verifiers 0.4.2.dev36__py3-none-any.whl → 0.4.2.dev45__py3-none-any.whl

verifiers/v1/clients/eval.py

@@ -9,6 +9,7 @@ from pydantic_core import from_json, to_json
 
 from verifiers.v1.clients.base import build_async_httpx, join_url
 from verifiers.v1.clients.client import SESSION_ID_HEADER, Client, RelayReply
+from verifiers.v1.clients.prompt_cache import maybe_add_prompt_cache_control
 from verifiers.v1.configs.client import (
     BaseClientConfig,
     resolve_api_key,
@@ -70,6 +71,7 @@ class EvalClient(Client):
     def __init__(self, config: BaseClientConfig) -> None:
         self.base_url = config.base_url
         self.api_key = resolve_api_key(config)
+        self.cache_retention = getattr(config, "cache_retention", None)
         # Keep endpoint headers separate so they can override intercepted request headers before
         # the dialect's provider authentication is applied.
         self.headers = resolve_headers(config)
@@ -84,6 +86,9 @@ class EvalClient(Client):
         turn: PendingTurn | None = None,
         headers: Mapping[str, str] | None = None,
     ) -> Response:
+        body = maybe_add_prompt_cache_control(
+            dialect, body, self.base_url, self.cache_retention
+        )
         resp = await self._request(
             join_url(self.base_url, dialect.upstream_path),
             body,
@@ -177,6 +182,9 @@ class EvalClient(Client):
     ) -> RelayReply:
         # Relay complete SSE events so the interception server can safely insert keepalives
         # between them. Error responses are mapped before any event is handed back.
+        body = maybe_add_prompt_cache_control(
+            dialect, body, self.base_url, self.cache_retention
+        )
         resp = await self._request(
             join_url(self.base_url, dialect.upstream_path),
             body,

