tsugite-claude-code 0.23.2__tar.gz → 0.24.0__tar.gz

{tsugite_claude_code-0.23.2 → tsugite_claude_code-0.24.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.5
 Name: tsugite-claude-code
-Version: 0.23.2
+Version: 0.24.0
 Summary: Tsugite plugin: Claude Code provider (routes LLM calls through the claude CLI)
 Author: Justyn Shull
 Requires-Python: >=3.11
-Requires-Dist: tsugite-cli==0.23.2
+Requires-Dist: tsugite-cli==0.24.0

{tsugite_claude_code-0.23.2 → tsugite_claude_code-0.24.0}/pyproject.toml

@@ -1,11 +1,11 @@
 [project]
 name = "tsugite-claude-code"
-version = "0.23.2"
+version = "0.24.0"
 description = "Tsugite plugin: Claude Code provider (routes LLM calls through the claude CLI)"
 authors = [{ name = "Justyn Shull" }]
 requires-python = ">=3.11"
 dependencies = [
-  "tsugite-cli==0.23.2",
+  "tsugite-cli==0.24.0",
 ]
 
 [project.entry-points."tsugite.providers"]

{tsugite_claude_code-0.23.2 → tsugite_claude_code-0.24.0}/tsugite_claude_code/process.py

@@ -19,6 +19,7 @@ from collections.abc import AsyncIterator
 from pathlib import Path
 
 from tsugite.cli.helpers import get_workspace_dir
+from tsugite_claude_code.usage import record_rate_limit
 
 logger = logging.getLogger(__name__)
 
@@ -368,7 +369,8 @@ class ClaudeCodeProcess:
                 }
                 return
 
-            # Skip: rate_limit_event, system, etc.
+            elif event_type == "rate_limit_event":
+                record_rate_limit(event.get("rate_limit_info") or {})
 
     async def stop(self) -> None:
         """Terminate subprocess and clean up temp files."""

{tsugite_claude_code-0.23.2 → tsugite_claude_code-0.24.0}/tsugite_claude_code/provider.py

@@ -14,62 +14,43 @@ logger = logging.getLogger(__name__)
 
 _CLAUDE_CODE_EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"]
 
+# --- BEGIN GENERATED MODELS (claude_code) --- #
+# Auto-generated by scripts/update_model_registry.py, do not edit manually
+# fmt: off
 _CLAUDE_CODE_MODELS: dict[str, ModelInfo] = {
-    "claude_code/claude-fable-5": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-fable-5-1": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-haiku-4-5": ModelInfo(
-        max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-haiku-4-5-20251001": ModelInfo(
-        max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-4-5": ModelInfo(
-        max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-4-5-20251101": ModelInfo(
-        max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-4-6": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-4-7": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-4-8": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-opus-5": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-sonnet-4-5": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-sonnet-4-5-20250929": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-sonnet-4-6": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
-    "claude_code/claude-sonnet-5": ModelInfo(
-        max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS
-    ),
+    "claude_code/claude-fable-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-fable-5-1": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-haiku-4-5": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-haiku-4-5-20251001": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-haiku-5-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-4-5": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-4-5-20251101": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-4-6": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-4-7": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-4-8": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-opus-5-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-sonnet-4-5": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-sonnet-4-5-20250929": ModelInfo(max_input_tokens=200_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-sonnet-4-6": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-sonnet-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
+    "claude_code/claude-sonnet-5-5": ModelInfo(max_input_tokens=1_000_000, supports_vision=True, supported_effort_levels=_CLAUDE_CODE_EFFORT_LEVELS),
 }
+# fmt: on
+# --- END GENERATED MODELS (claude_code) --- #
 
