Native Preservation Audit

This audit records how the old native LLLM runtime was restored inside the new protocol-first LLLM design.

The rule for this port is simple: preserve native behavior inside lllm.runtimes.native when practical, but do not let native concepts reshape the public Tactic protocol, FastAPI service contract, Pydantic AI adapter, or PsiHub package metadata.

Summary

Area	Status	Public Boundary
Prompt/dialog/message model	Ported	Native-scoped under lllm.runtimes.native.core.
Native Agent loop	Ported	Usable for native workflows; not part of protocol Tactic.
Native Runtime registry	Ported	Native resource registry only; PsiHub owns package manifests.
Native Tactic	Ported and adapted	Kept as native runtime object; bridged through NativeTacticAdapter.
Invokers	Ported	Optional dependencies; no eager provider import from lllm.
Proxies and builtin proxy assets	Ported	Optional/live API behavior stays out of normal tests.
Jupyter sandbox	Ported and guarded	Optional sandbox extra; lazy import with clear install hint.
Computer-use tool helper	Ported	Optional tools extra.
Shared parser utilities	Already shared	Runtime-agnostic parser module remains outside native.
Tactic-as-tool bridge	Adapted	Native can consume protocol tactics as Function objects.

Ported

The following v1 pieces are restored substantially as native runtime code:

v1 Module Area	Restored Location	Notes
Agent/dialog/prompt loop	lllm.runtimes.native.core.agent, dialog, prompt	Preserves named dialogs, fork lineage, parser retries, tool interrupts, and AgentCallSession.
Runtime/resource registry	lllm.runtimes.native.core.runtime, resource, config	Keeps native prompt/tool/proxy/tactic resource loading separate from PsiHub package ownership.
Native tactic machinery	lllm.runtimes.native.core.tactic, tactic_registry, tactic_tool, native.tactic	Keeps native call sessions, sub-tactic records, native agent groups, and @tactictool.
Native invokers	lllm.runtimes.native.invokers	LiteLLM remains optional through the native extra.
Proxy system	lllm.runtimes.native.proxies	Proxy managers, prompt templates, tactic endpoints, interpreter tools, and builtin proxy modules are restored.
Proxy assets	lllm.runtimes.native.proxies.builtin	Google Trends JSON files and Wolfram assumptions Markdown are included as package data.
Jupyter sandbox	lllm.runtimes.native.sandbox	Restored behind lazy optional imports.
Computer-use helper	lllm.runtimes.native.tools.cua	Restored as optional native tooling.
Utility helpers	lllm.runtimes.native.utils	Restored for native parser/proxy/runtime support.

Adapted

Some restored pieces were changed so they fit the protocol-first design:

Behavior	Adaptation	Reason
core.py collapsed primitive file	Replaced by restored core/ package.	Python cannot have both core.py and core/; v1 structure carries the real native runtime.
Native imports	Top-level lllm.runtimes.native and core use lazy exports.	Importing lllm should not import LiteLLM, Jupyter, Exa, OpenAI, or other optional integrations.
Native-to-protocol boundary	NativeTacticAdapter lives in lllm.runtimes.native.adapter.	Keeps protocol Tactic separate from native Tactic.
Runtime-owned kwargs and metadata	NativeTacticAdapter accepts run_kwargs and optional context-metadata forwarding.	Mirrors the useful Pydantic AI surrounding-features bridge without importing Pydantic AI.
Protocol tactic as native tool	tactic_as_function() wraps a protocol tactic as native Function.	Reuses the callable-tool idea without making native depend on Pydantic AI internals.
Native model validation	Tool names, public maps, metadata, usage, and mutable inputs use protocol boundary hardening.	Preserves native architecture while keeping public model behavior safe and copy-stable.
Optional extras	Restored native, sandbox, and tools extras.	Keeps native features available without making the public protocol heavy.
Builtin proxy module paths	Updated builtin proxy loader to lllm.runtimes.native.proxies.builtin.*.	Old top-level proxy paths no longer match the package layout.
Native Tactic bridge helpers	Added run, arun, tactic_name, and info() compatibility aliases.	Allows callable-tool and metadata helpers to wrap native tactics cleanly.

Cut

Nothing major was intentionally deleted from the v1 native runtime in this checkpoint. The port favors maximal preservation under the native boundary.

The important cut is architectural rather than file-level:

Old Assumption	Kept Out Of The Public Protocol	Reason
Native runtime as the center of LLLM	Public protocol layer	LLLM now centers on runtime-agnostic protocol Tactic.
Native resource/package registry owning reusable package metadata	PsiHub package protocol	PsiHub owns psi.toml, validation, cards, local hub lifecycle, and ref/config metadata.
Provider/tool/sandbox dependencies as core imports	Core lllm import path	Optional native integrations should not burden Pydantic AI, plain Python, or service-only users.

Deferred

The following areas are restored enough to import and test offline, but need more focused verification before they can be considered complete:

Area	Deferred Work
Live native invoker behavior	Add opt-in LiteLLM/live-provider tests that do not print or store secrets.
Builtin proxy integrations	Add pseudo-data tests for request/response/error mapping and opt-in live API examples where credentials exist.
Jupyter sandbox behavior	Add timeout/error/artifact/output-capture tests behind sandbox dependencies.
Live native-agent service examples	Add an opt-in live provider service example once live invoker behavior is tested.
Shared utility extraction	Move parser/proxy/sandbox helpers out of native only when they are clearly useful to Pydantic AI and plain Python tactics without adding complexity.
Native streaming/events	Expose only when a clean native implementation exists; otherwise report unsupported capability.

Verification

Current checkpoint coverage includes:

• native prompt rendering, extension, metadata isolation, and parser behavior;
• native function schema creation, execution, invalid-name rejection, and schema copy isolation;
• native message, function-call, logprob, usage, and dialog boundary validation;
• dialog fork lineage and serialization roundtrip;
• native runtime prompt registry;
• native agent named-dialog management;
• native Tactic.as_tool() compatibility with callable-tool wrapping;
• NativeTacticAdapter protocol boundary behavior;
• NativeTacticAdapter runtime kwargs, context forwarding, metadata injection, and native stream bridging;
• offline native service example through FastAPI /run;
• protocol tactic to native Function bridge.

The checkpoint also passed full LLLM tests, strict docs build, package build, twine check, and installed-wheel native smoke.