@arbiterforge/ca-pi 0.15.5 → 0.15.7

package/plugins/ca-pi/hooks/_usagelib.py

@@ -0,0 +1,484 @@
+#!/usr/bin/env python3
+# codeArbiter — Claude transcript usage normalization, identity, reconciliation
+# and pricing (spec: claude-statusline-accounting-integrity D-5..D-10, D-26, D-27).
+#
+# The ONE owner of what a Claude transcript usage record means. _ledgerlib persists
+# and schedules, and the statusline renders. Neither re-derives usage semantics.
+#
+# Design principles:
+#   - Stdlib only; pure functions over immutable tables; no I/O, no classes.
+#   - Money is exact: integer picodollars (1e-12 USD), with every rate and
+#     modifier held as an exact decimal/rational. There is no float math anywhere
+#     in pricing, so derived totals can be subtracted and re-added exactly.
+#   - A normalized fact merges with another fact of the same request by
+#     element-wise max (counters), min (event time) and union (days, modifier
+#     values). That is commutative, associative and idempotent, so neither scan
+#     order nor replay order can change the accepted state (spec D-6).
+#   - Unknown things are never priced as something else. They add a coverage
+#     reason and contribute no dollars (spec D-8, D-10, D-26, D-27).
+#
+# Public API:
+#   rates(model, speed="standard", geo="global") -> tuple|None   picodollars/token
+#   canonical_model(model) -> str|None
+#   known_models() -> list[str]
+#   classify_identity(record) -> (key|None, alias|None)
+#   normalize(record, day_fn) -> {"kind": "usage"|"malformed"|"ignore", ...}
+#   merge(fact_a, fact_b) -> fact
+#   price(fact) -> {"pd", "in", "out", "cr", "reasons"}
+#   event_day(fact) -> str|None ; days_seen(fact) -> list[str]
+#   usage_bearing(line_bytes) -> bool
+
+import math
+import os
+import re
+import sys
+from fractions import Fraction
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from _fmtlib import parse_iso  # noqa: E402
+
+# --------------------------------------------------------------------------- pricing registry
+# Anthropic API list prices, USD per million tokens, copied verbatim from the
+# official pricing page. Tuple order: input, output, 5-minute cache write,
+# 1-hour cache write, cache read. Strings keep every value an exact decimal.
+REGISTRY_CAPTURED = "2026-09-27"
+REGISTRY_SOURCE = "https://platform.claude.com/docs/en/about-claude/pricing"
+# Bump whenever a rate or rule changes: stored ledger state re-prices from its
+# retained evidence when the version it was priced under differs (spec D-21).
+REGISTRY_VERSION = "2026-09-27.1"
+
+_OPUS_45_TO_5 = ("5", "25", "6.25", "10", "0.50")
+_OPUS_4_41 = ("15", "75", "18.75", "30", "1.50")
+_SONNET_4X = ("3", "15", "3.75", "6", "0.30")
+_FABLE_51 = ("10", "50", "12.50", "20", "0.25")
+_FABLE_5 = ("10", "50", "12.50", "20", "1")
+_PRICES = {
+    "claude-fable-5-1": _FABLE_51,
+    "claude-mythos-5-1": _FABLE_51,
+    "claude-fable-5": _FABLE_5,
+    "claude-mythos-5": _FABLE_5,
+    "claude-opus-5-5": ("4", "20", "5", "8", "0.20"),
+    "claude-opus-5": _OPUS_45_TO_5,
+    "claude-opus-4-8": _OPUS_45_TO_5,
+    "claude-opus-4-7": _OPUS_45_TO_5,
+    "claude-opus-4-6": _OPUS_45_TO_5,
+    "claude-opus-4-5": _OPUS_45_TO_5,
+    "claude-opus-4-1": _OPUS_4_41,
+    "claude-opus-4": _OPUS_4_41,
+    "claude-sonnet-5": ("2", "10", "2.50", "4", "0.20"),
+    "claude-sonnet-4-6": _SONNET_4X,
+    "claude-sonnet-4-5": _SONNET_4X,
+    "claude-sonnet-4": _SONNET_4X,
+    "claude-haiku-4-5": ("1", "5", "1.25", "2", "0.10"),
+    "claude-3-5-haiku": ("0.80", "4", "1", "1.60", "0.08"),
+}
+# Explicit alternate spellings only. Dated snapshot suffixes (-YYYYMMDD) are
+# stripped by rule below; nothing else is guessed.
+_ALIASES = {
+    "claude-opus-4-0": "claude-opus-4",
+    "claude-sonnet-4-0": "claude-sonnet-4",
+}
