unique_toolkit 2026.32.0.dev5__py3-none-any.whl → 2026.32.0.dev7__py3-none-any.whl

unique_toolkit/agentic/debug_info_manager/debug_info_manager.py

@@ -43,14 +43,18 @@ class AnalyticsToolName(TypedDict):
     display_name: str
 
 
-class AnalyticsTokenUsage(TypedDict):
-    model_name: str
+class AnalyticsConsumption(TypedDict):
     completion_tokens: int | None
     prompt_tokens: int | None
     total_tokens: int | None
     reasoning_tokens: int | None
     cached_tokens: int | None
     cache_write_tokens: int | None
+    cost_usd: float | None
+
+
+class AnalyticsConsumptionByLlm(AnalyticsConsumption):
+    model_name: str
 
 
 class Analytics(TypedDict):
@@ -58,6 +62,10 @@ class Analytics(TypedDict):
     artifacts_created_count: int | None
     artifacts_created_filetype: list[str] | None
     context_memory_updated: bool | None
+    # Sum of all model usages this turn. None when no invocations reported usage.
+    consumption: AnalyticsConsumption | None
+    # Per-model usage totals, sorted by model_name.
+    consumption_by_llm: list[AnalyticsConsumptionByLlm]
     language_model: AnalyticsLanguageModel
     loop_iteration_count: int
     mcp_tool_names_used: list[AnalyticsToolName]
@@ -67,7 +75,6 @@ class Analytics(TypedDict):
     references_count: int
     skills_used: list[AnalyticsSkill]
     subagent_names_used: list[AnalyticsToolName]
-    tokens: list[AnalyticsTokenUsage]
     tool_call_count: int
     tools_used: list[AnalyticsTool]
     total_time_to_answer_ms: int | None
@@ -149,6 +156,7 @@ class DebugInfoManager:
         # DisplayCodeInterpreterFilesPostProcessor) into debug_info["artifacts"].
         # Absent when no Code Interpreter ran this turn → artifact fields stay
         # None, preserving the always-present, null-when-unknown contract.
+        consumption_by_llm = _aggregate_consumption_by_llm(invocations or [])
         analytics = Analytics(
             tools_used=analytics_tools,
             tool_call_count=len(analytics_tools),
@@ -167,7 +175,8 @@ class DebugInfoManager:
             loop_iteration_count=loop_iteration_count,
             subagent_names_used=_unique_tool_names(analytics_tools, "is_sub_agent"),
             mcp_tool_names_used=_unique_tool_names(analytics_tools, "is_mcp"),
-            tokens=_aggregate_tokens_by_model(invocations or []),
+            consumption_by_llm=consumption_by_llm,
+            consumption=_aggregate_consumption(invocations or []),
             total_time_to_answer_ms=total_time_to_answer_ms,
             artifacts_created_count=artifacts["count"] if artifacts else None,
             artifacts_created_filetype=artifacts["filetypes"] if artifacts else None,
@@ -179,27 +188,66 @@ class DebugInfoManager:
         self.add("analytics", analytics)
 
 
-def _aggregate_tokens_by_model(
+def _sum_invocation_costs(
+    invocations: list[LanguageModelInvocationStats],
+) -> float | None:
+    """Sum costs only when every invocation has a known cost; else None."""
+    costs = [invocation.cost_usd for invocation in invocations]
+    if not costs or any(cost is None for cost in costs):
+        return None
+    return sum(cost for cost in costs if cost is not None)
+
+
+def _to_analytics_consumption(
+    usage: LanguageModelTokenUsage,
+    cost_usd: float | None,
+) -> AnalyticsConsumption:
+    return AnalyticsConsumption(
+        completion_tokens=usage.completion_tokens,
+        prompt_tokens=usage.prompt_tokens,
+        total_tokens=usage.total_tokens,
+        reasoning_tokens=usage.reasoning_tokens,
+        cached_tokens=usage.cached_tokens,
+        cache_write_tokens=usage.cache_write_tokens,
+        cost_usd=cost_usd,
+    )
+
+
+def _aggregate_consumption(
     invocations: list[LanguageModelInvocationStats],
-) -> list[AnalyticsTokenUsage]:
-    usages_by_model: dict[str, list[LanguageModelTokenUsage]] = defaultdict(list)
+) -> AnalyticsConsumption | None:
+    """Aggregate token usage and cost across all models for this turn."""
+    usage = LanguageModelTokenUsage.sum_usages(
+        invocation.token_usage for invocation in invocations
+    )
+    if usage is None:
+        return None
+    return _to_analytics_consumption(usage, _sum_invocation_costs(invocations))
+
+
+def _aggregate_consumption_by_llm(
+    invocations: list[LanguageModelInvocationStats],
+) -> list[AnalyticsConsumptionByLlm]:
+    invocations_by_model: dict[str, list[LanguageModelInvocationStats]] = defaultdict(
+        list
+    )
     for invocation in invocations:
-        usages_by_model[str(invocation.model_name)].append(invocation.token_usage)
+        invocations_by_model[str(invocation.model_name)].append(invocation)
 