verifiers/v1/clients/prompt_cache.py

@@ -0,0 +1,350 @@
+"""Gateway-side prompt-cache markers for Anthropic models.
+
+Anthropic has no implicit prompt cache: a request only caches content when a
+block carries ``cache_control``. Verifiers relays harness-native bodies, and
+many harnesses send no marker, so Claude traffic through Prime Inference or
+OpenRouter pays full input price. When the caller sends no marker, mark the
+static prefix (last custom tool and last system/developer block) that repeats
+across rollouts; multi-turn requests also mark the latest message.
+
+Markers default to the five-minute cache (no ``ttl``). Set
+``cache_retention = "long"`` on the eval client, or ``PI_CACHE_RETENTION=long``,
+to emit ``ttl: "1h"`` instead; one-hour writes cost twice as much. ``"none"``
+disables injection entirely. Any caller marker wins over generation here.
+"""
+
+from __future__ import annotations
+
+import os
+from collections.abc import Mapping
+from typing import Any, Literal
+from urllib.parse import urlsplit
+
+from verifiers.v1.dialects.base import Dialect
+
+CacheRetention = Literal["none", "short", "long"]
+
+_CACHE_MARKERS = frozenset({"cache_control", "prompt_cache_breakpoint"})
+_SCHEMA_KEYS = frozenset({"input_schema", "parameters", "json_schema"})
+_ANTHROPIC_BLOCK_TYPES = frozenset({"text", "image", "tool_result", "document"})
+
+
+def _has_marker(value: Any) -> bool:
+    return isinstance(value, Mapping) and not _CACHE_MARKERS.isdisjoint(value)
+
+
+def _content_has_marker(content: Any) -> bool:
+    if isinstance(content, Mapping):
+        return _has_marker(content)
+    if not isinstance(content, list):
+        return False
+    return any(isinstance(block, Mapping) and _has_marker(block) for block in content)
+
+
+def _has_caller_marker(body: Mapping[str, Any]) -> bool:
+    """Only scan positions where Anthropic honors cache markers.
+
+    Tool-use inputs, tool arguments, metadata, and JSON schemas can contain a
+    ``cache_control`` key without being prompt-cache markers.
+    """
+    if _has_marker(body):
+        return True
+    if _content_has_marker(body.get("system")):
+        return True
+    messages = body.get("messages")
+    if isinstance(messages, list):
+        for message in messages:
+            if not isinstance(message, Mapping):
+                continue
+            if _has_marker(message):
+                return True
+            content = message.get("content")
+            if _content_has_marker(content):
+                return True
+            if isinstance(content, list):
+                for block in content:
+                    if (
+                        isinstance(block, Mapping)
+                        and block.get("type") == "tool_result"
+                        and _content_has_marker(block.get("content"))
+                    ):
+                        return True
+    tools = body.get("tools")
+    if isinstance(tools, list):
+        for tool in tools:
+            if _has_marker(tool):
+                return True
+    return False
+
+
+def resolve_cache_retention(
+    cache_retention: CacheRetention | None,
+) -> CacheRetention:
+    """Explicit setting wins, then ``PI_CACHE_RETENTION=long``, then short."""
+    if cache_retention is not None:
+        return cache_retention
+    env = os.environ.get("PI_CACHE_RETENTION", "").strip().casefold()
+    if env in {"none", "short", "long"}:
+        return env
+    return "short"
+
+
+def _cache_marker(retention: CacheRetention) -> dict[str, str]:
+    marker = {"type": "ephemeral"}
+    if retention == "long":
+        marker["ttl"] = "1h"
+    return marker
+
+
+def _is_anthropic_model(body: Mapping[str, Any]) -> bool:
+    model = body.get("model")
+    return isinstance(model, str) and model.casefold().startswith("anthropic/")
+
+
+def _is_gateway(base_url: str) -> bool:
+    hostname = urlsplit(base_url).hostname
+    if hostname is None:
+        return False
+    host = hostname.casefold()
+    return host in {"openrouter.ai", "pinference.ai"} or host.endswith(
+        (".openrouter.ai", ".pinference.ai")
+    )
+
+
+def _mark_last_tool(
+    body: dict[str, Any], *, anthropic: bool, marker: dict[str, str]
+) -> bool:
+    tools = body.get("tools")
+    if not isinstance(tools, list) or not tools:
+        return False
+    for index in range(len(tools) - 1, -1, -1):
+        tool = tools[index]
+        if not isinstance(tool, dict):
+            continue
+        if anthropic:
+            # Server tools (web search, computer use, ...) do not accept a
+            # cache marker; fall back to the last custom tool.
+            tool_type = tool.get("type")
+            if (tool_type is not None and tool_type != "custom") or "name" not in tool:
+                continue
+        elif tool.get("type") != "function":
+            continue
+        updated = list(tools)
+        updated[index] = {**tool, "cache_control": dict(marker)}
+        body["tools"] = updated
+        return True
+    return False
+
+
+def _mark_last_text_block(
+    message: dict[str, Any], marker: dict[str, str]
+) -> dict | None:
+    content = message.get("content")
+    if isinstance(content, str):
+        if not content:
+            return None
+        return {
+            **message,
+            "content": [
+                {
+                    "type": "text",
+                    "text": content,
+                    "cache_control": dict(marker),
+                }
+            ],
+        }
+    if not isinstance(content, list):
+        return None
+    for index in range(len(content) - 1, -1, -1):
+        block = content[index]
+        if (
+            isinstance(block, dict)
+            and block.get("type") == "text"
+            and isinstance(block.get("text"), str)
+            and block["text"]
+        ):
+            updated = list(content)
+            updated[index] = {**block, "cache_control": dict(marker)}
+            return {**message, "content": updated}
+    return None
+
+
+def _mark_last_anthropic_block(
+    message: dict[str, Any], marker: dict[str, str]
+) -> dict | None:
+    content = message.get("content")
+    if isinstance(content, str):
+        if not content:
+            return None
+        return {
+            **message,
+            "content": [
+                {
+                    "type": "text",
+                    "text": content,
+                    "cache_control": dict(marker),
+                }
+            ],
+        }
+    if not isinstance(content, list):
+        return None
+    for index in range(len(content) - 1, -1, -1):
+        block = content[index]
+        if not isinstance(block, dict):
+            continue
+        block_type = block.get("type")
+        if not isinstance(block_type, str) or block_type not in _ANTHROPIC_BLOCK_TYPES:
+            continue
+        if block_type == "text" and not (
+            isinstance(block.get("text"), str) and block["text"]
+        ):
+            continue
+        updated = list(content)
+        updated[index] = {**block, "cache_control": dict(marker)}
+        return {**message, "content": updated}
+    return None
+
+
+def _leading_instruction_indices(messages: list[Any]) -> list[int]:
+    indices: list[int] = []
+    for index, message in enumerate(messages):
+        if not isinstance(message, dict):
+            break
+        role = message.get("role")
+        if not isinstance(role, str) or role not in {"system", "developer"}:
+            break
+        indices.append(index)
+    return indices
+
+
+def _has_chat_history(messages: list[Any]) -> bool:
+    user_turns = 0
+    for message in messages:
+        if not isinstance(message, dict):
+            continue
+        role = message.get("role")
+        if not isinstance(role, str):
+            continue
+        if role == "user":
+            user_turns += 1
+        elif role in {"assistant", "tool"}:
+            return True
+    return user_turns > 1
+
+
+def _has_messages_history(messages: list[Any]) -> bool:
+    if len([message for message in messages if isinstance(message, dict)]) <= 1:
+        return False
+    return isinstance(messages[-1], dict) and messages[-1].get("role") == "user"
+
+
+def _mark_chat(body: dict[str, Any], marker: dict[str, str]) -> bool:
+    applied = _mark_last_tool(body, anthropic=False, marker=marker)
+    messages = body.get("messages")
+    if not isinstance(messages, list) or not messages:
+        return applied
+
+    updated_messages = list(messages)
+    messages_changed = False
+    for index in reversed(_leading_instruction_indices(messages)):
+        marked = _mark_last_text_block(messages[index], marker)
+        if marked is not None:
+            updated_messages[index] = marked
+            messages_changed = True
+            applied = True
+            break
+
+    last_message = messages[-1] if messages else None
+    last_role = last_message.get("role") if isinstance(last_message, dict) else None
+    if (
+        _has_chat_history(messages)
+        and isinstance(last_message, dict)
+        and isinstance(last_role, str)
+        and last_role in {"user", "tool"}
+    ):
+        marked = _mark_last_text_block(last_message, marker)
+        if marked is not None:
+            updated_messages[-1] = marked
+            messages_changed = True
+            applied = True
+
+    if messages_changed:
+        body["messages"] = updated_messages
+    return applied
+
+
+def _mark_anthropic_messages(body: dict[str, Any], marker: dict[str, str]) -> bool:
+    applied = _mark_last_tool(body, anthropic=True, marker=marker)
+
+    system = body.get("system")
+    if isinstance(system, str):
+        if system:
+            body["system"] = [
+                {
+                    "type": "text",
+                    "text": system,
+                    "cache_control": dict(marker),
+                }
+            ]
+            applied = True
+    elif isinstance(system, list):
+        for index in range(len(system) - 1, -1, -1):
+            block = system[index]
+            if (
+                isinstance(block, dict)
+                and block.get("type") == "text"
+                and isinstance(block.get("text"), str)
+                and block["text"]
+            ):
+                updated = list(system)
+                updated[index] = {**block, "cache_control": dict(marker)}
+                body["system"] = updated
+                applied = True
+                break
+
+    messages = body.get("messages")
+    if isinstance(messages, list) and messages and _has_messages_history(messages):
+        marked = _mark_last_anthropic_block(messages[-1], marker)
+        if marked is not None:
+            body["messages"] = [*messages[:-1], marked]
+            applied = True
+    return applied
+
+
+def maybe_add_prompt_cache_control(
+    dialect: Dialect,
+    body: dict[str, Any],
+    base_url: str,
+    cache_retention: CacheRetention | None = None,
+) -> dict[str, Any]:
+    """Return a request body with prompt-cache markers when eligible.
+
+    Only Anthropic models on Prime Inference/OpenRouter are touched, and only
+    when the caller has not already set any cache marker. The original body is
+    left untouched; edited lists and blocks are copied.
+    """
+    if dialect.name not in {"anthropic_messages", "chat_completions"}:
+        return body
+    if not _is_gateway(base_url) or not _is_anthropic_model(body):
+        return body
+    retention = resolve_cache_retention(cache_retention)
+    if retention == "none":
+        return body
+    if _has_caller_marker(body):
+        return body
+
+    marker = _cache_marker(retention)
+    updated = dict(body)
+    if isinstance(body.get("tools"), list):
+        updated["tools"] = list(body["tools"])
+    if isinstance(body.get("messages"), list):
+        updated["messages"] = list(body["messages"])
+    if isinstance(body.get("system"), list):
+        updated["system"] = list(body["system"])
+
+    applied = (
+        _mark_anthropic_messages(updated, marker)
+        if dialect.name == "anthropic_messages"
+        else _mark_chat(updated, marker)
+    )
+    return updated if applied else body