+# Fast mode (research preview): published input/output rates. Cache multipliers
+# stack on the fast input rate in the model's own standard ratios.
+_FAST = {
+    "claude-opus-5-5": ("8", "40"),
+    "claude-opus-5": ("10", "50"),
+    "claude-opus-4-8": ("10", "50"),
+}
+# Claude 4.6-and-later models: `inference_geo: "us"` is billed at 1.1x, and the
+# 1M context window is billed at standard rates (so a `[1m]` label is an alias).
+_MODERN = frozenset({
+    "claude-fable-5-1", "claude-mythos-5-1", "claude-fable-5", "claude-mythos-5",
+    "claude-opus-5-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7",
+    "claude-opus-4-6", "claude-sonnet-5", "claude-sonnet-4-6",
+})
+_GEO_US_MULTIPLIER = Fraction(11, 10)
+_STANDARD_SPEED = "standard"
+_STANDARD_GEO = frozenset({"global", "not_available"})   # not_available: see spec D-27
+_STANDARD_TIER = "standard"
+
+# Separately metered server-tool counters (spec D-10): picodollars per unit.
+# Web search $10 / 1,000 searches; web fetch has no charge beyond tokens. Any
+# other counter (incl. code execution, billed by unobservable container time)
+# is unknown and makes coverage partial when non-zero.
+TOOL_PRICES = {
+    "web_search_requests": 10 * 10**12 // 1000,
+    "web_fetch_requests": 0,
+}
+
+_PD_PER_USD_PER_MTOK = 10**6        # $/MTok -> picodollars per token
+_DATED = re.compile(r"-\d{8}$")
+
+
+def _pd(value):
+    v = Fraction(value) * _PD_PER_USD_PER_MTOK
+    if v.denominator != 1:           # a rate that is not a whole picodollar is a registry bug
+        raise ValueError(f"non-integral rate {value!r}")
+    return int(v)
+
+
+def known_models():
+    return sorted(_PRICES)
+
+
+def canonical_model(model):
+    """The registry id for a model string, or None when it is not known."""
+    if not isinstance(model, str):
+        return None
+    s = model.strip().lower()
+    long_ctx = s.endswith("[1m]")
+    if long_ctx:
+        s = s[:-4]
+    s = _ALIASES.get(s, s)
+    if s not in _PRICES and _DATED.search(s):
+        s = _DATED.sub("", s)
+        s = _ALIASES.get(s, s)
+    if s not in _PRICES:
+        return None
+    if long_ctx and s not in _MODERN:
+        return None                  # pre-4.6 long context had its own premium
+    return s
+
+
+def rates(model, speed=_STANDARD_SPEED, geo="global"):
+    """(input, output, 5m write, 1h write, cache read) in integer picodollars per
+    token, or None when the model or the modifier combination is not priceable."""
+    canon = canonical_model(model)
+    if canon is None:
+        return None
+    base = [Fraction(v) for v in _PRICES[canon]]
+    if speed == "fast":
+        if canon not in _FAST:
+            return None
+        fin, fout = (Fraction(v) for v in _FAST[canon])
+        base = [fin, fout, fin * base[2] / base[0], fin * base[3] / base[0],
+                fin * base[4] / base[0]]
+    elif speed != _STANDARD_SPEED:
+        return None
+    if geo == "us":
+        if canon not in _MODERN:
+            return None
+        base = [v * _GEO_US_MULTIPLIER for v in base]
+    elif geo not in _STANDARD_GEO:
+        return None
+    return tuple(_pd(v) for v in base)
+
+
+# --------------------------------------------------------------------------- identity
+_MAX_ID = 256
+
+
+def _valid_id(v):
+    return (isinstance(v, str) and 0 < len(v) <= _MAX_ID
+            and not any(ord(ch) < 32 or ord(ch) == 127 for ch in v))
+
+
+def classify_identity(record):
+    """(identity key, message-id alias) for a usage record (spec D-5).
+
+    Keys are domain-separated by class prefix. A record carrying both IDs
+    returns its `mid:` key as an alias, so a message-ID-only representation of
+    the same request can be merged in either arrival order. Anything missing
+    or structurally invalid is unproven: (None, None)."""
+    if not isinstance(record, dict):
+        return None, None
