ccx-native 0.0.1 → 0.2.0

package/ccx/claude/usage.py

@@ -0,0 +1,796 @@
+from __future__ import annotations
+
+import argparse
+from contextlib import contextmanager
+from email.utils import parsedate_to_datetime
+import fcntl
+import math
+from pathlib import Path
+from datetime import datetime, timezone
+import json
+import os
+import stat
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+
+from ccx.system import oslayer
+from ccx.system.private_json import (
+    _atomic_private_json,
+    _open_private_file,
+    _read_private_json,
+)
+from ccx.claude.oauth import oauth_headers, oauth_organization_uuid, oauth_token
+from ccx.constants import (
+    CAP_METRICS,
+    CLAUDE_HOME,
+    usage_source_path,
+    USAGE_MIN_FETCH_INTERVAL_S,
+    OAUTH_USAGE_MIN_FETCH_INTERVAL_S,
+    OAUTH_USAGE_STALE_LIMIT_S,
+    UNOFFICIAL_RESET_ENV,
+    CAP_POLL_INTERVAL_S,
+    usage_cache_path,
+    USAGE_STALE_LIMIT_S,
+    USAGE_URL,
+    opt_in_enabled,
+)
+from ccx.errors import InvalidRequest
+
+_SESSION_RESET_PROGRAM = "juniper_tide"
+_SESSION_RESET_RESULTS = frozenset({
+    "reset", "already_used", "not_limited", "ineligible", "unavailable",
+})
+_SESSION_RESET_REASONS = frozenset({
+    "tier", "tenure", "surface", "mobile", "cli_version", "not_at_wall",
+    "weekly_limit", "no_weekly_limit", "other_experiment", "extra_usage",
+    "unavailable", "unknown",
+})
+_SESSION_RESET_ARMS = frozenset({"control", "reset"})
+_SESSION_RESET_TIMEOUT_S = 25.0
+
+
+# --- machine-wide Claude plan usage source and shared cache -----------------
+
+_SOURCE_CHOICES = frozenset({"get_usage", "oauth"})
+_RETRY_ERROR_S = 120.0
+_RETRY_HELD_INITIAL_S = 300.0
+_RETRY_HELD_MAX_S = 1800.0
+
+
+def _open_usage_cache() -> int:
+    return _open_private_file(usage_cache_path())
+
+
+def _read_usage_cache() -> dict | None:
+    try:
+        return _read_private_json(usage_cache_path())
+    except (OSError, ValueError, UnicodeError):
+        return None
+
+
+def _utc_timestamp(epoch: float) -> str:
+    return datetime.fromtimestamp(epoch, tz=timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def _parse_epoch(value: object) -> float | None:
+    if not isinstance(value, str):
+        return None
+    try:
+        result = datetime.fromisoformat(value.replace("Z", "+00:00"))
+        if result.tzinfo is None:
+            return None
+        return result.timestamp()
+    except (ValueError, OverflowError):
+        return None
+
+
+def _cache_age(data: dict | None = None) -> float | None:
+    """Age of the last successful provider sample, not the metadata file mtime."""
+    if data is None:
+        data = _read_usage_cache()
+    epoch = _parse_epoch((data.get("_ccx") or {}).get("fetchedAt")) if isinstance(data, dict) else None
+    return max(0.0, time.time() - epoch) if epoch is not None else None
+
+
+def _write_usage_cache(data: dict) -> None:
+    _atomic_private_json(usage_cache_path(), data)
+
+
+def _usage_source() -> str:
+    """Persistent per-CCX_ROOT selector, read on every use (including live hosts)."""
+    try:
+        record = _read_private_json(usage_source_path())
+    except FileNotFoundError:
+        return "get_usage"
+    except (OSError, ValueError, UnicodeError) as exc:
+        raise InvalidRequest("invalid private Claude usage-source setting") from exc
+    if not isinstance(record, dict) or record.get("version") != 1 or record.get("source") not in _SOURCE_CHOICES:
+        raise InvalidRequest("invalid Claude usage-source setting; choose get_usage or oauth")
+    if record["source"] == "oauth" and not oslayer.direct_oauth_usage_supported():
+        raise InvalidRequest("Claude OAuth usage source is Linux-only; select get_usage")
+    return record["source"]
+
+
+def _lock_file_path() -> Path:
+    return Path(str(usage_cache_path()) + ".lock")
+
+
+@contextmanager
+def _usage_lock(timeout: float):
+    """Serialize source switches and refreshes across independent CCX processes."""
