cfg.repo,
            )
        except _pr_state_cache.PRStateCacheError as exc:
            _logger.warning("warmer-cache read failed; falling through: %s", exc)
            warmer_prs = []
            latest_at = None
        # Freshness is keyed on ``latest_at`` (the warmer's heartbeat),
        # NOT on whether ``warmer_prs`` is non-empty. A repo with zero
        # open PRs and a healthy warmer ``latest_at`` IS fresh — return
        # the empty list directly instead of falling through to live
        # fetch every cycle.
        fresh = bool(latest_at) and _warmer_cache_fresh(latest_at)
        if fresh:
            return warmer_prs
        if latest_at and not fresh:
            _logger.warning(
                "warmer cache for %s/%s is stale (latest write %s); "
                "warmer may be dead — falling through to live fetch",
                cfg.owner, cfg.repo, latest_at,
            )
        # Cold warmer cache (no rows ever written): fall through to the
        # live path below — it doubles as the bootstrap for cold starts.

    path = (
        f"/repos/{cfg.owner}/{cfg.repo}/pulls"
        f"?state=open&sort=newest&limit=50&page=1"
    )
    if _list_cache_disabled():
        return _do_live_pr_list_fetch(cfg, path)
    cached = _read_list_cache(cfg)
    now_dt = _dt.datetime.now(_dt.timezone.utc)
    if cached is not None and _list_backoff_active(cached, now_dt):
        _logger.info(
            "PR-list cache for %s/%s in backoff "
            "(failures=%s, next_attempt_after=%s); serving cached list "
            "(count=%s)",
            cfg.owner, cfg.repo,
            cached.get("consecutive_failures", 0),
            cached.get("next_attempt_after"),
            len(cached.get("prs") or []),
        )
        return list(cached.get("prs") or [])
    short_cfg = _cfg_with_short_timeout(cfg)
    try:
        live = _do_live_pr_list_fetch(short_cfg, path)
    except Exception as exc:  # noqa: BLE001
        # Live fetch failed (timeout / network). Bump failures + serve
        # cached list if we have one — caller still gets actionable
        # data. If no cache exists, re-raise so the dispatcher's
        # outer exception handler can fall back to the legacy TS
        # path (slow but a different code path that occasionally
        # succeeds when /pulls is flaky).
        prior_failures = int((cached or {}).get("consecutive_failures") or 0)
        next_failures = prior_failures + 1
        payload = {
            "schema_version": _LIST_CACHE_SCHEMA_VERSION,
            "fetched_at": _now_iso(),
            "prs": list((cached or {}).get("prs") or []),
            "consecutive_failures": next_failures,
            "next_attempt_after": _list_compute_next_attempt_after(
                next_failures, now_dt,
            ),
            "last_error": f"{type(exc).__name__}: {exc}",
        }
        _write_list_cache(cfg, payload)
        if cached is None:
            _logger.warning(
                "PR-list live fetch failed for %s/%s with no cached "
                "fallback; re-raising for legacy TS fallback: %s",
                cfg.owner, cfg.repo, exc,
            )
            raise
        _logger.warning(
            "PR-list live fetch failed for %s/%s (consecutive_failures=%s); "
            "serving cached list (count=%s); next attempt deferred "
            "until %s",
            cfg.owner, cfg.repo, next_failures,
            len(payload["prs"]), payload["next_attempt_after"],
        )
        return list(payload["prs"])
    # Live success — persist (full overwrite) and clear backoff.
    _write_list_cache(cfg, {
        "schema_version": _LIST_CACHE_SCHEMA_VERSION,
        "fetched_at": _now_iso(),
        "prs": live,
        "consecutive_failures": 0,
        "next_attempt_after": None,
    })
    return live


def _now_iso() -> str:
    return _dt.datetime.now(_dt.timezone.utc).isoformat()


def _do_live_pr_list_fetch(cfg: Any, path: str) -> list[dict[str, Any]]:
    """The raw Forgejo /pulls listing fetch — same shape as the
    pre-R1 code. Lifted out of ``_list_open_prs_sorted_by_updated``
    so the cache wrapper can call it once on the happy path and
    once for the disabled-mode bypass without duplicating logic."""
    resp = _claim_runtime.idempotent_get(path, cfg)
    status = int(resp.get("status") or 0)
    if status != 200:
        _logger.warning(
            "PR list fetch returned status=%s; returning empty list",
            status,
        )
        return []
    body = resp.get("body")
    if not isinstance(body, list):
        _logger.warning(
            "PR list fetch returned non-list body (type=%s); "
            "returning empty list",
            type(body).__name__,
        )
        return []
    return body