+    msg = record.get("message")
+    mid = msg.get("id") if isinstance(msg, dict) else None
+    mid_key = "mid:" + mid if _valid_id(mid) else None
+    rid = record.get("requestId")
+    if _valid_id(rid):
+        return "rid:" + rid, mid_key
+    return mid_key, None
+
+
+# --------------------------------------------------------------------------- normalization
+MAX_COUNTER = 10**15                 # sanity ceiling for any single usage counter
+_MIN_EPOCH, _MAX_EPOCH = 946684800, 4102444800   # 2000-01-01 .. 2100-01-01
+_TOP = ("input_tokens", "output_tokens", "cache_read_input_tokens",
+        "cache_creation_input_tokens")
+
+
+class _Malformed(Exception):
+    pass
+
+
+def _count(v):
+    """A usage counter as a non-negative int. Absent/None is 0. Booleans,
+    strings, negatives, non-finite, fractional and absurd values are malformed."""
+    if v is None:
+        return 0
+    if isinstance(v, bool):
+        raise _Malformed
+    if isinstance(v, int):
+        n = v
+    elif isinstance(v, float):
+        if not math.isfinite(v) or not v.is_integer():
+            raise _Malformed
+        n = int(v)
+    else:
+        raise _Malformed
+    if n < 0 or n > MAX_COUNTER:
+        raise _Malformed
+    return n
+
+
+def _split(obj):
+    """(c5, c1) from a cache_creation object, or None when absent."""
+    if obj is None:
+        return None
+    if not isinstance(obj, dict):
+        raise _Malformed
+    return (_count(obj.get("ephemeral_5m_input_tokens")),
+            _count(obj.get("ephemeral_1h_input_tokens")))
+
+
+def _vector(obj):
+    """[in, out, cache_read, cache_write_total, c5, c1, classified] for one
+    usage-shaped object (a top-level usage or one iteration)."""
+    inp, out, cr, cw = (_count(obj.get(k)) for k in _TOP)
+    split = _split(obj.get("cache_creation"))
+    if split is None:
+        return [inp, out, cr, cw, 0, 0, 0]
+    return [inp, out, cr, cw, split[0], split[1], 1]
+
+
+def _add(a, b):
+    """Sum two vectors of the SAME representation (several iterations of one kind)."""
+    out = [x + y for x, y in zip(a[:6], b[:6])]
+    return out + [min(a[6], b[6])]
+
+
+def _model_key(model):
+    canon = canonical_model(model)
+    if canon is not None:
+        return canon
+    return "?" + model.strip().lower() if isinstance(model, str) and model.strip() else "?"
+
+
+_MODIFIER_ALIASES = {"geo": {"not_available": "global"}}   # D-27: standard pricing
+
+
+def _modifier(value, kind):
+    """One representation's vote for a price modifier. An absent or empty field
+    casts NO vote (a streaming snapshot that omits it must not conflict with one
+    that states it); defaults are applied only when pricing the merged fact."""
+    if value is None:
+        return []
+    if isinstance(value, str):
+        v = value.strip().lower()
+        return [_MODIFIER_ALIASES.get(kind, {}).get(v, v)] if v else []
+    return ["invalid"]
+
+
+def _trusted_ts(ts):
+    if not isinstance(ts, str):
+        return None
+    e = parse_iso(ts)
+    if e is None or not (_MIN_EPOCH <= e < _MAX_EPOCH):
+        return None
+    return e
+
+
+def normalize(record, day_fn):
+    """Classify one decoded transcript record.
+
+    Returns {"kind": "ignore"} for anything that is not billable assistant usage
+    (non-objects, other record types, zero-usage records such as `<synthetic>`),
+    {"kind": "malformed", "reason": "malformed_record", "id": ...} when usage is
+    present but cannot be trusted, or {"kind": "usage", "id", "alias", "fact"}.
+    `day_fn(timestamp_string)` maps a trustworthy timestamp to its local day."""
+    if not isinstance(record, dict) or record.get("type") != "assistant":
+        return {"kind": "ignore"}
+    msg = record.get("message")
+    if not isinstance(msg, dict) or "usage" not in msg:
+        return {"kind": "ignore"}