+    path = _lock_file_path()
+    path.parent.mkdir(mode=0o700, parents=True, exist_ok=True)
+    fd = os.open(path, os.O_RDWR | os.O_CREAT | os.O_NOFOLLOW | os.O_NONBLOCK, 0o600)
+    try:
+        st = os.fstat(fd)
+        if not stat.S_ISREG(st.st_mode) or st.st_uid != os.getuid() or st.st_mode & 0o077:
+            raise PermissionError("usage cache lock must be private and owned by this user")
+        deadline = time.monotonic() + max(0.0, timeout)
+        while True:
+            try:
+                fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
+                break
+            except BlockingIOError:
+                if time.monotonic() >= deadline:
+                    raise TimeoutError("Claude usage cache lock busy")
+                time.sleep(min(0.025, max(0.0, deadline - time.monotonic())))
+        try:
+            yield
+        finally:
+            fcntl.flock(fd, fcntl.LOCK_UN)
+    finally:
+        os.close(fd)
+
+
+def set_usage_source(source: str) -> str:
+    """Explicit persistent selector update, without a provider read."""
+    if source not in _SOURCE_CHOICES:
+        raise InvalidRequest("usage source must be get_usage or oauth")
+    if source == "oauth" and not oslayer.direct_oauth_usage_supported():
+        raise InvalidRequest("direct OAuth usage is Linux-only; use get_usage")
+    try:
+        with _usage_lock(8.0):
+            try:
+                previous = _usage_source()
+            except InvalidRequest:
+                # The atomic writer still checks ownership and private-file mode.
+                previous = None
+            if previous != source or not usage_source_path().exists():
+                _atomic_private_json(usage_source_path(), {"version": 1, "source": source})
+                # Retire values from the previous source; never relabel them.
+                _write_usage_cache({"_ccx": _new_metadata(source)})
+    except (OSError, TimeoutError) as exc:
+        raise InvalidRequest(f"cannot set Claude usage source: {exc}") from exc
+    return source
+
+
+_USAGE_RESPONSE_KEYS = frozenset({
+    "five_hour", "seven_day", "seven_day_opus", "seven_day_sonnet",
+    "limits", "model_scoped", "extra_usage", _SESSION_RESET_PROGRAM,
+})
+
+
+def _is_usage_payload(value: object) -> bool:
+    return isinstance(value, dict) and any(key in value for key in _USAGE_RESPONSE_KEYS)
+
+
+def _usage_request_url(*, at_wall: bool, skip_spend: bool = False) -> str:
+    if not at_wall:
+        return USAGE_URL
+    query = {"at_wall": 1}
+    if skip_spend:
+        query["skip_spend"] = 1
+    return f"{USAGE_URL}?{urllib.parse.urlencode(query)}"
+
+
+def _oauth_sample(timeout: float = 8.0, *, at_wall: bool = True,
+                  skip_spend: bool = False) -> dict:
+    """The only read-only path that accesses the private OAuth identity."""