verifiers/v1/configs/client.py

@@ -75,6 +75,9 @@ class EvalClientConfig(BaseClientConfig):
     """The default (eval): forward each request to a matching endpoint via `EvalClient`."""
 
     type: Literal["eval"] = "eval"
+    cache_retention: Literal["none", "short", "long"] | None = None
+    """Anthropic prompt-cache lifetime. ``None`` follows ``PI_CACHE_RETENTION``
+    (short by default); ``long`` emits ``ttl: "1h"``; ``none`` disables markers."""
 
 
 class TrainClientConfig(BaseClientConfig):

verifiers/v1/harnesses/rlm/harness.py

@@ -101,6 +101,9 @@ class RLMHarnessConfig(HarnessConfig):
     exec_timeout: PositiveInt | None = None
     """IPython/native tool execution timeout in seconds; `None` uses nano-rlm's
     default (300). Separate from `tool_timeout`, which controls MCP calls."""
+    provider_max_retries: NonNegativeInt | None = None
+    """SDK retries per model call (429, 5xx, connection errors); `None` uses nano-rlm's
+    default (5). Raise it so long rollouts ride out extended rate-limit bursts."""
     allow_git: bool | None = None
     """Allow unrestricted Git history queries (e.g. `git log --all`); `None` uses
     nano-rlm's default (false). Ordinary Git commands remain allowed either way;
@@ -225,6 +228,11 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "provider": {
                 "base_url": endpoint,
                 "api_key": secret,
+                **(
+                    {"max_retries": self.config.provider_max_retries}
+                    if self.config.provider_max_retries is not None
+                    else {}
+                ),
             },
             # None = passthrough: the key stays off the wire and nano-rlm's own
             # default applies.

{verifiers-0.4.2.dev36.dist-info → verifiers-0.4.2.dev45.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.2.dev36
+Version: 0.4.2.dev45
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.4.2.dev36.dist-info → verifiers-0.4.2.dev45.dist-info}/RECORD

