Caveman Python middleware

Native framework adapters for the Caveman compression runtime. Your framework keeps its inference client, tools, retries, streams, and original conversation. Caveman projects eligible tool-result text into a copied outbound request. Inference stays with your provider.

This quickstart targets middleware 1.0.0 and SDK 1.2.0. Read the release notes and limitations before upgrading.

Run a complete example

Follow the LangChain quickstart for a fresh environment and runtime installation. Start the local runtime separately; the client package does not include it.

python -m pip install 'caveman-sdk==1.2.0' 'caveman-middleware[langchain]==1.0.1' 'langchain==1.4.1' 'langchain-core==1.6.3' 'langgraph==1.2.11'

The framework guide above is the runnable walkthrough. A separate downloadable quickstart script is not published on the docs site.

Choose an integration

• Framework guide: public entrypoints, native APIs, recovery ownership, transports, and limitations.
• Compatibility matrix: resolver ranges versus accepted ranges versus exact validation evidence. A range is not an exhaustive test result.
• Deployment: process/container lifecycle, remote TLS/authentication, persistence, session affinity, deadlines, and rollback.
• Recovery and scope: namespace/session/branch/cache epoch, exact originals, excerpts, and expiry.
• Troubleshooting: final reason codes and strict readiness versus normal inference fallback.
• Measurement: quality, latency, retries, recovery calls, cache effects, and provider usage.

Adapters, tiers, and what compresses

Certified adapters (langchain, openai, anthropic, litellm) gate every release. Experimental adapters get the same fail-open guard, logging, and behavioral tests, but their APIs may change in a minor release. caveman_middleware.COMPATIBILITY exposes each family's tier and tested version range.

Compression needs a recovery tool the model can call, so only some entry points compress. The record-only ones never change input: in compress mode each call reports recovery_unbound, logged once.

If your app already has a tool named caveman_retrieve, that tool keeps its name, recovery stays off for that registration, and the adapter logs recovery_name_conflict once.

Providers

• OpenAI: openai 2.x (on httpx) and 3.x (on httpx2). The adapter follows whichever HTTP library the installed SDK uses. A Responses API call that the provider would store (store defaults to true) is sent unchanged with reason provider_state_retained. Pass store=False, or allow_stored_responses=True to opt in.
• Anthropic: Anthropic, AnthropicBedrock, and AnthropicVertex, sync and async.
• LiteLLM: the async and proxy paths project any provider, because they edit LiteLLM's OpenAI-format input. The sync Router path edits the translated request and supports only the openai and anthropic providers; other providers pass through with reason unsupported_provider.
• LlamaIndex: the OpenAI and Anthropic LLMs, including subclasses such as AzureOpenAI. Bedrock Converse, Vertex, and other LLMs pass through with reason unsupported_provider.
• Pydantic AI: OpenAIChatModel (including Azure and OpenAI-compatible providers) and AnthropicModel (including Bedrock and Vertex clients). FallbackModel, Google/Gemini, Bedrock Converse, and other models pass through with reason unsupported_provider.
• Google: wrapping returns a clone, so your Client or Chat is never modified. Wrapping a wrapped object does not stack, and unwrap_google() returns the original. One Caveman transport serves every clone: each call carries the scope of the clone that made it.

Framework versions

Each adapter checks the installed framework version against its tested range:

• Outside the range: the adapter skips (original input, reason unsupported_version) and logs a warning once. Pass accept_framework_version=True after testing a newer release yourself.
• Unreadable version (vendored or bundled builds): the adapter runs and logs version_unverified once.
• Too old or too new to import: importing the adapter raises an ImportError that names the installed version and the tested range, with code == "unsupported_version".

Wrapping never raises for a version problem, even in strict mode. To fail fast at startup, call caveman_middleware.preflight(runtime, "langchain"), which reports unavailable for an untested framework. caveman_middleware.ready(...) raises instead.