+    token = oauth_token()
+    if not token:
+        return {"kind": "error", "error": "no Claude OAuth token (log in to Claude Code)"}
+    request = urllib.request.Request(
+        _usage_request_url(at_wall=at_wall, skip_spend=skip_spend),
+        headers=oauth_headers(token),
+    )
+    try:
+        with urllib.request.urlopen(request, timeout=timeout) as response:
+            if not 200 <= response.status < 300:
+                return {"kind": "error", "error": f"usage endpoint returned HTTP {response.status}"}
+            body = json.loads(response.read(2 * 1024 * 1024 + 1).decode())
+    except urllib.error.HTTPError as exc:
+        if exc.code == 429:
+            retry_at = None
+            headers = exc.headers
+            if headers is not None:
+                value = headers.get("Retry-After")
+                if value:
+                    try:
+                        retry_at = time.time() + max(0.0, float(value))
+                    except ValueError:
+                        try:
+                            retry_at = parsedate_to_datetime(value).timestamp()
+                        except (ValueError, TypeError, OverflowError):
+                            pass
+            return {"kind": "held", "error": "usage endpoint returned HTTP 429",
+                    "retry_at": retry_at}
+        return {"kind": "error", "error": f"usage endpoint returned HTTP {exc.code}"}
+    except (urllib.error.URLError, TimeoutError, OSError) as exc:
+        return {"kind": "error", "error": f"usage endpoint unreachable: {exc.reason if isinstance(exc, urllib.error.URLError) else type(exc).__name__}"}
+    except (ValueError, UnicodeError):
+        return {"kind": "error", "error": "usage endpoint returned unreadable JSON"}
+    if _is_usage_payload(body):
+        return {"kind": "success", "body": body}
+    return {"kind": "unparsed", "raw": body,
+            "error": "usage endpoint returned an unrecognized payload"}
+
+
+def _request_usage_live_detail(
+    timeout: float = 8.0, *, at_wall: bool = True, skip_spend: bool = False,
+) -> tuple[dict | None, str | None]:
+    """Private direct-OAuth read, kept separate for reset's authoritative preflight."""
+    result = _oauth_sample(timeout, at_wall=at_wall, skip_spend=skip_spend)
+    return (result.get("body"), None) if result["kind"] == "success" else (None, result["error"])
+
+
+def _request_usage_live(
+    timeout: float = 8.0, *, at_wall: bool = True, skip_spend: bool = False,
+) -> dict | None:
+    return _request_usage_live_detail(timeout, at_wall=at_wall, skip_spend=skip_spend)[0]
+
+
+def _new_metadata(source: str) -> dict:
+    """Only the owner-approved persisted CCX metadata fields."""
+    return {"version": 1, "source": source, "fetchedAt": None,
+            "lastAttemptAt": None, "nextAttemptAt": None, "held": False,
+            "heldUntil": None, "heldUntilEstimated": False, "unparsed": False,
+            "raw": None, "error": None}
+
+
+def _matching_cache(source: str) -> dict | None:
+    data = _read_usage_cache()
+    meta = data.get("_ccx") if isinstance(data, dict) else None
+    if isinstance(meta, dict) and meta.get("version") == 1 and meta.get("source") == source:
+        return data
+    return None
+
+
+def _is_stale(data: dict | None, stale_limit: float = USAGE_STALE_LIMIT_S) -> bool:
+    if not isinstance(data, dict):
+        return True
+    meta = data.get("_ccx") or {}
+    if not isinstance(meta, dict) or meta.get("held") or meta.get("unparsed") or meta.get("error"):
+        return True
+    if meta.get("source") == "oauth":
+        stale_limit = max(stale_limit, OAUTH_USAGE_STALE_LIMIT_S)
+    age = _cache_age(data)
+    return age is None or age > stale_limit
+
+
+def _refresh_due(data: dict | None, *, force: bool, max_age: float, now: float) -> bool:
+    if isinstance(data, dict):
+        meta = data.get("_ccx") or {}
+        deadline = _parse_epoch(meta.get("nextAttemptAt"))
+        if deadline is not None and now < deadline:
+            return False
+        if not force:
+            age = _cache_age(data)
+            if age is not None and age < max_age and not meta.get("held") and not meta.get("error") and not meta.get("unparsed"):
+                return False
+    return True
+
+
+def _next_held_interval(old_meta: dict) -> float:
+    if old_meta.get("held"):
+        start = _parse_epoch(old_meta.get("lastAttemptAt"))
+        end = _parse_epoch(old_meta.get("heldUntil"))
+        if start is not None and end is not None and end > start:
+            return min(_RETRY_HELD_MAX_S, max(_RETRY_HELD_INITIAL_S, (end - start) * 2))
+    return _RETRY_HELD_INITIAL_S
+
+
+def _apply_sample(source: str, prior: dict | None, sample: dict, now: float) -> dict:
+    """Last-known provider values are retained; failed attempts never freshen them."""