@@ -29,11 +29,12 @@ verifiers/v1/cli/validate.py,sha256=IXvY9Ey7snPc6pJCX9QlGO_t9KFYRwcIVj0G5TFDanU,
 verifiers/v1/clients/__init__.py,sha256=Ysig0tE_0E4Jsfgfes1XHN-fK1s_RCXdqZD6E7lK4PU,507
 verifiers/v1/clients/base.py,sha256=C0G5lRm3nxt1xmPoA2cweBgCnN2KcLepmCYq1vdDrnU,1691
 verifiers/v1/clients/client.py,sha256=HT3aQiOyYYjvz_cKEXchtUSD07-4VehebWUJL9AyON8,3166
-verifiers/v1/clients/eval.py,sha256=pWHgnRU8QQE6jnjJdwJNdvR8BeJ-z_OfjIB1tcwLpkc,8340
+verifiers/v1/clients/eval.py,sha256=GgpHZGh4IW3gC-tOZiFPp62MCL4aHik0b4afz2g3-Qo,8729
+verifiers/v1/clients/prompt_cache.py,sha256=shJhbTUwKwao6m3DsYDyHPWmhXSJJnXk3wYQhzXFsnY,11849
 verifiers/v1/clients/train.py,sha256=v1YCVNCEXw2FNQw4lpbLoq18u8TQeo9TnsYOoMWr4zE,20277
 verifiers/v1/configs/__init__.py,sha256=X7u6X7B3ieD1HZl79Tv1kg-yPHbT0QLvAebaGqxVtvA,317
 verifiers/v1/configs/agent.py,sha256=xn9j1WcBIGz7LjgW8WOWDvSvk7vPLztisSLVcT2foxk,7530
-verifiers/v1/configs/client.py,sha256=IvYQuPLVWr8qZNdMbQ9lRXK0wZ4vOjK67lO1lgD2fmM,6149
+verifiers/v1/configs/client.py,sha256=nFUI5oCih6L7by6M2l3wevghzpct8E_grmabT8lYWrA,6381
 verifiers/v1/configs/deployment.py,sha256=Eggd8phUNVcq_f56Haw-OT5U6fTBLDhBnXFJ5AEb7MI,389
 verifiers/v1/configs/env.py,sha256=DgjIbU5NxLGV7h4v-2lcDJF-y4S_VqnEqNNrNt7Ztrk,7532
 verifiers/v1/configs/harness.py,sha256=I9DcHzml17T7SYHobP7E6jHwNSL2eSqEtckioQ2pGNQ,3356
@@ -104,7 +105,7 @@ verifiers/v1/harnesses/pi/harness.py,sha256=K2FtvDhoAUU_CdKVqE_NmDwsw3vqlSaZuLJW
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
 verifiers/v1/harnesses/prime_agent/harness.py,sha256=sV82MCi7uZTKlOH8Vg9_WSytgZNSbknShOhmYcdPPaE,12354
 verifiers/v1/harnesses/rlm/__init__.py,sha256=hwx51xTdEWzzYgMj-p4ETNcptoub3IPMDbEdJ4uB8Ns,122
-verifiers/v1/harnesses/rlm/harness.py,sha256=4Q6g93YJp9JQQWoCaNRYX6ojaMzoBGEuSHbChnJYprg,14873
+verifiers/v1/harnesses/rlm/harness.py,sha256=ehWNA6MgSlKJ8Ee9pC_7WAxIVKUzAImAJn1SJYj5K7M,15304
 verifiers/v1/harnesses/stirrup/__init__.py,sha256=a-NQxX1bOwaGx8QHj2jWcRJEGiLIxdY6XUMbojr9dOk,142
 verifiers/v1/harnesses/stirrup/harness.py,sha256=ypWsg89r4CTeMy5hQVS8z8CUt9AUWgSVDidKjxZDqag,3139
 verifiers/v1/harnesses/stirrup/program.py,sha256=ks7hxFVc1Y_wQGhlUdFI62LeJBrAYgczZ0aioXSynm8,14939
@@ -186,8 +187,8 @@ verifiers/v1/utils/scope.py,sha256=zbs4pj6AVk4jywlAJ3AX0Q0_kzB9dBB02nY-9H4ELdQ,8
 verifiers/v1/utils/score.py,sha256=493yJVMw8teCu9JxapxMFPFzI0hNUqdo0Y2nGW4kckk,6200
 verifiers/v1/utils/trace_store.py,sha256=Qo9qvsJAl8PoRUavjQAfUB2VNxmUtYQtU2ejLKPzY9c,3382
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.4.2.dev36.dist-info/METADATA,sha256=rDjstLcPqgBJzJiro5gbMxVFQpKQKkRDdXSSdkGgqlM,5115
-verifiers-0.4.2.dev36.dist-info/WHEEL,sha256=2iZrJkDIkeDh0k9SikKMqFuRP-ua6mb4Ir1AsER4BEw,87
-verifiers-0.4.2.dev36.dist-info/entry_points.txt,sha256=qdMg7ajiFpDsHf7tuPMRclJXizu1J6LTWawS5YS0B0I,137
-verifiers-0.4.2.dev36.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.4.2.dev36.dist-info/RECORD,,
+verifiers-0.4.2.dev45.dist-info/METADATA,sha256=wZz6cLSWz0rOYnRlEYL5sc7bBKaNpS6VzjnSaS6RXTk,5115
+verifiers-0.4.2.dev45.dist-info/WHEEL,sha256=2iZrJkDIkeDh0k9SikKMqFuRP-ua6mb4Ir1AsER4BEw,87
+verifiers-0.4.2.dev45.dist-info/entry_points.txt,sha256=qdMg7ajiFpDsHf7tuPMRclJXizu1J6LTWawS5YS0B0I,137
+verifiers-0.4.2.dev45.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.4.2.dev45.dist-info/RECORD,,