CI tests the oldest versions on Python 3.11 and the newest on Python 3.13, using constraints/floor.txt and constraints/latest.txt. It also runs the certified families on 3.12 and 3.14. Every lane tests the built wheel, installed. A nightly run installs each extra's newest releases with no constraints and opens an issue when one breaks an adapter. The extras stop at the next major version, so a new major is only tested once its range is widened.

[litellm] excludes 1.106.0, 1.106.1, and 1.106.2: ordinary tool calls fail without the optional mcp package (upstream issue). It also excludes 1.107.0, the compromised release named in LiteLLM's security notice. These are installation exclusions, not a narrower runtime gate; the gate checks adapter compatibility, not package security.

[openai] accepts openai 2.20 through 3.x. LiteLLM and CrewAI require openai<3, so [openai,litellm] or [openai,crewai] installs openai 2.x. The adapter supports 2.x, but you can't get openai 3.x in the same environment as those two frameworks. [asgi] has no dependencies; it works with any ASGI 3 server or framework.

The [litellm] extra requires filelock>=4.2.1 to exclude vulnerable lock-file and timeout implementations. The current latest test lane uses LiteLLM 1.103.10 with filelock 4.2.4. LiteLLM releases that require filelock<4 cannot satisfy this extra; pip selects an earlier compatible LiteLLM release until upstream permits patched filelock versions. This floor is published in the wheel metadata, not only applied to CI.

Failures, scopes, and logging

• Fail-open: outside strict mode, any adapter or runtime failure sends your original request and reports a reason (adapter_error for a bug in adapter code). In strict mode those reasons raise MiddlewareError, except version problems (see above).
• One warning per problem: every adapter logs one WARNING per adapter and reason on the caveman.middleware logger. The line never contains content, scope values, or credentials. Calls that are never LLM calls (embeddings, token counting, other routes) pass through with no report and no warning.
• Recovery errors: when the model calls caveman_retrieve with an unknown or expired handle, or the runtime is unavailable, the tool answers {"error": "<code>"} through the framework's own tool-error result (an error ToolMessage in LangChain, ToolFailed in Pydantic AI, is_error in Anthropic, MCP and AutoGen, an error ToolResult in Strands) and the run continues, even in strict mode. Cancellation still propagates.
• Scopes: thread and session IDs may be free-form. An email or "user 42 / chat #7" is hashed into a valid scope token. A missing scope, such as scope_from_config without a thread_id, passes through with reason invalid_scope; strict mode raises.
• Runtimes: every adapter accepts either MiddlewareRuntime or AsyncMiddlewareRuntime and uses the matching view on sync and async paths.
• Long histories: history hashing stops at 2 MiB (manifest_bytes) and images or bytes are hashed, so a long or multimodal history never skips the whole call. The ASGI adapter's body limit is max_body_bytes (default 2 MiB); a larger body passes through with reason payload_budget.

Contracts to keep

Keep original stored history. Register the actual recovery executor through the native helper; a tool schema alone does not attest recovery. Handles are scope-bound and expire according to runtime retention. Recoverability does not guarantee model quality.

Client modes are off, record, and compress. The client defaults to compression; the standalone runtime defaults to recording. Set both deliberately. Runtime unavailability normally retains original inference input. Strict mode, startup ready(), cancellation, and requested recovery failures have different error contracts.

Final decision reports contain status, reason, transform IDs, replacement/reuse counts, and call IDs. They contain no token counters. Local segment estimates are inferred; provider usage and billed savings are separate evidence. Nothing in this local example verifies billing savings.

Close the runtime client and native framework/provider resources at shutdown. Closing the client does not stop the runtime process. See the deployment guide before sharing a runtime across workers or tenants.

Licence and support

The client, adapters, and Engine runtime are all Apache-2.0. Read LICENSING.md. This package is separate from Caveman Agent SDK. File sanitized reproducible issues in Caveman.