+    old_meta = (prior or {}).get("_ccx") or {}
+    meta = _new_metadata(source)
+    meta["fetchedAt"] = old_meta.get("fetchedAt")
+    meta["lastAttemptAt"] = _utc_timestamp(now)
+    kind = sample["kind"]
+    if kind == "success" and (
+        not _is_usage_payload(sample.get("body")) or "_ccx" in sample["body"]
+    ):
+        sample = {"kind": "unparsed", "raw": sample.get("body"),
+                  "error": "unrecognized Claude usage response"}
+        kind = "unparsed"
+    if kind == "success":
+        body = dict(sample["body"])
+        meta["fetchedAt"] = _utc_timestamp(now)
+        floor = OAUTH_USAGE_MIN_FETCH_INTERVAL_S if source == "oauth" else USAGE_MIN_FETCH_INTERVAL_S
+        meta["nextAttemptAt"] = _utc_timestamp(now + floor)
+    else:
+        body = {k: v for k, v in (prior or {}).items() if k != "_ccx"}
+        meta["error"] = sample.get("error") or "Claude usage unavailable"
+        if kind == "held":
+            interval = _next_held_interval(old_meta)
+            provider_retry = sample.get("retry_at")
+            valid_retry = isinstance(provider_retry, (int, float)) and now < provider_retry <= now + _RETRY_HELD_MAX_S
+            until = provider_retry if valid_retry else now + interval
+            meta["held"] = True
+            meta["heldUntil"] = _utc_timestamp(until)
+            meta["heldUntilEstimated"] = not valid_retry
+            meta["nextAttemptAt"] = meta["heldUntil"]
+        else:
+            delay = 300.0 if kind == "unavailable" else _RETRY_ERROR_S
+            meta["nextAttemptAt"] = _utc_timestamp(now + delay)
+            if kind == "unparsed":
+                meta["unparsed"] = True
+                meta["raw"] = sample.get("raw")
+    body["_ccx"] = meta
+    return body
+
+
+def _origin(data: dict | None, source: str | None, live: bool,
+            error: str | None = None) -> dict:
+    meta = (data or {}).get("_ccx") or {}
+    age = _cache_age(data)
+    return {"source": "live" if live else "cache" if data is not None else None,
+            "provider": source, "age_s": age,
+            "fetched_at": meta.get("fetchedAt"),
+            "held": bool(meta.get("held")), "held_until": meta.get("heldUntil"),
+            "held_until_estimated": bool(meta.get("heldUntilEstimated")),
+            "stale": _is_stale(data), "unparsed": bool(meta.get("unparsed")),
+            "next_attempt_at": meta.get("nextAttemptAt"),
+            "live_error": error or meta.get("error"), "cache": str(usage_cache_path())}
+
+
+def _fetch_usage_common(*, timeout: float, max_age: float, force: bool) -> tuple[dict | None, dict]:
+    try:
+        source = _usage_source()
+    except InvalidRequest as exc:
+        return None, _origin(None, None, False, str(exc))
+    cached = _matching_cache(source)
+    now = time.time()
+    if not _refresh_due(cached, force=force, max_age=max_age, now=now):
+        return cached, _origin(cached, source, False)
+    began = time.monotonic()
+    try:
+        with _usage_lock(timeout):
+            # A different process might have fetched or changed the source.
+            source = _usage_source()
+            cached = _matching_cache(source)
+            now = time.time()
+            if not _refresh_due(cached, force=force, max_age=max_age, now=now):
+                return cached, _origin(cached, source, False)
+            left = max(0.0, timeout - (time.monotonic() - began))
+            if left <= 0:
+                return cached, _origin(cached, source, False, "Claude usage fetch deadline exceeded")
+            if source == "oauth":
+                sample = _oauth_sample(left)
+            else:
+                from ccx.claude.usage_control import request_usage
+                sample = request_usage(left)
+            fresh = _apply_sample(source, cached, sample, time.time())
+            _write_usage_cache(fresh)
+            return fresh, _origin(fresh, source, sample["kind"] == "success")
+    except (OSError, TimeoutError, InvalidRequest) as exc:
+        return cached, _origin(cached, source, False, f"Claude usage fetch failed: {exc}")
+
+
+def fetch_usage(timeout: float = 8.0, max_age: float = CAP_POLL_INTERVAL_S,
+                force: bool = False, stale_limit: float | None = None) -> dict | None:
+    """One source-aware, process-shared cached Claude account usage read.