def _ensure_classified(
    cache: Any,
    cfg: Any,
    pr: dict[str, Any],
    ttl_seconds: int,
) -> dict[str, Any]:
    """Cache-aware classification: return the cached row when fresh,
    else re-classify and write back."""
    pr_number = int(pr.get("number") or 0)
    cached = cache.get_pr_classification(pr_number)
    if cached and _is_cache_fresh(cached, pr, ttl_seconds):
        return cached
    fresh = _classify_pr(cfg, pr)
    cache.upsert_pr_classification(fresh)
    return fresh


def _is_cache_fresh(
    cached: dict[str, Any],
    pr: dict[str, Any],
    ttl_seconds: int,
) -> bool:
    """Four-axis freshness check (per the plan + module docstring)."""
    head_sha = (pr.get("head") or {}).get("sha") or ""
    pr_updated = pr.get("updated_at") or ""
    if cached.get("head_sha") != head_sha:
        return False
    # ``>=`` because the cache may have been refreshed AFTER the PR's
    # updated_at (e.g. two cycles back-to-back). Only when the PR's
    # updated_at moves strictly past the cache's recorded value do
    # we re-classify.
    if (cached.get("updated_at") or "") < pr_updated:
        return False
    expected_version = (
        _forgejo_cache.ForgejoCache.PR_CLASSIFICATION_SCHEMA_VERSION
    )
    if cached.get("classification_schema_version") != expected_version:
        return False
    last = cached.get("last_checked_at") or ""
    if not last:
        return False
    try:
        last_dt = datetime.fromisoformat(last.replace("Z", "+00:00"))
    except (TypeError, ValueError):
        return False
    age = datetime.now(timezone.utc) - last_dt
    if age > timedelta(seconds=ttl_seconds):
        return False
    return True


def _classify_pr(cfg: Any, pr: dict[str, Any]) -> dict[str, Any]:
    """Per-PR fetch + classification. Returns a dict with all the
    schema fields populated."""
    pr_number = int(pr.get("number") or 0)
    head_sha = (pr.get("head") or {}).get("sha") or ""

    ci_status = _classify_ci_status(cfg, head_sha)

    reviews, _completed = _review_fetch.fetch_existing_reviews(cfg, pr_number)
    approvals_count = _count_active_approvals(reviews)
    has_active_rc = (
        _review_fetch.count_active_request_changes(reviews) > 0
    )
    has_unaddressed_rc = _has_unaddressed_request_changes(
        cfg, pr_number, reviews,
    )

    labels = [(l.get("name") or "") for l in (pr.get("labels") or [])]
    is_claimed = any(name in CLAIM_LABELS for name in labels)

    mergeable = pr.get("mergeable")  # bool or None

    # stale_state: not used by any of the 5 reviewer filters today,
    # so we punt — set to "not_stale" without fetching the base branch.
    # When a future filter needs real stale-state, add a one-call-per-
    # base-ref lookup (typically just origin/master once per cycle).
    stale_state = "not_stale"

    return {
        "pr_number": pr_number,
        "head_sha": head_sha,
        "updated_at": pr.get("updated_at") or "",
        "last_checked_at": datetime.now(timezone.utc).isoformat(),
        "ci_status": ci_status,
        "approvals_count": approvals_count,
        "has_active_request_changes": has_active_rc,
        "has_unaddressed_request_changes": has_unaddressed_rc,
        "is_claimed": is_claimed,
        "is_mergeable": mergeable,
        "stale_state": stale_state,
        "labels_json": json.dumps(labels),
        "classification_schema_version": (
            _forgejo_cache.ForgejoCache.PR_CLASSIFICATION_SCHEMA_VERSION
        ),
    }


def _classify_ci_status(cfg: Any, head_sha: str) -> str:
    """Map Forgejo's combined-status state into the 4-value
    classification ``list_prs.ts`` uses. Match its mapping exactly so
    the Python output is parity with the TS output."""
    if not head_sha:
        return "unknown"
    status = _review_fetch.fetch_ci_status(cfg, head_sha)
    if not status:
        return "unknown"
    state = (status.get("state") or "").lower()
    if state == "success":
        return "passing"
    # list_prs.ts treats failure/error/warning all as "failing".
    if state in ("failure", "error", "warning"):
        return "failing"
    if state == "pending":
        return "pending"
    # Empty string or anything else is "no CI checks reported yet".
    return "unknown"


def _count_active_approvals(reviews: list[dict[str, Any]]) -> int:
    """Count non-dismissed APPROVE reviews using one-per-author
    semantics (only the author's latest review counts).

