Skip to main content

Changelog

Breaking changes, new features, and migration notes for each Developer Preview release.

Info

Edge products are in Developer Preview. Expect breaking changes between releases. Pin your package versions and validate after each upgrade.

Versioning policy​

During Developer Preview, Swift package releases follow 1.0.0-rcN tags and the Edge Studio Python package releases as 0.0.1rcN on PyPI. The current pins for every surface live on Current Versions. Breaking changes are documented here with migration steps. After general availability, we will follow semantic versioning.

PyPI retention note: Edge Studio preview wheels before 0.0.1rc19 have been removed from PyPI. Changelog entries before v0.0.1rc19 remain as release history, not as currently installable package pins.

How to upgrade​

  1. Update the version pin in your Package.swift.
  2. Read the changelog entry for any breaking changes.
  3. Build and fix compiler errors.
  4. Run your test suite on a real device.
  5. Verify first-launch, multi-turn, and memory behavior.

Developer Preview boundaries​

Developer Preview is a limited preview channel. The changelog documents what is shipped, what is public, and what is deliberately not enabled yet.

Current versions​

Current version pins and the compatibility matrix live on one page: Current Versions. This changelog records release history and per-version notes; it is not the source of truth for what to install today.

Generic builds and simulator checks are not enough for runtime claims. Re-run real-device validation after changing any preview tag.

Known limitations​

The B2/B4/B5/B6/B7 CLI commands listed below are shipped in current preview; the limitations describe their safety boundaries.

  • edge studio launches the local Studio UI and API server on localhost by default. It is a local developer workbench entrypoint, not a hosted service.
  • edge doctor is a read-only environment check. It does not download models, load models, start the backend, or run Neural Imprint workflows.
  • edge models list, edge models where, and edge models doctor are read-only model readiness checks. They do not download models, write receipts, or probe the network.
  • edge models fetch is an explicit model preparation command with --dry-run, source selection, local receipts, and no silent demo download.
  • edge demo chat is a base-model chat command. It uses an explicitly prepared local model and writes a hash-only edge.demo.chat.receipt.v1 receipt by default.
  • edge demo receipt and edge demo local-only are receipt inspection tools. They validate edge.demo.receipt.v1 local-only invariants and do not generate Neural Imprint artifacts or call model runtimes.
  • edge demo imprint run --dry-run is a pre-flight planner. It emits edge.demo.imprint.plan.v1 with hash-only sample/question metadata and local model prerequisite status.
  • edge demo imprint run without --dry-run is the Neural Imprint demo. It generates and restores a local artifact and writes a comparison receipt.
  • edge demo imprint compare is a receipt-only comparison inspector. It reads a completed edge.demo.receipt.v1 receipt and emits edge.demo.imprint.compare.v1 without loading models, restoring artifacts, generating answers, or using the network.
  • edge demo learn run --dry-run is a correction-learning pre-flight planner. It emits edge.demo.learn.plan.v1 with hash-only synthetic correction metadata and isolated-state paths; it does not write correction ledgers, call regen, load models, or write a learn receipt.
  • edge demo learn run without --dry-run is the correction-learning demo. It writes synthetic Persona/RPP input and correction ledger entries under the demo run state, triggers correction regen, restores the regenerated local Neural Imprint artifact, compares before/after answer hashes, and writes edge.demo.learn.receipt.v1.
  • edge demo learn run --prepare-model combines model preparation and the learning demo in one command. It may explicitly prepare a compatible local model first, then records model-preparation network use separately as network_used_during_model_prepare from the local learning demo.
  • edge demo facts import-url is a bounded single-URL material import path. It accepts HTTP(S) text content, records network_used=true, writes hash-only receipts by default, and does not crawl linked pages. In rc22 the optional host-model extractor can propose candidate facts from one page with a local Mac model; Edge deterministically validates the structured output before writing facts.
  • edge demo facts crawl-url is a bounded static HTTP(S) material import path for a small same-origin documentation set. It requires explicit depth, URL, byte, total-byte, and timeout bounds; it does not execute JavaScript, use a browser, consult robots.txt, or leave the origin.
  • edge demo tools validate accepts only local read-only facts lookup manifests in this preview. It does not execute processes, perform network access, or write demo state.
  • edge tools validate / edge tools inspect <tools.py> (rc21+) import the tool file inside an isolated Edge-owned runner subprocess — top-level code executes. They never run developer code in the Edge CLI process, and they are not static safety scans for untrusted files.
  • edge demo chat --tools <tools.py> (rc21+) enables Edge-managed custom Python tools; it is mutually exclusive with --tools-manifest and --facts-store. Tool code executes only in the runner subprocess; the model can only emit JSON calls.
  • edge demo chat --tools-manifest <tools.json> enables developer-named read-only tools backed by local facts stores. It is mutually exclusive with the --facts-store shortcut.
  • edge demo reuse is an artifact reuse smoke check. It reads a completed receipt and writes per-app edge.demo.reuse.receipt.v1 manifests without copying artifacts, syncing devices, restoring artifacts, loading models, or using the network.
  • Product-default paired-device route is not enabled by this preview documentation or changelog. Broad live routing still requires separate explicit policy, opt-in, and real-device evidence.
  • Background automation scheduler is not shipped. The bounded automation API remains explicit, dry-run by default, and fail-closed.
  • Generic capsule apply-status UI reference is shipped in Edge Scaffold and the dogfood validation app. Product-specific placement, layout, or copy remains outside the preview baseline.
  • Model push and product-default Neural Imprint regen execution remain unsupported without separate explicit policy/design. The shipped edge demo learn run path is an explicit local synthetic demo.
  • EdgeMesh capsule auto-restore SDK orchestration is already shipped through HaloCapsuleAutoRestoreCoordinator in Edge Kit 1.0.0-rc94; it is not a current limitation.