+    ident, alias = classify_identity(record)
+    try:
+        fact = _fact(record, msg, msg.get("usage"), day_fn)
+    except _Malformed:
+        return {"kind": "malformed", "reason": "malformed_record", "id": ident}
+    if fact is None:
+        return {"kind": "ignore"}
+    return {"kind": "usage", "id": ident, "alias": alias, "fact": fact}
+
+
+def _fact(record, msg, usage, day_fn):
+    if not isinstance(usage, dict):
+        raise _Malformed
+    comps = {}
+    top = _vector(usage)
+    iters = usage.get("iterations")
+    if iters is None:
+        iters = []
+    if not isinstance(iters, list):
+        raise _Malformed
+    msg_split = None
+    for it in iters:
+        if not isinstance(it, dict):
+            raise _Malformed
+        kind = it.get("type")
+        vec = _vector(it)
+        if kind == "message":
+            msg_split = vec if msg_split is None else _add(msg_split, vec)
+            continue
+        if kind == "advisor_message":
+            key = "a|" + _model_key(it.get("model"))
+        else:
+            key = "x|" + (kind if isinstance(kind, str) else "?") + "|" + _model_key(it.get("model"))
+        comps[key] = vec if key not in comps else _add(comps[key], vec)
+    # The top-level counters are the message iterations (spec observation 2). Their
+    # TTL split is taken from the per-iteration splits when those account for the
+    # whole aggregate; the top-level split can be stale on multi-iteration records.
+    if msg_split is not None and msg_split[6] and msg_split[4] + msg_split[5] == top[3]:
+        top[4], top[5], top[6] = msg_split[4], msg_split[5], 1
+    comps["m|" + _model_key(msg.get("model"))] = top
+    tools = {}
+    stu = usage.get("server_tool_use")
+    if stu is not None:
+        if not isinstance(stu, dict):
+            raise _Malformed
+        for name, v in stu.items():
+            tools[str(name)] = _count(v)
+    if not any(any(v[:6]) for v in comps.values()) and not any(tools.values()):
+        return None                  # non-billable (e.g. <synthetic>, all zero)
+    t = _trusted_ts(record.get("timestamp"))
+    day = day_fn(record["timestamp"]) if t is not None else None
+    return {
+        "t": t,
+        "td": day,
+        "d": [day] if day else [],
+        "c": comps,
+        "tl": tools,
+        "md": {"speed": _modifier(usage.get("speed"), "speed"),
+               "geo": _modifier(usage.get("inference_geo"), "geo"),
+               "tier": _modifier(usage.get("service_tier"), "tier")},
+    }
+
+
+# --------------------------------------------------------------------------- reconciliation
+def merge(a, b):
+    """Reconcile two facts of the SAME request (spec D-6). Commutative,
+    associative and idempotent: counters by max, event time by min, day and
+    modifier sets by union. Cache-write classification is a max-join whose
+    consistency is judged at pricing time (spec D-9)."""
+    if a is None:
+        return b
+    if b is None:
+        return a
+    if a["t"] is None:
+        t, td = b["t"], b["td"]
+    elif b["t"] is None or a["t"] < b["t"]:
+        t, td = a["t"], a["td"]
+    elif b["t"] < a["t"]:
+        t, td = b["t"], b["td"]
+    else:
+        t, td = a["t"], min(x for x in (a["td"], b["td"]) if x) if (a["td"] or b["td"]) else None
+    comps = {}
+    for key in set(a["c"]) | set(b["c"]):
+        x, y = a["c"].get(key), b["c"].get(key)
+        comps[key] = list(x) if y is None else list(y) if x is None else \
+            [max(p, q) for p, q in zip(x, y)]
+    tools = {k: max(a["tl"].get(k, 0), b["tl"].get(k, 0)) for k in set(a["tl"]) | set(b["tl"])}
+    mods = {k: sorted(set(a["md"].get(k, [])) | set(b["md"].get(k, [])))
+            for k in set(a["md"]) | set(b["md"])}
+    return {"t": t, "td": td, "d": sorted(set(a["d"]) | set(b["d"])), "c": comps,
