v0.3.0 — Generalized integrations and verified IQ3_S agent task

Adaptive Inference 0.3.0 makes the shared integration work reproducible from this repository and adds a verified real IQ3_S agent-task result.

The release includes configurable Python client routing, checks that prevent generation options from replacing the selected model or gateway credential, an optional Hermes plugin using public hooks, an OpenClaw configuration composer, and executable client/gateway examples. The shared implementation and contracts remain here; upstream PRs deliver focused configuration guides, catalog entries or reproduced fixes.

The source archive includes local native baselines, licensed reference inputs, authored proposal patches, dependency locks, reproduction scripts and evidence. A new source delta preserves the exact Hermes revision used by the plugin and real task: all 20,897 tracked files restore without fetching upstream. The earlier 25,177-file Strata/native-dependency/Hermes reconstruction remains separately preserved. Model weights, compilers and dependency package binaries are not included.

Validation for this release:

• 110 core tests, 17 source-preservation checks and 12 task-boundary checks across Windows, Linux and macOS with Python 3.11 and 3.12.
• A separate offline-source CI job restores and verifies both the original integration sources and the current Hermes snapshot, with network-denial positive controls.
• Three installed Python SDKs pass 15 positive HTTP fixture cases and six nested-override refusals. Additional pinned host/client fixtures have their own scoped receipts.
• Complete source/runtime artifacts are checked against the tagged source, and the exact wheel receives a fresh non-editable installation outside the checkout. Attached receipts identify versions, source trees, artifact hashes and validation limits.

The final seven-job CI run passed at source commit caab01454f5905f40ccf559eee05c5532b34dff7 (tree ad4f7310b75c591e0c5aa3400d6bcfa96e219210). The source archive is 198,600,972 bytes and the wheel is 54,630 bytes; attached manifests verify all 340 source files, 15 installed runtime modules and five required license/notice files.

The included real-model evidence used the released v0.2.0 gateway with the newer optional plugin, exact current Hermes source, and the existing IQ3_S Strata backend. A parent read a synthetic orders file and wrote its totals; a host-created production child independently verified the report. Six real requests and five native file operations produced the correct 17 units and 4,350 cents. All requests retained context 65,536, output allowance 10,923, high reasoning with 8,192 reasoning tokens and disabled speed projection. The owned gateway and backend shut down cleanly. The later parameterized runner has separate fixture-only qualification; it is not claimed as a second model run or a model test of a different release composition.

The delivery ledger records eleven newly opened upstream PRs plus the updated Strata draft across eleven of the eighteen matrix targets. Archived/unmaintained projects and remaining human-author, contributor-agreement, maintainer-approval and hardware requirements are explicit. Publication is not upstream acceptance.

This remains a foundation for an always-on agent. The serial accepted task does not establish foreground gaming/video responsiveness, native resource handoff, idle unload/reload recovery, OS reboot recovery, comparative speed or native qualification of all engines.

Use the complete source archive for a self-contained copy of code, integration inputs and evidence. The much smaller wheel installs the portable runtime. Verify downloaded assets with SHA256SUMS. Python and separately acquired dependency packages are required; actual inference also requires an existing compatible model server and the separately acquired IQ3_S assets used for this qualification.

See the integration contract, delivery ledger, evidence and limits, and real IQ3_S receipt.

v0.2.0: standalone source and durable task continuation

Adaptive Inference now preserves its implementation and the source needed to reproduce the shipped Strata/Hermes integrations in this repository and in the complete source download. Old upstream PRs remain useful history, but are no longer required to reconstruct that source.

This release adds durable task checkpoints and a resumable feedback runner, optional sustained-memory-headroom admission, explicit context planning, and an offline preservation workflow. The long-term goal is an always-on agent that yields resources to the user and its tools and resumes its original task. Full machine-wide coordination and installed-agent restart integration remain in development.

Downloads

• adaptive_inference-0.2.0.tar.gz is the complete source archive: portable code, tests, documentation, CI, complete pinned integration baselines, overlays, licenses and the original reference files actually used. It is approximately 148 MB.
• adaptive_inference-0.2.0-py3-none-any.whl is the small installable runtime. It does not contain the integration source archive.
• SHA256SUMS identifies the exact downloads. release-validation.json, source-artifact-validation.json and wheel-validation.json record their verification.