 # Bare tier aliases track the newest model in that tier; version-pinned aliases
 # stay for reproducible configs. Exception: opus stays on 4-8. Opus 5's replies
 # through Claude Code's resume path can leave an empty assistant block that wedges
 # the sidecar transcript (400 "text content blocks must be non-empty"), and the
-# fresh-session fallback only covers the first send, so opus-5 is opt-in via its
-# pinned alias until a mid-conversation poison recovers cleanly.
+# fresh-session fallback only covers the first send, so opus-5 and opus-5-5 are
+# opt-in via their pinned aliases until a mid-conversation poison recovers cleanly.
 _ALIASES = {
     "fable": "claude-fable-5-1",
     "fable-5-1": "claude-fable-5-1",
     "fable-5": "claude-fable-5",
     "opus": "claude-opus-4-8",
+    "opus-5-5": "claude-opus-5-5",
     "opus-5": "claude-opus-5",
     "opus-4-8": "claude-opus-4-8",
     "opus-4-7": "claude-opus-4-7",
@@ -77,11 +58,13 @@ _ALIASES = {
     "opus-4-5": "claude-opus-4-5",
     "opus-4-5-20251101": "claude-opus-4-5-20251101",
     "sonnet": "claude-sonnet-5",
+    "sonnet-5-5": "claude-sonnet-5-5",
     "sonnet-5": "claude-sonnet-5",
     "sonnet-4-6": "claude-sonnet-4-6",
     "sonnet-4-5": "claude-sonnet-4-5",
     "sonnet-4-5-20250929": "claude-sonnet-4-5-20250929",
-    "haiku": "claude-haiku-4-5-20251001",
+    "haiku": "claude-haiku-5-5",
+    "haiku-5-5": "claude-haiku-5-5",
     "haiku-4-5": "claude-haiku-4-5",
     "haiku-4-5-20251001": "claude-haiku-4-5-20251001",
 }
@@ -150,6 +133,8 @@ class ClaudeCodeProvider:
         self._resume_session = kwargs.get("resume_session")
         self._resume_after_compaction = kwargs.get("resume_after_compaction", False)
         self._previous_messages = kwargs.get("previous_messages", [])
+        # total_cost_usd is cumulative for the whole --resume'd session, not this process.
+        self._cumulative_cost = kwargs.get("resume_cost_baseline") or 0.0
 
     def get_state(self) -> dict | None:
         return {
@@ -158,6 +143,7 @@ class ClaudeCodeProvider:
             "context_window": self._context_window,
             "cache_creation_tokens": self._cache_creation_tokens,
             "cache_read_tokens": self._cache_read_tokens,
+            "cumulative_cost": self._cumulative_cost,
         }
 
     async def acompletion(
@@ -236,6 +222,7 @@ class ClaudeCodeProvider:
         await self._process.stop()
         self._resume_session = None
         self._resume_after_compaction = False
+        self._cumulative_cost = 0.0
         self._process = ClaudeCodeProcess()
         await self._process.start(
             model=self._resolved_model,
@@ -350,6 +337,9 @@ class ClaudeCodeProvider:
             header += ">"
             parts.append(header + "\n" + "\n\n".join(history_lines) + "\n</conversation_history>\n")
 
+        if self._resume_session:
+            parts.extend(self._unseen_delivery_cards())
+
         # Add the task (last user message — earlier user messages are context/history)
         for msg in reversed(messages):
             if msg.get("role") == "user":
@@ -379,6 +369,22 @@ class ClaudeCodeProvider:
             blocks.insert(0, {"type": "text", "text": text})
         return blocks
 
+    def _unseen_delivery_cards(self) -> list[str]:
+        """FYI delivery cards recorded after the last assistant reply.
+
+        A resumed session skips stored history and the CLI transcript never saw
+        these cards. needs_ack cards already render in <pending_deliveries>.
+        """
+        last_reply = max((i for i, m in enumerate(self._previous_messages) if m.get("role") == "assistant"), default=-1)
+        cards = []
+        for msg in self._previous_messages[last_reply + 1 :]:
+            content = msg.get("content")
+            if not isinstance(content, str) or not content.startswith("<tsugite_delivery "):
+                continue
+            if 'kind="needs_ack"' not in content.split("\n", 1)[0]:
+                cards.append(content + "\n")
+        return cards
+
     def _image_blocks(self) -> list[dict]:
         """base64 image `source` payloads for the API-supported image attachments.
 

{tsugite_claude_code-0.23.2 → tsugite_claude_code-0.24.0}/tsugite_claude_code/usage.py

@@ -1,9 +1,11 @@
 """Claude Code subscription limits as a tsugite quota source.
 
 Read from ``limits[]`` in the usage endpoint Claude Code calls, with its OAuth
-token. When that fails, the usage copy cached in Claude Code's config is used
-while it is under an hour old. The list is self-describing, so a limit kind
-tsugite has never seen still shows up.
+token. The endpoint rate-limits hard, so it is called at most hourly and backs
+off after a failure, and each turn's ``rate_limit_event`` keeps the session and
+weekly windows current in between. With no endpoint reading, the usage copy
+cached in Claude Code's config is used while it is under an hour old. The list
+is self-describing, so a limit kind tsugite has never seen still shows up.
 """
 
 from __future__ import annotations
@@ -12,6 +14,7 @@ import json
 import logging
 import os
 import time
+from datetime import datetime, timezone
 from pathlib import Path
 
 import httpx
@@ -25,9 +28,14 @@ PROVIDER = "claude_code"
 LABEL = "Claude Code"
 USAGE_URL = "https://api.anthropic.com/api/oauth/usage"
 CACHE_MAX_AGE_MS = 60 * 60 * 1000
-LIVE_TTL_S = 60
+LIVE_TTL_S = 60 * 60
+BACKOFF_S = (5 * 60, 10 * 60, 20 * 60, 60 * 60)
+TURN_WINDOWS = {"five_hour": "session", "seven_day": "weekly_all"}
 
