Cognia 7.2.0 — Agent Design Kit and machine-readable CLI contracts

Applies to: all

Prepared October 16, 2026. This release is for AI agents (and people) who design applications on top of a packaged Sensory without access to the source code. Every package now has one entry point for agents, a card for each engine, a capability matrix, recipes and a complete reference of the public CLIs, generated from the same descriptors that parse the command line. The engines and their numerical behaviour are unchanged from 7.1.0: saved networks, checkpoints, snapshots and DPSH event logs stay compatible.

Added

  • Agent Design Kit in the package. AGENT-GUIDE.md and llms.txt in the package root point to Docs/agents/: a component inventory (also as JSON), engine cards for all runtimes and operational layers (English and Czech), a guide to choosing an engine, an evidence-backed capability matrix, a document applicability table that says which documents apply to which engine, application design and training guides, end-to-end recipes (image and short-text classification, sequence and language, regime change, causal verification, closed-loop robot, three-engine composition), a generated artifact format reference, pitfalls and a glossary. The planned ONNX engine is listed with an explicit "unavailable" status.
  • Machine-readable CLI contracts. All public command-line tools, including Sensory, describe themselves through a shared cognia-cli-v1 descriptor: --help prints the usage and --help --json prints every positional argument, option, default, allowed value, exit code and usage form as JSON. The Markdown reference in Docs/agents/reference/cli/ is generated from the same descriptors, and a drift test fails the build gate when a tool and its reference disagree.
  • cognia_capabilities --json lists the engine registry and the capabilities of each engine.
  • Named-channel datasets for cognia_learning: dataset-validate checks a cognia-named-channels-v1 JSON manifest with its JSONL records, and dataset-split makes a deterministic, seeded train/validation/test split by source group, so records of one group never end up in two sets.
  • cognia_learning evaluate-classification turns multi-seed held-out predictions (cognia-classification-predictions-v1) into a JSON and Markdown report (cognia-classification-evaluation-v1) that always compares against a mandatory baseline.
  • Runtime Docker image for Linux (packaging/linux/Dockerfile.runtime): sensory:<version> installs the .deb into /opt/Sensory, the cublas target adds the cuBLAS runtime for the transformer CUDA tools.

Changed

  • Sensory parses its command line through the shared descriptor. The options are the same as in 7.1.0; the help text is now in English and an invalid option ends with exit code 2 and the usage on standard error.
  • Examples 18, 20 and 21 no longer contain C++ sources or their own CMake projects: the programs they need ship precompiled in bin/, and the example scripts call them from the package. Distributed examples contain only declarations, data and scripts.
  • The release scripts (release.ps1, release.sh) additionally check that the package contains the agent documentation (inventory, engine cards, capability matrix, engine selection guide); the website sources in Docs/website/ are never packaged.

Fixed

  • On Linux, cognia_validate* gates and tools that start sibling programs located them through argv[0]. Called by bare name through PATH (for example after installing the .deb), the gates reported BLOCKED; they now locate their directory through /proc/self/exe.
  • The Linux release README pointed cognia-desktop.sh at 127.1.0.1 instead of 127.0.0.1.
  • Document-routing and website-route checks gave different results on Linux (path separators, culture-dependent sorting).

Known limitations

The limitations of 7.1.0 still apply: Linux camera input was not verified on real hardware, cuBLAS is not bundled, the Windows package targets the RTX 50 series only, and on Linux the desktop client is a browser launcher. The ONNX engine described in the Agent Design Kit is planned, not implemented.