+
+    force ignores the caller's age preference, never the 60s global floor or
+    provider hold/backoff. stale_limit additionally fails open for caps.
+    """
+    data, _ = _fetch_usage_common(timeout=8.0 if timeout is None else max(0.0, timeout),
+                                  max_age=max(0.0, max_age), force=force)
+    if data is None or (stale_limit is not None and _is_stale(data, stale_limit)):
+        return None
+    if not _is_usage_payload(data) and not (data.get("_ccx") or {}).get("unparsed"):
+        return None
+    return data
+
+
+def fetch_usage_for_display(
+    *, force: bool, timeout: float = 8.0, max_age: float = CAP_POLL_INTERVAL_S,
+) -> tuple[dict | None, dict]:
+    data, origin = _fetch_usage_common(timeout=max(0.0, timeout),
+                                       max_age=max(0.0, max_age), force=force)
+    if data is not None and not _is_usage_payload(data) and not origin["unparsed"]:
+        return None, origin
+    return data, origin
+
+
+def describe_usage_age(age_s: float | None) -> str:
+    if age_s is None:
+        return "age unknown"
+    if age_s < 90:
+        return f"{max(0, int(age_s))}s old"
+    if age_s < 90 * 60:
+        return f"{int(age_s // 60)}m old"
+    return f"{int(age_s // 3600)}h old"
+
+
+def _strict_provider_time(value: object) -> str | None:
+    if value is None:
+        return None
+    if not isinstance(value, str) or not value.strip():
+        raise ValueError("provider time must be a non-empty ISO string or null")
+    try:
+        parsed = datetime.fromisoformat(value.strip().replace("Z", "+00:00"))
+    except ValueError as exc:
+        raise ValueError("invalid provider time") from exc
+    if parsed.tzinfo is None:
+        raise ValueError("provider time must include a timezone")
+    return value
+
+
+def _parse_session_reset_status(data: dict | None) -> dict | None:
+    """Strictly parse Anthropic's private session-reset status, else fail closed."""
+    if not isinstance(data, dict):
+        return None
+    block = data.get(_SESSION_RESET_PROGRAM)
+    if not isinstance(block, dict):
+        return None
+    eligible = block.get("eligible")
+    available = block.get("available")
+    if not isinstance(eligible, bool) or not isinstance(available, bool):
+        return None
+    reason = block.get("ineligible_reason")
+    if reason is not None and (
+        not isinstance(reason, str) or reason not in _SESSION_RESET_REASONS
+    ):
+        return None
+    arm = block.get("arm")
+    if arm is not None and (not isinstance(arm, str) or arm not in _SESSION_RESET_ARMS):
+        return None
+    in_experiment = block.get("in_experiment")
+    if in_experiment is not None and not isinstance(in_experiment, bool):
+        return None
+    resets_per_week = block.get("resets_per_week")
+    if resets_per_week is not None and (
+        isinstance(resets_per_week, bool)
+        or not isinstance(resets_per_week, (int, float))
+        or not float(resets_per_week).is_integer()
+        or resets_per_week < 0
+    ):
+        return None
+    try:
+        next_available_at = _strict_provider_time(block.get("next_available_at"))
+        weekly_resets_at = _strict_provider_time(block.get("weekly_resets_at"))
+    except ValueError:
+        return None
+    return {
+        "eligible": eligible,
+        "available": available,
+        "claimable": bool(eligible and available and arm == "reset"),
+        "ineligible_reason": reason,
+        "next_available_at": next_available_at,
+        "weekly_resets_at": weekly_resets_at,
+        "resets_per_week": int(resets_per_week) if resets_per_week is not None else None,
+        # Private protocol detail used only inside ccx's mutation gate.
+        "_arm": arm,
+    }
+
+
+def session_reset_status(data: dict | None) -> dict | None:
+    """Stable public projection of the private provider reset entitlement."""
