Agent Zero AI framework
Add provider accounting to context window
Alessandro4fb3d2555f66
diff --git a/plugins/_context_window/AGENTS.md b/plugins/_context_window/AGENTS.mdindex a8feedb35..c7dee5a52 100644--- a/plugins/_context_window/AGENTS.md+++ b/plugins/_context_window/AGENTS.md@@ -8,7 +8,8 @@## Ownership- `helpers/usage.py` owns per-prompt bucket measurement and reconciliation.-- `extensions/python/` records prompt parts at their source extension points.+- `extensions/python/` records prompt parts at their source extension points,+ preserves terminal streamed usage, and captures optional provider usage.- `api/context_window.py` exposes the active chat's token usage and effectivemodel limit without returning prompt content.- `webui/` and `extensions/webui/` own the Alpine store, indicator, popover,@@ -26,6 +27,19 @@fragments use a bounded, content-addressed, runtime-only cache.- Bucket totals reconcile to the already-stored prompt token total; theunclaimed remainder belongs to System prompt.+- If the history ledger would consume the whole prompt estimate, recompute only+ the rendered message portion before reconciliation; ordinary prompt builds+ keep the fast ledger path.+- The prompt estimate never guesses provider-specific image token costs or+ counts embedded image bytes as text.+- Provider price, cache hit, and input/output tokens form a flat summary without+ diagnostic detail rows.+- Provider rows are exposed only when the provider or transport reports their+ values; unavailable price and cache data render no row.+- Streamed OpenRouter main turns request LiteLLM's terminal usage event. The+ response callback still runs normally; only an actual Chat Completions result+ restores the accepted response after the accounting tail is drained.+- Responses API turns keep their native result and callback behavior unchanged.- Older chats without a stored breakdown show the explanatory empty state.- `_model_config` supplies the effective model limit and the`model-context-strip-end` WebUI slot; it does not own this feature's state.@@ -35,6 +49,7 @@## Work Guidance- Keep prompt accounting out of rendered-text heuristics.+- Keep provider-reported usage separate from the six estimated context buckets.- Keep the API response limited to counts needed by the UI.- Preserve the upward, right-aligned popover geometry used beside the model andprofile selectors.diff --git a/plugins/_context_window/README.md b/plugins/_context_window/README.mdindex 2f7859f0b..bdb1b293a 100644--- a/plugins/_context_window/README.md+++ b/plugins/_context_window/README.md@@ -7,3 +7,11 @@ Extras, and Free space.Older chats gain the detailed breakdown after their next model turn. Mobile anddesktop visibility can be changed under **Settings > Interface**.++When a model provider reports usage, the popover shows price, cache-hit rate,+and input/output tokens. Unreported price and cache data are omitted. The+context breakdown does not guess model-specific image token costs.++For streamed OpenRouter chat calls, the plugin requests and drains LiteLLM's+terminal usage event after Agent Zero has accepted the response. Price remains+hidden when LiteLLM does not report a cost or map the selected model.diff --git a/plugins/_context_window/api/context_window.py b/plugins/_context_window/api/context_window.pyindex e42da181d..ed701be5e 100644--- a/plugins/_context_window/api/context_window.py+++ b/plugins/_context_window/api/context_window.py@@ -1,5 +1,8 @@from helpers.api import ApiHandler, Input, Output, Request-from plugins._context_window.helpers.usage import usage_snapshot+from plugins._context_window.helpers.usage import (+ latest_provider_usage,+ usage_snapshot,+)from plugins._model_config.helpers.model_config import get_chat_model_config@@ -15,4 +18,5 @@ class ContextWindow(ApiHandler):"tokens": max(int(window.get("tokens") or 0), 0),"context_window": max(int(config.get("ctx_length") or 0), 0),"usage": usage_snapshot(window.get("usage")),+ "provider_usage": latest_provider_usage(agent),}diff --git a/plugins/_context_window/extensions/python/_functions/agent/Agent/call_chat_model_turn/end/_90_record_provider_usage.py