edge-studio​

v0.0.1rc23​

  • PyPI release candidate version: 0.0.1rc23. Deterministic install: python -m pip install edge-studio==0.0.1rc23.
  • Adds generic read, prepare, and commit execution levels for custom Python tools. commit calls become local pending actions and run only after an explicit edge tools confirm command validates the expiring confirmation token and the frozen tool contract.
  • Requires every return_direct=True tool to declare a closed output_schema. Edge validates the actual tool result before direct delivery and fails closed with tool_result_schema_mismatch for missing fields, wrong types, or undeclared fields.
  • Confirmation is bound to the tools file, active set, tool schema, arguments, and execution level. Concurrent or repeated confirmation does not re-run a claimed action.
  • Known limits: expired pending files are not automatically removed, and a process crash after a side effect but before receipt persistence can leave an audit gap. Edge preserves at-most-once execution and does not retry that action automatically.

v0.0.1rc22​

  • PyPI release candidate version: 0.0.1rc22. Deterministic install: python -m pip install edge-studio==0.0.1rc22.
  • Adds edge demo facts import-url --extractor host-model --extractor-model <model> for local Mac host-model facts extraction. The model output is treated as candidate facts only; Edge validates the structured payload before writing the local facts store. Receipts include hashes for the model, prompt, schema, model input, raw model output, and validated payload, plus truncation metadata and non_deterministic_extraction=true.
  • Adds edge demo facts crawl-url <url> for bounded static same-origin material import. Required bounds are --max-depth, --max-urls, --max-bytes, --max-bytes-total, and --timeout; receipts capture redirect-chain hashes, per-page statuses, total bytes, same-origin policy decisions, and network_used=true.
  • Keeps crawling deliberately narrow: no browser, no JavaScript execution, no cross-origin traversal, and no robots.txt fetch. You remain responsible for using sources you are allowed to access.
  • Updates the Ethereum example as a generic domain-material workflow: URL import, optional local host-model extraction, bounded crawl, custom Python tools, and Neural Imprint learning. Ethereum is an example domain, not a special Edge runtime path.
  • Keeps the rc21 custom Python tool path unchanged: @edge_tool, edge tools validate, edge tools inspect, edge demo chat --tools, and edge demo learn run --tools.