+    parsed = _parse_session_reset_status(data)
+    if parsed is None:
+        return None
+    return {k: v for k, v in parsed.items() if not k.startswith("_")}
+
+
+def _parse_reset_result(value: object) -> dict | None:
+    if not isinstance(value, dict):
+        return None
+    result = value.get("result")
+    if not isinstance(result, str) or result not in _SESSION_RESET_RESULTS:
+        return None
+    try:
+        next_available_at = _strict_provider_time(value.get("next_available_at"))
+        weekly_resets_at = _strict_provider_time(value.get("weekly_resets_at"))
+    except ValueError:
+        return None
+    return {
+        "result": result,
+        "next_available_at": next_available_at,
+        "weekly_resets_at": weekly_resets_at,
+    }
+
+
+def reset_five_hour_limit(*, timeout: float = _SESSION_RESET_TIMEOUT_S) -> dict:
+    """Claim Anthropic's provider-reported weekly 5-hour reset exactly once.
+
+    Eligibility is re-read live immediately before the mutation. A stale cache,
+    the locally displayed 5-hour percentage, or a locally calculated date can
+    never authorize this call. The mutation itself is never retried. After any
+    attempted mutation, one read-only usage fetch reconciles current state; an
+    ambiguous mutation error stays an error even if that reconciliation succeeds.
+
+    This claims Claude Code's own reset offer (a limited rollout) through its
+    account endpoint, which isn't a public API; disabled without
+    CCX_ENABLE_UNOFFICIAL_RESET=1. The opt-in check precedes all network I/O.
+    """
+    if not opt_in_enabled(UNOFFICIAL_RESET_ENV):
+        raise InvalidRequest(
+            "reset_five_hour_limit: this claims Claude Code's five-hour reset "
+            "offer through its account endpoint, which isn't a public API and can "
+            f"change without notice; it is disabled by default. Set {UNOFFICIAL_RESET_ENV}=1 to "
+            "opt in."
+        )
+    if _usage_source() != "oauth":
+        raise InvalidRequest(
+            "reset_five_hour_limit: also select the machine-wide oauth usage "
+            "source with ccx usage --engine claude --set-source oauth"
+        )
+    preflight = _request_usage_live(
+        timeout=min(8.0, timeout), at_wall=True, skip_spend=True
+    )
+    parsed_status = _parse_session_reset_status(preflight)
+    public_status = session_reset_status(preflight)
+    if parsed_status is None or not parsed_status["claimable"]:
+        return {
+            "result": "not_available",
+            "status": public_status,
+            "refreshed": False,
+        }
+
+    token = oauth_token()
+    organization = oauth_organization_uuid()
+    if not token or not organization:
+        return {"result": "auth_error", "status": public_status, "refreshed": False}
+
+    encoded_org = urllib.parse.quote(organization, safe="")
+    url = f"https://api.anthropic.com/api/organizations/{encoded_org}/reset_rate_limits"
+    body = json.dumps({"program": _SESSION_RESET_PROGRAM}, separators=(",", ":")).encode()
+    req = urllib.request.Request(
+        url,
+        data=body,
+        method="POST",
+        headers=oauth_headers(token),
+    )
+    outcome: dict
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as response:
+            if not (200 <= response.status < 300):
+                outcome = {"result": "error"}
+            else:
+                try:
+                    decoded = json.loads(response.read().decode())
+                except ValueError:
+                    decoded = None
+                outcome = _parse_reset_result(decoded) or {"result": "error"}
+    except urllib.error.HTTPError as exc:
+        if exc.code == 429:
+            outcome = {"result": "rate_limited"}
+        elif exc.code in (401, 403):
+            outcome = {"result": "auth_error"}
+        else:
+            outcome = {"result": "error"}
+    except (OSError, urllib.error.URLError, TimeoutError):
+        outcome = {"result": "error"}
+
+    # Reconcile with one safe read. This is not a mutation retry and never
+    # upgrades an ambiguous POST failure into a guessed success.