b/plugins/_context_window/extensions/python/_functions/agent/Agent/call_chat_model_turn/end/_90_record_provider_usage.pynew file mode 100644index 000000000..5efc7a437--- /dev/null+++ b/plugins/_context_window/extensions/python/_functions/agent/Agent/call_chat_model_turn/end/_90_record_provider_usage.py@@ -0,0 +1,9 @@+from helpers.extension import Extension+from plugins._context_window.helpers.usage import capture_provider_usage+++class RecordProviderUsage(Extension):+ def execute(self, data: dict | None = None, **kwargs):+ payload = data if isinstance(data, dict) else {}+ if not payload.get("exception"):+ capture_provider_usage(self.agent, payload.get("result"))diff --git a/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/end/_90_restore_provider_response.py b/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/end/_90_restore_provider_response.pynew file mode 100644index 000000000..53baa406b--- /dev/null+++ b/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/end/_90_restore_provider_response.py@@ -0,0 +1,24 @@+from helpers.extension import Extension+from helpers.llm_result import LLMResult+++RESPONSE_KEY = "_context_window_accepted_response"+++class RestoreProviderResponse(Extension):+ def execute(self, data: dict, **kwargs):+ response = data.pop(RESPONSE_KEY, None)+ result = data.get("result")+ if (+ response is None+ or data.get("exception")+ or not isinstance(result, LLMResult)+ or result.mode != "chat_completions"+ ):+ return++ result.response = response+ if not result.output_items:+ result.output_items = LLMResult.from_chat(+ response=response, reasoning=result.reasoning+ ).output_itemsdiff --git a/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/start/_10_drain_provider_usage.py b/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/start/_10_drain_provider_usage.pynew file mode 100644index 000000000..cbc90db4c--- /dev/null+++ b/plugins/_context_window/extensions/python/_functions/models/LiteLLMChatWrapper/unified_turn/start/_10_drain_provider_usage.py@@ -0,0 +1,30 @@+from helpers.extension import Extension+++RESPONSE_KEY = "_context_window_accepted_response"+++class DrainProviderUsage(Extension):+ def execute(self, data: dict, **kwargs):+ model = data["args"][0]+ callback = data["kwargs"].get("response_callback")+ if callback is None:+ return++ if (+ model.provider == "openrouter"+ and data["kwargs"].get("explicit_caching")+ ):+ data["kwargs"]["stream_options"] = {+ **model.kwargs.get("stream_options", {}),+ **data["kwargs"].get("stream_options", {}),+ "include_usage": True,+ }++ async def drain_callback(chunk: str, full: str):+ response = await callback(chunk, full)+ if response is not None and RESPONSE_KEY not in data:+ data[RESPONSE_KEY] = response+ return None++ data["kwargs"]["response_callback"] = drain_callbackdiff --git a/plugins/_context_window/extensions/webui/model-context-strip-end/context-window.html b/plugins/_context_window/extensions/webui/model-context-strip-end/context-window.htmlindex b7a4f9e6f..fa8150452 100644--- a/plugins/_context_window/extensions/webui/model-context-strip-end/context-window.html+++ b/plugins/_context_window/extensions/webui/model-context-strip-end/context-window.html@@ -30,7 +30,10 @@aria-label="Context window usage"><div class="context-window-header"><strong>Context window</strong>- <span x-text="$store.contextWindow.usage.summary"></span>+ <div class="context-window-summary">+ <span class="context-window-summary-tokens" x-text="$store.contextWindow.usage.summaryTokens"></span>+ <span class="context-window-summary-percent" x-text="$store.contextWindow.usage.summaryPercent"></span>+ </div></div><div class="context-window-meter" aria-hidden="true"><span :style="$store.contextWindow.usage.meterStyle"></span>@@ -39,7 +42,6 @@<div class="context-window-rows"><template x-for="row in $store.contextWindow.usage.rows" :key="row.key"><div class="context-window-row">- <span class="context-window-dot" :style="row.dotStyle"></span><span x-text="row.label"></span><span class="context-window-value context-window-tokens" x-text="row.tokensLabel"></span><span class="context-window-value" x-text="row.percentLabel"></span>@@ -50,6 +52,28 @@<template x-if="$store.contextWindow.usage.missingBreakdown"><div class="context-window-empty">Breakdown available after the next message.