+            "tl": tools, "md": mods}
+
+
+def event_day(fact):
+    return fact.get("td") if fact else None
+
+
+def days_seen(fact):
+    return list(fact.get("d") or []) if fact else []
+
+
+# --------------------------------------------------------------------------- pricing a fact
+def _vec_cost(vec, r, reasons):
+    """Picodollars for one component vector at rates `r` (None = unpriced)."""
+    inp, out, cr, cw, c5, c1, classified = vec
+    if not (classified and c5 + c1 <= cw):
+        c5 = c1 = 0                  # no classification, or one that cannot be trusted
+    if cw - c5 - c1 > 0:
+        reasons.add("unknown_cache_ttl")
+    if r is None:
+        return 0
+    return inp * r[0] + out * r[1] + c5 * r[2] + c1 * r[3] + cr * r[4]
+
+
+def price(fact):
+    """Tokens and exact picodollars for one accepted request, plus the set of
+    coverage reasons that keep it from being completely priced. Display input
+    counts uncached input plus every cache write; cache reads are priced but not
+    counted as input."""
+    reasons = set()
+    total = tin = tout = tcr = 0
+    comps = fact["c"]
+    mods = fact.get("md", {})
+    speeds = mods.get("speed") or [_STANDARD_SPEED]
+    geos = mods.get("geo") or ["global"]
+    tiers = mods.get("tier") or [_STANDARD_TIER]
+    modifiers_ok = len(speeds) == 1 and len(geos) == 1 and tiers == [_STANDARD_TIER]
+
+    top_keys = sorted(k for k in comps if k.startswith("m|"))
+    if len(top_keys) > 1:
+        # One request cannot have run on two models. Count tokens once (the
+        # component-wise max, never a sum) and price nothing until proven.
+        reasons.add("model_conflict")
+        vec = [max(col) for col in zip(*(comps[k] for k in top_keys))]
+        _vec_cost(vec, None, reasons)
+        tin, tout, tcr = vec[0] + vec[3], vec[1], vec[2]
+    elif top_keys:
+        vec = comps[top_keys[0]]
+        model = top_keys[0][2:]
+        r = None
+        if model.startswith("?"):
+            reasons.add("unknown_model")
+        elif modifiers_ok:
+            r = rates(model, speed=speeds[0], geo=geos[0])
+            if r is None:
+                reasons.add("unpriced_modifier")
+        else:
+            reasons.add("unpriced_modifier")
+        total += _vec_cost(vec, r, reasons)
+        tin, tout, tcr = vec[0] + vec[3], vec[1], vec[2]
+
+    for key in sorted(comps):
+        if key.startswith("m|"):
+            continue
+        vec = comps[key]
+        tin += vec[0] + vec[3]
+        tout += vec[1]
+        tcr += vec[2]
+        if key.startswith("a|"):
+            model = key[2:]
+            r = None if model.startswith("?") else rates(model)
+            if r is None and any(vec[:6]):
+                reasons.add("unknown_model")
+            total += _vec_cost(vec, r, reasons)
+        elif any(vec[:6]):
+            reasons.add("unknown_iteration")
+
+    for name, n in sorted(fact.get("tl", {}).items()):
+        if n <= 0:
+            continue
+        if name in TOOL_PRICES:
+            total += n * TOOL_PRICES[name]
+        else:
+            reasons.add("unknown_charge")
+    return {"pd": total, "in": tin, "out": tout, "cr": tcr, "reasons": reasons}
+
+
+# --------------------------------------------------------------------------- byte-level classification
+TAIL_WINDOW = 65536
+_USAGE_KEY = b'"usage":'
+_ASSISTANT = (b'"type":"assistant"', b'"type": "assistant"')
+
+
+def usage_bearing(line):
+    """True when a raw JSONL line may carry billable assistant usage: its final
+    64 KiB contain both a `"usage":` key and an assistant type marker (spec
+    observation 10). Used to skip decoding every other line, and to classify an
+    undecodable or oversized line."""
+    tail = line[-TAIL_WINDOW:]
+    return _USAGE_KEY in tail and any(m in tail for m in _ASSISTANT)