v0.0.1rc21​

  • PyPI release candidate version: 0.0.1rc21. Deterministic install: python -m pip install edge-studio==0.0.1rc21.
  • Adds custom Python tools. Mark plain functions with @edge_tool (from edgestudio.tools import edge_tool), then run edge tools validate <tools.py>, edge tools inspect <tools.py>, and edge demo chat --tools <tools.py>. Up to 8 tools are active per run; select with repeatable --tool <name> or --tool-tag <tag>. See the Custom Python Tools guide.
  • Developer tool code never runs in the Edge CLI or model process. Discovery and every call execute in a fixed Edge-owned runner subprocess that verifies the tools file is byte-identical to the frozen active set before importing it (tools_file_changed fails closed). Per-call timeout kills the runner; results are capped at 64 KB of canonical JSON.
  • Tool schemas are generated deterministically from type hints (edge.tools.schema_gen.v1). Supported hints: str, int, float, bool, Literal, Optional, list[T]. Unsupported hints fail validation instead of degrading to Any.
  • Receipts for Python tool runs are hash-first and runner-aware: tools_file_sha256, active_set_sha256, schema_generator_version, and per-call args_sha256, result_sha256, tool_schema_sha256, runner_secret_verified, tools_file_sha256_verified, network_used_by_edge: false.
  • Adds edge demo learn run --tools <tools.py>: bakes the Python tool contract into the Neural Imprint so a restored Agent already carries the tool schemas. edge demo chat --with-imprint --tools gates restore at the schema level: implementation-only edits keep the artifact valid, while schema or active-set changes fail closed and require relearning (imprint_requires_tools, imprint_tool_active_set_mismatch, imprint_tool_schema_mismatch).
  • --tools, --tools-manifest, and --facts-store are pairwise mutually exclusive per run; --tool / --tool-tag require --tools.

v0.0.1rc20​

  • PyPI release candidate version: 0.0.1rc20. Deterministic install: python -m pip install edge-studio==0.0.1rc20.
  • Adds edge demo facts import-url <url> for bounded HTTP(S) material import into the local facts store. The importer supports --split page and --split html-table-rows, records source/final URL hashes, content type, status, raw/extracted hashes, truncation state, and network_used=true.
  • html-table-rows captures links inside each row as data and absolutizes relative href values against the final URL. It stores those links in fact text and does not follow them.
  • Adds edge demo tools validate <tools.json> for edge.demo.tools.manifest.v1. The rc20 runtime supports developer-named read-only tools with kind: "local_facts_lookup" only.
  • Adds edge demo chat --tools-manifest <tools.json>. Chat can now use developer-named read-only local fact tools; receipts record tools_manifest_sha256, tool summaries, tool calls, and network_used=false.
  • Adds edge demo tools validate --learn-sample <sample.json> mismatch warnings when a learn sample's tool_schema_export.tools[].name does not match the runtime tools manifest. The warning is non-blocking so you can audit sample/runtime drift before a run.
  • Keeps edge demo chat --facts-store <store> as the shortcut for the built-in local_facts_lookup tool. Use --tools-manifest when the carrier wants stable developer-owned tool names.
  • Updates model catalog readiness for local Mac-class workstations so qwen3.5-27b-4bit resolves as an LLM-capable local model for learn/chat workflows. The strongest available local model remains a developer choice, not a hard dependency.

v0.0.1rc19​

  • PyPI release candidate version: 0.0.1rc19. Deterministic install: python -m pip install edge-studio==0.0.1rc19.
  • Adds local facts workflows through edge demo facts import, list, and inspect. The store is local-only and defaults to hash-only inspection output.
  • Adds edge demo chat --facts-store <store> for read-only local_facts_lookup tool use. The tool loop is fail-closed and records network_used=false.
  • Supports combining --with-imprint and --facts-store, so Neural Imprint behavior and local fact lookup can run together in one chat path.
  • Extends chat receipts with tool_calls[], tool_instruction_mode, and tool_instruction_sha256 for auditable local tool behavior.
  • Updates the Neural Imprint prefix renderer to the v2 JSON tool-call contract, keeping baked and runtime tool instructions aligned.
  • Lets you refresh local knowledge by re-importing facts into a store without re-running the learning flow.
  • Renames the installed Python package namespace from backend to edgestudio. The public CLI remains edge; package-internal source references now use paths such as edgestudio/cli/demo_samples.py.
  • Keeps --sample-file as the customization path for local edge.demo.learn.sample.v1 JSON learning samples.
  • Aligns release metadata with the AtomGradient Proprietary License. The public AtomGradient/edge-studio repository is an issue and support shell, not an open-source source distribution.

v0.0.1rc18​

  • Historical PyPI release candidate version: 0.0.1rc18 (removed from PyPI; kept here as release history).
  • Adds edge demo learn run --sample-file for local edge.demo.learn.sample.v1 JSON samples. The file path overrides built-in --sample fixtures and is validated before dry-run or execution.

v0.0.1rc9​

  • Historical PyPI release candidate version: 0.0.1rc9 (removed from PyPI; kept here as release history).
  • Adds expected_tool_policy to the finance learning demo dry-run, receipt, and text preview. The field is a deterministic preview, not a live tool-call trace.
  • Updates the first-run developer path into a one-page device Agent flow: preference learning, tool policy inspection, and Edge Studio carrier export.