</div></template>+ <template x-if="$store.contextWindow.usage.provider.hasData">+ <div class="context-window-provider">+ <template x-if="$store.contextWindow.usage.provider.price.hasData">+ <div class="context-window-provider-row context-window-provider-price">+ <span>Price</span>+ <span class="context-window-value" x-text="$store.contextWindow.usage.provider.price.label"></span>+ </div>+ </template>+ <template x-if="$store.contextWindow.usage.provider.cache.hasData">+ <div class="context-window-provider-row">+ <span>Cache hit</span>+ <span class="context-window-value" x-text="$store.contextWindow.usage.provider.cache.label"></span>+ </div>+ </template>+ <template x-if="$store.contextWindow.usage.provider.tokens">+ <div class="context-window-provider-row">+ <span>Tokens In/Out</span>+ <span class="context-window-value" x-text="$store.contextWindow.usage.provider.tokens"></span>+ </div>+ </template>+ </div>+ </template></div></div></template>@@ -99,7 +123,7 @@.context-window-button > span {position: relative;z-index: 1;- font-family: var(--font-family-code);+ font-family: var(--font-family-main);font-size: .42rem;font-weight: 600;letter-spacing: -.03em;@@ -117,27 +141,59 @@box-shadow: 0 16px 38px rgba(0, 0, 0, .3);}.context-window-header,- .context-window-row {+ .context-window-row,+ .context-window-provider-row {display: grid;align-items: center;}.context-window-header {- grid-template-columns: minmax(0, 1fr) auto;- gap: .75rem;+ gap: .4rem;font-size: .86rem;}- .context-window-header > span,+ .context-window-summary,.context-window-value {- font-family: var(--font-family-code);+ font-family: var(--font-family-main);font-variant-numeric: tabular-nums;}- .context-window-header > span {+ .context-window-summary {+ display: flex;+ align-items: baseline;+ justify-content: space-between;+ gap: .75rem;+ color: color-mix(in srgb, var(--color-text) 72%, transparent);+ font-size: .76rem;+ }+ .context-window-summary-percent {+ color: var(--color-text);+ }+ .context-window-summary-tokens,+ .context-window-tokens {+ font-family: var(--font-family-code);+ }+ .context-window-provider {+ margin-top: .7rem;+ padding-top: .7rem;+ border-top: 1px solid var(--color-border);+ }+ .context-window-provider-row {+ grid-template-columns: minmax(0, 1fr) auto;+ gap: .75rem;+ min-height: 1.35rem;color: color-mix(in srgb, var(--color-text) 72%, transparent);font-size: .76rem;}+ .context-window-provider-price {+ margin-bottom: .25rem;+ color: var(--color-text);+ font-size: .82rem;+ }+ .context-window-provider-price .context-window-value {+ font-size: 1rem;+ font-weight: 600;+ }.context-window-meter {height: .45rem;- margin: .85rem 0 .75rem;+ margin: .4rem 0 .75rem;overflow: hidden;border-radius: 999px;background: color-mix(in srgb, var(--color-text) 9%, transparent);@@ -159,7 +215,7 @@font-size: .8rem;}.context-window-row {- grid-template-columns: .55rem minmax(0, 1fr) auto 3rem;+ grid-template-columns: minmax(0, 1fr) auto 3rem;gap: .5rem;min-height: 1.5rem;color: color-mix(in srgb, var(--color-text) 78%, transparent);@@ -172,12 +228,6 @@.context-window-tokens {color: color-mix(in srgb, var(--color-text) 62%, transparent);}- .context-window-dot {- width: .55rem;- height: .55rem;- border-radius: 50%;- background: var(--color-highlight);- }@media (max-width: 25rem) {.context-window-popover {right: 1.25rem;diff --git a/plugins/_context_window/helpers/usage.py b/plugins/_context_window/helpers/usage.pyindex 66a9942f0..ba0cc69c7 100644--- a/plugins/_context_window/helpers/usage.py+++ b/plugins/_context_window/helpers/usage.py@@ -1,11 +1,14 @@import hashlib+import mathfrom typing import Anyfrom helpers import files, history, skills, tokens+from helpers.llm_result import result_from_metadataPARTS_KEY = "context_window_usage"CACHE_KEY = "_context_window_usage_cache"+PROVIDER_USAGE_KEY = "context_window_provider_usage"USAGE_KEYS = ("messages","system_tools",@@ -43,6 +46,7 @@ def capture_context(agent: Any, loop_data: Any) -> None:returnoutput = list(getattr(loop_data, "history_output", None) or [])+ parts["_history_output"] = outputskill_output = [message for message in output if skills.skill_instruction_name(message)]skill_tokens = _output_tokens(agent, "history_skills", skill_output)parts["messages"] = max(_history_tokens(agent, output) - skill_tokens, 0)@@ -78,9 +82,20 @@ def finalize(agent: Any) -> None:if not isinstance(parts, dict) or not isinstance(window, dict):return+ history_output = parts.pop("_history_output", None)total = _non_negative_int(window.get("tokens"))usage = {key: _non_negative_int(parts.get(key)) for key in MEASURED_KEYS}measured_total = sum(usage.values())+ if total and measured_total >= total and isinstance(history_output, list):+ message_output = [+ message+ for message in history_output+ if not skills.skill_instruction_name(message)+ ]+ usage["messages"] = _output_tokens(+ agent, "history_messages", message_output+ )+ measured_total = sum(usage.values())if measured_total > total and measured_total:usage = _scale_to_total(usage, total, measured_total)measured_total = total@@ -98,7 +113,73 @@ def usage_snapshot(value: Any) -> dict[str, int]:return {key: _non_negative_int(value.get(key)) for key in USAGE_KEYS}-def _parts(agent: Any) -> dict[str, int] | None:+def capture_provider_usage(agent: Any, result: Any) -> None:+ if agent is None or result is None or not hasattr(result, "usage"):+ return++ snapshot = provider_usage_snapshot(getattr(result, "usage", None))+ agent.set_data(+ PROVIDER_USAGE_KEY,+ snapshot if snapshot else {"available": False},+ )+++def latest_provider_usage(agent: Any) -> dict[str, int | float]:+ data = getattr(agent, "data", None)+ if isinstance(data, dict) and PROVIDER_USAGE_KEY in data:+ stored = data.get(PROVIDER_USAGE_KEY)+ if isinstance(stored, dict) and stored.get("available") is False:+ return {}+ return provider_usage_snapshot(stored)++ all_messages = getattr(getattr(agent, "history", None), "all_messages", None)+ if not callable(all_messages):+ return {}+ for message in reversed(all_messages()):+ if not getattr(message, "ai", False):+ continue+ result = result_from_metadata(getattr(message, "metadata", None))+ if result:+ return provider_usage_snapshot(result.usage)+ return {}+++def provider_usage_snapshot(value: Any) -> dict[str, int | float]:+ if not isinstance(value, dict):+ return {}++ input_details = {+ **_mapping(value.get("prompt_tokens_details")),+ **_mapping(value.get("input_tokens_details")),+ }+ result: dict[str, int | float] = {}+ fields = {+ "input_tokens": (value.get("input_tokens"), value.get("prompt_tokens")),+ "cached_tokens": (+ input_details.get("cached_tokens"),+ input_details.get("cache_read_tokens"),+ value.get("cache_read_input_tokens"),+ value.get("cached_tokens"),+ ),+ "output_tokens": (+ value.get("output_tokens"),+ value.get("completion_tokens"),+ ),+ }+ for key, values in fields.items():+ number = _optional_non_negative_int(*values)+ if number is not None:+ result[key] = number++ cost = _optional_non_negative_float(+ value.get("cost"), value.get("response_cost")+ )+ if cost is not None:+ result["cost"] = cost+ return result+++def _parts(agent: Any) -> dict[str, Any] | None:params = _temporary_params(agent)value = params.get(PARTS_KEY) if params is not None else Nonereturn value if isinstance(value, dict) else None@@ -160,6 +241,34 @@ def _non_negative_int(value: Any) -> int:return 0+def _mapping(value: Any) -> dict[str, Any]:+ return value if isinstance(value, dict) else {}+++def _optional_non_negative_int(*values: Any) -> int | None:+ for value in values:+ if value is None:+ continue+ try:+ return max(int(value), 0)+ except (TypeError, ValueError):+ continue+ return None+++def _optional_non_negative_float(*values: Any) -> float | None:+ for value in values:+ if value is None:+ continue+ try:+ number = float(value)+ except (TypeError, ValueError):+ continue+ if math.isfinite(number):+ return max(number, 0)+ return None++def _scale_to_total(values: dict[str, int], total: int, current: int) -> dict[str, int]:scaled = {key: value * total // current for key, value in values.items()}remainder = total - sum(scaled.values())diff --git a/plugins/_context_window/tests/test_context_window.py b/plugins/_context_window/tests/test_context_window.pyindex 01a7e3cf0..f629f3881 100644--- a/plugins/_context_window/tests/test_context_window.py+++ b/plugins/_context_window/tests/test_context_window.py@@ -1,17 +1,23 @@+import sysfrom pathlib import Pathfrom types import SimpleNamespaceimport pytest+from langchain_core.messages import HumanMessage, SystemMessage++ROOT = Path(__file__).resolve().parents[3]+if str(ROOT) not in sys.path:+ sys.path.insert(0, str(ROOT))++import modelsfrom agent import Agent, LoopData-from helpers import extension, history+from helpers import extension, extract_tools, history, litellm_transport+from helpers.llm_result import LLMResultfrom plugins._context_window.api.context_window import ContextWindowfrom plugins._context_window.helpers import usage-ROOT = Path(__file__).resolve().parents[3]--class _Log:def set_progress(self, _message: str) -> None:pass@@ -80,6 +86,7 @@ async def test_usage_follows_prompt_sources_and_reconciles_to_total(monkeypatch)assert sum(breakdown.values()) == window["tokens"]assert all(breakdown[key] > 0 for key in usage.USAGE_KEYS)assert usage.PARTS_KEY not in loop_data.params_temporary+ assert "history_messages" not in agent.data[usage.CACHE_KEY]@pytest.mark.asyncio@@ -115,6 +122,7 @@ async def test_api_returns_only_counts_and_effective_limit(monkeypatch):"system_prompt": 0,"extras": 0,},+ "provider_usage": {},}assert "text" not in result@@ -147,6 +155,34 @@ def test_webui_and_accounting_are_plugin_owned():assert "right: 1.25rem" in componentassert "width: min(17rem, calc(100vw - 3rem))" in componentassert 'label: "Free space"' in context_store+ assert "Last model call" not in component+ assert ">Price<" in component+ assert ">Cache hit<" in component+ assert ">Tokens In/Out<" in component+ assert "context-window-cache-meter" not in component+ assert "price: {" in context_store+ assert "hasData: cost !== null" in context_store+ assert 'label: cost === null ? "" : formatCost(cost)' in context_store+ assert "usage.provider.price.hasData" in component+ assert "usage.provider.price.label" in component+ assert 'value < 0.001 ? "<$0.001"' in context_store+ assert "maximumSignificantDigits: 3" in context_store+ assert "border-top: 1px solid var(--color-border)" in component+ assert " → " in context_store+ assert "summaryTokens" in context_store+ assert 'summaryPercent: `${percentLabel} used`' in context_store+ assert "formatTokens(output)} tok" not in context_store+ assert "context-window-summary-tokens" in component+ assert "context-window-summary-percent" in component+ assert "font-family: var(--font-family-main)" in component+ assert "<details" not in component+ assert "Reasoning" not in component+ assert "Images sent" not in component+ assert "provider did not split out their token cost" not in context_store+ assert "cached / input" in context_store+ assert "Math.round(cachePercent)" in context_store+ assert "context-window-dot" not in component+ assert "dotStyle" not in context_storeassert "Breakdown available after the next message." in componentassert "startswith(" not in helperassert "rpartition(" not in helper@@ -156,6 +192,9 @@ def test_source_prompt_extensions_are_registered():expected = {"_functions/agent/Agent/prepare_prompt/start": "ResetContextUsage","_functions/agent/Agent/prepare_prompt/end": "StoreContextUsage",+ "_functions/agent/Agent/call_chat_model_turn/end": "RecordProviderUsage",+ "_functions/models/LiteLLMChatWrapper/unified_turn/start": "DrainProviderUsage",+ "_functions/models/LiteLLMChatWrapper/unified_turn/end": "RestoreProviderResponse","message_loop_prompts_after": "CaptureContextUsage",}for point, class_name in expected.items():@@ -178,6 +217,75 @@ def test_source_prompt_extensions_are_registered():assert any(cls.