 _last_live: tuple[float, list] | None = None
+_failures = 0
+_retry_at = 0.0
+_turn: dict[str, tuple[float, QuotaWindow]] = {}
 
 
 def _config_path() -> Path:
@@ -40,25 +48,35 @@ def _credentials_path() -> Path:
     return Path(base) / ".credentials.json"
 
 
-def _live_limits() -> list:
-    global _last_live
-    if _last_live and time.monotonic() - _last_live[0] < LIVE_TTL_S:
-        return _last_live[1]
+def _request_limits() -> list:
     token = json.loads(_credentials_path().read_text())["claudeAiOauth"]["accessToken"]
     headers = {"Authorization": f"Bearer {token}", "anthropic-beta": "oauth-2025-04-20"}
     set_user_agent_header(headers)
     response = httpx.get(USAGE_URL, headers=headers, timeout=10)
     response.raise_for_status()
-    limits = response.json()["limits"]
-    _last_live = (time.monotonic(), limits)
-    return limits
+    return response.json()["limits"]
 
 
-def _cached_limits() -> list:
+def _refresh_live() -> None:
+    global _last_live, _failures, _retry_at
+    now = time.time()
+    if now < _retry_at:
+        return
+    try:
+        _last_live = (now, _request_limits())
+    except Exception as e:
+        logger.warning("Claude Code usage endpoint failed: %r", e)
+        _retry_at = now + BACKOFF_S[min(_failures, len(BACKOFF_S) - 1)]
+        _failures += 1
+        return
+    _failures, _retry_at = 0, now + LIVE_TTL_S
+
+
+def _cached_limits() -> tuple[float, list]:
     cache = json.loads(_config_path().read_text())["cachedUsageUtilization"]
     if time.time() * 1000 - cache["fetchedAtMs"] > CACHE_MAX_AGE_MS:
         raise ValueError("cached usage is over an hour old")
-    return cache["utilization"]["limits"]
+    return cache["fetchedAtMs"] / 1000, cache["utilization"]["limits"]
 
 
 def _label(limit: dict) -> str:
@@ -72,22 +90,41 @@ def _label(limit: dict) -> str:
     return kind
 
 
+def record_rate_limit(info: dict) -> None:
+    """Keep the windows from a turn's ``rate_limit_event`` (``rate_limit_info``)."""
+    windows = info.get("unifiedWindows") or {}
+    for name, key in TURN_WINDOWS.items():
+        window = windows.get(name)
+        if not window or window.get("utilization") is None:
+            continue
+        resets_at = window.get("resetsAt")
+        _turn[key] = (
+            time.time(),
+            QuotaWindow(
+                key=key,
+                label=_label({"kind": key}),
+                used_pct=round(window["utilization"] * 100, 1),
+                resets_at=datetime.fromtimestamp(resets_at, timezone.utc).isoformat() if resets_at else None,
+            ),
+        )
+
+
 def _error(message: str) -> QuotaReport:
     return QuotaReport(provider=PROVIDER, label=LABEL, error=message)
 
 
 def fetch() -> QuotaReport:
     try:
-        limits = _live_limits()
-    except Exception as live_error:
-        try:
-            limits = _cached_limits()
-        except Exception as cache_error:
-            logger.warning("Claude Code usage unavailable: live %r, cache %r", live_error, cache_error)
+        _refresh_live()
+        fetched_at, limits = _last_live or _cached_limits()
+    except Exception as e:
+        if not _turn:
+            logger.warning("Claude Code usage unavailable: %r", e)
             return _error("Claude Code's usage limits could not be fetched, and no recent cached copy exists")
+        fetched_at, limits = 0.0, []
 
     try:
-        windows = tuple(
+        windows = [
             QuotaWindow(
                 key=limit["kind"],
                 label=_label(limit),
@@ -96,11 +133,13 @@ def fetch() -> QuotaReport:
             )
             for limit in limits
             if limit.get("percent") is not None
-        )
+        ]
     except Exception:
         return _error("Claude Code's usage limits are not in a shape tsugite understands")
 
-    return QuotaReport(provider=PROVIDER, label=LABEL, windows=windows)
+    turn = {key: window for key, (read_at, window) in _turn.items() if read_at > fetched_at}
+    windows = [*turn.values(), *(w for w in windows if w.key not in turn)]
+    return QuotaReport(provider=PROVIDER, label=LABEL, windows=tuple(windows))
 
 
 register_quota_source(PROVIDER, LABEL, fetch)