package/plugins/ca-pi/hooks/artifact-reconcile.py

@@ -54,4 +54,8 @@ def main(argv=None):
 
 
 if __name__ == "__main__":
-    raise SystemExit(main())
+    try:
+        raise SystemExit(main())
+    except (_reconciliationlib.ReconciliationError, _artifactlib.ArtifactError, OSError) as exc:
+        print(f"codeArbiter: reconciliation failed: {exc}", file=sys.stderr)
+        raise SystemExit(1) from exc

package/plugins/ca-pi/hooks/hostapi.py

@@ -313,7 +313,7 @@ class Host:
 
     name = "claude"
     adapter_name = "ca"
-    adapter_version = "2.22.5"
+    adapter_version = "2.23.0"
 
     # Update-notifier descriptor. Each independently versioned host overrides
     # these three values in its per-plugin _host.py. Keeping the target,

package/plugins/ca-pi/hooks/statusline.py

@@ -13,19 +13,21 @@
 #   - native-font glyphs only: box-drawing, block elements, arrows, ASCII labels.
 #   - top line = active folder; second line = git project (owner/name + branch),
 #     with a no-git fallback.
-#   - token COUNTS come from the session TRANSCRIPT (deduped per requestId; the
-#     host sends no cumulative counts). The COST is the host's authoritative
-#     cost.total_cost_usd — it already prices every call (incl. subagents, which
-#     live in separate transcripts) the way your bill does; a token*price table
-#     would miss subagents and drift, so it is used only as a fallback.
+#   - token counts AND cost are reconstructed from the transcripts (parent plus
+#     every subagent transcript, each request counted once) and priced at the
+#     list rates pinned in _usagelib: api≈ when complete, api≥ as a lower bound,
+#     api≈? when no defensible figure exists. The host's cost.total_cost_usd is a
+#     separate estimate, shown as host≈ for Session only when reconstruction is
+#     unavailable, and never added to or substituted for an api figure.
 #   - context trusts context_window.used_percentage + context_window_size
 #     (1M for million-token models, 200K otherwise) — never exceeds 100%.
 #
-# A small user-level ledger (~/.codearbiter/ledger.json) tails each session's
-# transcript from a stored byte offset (append-only -> O(new lines) per render),
-# accumulating true cumulative tokens + API-equivalent cost per session, so the
-# box shows this session, today's totals across sessions, and a sparkline of the
-# real per-message token burn.
+# A small user-level ledger (per-session state under
+# ~/.codearbiter/ledger.json.sessions/) tails each transcript from a stored byte
+# offset within a fixed per-render budget, so the box shows this session, today's
+# totals across sessions, and a sparkline of the parent session's per-request