    Matches ``list_prs.ts``'s approvals_count: a reviewer who first
    APPROVEs then later REQUEST_CHANGES gives zero approvals (the
    APPROVE is superseded). Stale reviews (those marked ``stale``
    after a new commit) still count if not dismissed — that's the
    Forgejo default and matches the TS semantic.
    """
    latest_per_user: dict[str, dict[str, Any]] = {}
    for r in reviews:
        if not isinstance(r, dict):
            continue
        user_obj = r.get("user")
        login = (user_obj.get("login") if isinstance(user_obj, dict) else None) or ""
        submitted_at = r.get("submitted_at") or ""
        current = latest_per_user.get(login)
        if current is None or submitted_at > (current.get("submitted_at") or ""):
            latest_per_user[login] = r
    count = 0
    for r in latest_per_user.values():
        if r.get("dismissed"):
            continue
        if (r.get("state") or "").upper() == "APPROVED":
            count += 1
    return count


def _has_unaddressed_request_changes(
    cfg: Any,
    pr_number: int,
    reviews: list[dict[str, Any]],
) -> bool:
    """True iff there's at least one active REQUEST_CHANGES review
    that has NOT been followed by a new commit.

    Pulls the PR's commit list to compare timestamps against the
    latest active RC review. ``list_prs.ts`` uses the same shape.
    The commit fetch is the most expensive part of classification (PR
    commit list can be long), so the cache-hit path SKIPS this entirely.
    """
    rc_timestamps: list[str] = []
    for r in reviews:
        if not isinstance(r, dict):
            continue
        if r.get("dismissed"):
            continue
        if (r.get("state") or "").upper() != "REQUEST_CHANGES":
            continue
        ts = r.get("submitted_at") or ""
        if ts:
            rc_timestamps.append(ts)
    if not rc_timestamps:
        return False
    latest_rc = max(rc_timestamps)
    commits, _completed = _review_fetch.fetch_pr_commits(cfg, pr_number)
    for c in commits:
        if not isinstance(c, dict):
            continue
        commit = c.get("commit") or {}
        committer = commit.get("committer") if isinstance(commit, dict) else {}
        commit_ts = (
            committer.get("date") if isinstance(committer, dict) else ""
        ) or ""
        if commit_ts > latest_rc:
            return False
    return True


def _evaluate_filter(filter_name: str, row: dict[str, Any]) -> bool:
    """Apply the named filter predicate to a classification row.

    Predicates are mechanical translations of the 5 TS wrapper scripts'
    hard-coded args in
    ``.opencode/skills/auto-agents-system/scripts/list_prs_*.ts``.
    Verified against the scripts' comment blocks 2026-05-16.
    """
    if filter_name == "addressed_changes_ci_passing":
        return (
            row["ci_status"] == "passing"
            and row["approvals_count"] == 0
            and bool(row["has_active_request_changes"])
            and not bool(row["has_unaddressed_request_changes"])
            and not bool(row["is_claimed"])
        )
    if filter_name == "addressed_changes_ci_failing":
        return (
            row["ci_status"] == "failing"
            and row["approvals_count"] == 0
            and bool(row["has_active_request_changes"])
            and not bool(row["has_unaddressed_request_changes"])
            and not bool(row["is_claimed"])
        )
    if filter_name == "no_active_review_ci_passing":
        return (
            row["ci_status"] == "passing"
            and row["approvals_count"] == 0
            and not bool(row["has_active_request_changes"])
            and not bool(row["is_claimed"])
        )
    if filter_name == "no_active_review_ci_failing":
        return (
            row["ci_status"] == "failing"
            and row["approvals_count"] == 0
            and not bool(row["has_active_request_changes"])
            and not bool(row["is_claimed"])
        )
    if filter_name == "missing_ci_checks":
        return (
            row["ci_status"] == "unknown"
            and row["approvals_count"] == 0
            and not bool(row["has_unaddressed_request_changes"])
            and not bool(row["is_claimed"])
        )
    # Unreachable — refresh_then_filter validates filter_name first.
    raise ValueError(f"unknown filter {filter_name!r}")


def _project_pr(
    pr: dict[str, Any],
    cls: dict[str, Any],
) -> dict[str, Any]:
    """Trim the Forgejo PR object to the fields downstream callers
    actually use, plus the classification flags. Matches the shape
    the TS scripts emitted so the dispatcher cutover (Phase 2) is a
    drop-in."""
    return {
        "number": pr.get("number"),
        "title": pr.get("title"),
        "head_sha": cls["head_sha"],
        "head_ref": (pr.get("head") or {}).get("ref"),
        "base_ref": (pr.get("base") or {}).get("ref"),
        "updated_at": pr.get("updated_at"),
        "labels": [(l.get("name") or "") for l in (pr.get("labels") or [])],
        "ci_status": cls["ci_status"],
        "approvals_count": cls["approvals_count"],
        "has_active_request_changes": bool(cls["has_active_request_changes"]),
        "has_unaddressed_request_changes": bool(
            cls["has_unaddressed_request_changes"]
        ),
        "is_claimed": bool(cls["is_claimed"]),
    }


__all__ = (
    "FILTER_NAMES",
    "DEFAULT_TTL_SECONDS",
    "CLAIM_LABELS",
    "refresh_then_filter",
)