-    totals: list[AnalyticsTokenUsage] = []
-    for model_name in sorted(usages_by_model):
-        usage = LanguageModelTokenUsage.sum_usages(usages_by_model[model_name])
+    totals: list[AnalyticsConsumptionByLlm] = []
+    for model_name in sorted(invocations_by_model):
+        model_invocations = invocations_by_model[model_name]
+        usage = LanguageModelTokenUsage.sum_usages(
+            invocation.token_usage for invocation in model_invocations
+        )
         if usage is None:
             continue
         totals.append(
-            AnalyticsTokenUsage(
+            AnalyticsConsumptionByLlm(
                 model_name=model_name,
-                completion_tokens=usage.completion_tokens,
-                prompt_tokens=usage.prompt_tokens,
-                total_tokens=usage.total_tokens,
-                reasoning_tokens=usage.reasoning_tokens,
-                cached_tokens=usage.cached_tokens,
-                cache_write_tokens=usage.cache_write_tokens,
+                **_to_analytics_consumption(
+                    usage, _sum_invocation_costs(model_invocations)
+                ),
             )
         )
     return totals

unique_toolkit/language_model/infos.py

@@ -93,6 +93,7 @@ class LanguageModelName(StrEnum):
     ANTHROPIC_CLAUDE_OPUS_4_6 = "litellm:anthropic-claude-opus-4-6"
     ANTHROPIC_CLAUDE_OPUS_4_7 = "litellm:anthropic-claude-opus-4-7"
     ANTHROPIC_CLAUDE_OPUS_4_8 = "litellm:anthropic-claude-opus-4-8"
+    ANTHROPIC_CLAUDE_OPUS_5 = "litellm:anthropic-claude-opus-5"
     ANTHROPIC_CLAUDE_FABLE_5 = "litellm:anthropic-claude-fable-5"
     GEMINI_2_0_FLASH = "litellm:gemini-2-0-flash"
     GEMINI_2_5_FLASH = "litellm:gemini-2-5-flash"
@@ -144,6 +145,7 @@ class LanguageModelName(StrEnum):
     VERTEX_CLAUDE_OPUS_4_6 = "litellm:vertex-claude-opus-4-6"
     VERTEX_CLAUDE_OPUS_4_7 = "litellm:vertex-claude-opus-4-7"
     VERTEX_CLAUDE_OPUS_4_8 = "litellm:vertex-claude-opus-4-8"
+    VERTEX_CLAUDE_OPUS_5 = "litellm:vertex-claude-opus-5"
     VERTEX_CLAUDE_FABLE_5 = "litellm:vertex-claude-fable-5"
 
 
@@ -1812,6 +1814,30 @@ class LanguageModelInfo(BaseModel):
                     published_at=date(2026, 5, 28),
                     supported_reasoning_efforts=[],
                 )
+            case (
+                LanguageModelName.ANTHROPIC_CLAUDE_OPUS_5
+                | LanguageModelName.VERTEX_CLAUDE_OPUS_5
+            ):
+                return cls(
+                    name=model_name,
+                    capabilities=[
+                        ModelCapabilities.FUNCTION_CALLING,
+                        ModelCapabilities.STREAMING,
+                        ModelCapabilities.VISION,
+                        ModelCapabilities.REASONING,
+                    ],
+                    provider=LanguageModelProvider.LITELLM,
+                    family=ModelFamily.ANTHROPIC,
+                    version="claude-opus-5",
+                    encoder_name=EncoderName.O200K_BASE,  # TODO: Update encoder with litellm
+                    token_limits=LanguageModelTokenLimits(
+                        token_limit_input=1_000_000,
+                        token_limit_output=128_000,
+                    ),
+                    info_cutoff_at=date(2026, 5, 1),
+                    published_at=date(2026, 7, 24),
+                    supported_reasoning_efforts=[],
+                )
             case (
                 LanguageModelName.ANTHROPIC_CLAUDE_FABLE_5
                 | LanguageModelName.VERTEX_CLAUDE_FABLE_5