+    refreshed = _request_usage_live(
+        timeout=min(8.0, timeout), at_wall=True, skip_spend=False
+    )
+    outcome["status"] = (
+        session_reset_status(refreshed) if refreshed is not None else public_status
+    )
+    outcome["refreshed"] = refreshed is not None
+    return outcome
+
+
+def read_usage_cached(stale_limit: float | None = None) -> dict | None:
+    """Nonblocking, cache-only source-aware read for the host's select loop."""
+    try:
+        data = _matching_cache(_usage_source())
+    except InvalidRequest:
+        return None
+    if data is None or (stale_limit is not None and _is_stale(data, stale_limit)):
+        return None
+    return data if _is_usage_payload(data) else None
+
+
+def _valid_pct(value: object) -> float | None:
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        return None
+    number = float(value)
+    return number if math.isfinite(number) and 0.0 <= number <= 100.0 else None
+
+
+def _metric_util(data: dict, key: str) -> float | None:
+    """A missing/null/invalid metric is unknown, never an assumed zero."""
+    field = CAP_METRICS.get(key)
+    if field is None or not data:
+        return None
+    block = data.get(field)
+    util = _valid_pct(block.get("utilization")) if isinstance(block, dict) else None
+    if util is not None:
+        return util
+    if key == "opus":
+        for quota in scoped_weekly_limits(data):
+            if quota["model"].casefold() == "opus":
+                return quota["utilization"]
+    return None
+
+
+def unknown_cap_metrics(data: dict | None, caps: dict | None) -> list[str]:
+    """Explain fail-open omissions without exposing an arbitrary private payload."""
+    if not caps:
+        return []
+    if data is None or (isinstance(data.get("_ccx"), dict) and _is_stale(data)):
+        return [key for key, limit in caps.items() if limit is not None]
+    return [key for key, limit in caps.items()
+            if limit is not None and _metric_util(data, key) is None]
+
+
+def cap_tripped(data: dict | None, caps: dict | None):
+    """Return first known exceeded cap; held/stale/unknown signals fail open."""
+    if not data or not caps:
+        return None
+    if isinstance(data.get("_ccx"), dict) and _is_stale(data):
+        return None
+    for key, cap in caps.items():
+        if cap is None:
+            continue
+        util = _metric_util(data, key)
+        if util is not None and util >= float(cap):
+            return (key, util, float(cap))
+    return None
+
+
+_SUMMARY_QUOTA_FIELDS = ("five_hour", "seven_day")
+_LEGACY_MODEL_FIELDS = ("seven_day_opus", "seven_day_sonnet")
+
+def _summary_block(block) -> dict | None:
+    if block is None:
+        return None
+    if not isinstance(block, dict):
+        return None
+    out = {}
+    if "utilization" in block:
+        util = block.get("utilization")
+        validated = _valid_pct(util)
+        if validated is not None:
+            out["utilization"] = validated
+    resets_at = block.get("resets_at")
+    if _parse_epoch(resets_at) is not None:
+        out["resets_at"] = resets_at
+    return out or None
+
+
+def scoped_weekly_limits(data: dict) -> list[dict]:
+    """Normalize the currently relevant per-model weekly quota catalog.
+
+    The ``limits`` catalog describes model-specific quotas. Consumers follow
+    that catalog so their UI does not depend on a fixed set of model names.
+    """
+    out = []
+    limits = data.get("limits") if isinstance(data, dict) else None
+    for item in limits if isinstance(limits, list) else ():
+        if not isinstance(item, dict) or item.get("kind") != "weekly_scoped":
+            continue
+        scope = item.get("scope")
+        model = scope.get("model") if isinstance(scope, dict) else None
+        name = model.get("display_name") if isinstance(model, dict) else None
+        percent = _valid_pct(item.get("percent"))
+        if not isinstance(name, str) or not name or percent is None:
+            continue
+        quota = {
+            "model": name,
+            "utilization": percent,
+        }
+        model_id = model.get("id")
+        if isinstance(model_id, str) and model_id:
+            quota["model_id"] = model_id
+        resets_at = item.get("resets_at")
+        if _parse_epoch(resets_at) is not None:
+            quota["resets_at"] = resets_at
+        if isinstance(item.get("is_active"), bool):
+            quota["is_active"] = item["is_active"]
+        out.append(quota)
+    # The control reply's model_scoped[] is additive. Never synthesize limits[]
+    # in the cache; only add missing models to the in-memory presentation.