+# burn. Deferred work is never lost: it resumes on the next host redraw (the
+# statusline has no timer of its own), and until then the figure reads api≥.
 #
 # Robustness contract: this script must NEVER print a traceback. The host pipes
 # arbitrary JSON on stdin and the output lands in the user's terminal. Every
@@ -106,7 +108,6 @@ except Exception:  # pragma: no cover — never let an import break the statusli
 try:
     import _ledgerlib
     ledger_update = _ledgerlib.ledger_update
-    _tx_accumulate = _ledgerlib._tx_accumulate
     persist_sess_start = _ledgerlib.persist_sess_start
 except Exception:  # pragma: no cover — never let an import break the statusline
     _ledgerlib = None
@@ -295,17 +296,15 @@ try:
                                                _subagentslib.read_subagents,
                                                _subagentslib.sub_label)
     ACTIVE_WINDOW, SHOW_WINDOW = _subagentslib.ACTIVE_WINDOW, _subagentslib.SHOW_WINDOW
-    MAX_SUB_ROWS, MAX_SUB_FILES, MAX_SUB_LINES = (_subagentslib.MAX_SUB_ROWS,
-                                                  _subagentslib.MAX_SUB_FILES,
-                                                  _subagentslib.MAX_SUB_LINES)
+    MAX_SUB_ROWS, MAX_SUB_FILES = _subagentslib.MAX_SUB_ROWS, _subagentslib.MAX_SUB_FILES
 except Exception:  # pragma: no cover — never let an import break the statusline
     _subagentslib = None
-    ACTIVE_WINDOW = SHOW_WINDOW = MAX_SUB_ROWS = MAX_SUB_FILES = MAX_SUB_LINES = 0
+    ACTIVE_WINDOW = SHOW_WINDOW = MAX_SUB_ROWS = MAX_SUB_FILES = 0
 
     def subagent_dir(data, root, sid):
         return None
 
-    def read_subagents(sdir):
+    def read_subagents(sdir, rec=None):
         return 0, 0, [], (0, 0)
 
     def sub_label(content):
@@ -354,6 +353,7 @@ try:
     EFF_DISP, model_pill, usage_row = _segmentslib.EFF_DISP, _segmentslib.model_pill, _segmentslib.usage_row
     seg_lines, seg_pr, seg_prune = _segmentslib.seg_lines, _segmentslib.seg_pr, _segmentslib.seg_prune
     redshift = _segmentslib.redshift
+    cost_cell, USAGE_ROW_FIXED = _segmentslib.cost_cell, _segmentslib.USAGE_ROW_FIXED
 except Exception:  # pragma: no cover — never let an import break the statusline
     _segmentslib = None
     EFF_DISP = {}
@@ -371,7 +371,12 @@ except Exception:  # pragma: no cover — never let an import break the statusli
         return f"{V2}{model}{RESET}"
 
     def usage_row(label, tin, tout, cost, trail=""):
-        return f"{label} {fmt_tok(tin)}/{fmt_tok(tout)} {usd_fine(cost)}"
+        return f"{label} {fmt_tok(tin)}/{fmt_tok(tout)} {cost}"
+
+    USAGE_ROW_FIXED = 30
+
+    def cost_cell(scope, width, allow_host=False):
+        return "api≈?"   # no segment lib: never claim a figure it cannot label
 