unique_toolkit/language_model/invocation_stats.py

@@ -11,6 +11,7 @@ from humps import camelize
 from pydantic import BaseModel, BeforeValidator, ConfigDict
 
 from unique_toolkit.language_model.infos import LanguageModelName
+from unique_toolkit.language_model.model_costs import calculate_invocation_cost_usd
 from unique_toolkit.language_model.schemas import LanguageModelTokenUsage
 
 # `protected_namespaces=()` allows the `model_name` field name.
@@ -59,6 +60,7 @@ class LanguageModelInvocationStats(BaseModel):
     model_name: ModelName
     token_usage: LanguageModelTokenUsage
     source: Source  # e.g. "main_loop", tool/evaluation/postprocessor name
+    cost_usd: float | None = None
 
     @classmethod
     def from_usage(
@@ -67,4 +69,9 @@ class LanguageModelInvocationStats(BaseModel):
         token_usage: LanguageModelTokenUsage,
         source: str,
     ) -> Self:
-        return cls(model_name=model_name, token_usage=token_usage, source=source)
+        return cls(
+            model_name=model_name,
+            token_usage=token_usage,
+            source=source,
+            cost_usd=calculate_invocation_cost_usd(str(model_name), token_usage),
+        )

unique_toolkit/language_model/model_costs.py

@@ -0,0 +1,118 @@
+"""Load model prices and calculate the USD cost of LLM invocations."""
+
+import logging
+import os
+import time
+from pathlib import Path
+
+import yaml
+from pydantic import BaseModel, ConfigDict, Field
+
+from unique_toolkit.language_model.schemas import LanguageModelTokenUsage
+
+logger = logging.getLogger(f"toolkit.language_model.{__name__}")
+
+MODEL_COSTS_FILE_ENV = "MODEL_COSTS_FILE"
+_LITELLM_PREFIX = "litellm:"
+_TOKENS_PER_MILLION = 1_000_000
+
+# Safety-only bound on how long a process may keep a loaded catalog. Not the
+# primary refresh mechanism — just ensures a long-lived worker eventually
+# re-reads a remounted price sheet.
+_CACHE_MAX_AGE_SECONDS = 5 * 60
+
+
+class ModelCost(BaseModel):
+    """Per-million-token prices for one language model."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    input: float
+    completion: float
+    currency: str = Field(default="USD", min_length=1)
+
+
+class ModelCostCatalog(BaseModel):
+    """Versioned model-price catalog rendered by the platform Helm charts."""
+
+    model_config = ConfigDict(extra="forbid", populate_by_name=True)
+
+    cost_schema_version: int = Field(alias="costSchemaVersion")
+    models: dict[str, ModelCost] = Field(min_length=1)
+
+
+_catalog_cache: dict[Path, tuple[ModelCostCatalog, float]] = {}
+
+
+def _parse_model_cost_catalog(path: Path) -> ModelCostCatalog:
+    payload = yaml.safe_load(path.read_text(encoding="utf-8"))
+    catalog = ModelCostCatalog.model_validate(payload)
+    if catalog.cost_schema_version != 1:
+        raise ValueError(
+            f"Unsupported model cost schema version: {catalog.cost_schema_version}"
+        )
+    return catalog
+
+
+def _load_model_cost_catalog(path: Path) -> ModelCostCatalog:
+    now = time.monotonic()
+    cached = _catalog_cache.get(path)
+    if cached is not None:
+        catalog, loaded_at = cached
+        if now - loaded_at < _CACHE_MAX_AGE_SECONDS:
+            return catalog
+
+    catalog = _parse_model_cost_catalog(path)
+    _catalog_cache[path] = (catalog, now)
+    return catalog
+
+
+def load_model_cost_catalog(path: str | Path | None = None) -> ModelCostCatalog | None:
+    """Load and cache the configured catalog, or return None when unavailable.
+
+    Returns None when the catalog path is unset, or when the file cannot be
+    read/parsed. Cost capture must never abort a successful LLM call over a
+    missing or invalid price sheet.
+    """
+
+    configured_path = path or os.getenv(MODEL_COSTS_FILE_ENV)
+    if configured_path is None or not str(configured_path).strip():
+        return None
+    try:
+        return _load_model_cost_catalog(Path(configured_path))
+    except Exception as exc:
+        logger.warning(
+            "Failed to load model cost catalog from %s: %s",
+            configured_path,
+            exc,
+        )
+        return None
+
+
+def calculate_invocation_cost_usd(
+    model_name: str,
+    token_usage: LanguageModelTokenUsage,
+    catalog: ModelCostCatalog | None = None,
+) -> float | None:
+    """Calculate invocation cost using prompt and completion token totals."""
+
+    if token_usage.prompt_tokens is None or token_usage.completion_tokens is None:
+        return None
+
+    active_catalog = catalog if catalog is not None else load_model_cost_catalog()
+    if active_catalog is None:
+        return None
+
+    normalized_model_name = str(model_name)
+    model_cost = active_catalog.models.get(normalized_model_name)
+    if model_cost is None and normalized_model_name.startswith(_LITELLM_PREFIX):
+        model_cost = active_catalog.models.get(
+            normalized_model_name.removeprefix(_LITELLM_PREFIX)
+        )
+    if model_cost is None:
+        return None
+
+    return (
+        token_usage.prompt_tokens * model_cost.input
+        + token_usage.completion_tokens * model_cost.completion
+    ) / _TOKENS_PER_MILLION