__name__ == recorder for cls in classes)+async def test_chat_stream_drains_terminal_provider_usage(monkeypatch):+ response = '{"tool_name":"response","tool_args":{"text":"done"}}'+ chunks = [+ {"choices": [{"delta": {"content": response}, "message": {}}]},+ {"choices": [{"delta": {"content": " ignored"}, "message": {}}]},+ {+ "choices": [],+ "usage": {+ "prompt_tokens": 12_000,+ "prompt_tokens_details": {"cached_tokens": 9_000},+ "completion_tokens": 80,+ },+ "_hidden_params": {"response_cost": 0.0042},+ },+ ]+ consumed = []++ async def stream():+ for chunk in chunks:+ consumed.append(chunk)+ yield chunk++ async def fake_acompletion(*args, **kwargs):+ assert kwargs["stream_options"] == {"include_usage": True}+ return stream()++ async def fake_rate_limiter(*args, **kwargs):+ return None++ callback_calls = []++ async def response_callback(chunk: str, full: str):+ callback_calls.append((chunk, full))+ return full if extract_tools.extract_tool_request(full) else None++ monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion)+ monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter)+ wrapper = models.LiteLLMChatWrapper(+ model="test-model",+ provider="openrouter",+ model_config=None,+ api_base="https://openrouter.ai/api/v1",+ )++ result = await wrapper.unified_turn(+ messages=[+ SystemMessage(content="stable instructions"),+ HumanMessage(content="question"),+ ],+ response_callback=response_callback,+ explicit_caching=True,+ )++ assert consumed == chunks+ assert callback_calls == [+ (response, response),+ (" ignored", response + " ignored"),+ ]+ assert result.response == response+ assert result.output_items[0].type == "message"+ assert result.usage == {+ "prompt_tokens": 12_000,+ "prompt_tokens_details": {"cached_tokens": 9_000},+ "completion_tokens": 80,+ "cost": 0.0042,+ }++def test_prompt_fragment_cache_is_bounded_and_content_addressed(monkeypatch):calls = []agent = SimpleNamespace(data={}, loop_data=LoopData())@@ -248,3 +356,72 @@ def test_history_ledger_changes_without_invalidating_fragment_cache(monkeypatch)assert second["messages"] == 400 - second["skills"]assert calls.count("stable tools") == 1assert len(agent.data[usage.CACHE_KEY]) == 3+++def test_rendered_history_fallback_preserves_system_prompt_bucket(monkeypatch):+ data = {}+ output = [{"ai": False, "content": "short message"}]+ agent = SimpleNamespace(+ DATA_NAME_CTX_WINDOW="ctx_window",+ data=data,+ loop_data=LoopData(),+ history=SimpleNamespace(get_tokens=lambda: 10_000),+ _build_context_message=lambda *args, **kwargs: [],+ get_data=lambda key: data.get(key),+ set_data=lambda key, value: data.__setitem__(key, value),+ )+ loop_data = SimpleNamespace(+ history_output=output,+ protocol_persistent={},+ protocol_temporary={},+ extras_persistent={},+ extras_temporary={},+ )+ monkeypatch.setattr(+ usage.tokens,+ "approximate_prompt_tokens",+ lambda text: len(text),+ )++ usage.reset(agent)+ usage.capture_context(agent, loop_data)+ data[agent.DATA_NAME_CTX_WINDOW] = {"tokens": 100}+ usage.finalize(agent)++ breakdown = data[agent.DATA_NAME_CTX_WINDOW]["usage"]+ assert breakdown["messages"] == len("user: short message")+ assert breakdown["system_prompt"] > 0+ assert sum(breakdown.values()) == 100+++def test_provider_usage_is_optional():+ data = {}+ agent = SimpleNamespace(+ data=data,+ history=SimpleNamespace(all_messages=lambda: []),+ set_data=lambda key, value: data.__setitem__(key, value),+ )+ result = LLMResult.from_chat(+ response="done",+ usage={+ "prompt_tokens": 12_000,+ "prompt_tokens_details": {"cached_tokens": 9_000},+ "completion_tokens": 80,+ "cost": 0.0123,+ },+ )++ usage.capture_provider_usage(agent, result)++ assert usage.latest_provider_usage(agent) == {+ "input_tokens": 12_000,+ "cached_tokens": 9_000,+ "output_tokens": 80,+ "cost": 0.0123,+ }++ usage.capture_provider_usage(agent, LLMResult.from_chat(response="no usage"))+ assert usage.latest_provider_usage(agent) == {}+ assert usage.provider_usage_snapshot(+ {"input_tokens": 100, "cached_tokens": None, "cost": None}+ ) == {"input_tokens": 100}diff --git a/plugins/_context_window/webui/context-window-store.js b/plugins/_context_window/webui/context-window-store.jsindex ad80f6cc0..30349ea02 100644--- a/plugins/_context_window/webui/context-window-store.js+++ b/plugins/_context_window/webui/context-window-store.js@@ -5,13 +5,18 @@ import { store as preferencesStore } from "/components/sidebar/bottom/preferenceconst API_PATH = "/plugins/_context_window/context_window";const ROWS = [- { key: "messages", label: "Messages", opacity: 1 },- { key: "system_tools", label: "System tools", opacity: 0.88 },- { key: "skills", label: "Skills", opacity: 0.76 },- { key: "mcp_tools", label: "MCP tools", opacity: 0.64 },- { key: "system_prompt", label: "System prompt", opacity: 0.52 },- { key: "extras", label: "Extras", opacity: 0.4 },+ { key: "messages", label: "Messages" },+ { key: "system_tools", label: "System tools" },+ { key: "skills", label: "Skills" },+ { key: "mcp_tools", label: "MCP tools" },+ { key: "system_prompt", label: "System prompt" },+ { key: "extras", label: "Extras" },];+const COST_FORMATTER = new Intl.NumberFormat("en-US", {+ style: "currency",+ currency: "USD",+ maximumSignificantDigits: 3,+});preferencesStore.registerUiControlVisibility("contextWindowUsage", {mobile: true,@@ -31,6 +36,44 @@ function formatPercent(value) {return `${Number.isInteger(rounded) ? rounded.toFixed(0) : rounded.toFixed(1)}%`;}+function optionalNumber(value, key) {+ if (!value || !Object.prototype.hasOwnProperty.call(value, key)) return null;+ const number = Number(value[key]);+ return Number.isFinite(number) && number >= 0 ? number : null;+}++function formatCost(value) {+ if (value === 0) return "$0";+ return value < 0.001 ? "<$0.001" : COST_FORMATTER.format(value);+}++function buildProviderUsage(value = {}) {+ const input = optionalNumber(value, "input_tokens");+ const cached = optionalNumber(value, "cached_tokens");+ const output = optionalNumber(value, "output_tokens");+ const cost = optionalNumber(value, "cost");++ const tokenSummary = input === null && output === null+ ? ""+ : `${input === null ? "–" : formatTokens(input)} → ${output === null ? "–" : formatTokens(output)}`;++ const cachePercent = input > 0 && cached !== null+ ? Math.min((cached / input) * 100, 100)+ : null;+ return {+ hasData: cost !== null || cachePercent !== null || Boolean(tokenSummary),+ price: {+ hasData: cost !== null,+ label: cost === null ? "" : formatCost(cost),+ },+ cache: {+ hasData: cachePercent !== null,+ label: cachePercent === null ? "" : `${Math.round(cachePercent)}%`,+ },+ tokens: tokenSummary,+ };+}+function buildUsage(data = {}) {const tokens = Math.max(Number(data.tokens) || 0, 0);const contextWindow = Math.max(Number(data.context_window) || 0, 0);@@ -43,7 +86,6 @@ function buildUsage(data = {}) {...row,tokensLabel: formatTokens(rowTokens),percentLabel: formatPercent(rowPercent),- dotStyle: `opacity:${row.opacity}`,};});const hasBreakdown = rows.some(row => Number(breakdown[row.key]) > 0);@@ -55,7 +97,6 @@ function buildUsage(data = {}) {label: "Free space",tokensLabel: formatTokens(freeTokens),percentLabel: formatPercent(freePercent),- dotStyle: "opacity:0.24",});}const percentLabel = formatPercent(percent);@@ -66,8 +107,10 @@ function buildUsage(data = {}) {ariaLabel: `Context window ${percentLabel} used`,ringLabel: contextWindow ? `${Math.round(percent)}%` : "–",ringDasharray: `${Math.min(percent, 100)} 100`,- summary: `${formatTokens(tokens)}/${contextWindow ? formatTokens(contextWindow) : "–"} (${percentLabel})`,+ summaryTokens: `${formatTokens(tokens)}/${contextWindow ? formatTokens(contextWindow) : "–"} tokens`,+ summaryPercent: `${percentLabel} used`,meterStyle: `width:${Math.min(percent, 100)}%`,+ provider: buildProviderUsage(data.provider_usage),};}diff --git a/tests/test_stream_tool_early_stop.py b/tests/test_stream_tool_early_stop.pyindex 91f61b627..3368bfc8d 100644--- a/tests/test_stream_tool_early_stop.py+++ b/tests/test_stream_tool_early_stop.py@@ -449,7 +449,8 @@ async def test_unified_turn_stops_chat_stream_after_text_tool_request(monkeypatcasync def response_callback(chunk: str, full: str):return full if extract_tools.extract_tool_request(full) else None- result = await wrapper.unified_turn(+ result = await wrapper.unified_turn.__wrapped__(+ wrapper,messages=[],response_callback=response_callback,)