     def seg_lines(data):
         return None
@@ -401,9 +406,10 @@ except Exception:  # pragma: no cover — never let an import break the statusli
 # the lib's numeric samples into a colored sparkline, so it keeps the ANSI
 # dependency out of the lib.
 def burn_spark(rec):
-    """Sparkline of recent per-message token burn — real per-API-call totals
-    accumulated from the transcript (via _ledgerlib.burn_samples), not a
-    time-extrapolated estimate."""
+    """Sparkline of recent per-request token burn, parent session only (child
+    usage counts toward totals, not burn): real per-API-call totals accumulated
+    from the transcript (via _ledgerlib.burn_samples), not a time-extrapolated
+    estimate."""
     samples = _ledgerlib.burn_samples(rec) if _ledgerlib is not None else []
     return sparkline(samples) if len(samples) >= 2 else ""
 
@@ -577,8 +583,11 @@ def _render_active_palette(raw):
 
     led = safe(ledger_update, data, sid)
     if not (isinstance(led, tuple) and len(led) == 3):
-        led = ({}, {"in": 0.0, "out": 0.0, "cost": 0.0}, {"in": 0.0, "out": 0.0, "cost": 0.0})
+        blank = {"in": 0.0, "out": 0.0, "cost": 0.0, "pd": 0, "state": "unavailable",
+                 "reasons": [], "stale": False, "host": None}
+        led = ({}, dict(blank), dict(blank))
     rec, sess, day = led
+    rec = rec if isinstance(rec, dict) else {}
     spark = safe(burn_spark, rec) or ""
     # True session age from Claude Code's session metadata (the wall clock /usage
     # shows, incl. idle gaps); fall back to the current transcript's first message.
@@ -649,8 +658,15 @@ def _render_active_palette(raw):
     # double-height left table; the right column carries the context-window
     # detail. Rate limits live on the header line now, not their own row.
     LW = max(34, min(50, inner // 2))   # proportional: don't starve the ctx panel at narrow widths
-    srow = safe(usage_row, "Session", sess["in"], sess["out"], sess["cost"], s_trail) or ""
-    trow = safe(usage_row, "Today", day["in"], day["out"], day["cost"]) or ""
+    # The cost cell carries its provenance label (api≈ / api≥ / api≈? / host≈);
+    # widen the usage column just enough that a figure is never clipped
+    # mid-number, while the ctx panel keeps at least 8 columns.
+    avail = (inner - 11) - USAGE_ROW_FIXED
+    scell = safe(cost_cell, sess, avail, allow_host=True) or "api≈?"
+    tcell = safe(cost_cell, day, avail) or "api≈?"
+    LW = max(LW, min(USAGE_ROW_FIXED + max(vlen(scell), vlen(tcell)), inner - 11))
+    srow = safe(usage_row, "Session", sess.get("in"), sess.get("out"), scell, s_trail) or ""
+    trow = safe(usage_row, "Today", day.get("in"), day.get("out"), tcell) or ""
     ctxl = safe(seg_ctx_lines, data, max(8, inner - LW - 3)) or [f"{GREY}ctx --{RESET}", ""]
     div = f"{V0}{V}{RESET}"
     box.row(f"{pad(srow, LW)} {div} {ctxl[0]}")
@@ -663,7 +679,7 @@ def _render_active_palette(raw):
     if not compact:
         sdir = safe(subagent_dir, data, root, sid)
         if sdir:
-            res = safe(read_subagents, sdir)
+            res = safe(read_subagents, sdir, rec)
             if res:
                 active, recent, shown, (tin, tout) = res
                 if shown:

package/plugins/ca-pi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arbiterforge/ca-pi",
-  "version": "0.15.5",
+  "version": "0.15.7",
   "private": true,
   "license": "AGPL-3.0-only",
   "type": "module",