{unique_toolkit-2026.32.0.dev5.dist-info → unique_toolkit-2026.32.0.dev7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.32.0.dev5
+Version: 2026.32.0.dev7
 Summary: 
 Author: Cedric Klinkert, Martin Fadler, Sadique Sheik, Fabian Schläpfer, Pascal Hauri, Ejiroghene Oki-Peter, Rami Azouz, Ricardo Visini, Sanzio Monti
 Author-email: Cedric Klinkert <cedric.klinkert@unique.ch>, Martin Fadler <martin.fadler@unique.ch>, Sadique Sheik <sadique@unique.ch>, Fabian Schläpfer <fabian@unique.ch>, Pascal Hauri <pascal@unique.ch>, Ejiroghene Oki-Peter <oki-peter.ejiroghene@unique.ai>, Rami Azouz <rami.ext@unique.ch>, Ricardo Visini <ricardo.visini@unique.ch>, Sanzio Monti <sanzio.monti@unique.ai>
@@ -33,6 +33,7 @@ Requires-Dist: tenacity>=9.1.4
 Requires-Dist: pydantic-core>=2.41.5
 Requires-Dist: cachetools>=5.0,<6
 Requires-Dist: rapidfuzz>=3.0,<4
+Requires-Dist: pyyaml>=6.0.3
 Requires-Dist: fastapi>=0.133.0,<1.0.0 ; extra == 'fastapi'
 Requires-Dist: starlette>=1.3.1,<2.0.0 ; extra == 'fastapi'
 Requires-Dist: uvicorn>=0.37.0,<0.38 ; extra == 'fastapi'

{unique_toolkit-2026.32.0.dev5.dist-info → unique_toolkit-2026.32.0.dev7.dist-info}/RECORD