+    scoped = data.get("model_scoped") if isinstance(data, dict) else None
+    seen_names = {item["model"].casefold() for item in out}
+    if isinstance(scoped, list):
+        for item in scoped:
+            if not isinstance(item, dict):
+                continue
+            name, util = item.get("display_name"), _valid_pct(item.get("utilization"))
+            if not isinstance(name, str) or not name or util is None or name.casefold() in seen_names:
+                continue
+            quota = {"model": name, "utilization": util}
+            if _parse_epoch(item.get("resets_at")) is not None:
+                quota["resets_at"] = item["resets_at"]
+            out.append(quota)
+            seen_names.add(name.casefold())
+    return out
+
+
+def usage_summary(data: dict) -> dict:
+    """Compact public quota summary: stable actionable fields only.
+
+    Includes five_hour, seven_day, provider-announced model_quotas, and
+    extra_usage (when enabled). Non-null legacy per-model fields are retained
+    for cached payload compatibility. Omits all other raw API keys."""
+    meta = data.get("_ccx") or {}
+    if isinstance(meta, dict) and meta.get("unparsed"):
+        return {"unparsed": True, "raw": meta.get("raw"), "error": meta.get("error")}
+    out = {}
+    for key in _SUMMARY_QUOTA_FIELDS:
+        if key not in data:
+            continue
+        block = _summary_block(data.get(key))
+        if block is not None:
+            out[key] = block
+    model_quotas = scoped_weekly_limits(data)
+    if model_quotas:
+        out["model_quotas"] = model_quotas
+    else:
+        # Old caches may predate the dynamic limits catalog.  Preserve actual
+        # legacy quota blocks, but omit null retired fields rather than showing
+        # misleading Opus/Sonnet 0% gauges.
+        for key in _LEGACY_MODEL_FIELDS:
+            block = data.get(key)
+            if not isinstance(block, dict):
+                continue
+            summary = _summary_block(block)
+            if summary is not None:
+                out[key] = summary
+    xu = data.get("extra_usage")
+    if isinstance(xu, dict) and xu.get("is_enabled"):
+        extra = {"enabled": True}
+        for key in ("utilization", "used_credits", "monthly_limit", "currency"):
+            value = xu.get(key)
+            if value is not None:
+                extra[key] = value
+        out["extra_usage"] = extra
+    return out
+
+
+def _pct(s: str) -> float:
+    """argparse type for a 0–100 utilization percentage."""
+    try:
+        v = float(s)
+    except ValueError:
+        raise argparse.ArgumentTypeError(f"{s!r} is not a number")
+    if not (0.0 <= v <= 100.0):
+        raise argparse.ArgumentTypeError(f"cap must be between 0 and 100 (got {v})")
+    return v
+
+def _add_cap_flags(p):
+    p.add_argument("--cap-5h", type=_pct, default=None, metavar="PCT",
+                   help="Refuse to start when 5-hour usage is already at PCT%% (0–100), and stop "
+                        "the turn if it reaches PCT%% while running (exit 7).")
+    p.add_argument("--cap-7d", type=_pct, default=None, metavar="PCT",
+                   help="The same for 7-day usage.")
+    p.add_argument("--cap-opus", type=_pct, default=None, metavar="PCT",
+                   help="Claude only: the same for 7-day Opus usage.")
+
+def _caps_from_args(args) -> dict | None:
+    caps = {}
+    for k, attr in (("5h", "cap_5h"), ("7d", "cap_7d"), ("opus", "cap_opus")):
+        v = getattr(args, attr, None)
+        if v is not None:
+            caps[k] = v
+    return caps or None
+
+def _merge_caps(caps_list) -> dict | None:
+    """Combine the caps of several sends that batch into one turn — strictest wins
+    (lowest threshold per metric, union of metrics)."""
+    merged = {}
+    for c in caps_list:
+        if not c:
+            continue
+        for k, v in c.items():
+            if v is None:
+                continue
+            merged[k] = min(v, merged[k]) if k in merged else v
+    return merged or None