The complete source restores 25,177 resulting integration files without fetching upstream repositories. Python network access and Git transports were denied during the check, including a positive control that verifies HTTPS was refused. This is a controlled reproduction check, not an operating-system network sandbox. A fresh unpacked download also passed restoration beneath a deep Windows path. The preserved references include their original licenses and source revisions; attribution remains beside actual reuse.

Validation

• 95 portable software tests and 16 preservation checks pass on Windows, Linux and macOS with Python 3.11 and 3.12.
• An additional Linux CI job restores and independently verifies the complete preserved source.
• The exact source download passes the archive-owned offline restoration and verifies all 13 archived reference files.
• The exact wheel installs independently of the checkout, preserves all 14 runtime modules and five license/notice files, and passes dependency, HTTP demo and separate-process task-resume checks.
• Process-restart tests retain the original objective and corrective feedback, reject a competing owner, and refuse to repeat an interrupted ambiguous effect automatically.

These checks use software fixtures and deterministic callbacks. Earlier real-model evidence remains explicitly scoped to the existing IQ3_S model on the recorded Strata composition; this release does not claim a new model benchmark, automatic recovery after physical power loss, native qualification of all eight engines, or installed qualification of all ten agent recipes. Full inference model weights and machine-specific native executables are not bundled. Exact upstream source snapshots retain their original small support/test assets.

Source: 45fe39836193b1fb3dc4fa375d6662410a70ad7b. Artifacts were built from canonical Git bytes at 489270dc1e319b32e7450473f9a9845c30cc076f, whose source tree is identical to the release merge commit (dba03a66873813d1067294000fffbd6e3c70bbbd).

CI for the release commit · Preservation contract · Reproduction instructions · Evidence and limits · Always-on goal

This is an alpha prerelease. Engine-specific and agent-specific contributions will consume this canonical project through their supported extension paths.

v0.1.0 — Portable adaptive gateway

Adaptive Inference is now the canonical home for the shared adaptive engine work. This initial release provides an installable gateway in front of existing model servers, while retaining earlier Strata and Hermes integrations as pinned, reconstructable source overlays.

Included

• A bounded shared queue for user and background agent requests, learned execution-time feedback, and optional observed RAM/Windows-commit admission.
• Transparent Chat Completions, supported text/embedding routes, model listing and SSE; separate gateway/backend authentication, bounded cancellation and uncertain-completion handling.
• A caller-validated correction helper that retains the original task and bounds attempts.
• Eight engine profiles: Strata, llama.cpp, vLLM, SGLang, Ollama, LM Studio, MLX-LM and TGI.
• Ten agent/client recipes: Hermes, OpenClaw, Aider, OpenHands, Continue, Cline, Roo Code, LangGraph, AutoGen and CrewAI.
• Pinned dependencies, a no-model HTTP demo, wheel/source packages, cross-platform CI, source provenance, licenses, 33 primary-source bibliography entries and a staged development/upstream plan.

Verified

All 45 gateway/profile/CLI tests and four reconstruction-admission tests passed on Windows, Linux and macOS with Python 3.11 and 3.12. Each CI job also ran the demo and built both distributions. A separate clean non-editable wheel installation passed its console commands.

The portable gateway completed eight actual Strata model requests. All eight returned the exact expected integer, both request classes used learned timing on their fourth request, and the gateway drained. The separately operated backend shut down with all six retained process identities confirmed exited. This is a serial functional check, not a speed benchmark or qualification of every application.

Fresh public source reconstructions matched 88 Strata and seven Hermes changed files. The reconstructed code passed 54 Strata CPU tests and two Hermes intent tests.

Reproduce

Clone the tag, install requirements.lock and the package, then run adaptive-inference demo. See the quick start, integration guide, and evidence ledger. The wheel installs the gateway; clone or extract the source distribution for examples, citations and optional reconstruction scripts. SHA256SUMS.txt covers the attached wheel and source archive.

Known scope and next work

The generic gateway does not preempt GPU kernels, resize arbitrary native caches, own backend processes or grant tool authority. The other seven engine profiles and ten client recipes have contract/configuration evidence, not universal installed-application qualification.

The first real-backend attempt timed out after Strata idle-unloaded and its recorded reload footprint exceeded available RAM. That failure is retained in the ledger. The accepted run exercised a freshly ready backend; idle reload coordination remains future native-adapter work. Historical Windows owner software is not bundled in the source overlays.

Continue shared development here, then port focused capability changes upstream. The plan, 43-PR registry, and research attribution preserve that relationship.