@@ -83,7 +83,7 @@ unique_toolkit/_common/utils/write_configuration.py,sha256=fzvr4C-XBL3OSM3Od9Tbq
 unique_toolkit/_common/validate_required_values.py,sha256=cA8GPwkV1EPsG6ffIkmAn3Dx4LRRMC1EAURmT7PyuCg,551
 unique_toolkit/_common/validators.py,sha256=ElnkMsyEY24TfzfTVHvireyT39EnZgW5N40T0P4b6gE,2638
 unique_toolkit/agentic/__init__.py,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
-unique_toolkit/agentic/debug_info_manager/debug_info_manager.py,sha256=VZLDOTAJJ_jVOZ4Zw4C2F-uCv3vsjZTcvbljGBYKh-A,10501
+unique_toolkit/agentic/debug_info_manager/debug_info_manager.py,sha256=fqhbKe0WWB8nZ7CzeEcgQGMHJajIMVOpKkUWOqiDdUE,12149
 unique_toolkit/agentic/debug_info_manager/test/test_debug_info_manager.py,sha256=bDjrHLqfwtd7ewdIdyTFHabG-15t3iXvs34ywoYXcik,11888
 unique_toolkit/agentic/evaluation/config.py,sha256=u5-iuT-4mfA2_9UZjZ-TN3YutVf35aR4EsF9l61Odnk,1696
 unique_toolkit/agentic/evaluation/context_relevancy/prompts/__init__.py,sha256=IKGvHW2viBylTUjEJNKvIXO4YrYwdJXBHSxA9rKtiRI,355
@@ -392,8 +392,9 @@ unique_toolkit/language_model/builder.py,sha256=KwBzsD6-fvQnB4p2I7M-5Tk6j4tmQ973
 unique_toolkit/language_model/constants.py,sha256=B-topqW0r83dkC_25DeQfnPk3n53qzIHUCBS7YJ0-1U,119
 unique_toolkit/language_model/default_language_model.py,sha256=mcyuFs_V6tp52C8xJSHiEHO4kFGj6xed1gvozvCRD7k,1181
 unique_toolkit/language_model/functions.py,sha256=uYIsVQfF4uwW6cftlKhQElMLpXNaVe_57FFobhC7R4A,23503
-unique_toolkit/language_model/infos.py,sha256=y1JLszdCyKBUTCjPobYsB02G2WXZOGoHzQtjNqJ7EOc,140831
-unique_toolkit/language_model/invocation_stats.py,sha256=ApRy7YzlcWd5so3tm8RhkrJn6Ip5_rjZwYcyh92Wc_4,2244
+unique_toolkit/language_model/infos.py,sha256=zElSKg539eX0X-WWCAs4UO_SL1PxJ2epynut0vY4jRA,142052
+unique_toolkit/language_model/invocation_stats.py,sha256=XAM_qckI-BoMBSgDpVcxFD1cFSdi1-GHZToHz_jLynM,2491
+unique_toolkit/language_model/model_costs.py,sha256=WM1u2fWAKpPB_creS-ew8YAzc1V8oiaCbvhpSJaO8B0,3789
 unique_toolkit/language_model/prompt.py,sha256=JSawaLjQg3VR-E2fK8engFyJnNdk21zaO8pPIodzN4Q,3991
 unique_toolkit/language_model/reference.py,sha256=T3wr2jP_uRg5wKVIbKqLbxJlcgxfbZQGq3XdjbQ2kSo,6984
 unique_toolkit/language_model/schemas.py,sha256=IUmp_JMHx0czadJlmsLL9nmc-nCPmVRVLp2T5DSYsqY,35402
@@ -418,6 +419,6 @@ unique_toolkit/short_term_memory/service.py,sha256=2JMBEBn6GoyK-a0xvdcWpPft-3C74
 unique_toolkit/smart_rules/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 unique_toolkit/smart_rules/compile.py,sha256=vRe0aMJWuVIK-VsqomgxQImzke7MDmwgSxxJfYCa7zw,1386
 unique_toolkit/test_utilities/events.py,sha256=FLrhMh0A_mQutAUjoD0FdMcVQ4Gq31r35OVGzRjvLro,6760
-unique_toolkit-2026.32.0.dev5.dist-info/WHEEL,sha256=eh7sammvW2TypMMMGKgsM83HyA_3qQ5Lgg3ynoecH3M,79
-unique_toolkit-2026.32.0.dev5.dist-info/METADATA,sha256=1bEknHHM6C2M0YX9ZPdElqUVEOA1AnI_u-_zmXVKrQM,7465
-unique_toolkit-2026.32.0.dev5.dist-info/RECORD,,
+unique_toolkit-2026.32.0.dev7.dist-info/WHEEL,sha256=eh7sammvW2TypMMMGKgsM83HyA_3qQ5Lgg3ynoecH3M,79
+unique_toolkit-2026.32.0.dev7.dist-info/METADATA,sha256=NPHWeljbN2U6iJuo-_bPj7Fiw7l1ZxiBiCVJTGBnLzs,7494
+unique_toolkit-2026.32.0.dev7.dist-info/RECORD,,