v0.0.1rc8​

  • Historical PyPI release candidate version: 0.0.1rc8 (removed from PyPI; kept here as release history).
  • Fixes package version consistency so Python package metadata and edgestudio_core.__version__ both report 0.0.1rc8.

v0.0.1rc7​

  • Historical PyPI release candidate version: 0.0.1rc7 (removed from PyPI; kept here as release history).
  • Adds finance_conservative_cashflow_v1 as the default learning demo sample. The sample is synthetic, inspectable, finance-shaped, and designed for the device Agent quickstart.
  • Updates the public first-run path around raw local signal inspection, base model chat, RPP self-learning, Neural Imprint generation, base model + Neural Imprint chat, and Edge Studio carrier export.

v0.0.1rc6​

  • Historical PyPI release candidate version: 0.0.1rc6 (removed from PyPI; kept here as release history).
  • Improves Edge Scaffold export documentation. Generated apps now receive an instance-specific README that includes the app name, model name, model path, ODR tag, key files, model loading notes, and troubleshooting.
  • Keeps the exported app structure flat: MyApp/MyApp/App/ScaffoldConfig.swift, not a triple-nested app path.

v0.0.1rc5​

  • Current public Python package distribution name: edge-studio.
  • Historical PyPI release candidate version: 0.0.1rc5 (removed from PyPI; kept here as release history).
  • edge demo chat streams tokens in interactive mode and supports Neural Imprint restore through --with-imprint <learn_receipt.json>.
  • edge demo learn run --include-text prints the completed learning receipt and a ready-to-run next: command for after-learning chat.
  • edge studio now reports an already-running EdgeMesh service as an actionable warning while keeping the Studio UI available.

v0.0.1rc1​

  • Public Python package distribution name: edge-studio.
  • Historical PyPI release candidate version: 0.0.1rc1 (removed from PyPI; kept here as release history).
  • Installed command surface is intentionally a single edge entry point.
  • edge studio launches the local Studio UI/API server at http://127.0.0.1:18842 by default.
  • edge demo chat, edge demo learn, model readiness, explicit model fetch, receipt inspection, and install docs are aligned with the public package path.
  • The repository history was compacted to a single root commit before public release and tagged v0.0.1rc1.

edge-kit​

The current Edge Kit pin lives on Current Versions. Preview capability summary: modules EdgeInference, EdgeModelKit, EdgeVoice, EdgeMesh, EdgeData, EdgeDataMeshBridge, EdgeUI, EdgeSession; LLM, VLM, TTS, STT engine support; DSR Attention for long-context multi-turn sessions; automatic KV cache memory policy; Neural Imprint runtime restore primitives and EdgeMesh capsule auto-restore coordinator APIs; production app builds can embed generic EdgeBuildCommit metadata for snapshot traceability.

1.0.0-rc103​

  • SDK release gate for this release.
  • Aligns EdgeKitRuntime.version, edge-kit/.dependency_versions, Package.swift, and public SPM metadata to Edge Engine 1.0.0-rc143.
  • Keeps the public package graph on edge-engine, swift-transformers, and GRDB; historical mlx-swift / mlx-swift-lm packages remain out of the default dependency graph.

1.0.0-rc98​

  • Historical SDK release gate.
  • Aligns package, tests, and public metadata to Edge Engine 1.0.0-rc138.
  • Keeps RPP orchestration and activation steering implementation out of the public SDK source; those internals live behind Edge Halo.

1.0.0-rc94​

  • Added EdgeMesh capsule auto-restore coordinator APIs.
  • Depends on Edge Engine 1.0.0-rc137.

edge-halo​

The current Edge Halo binary pin lives on Current Versions. Preview capability summary: lifecycle for local profile jobs and Neural Imprint capsule compatibility; fail-closed validation for model, tokenizer, runtime, and tool-schema identity; HaloTextGenerator and HaloEngineSession protocols; EdgeHaloRuntime public actor; RPP A-library provenance validation and profile artifact lifecycle helpers. Public apps consume the binary package edge-halo-binary; the source repository remains private.

edge-engine​

The current Edge Engine dependency tag lives on Current Versions; it is resolved by Edge Kit rather than imported directly. Preview capability summary: native Metal inference runtime; DSR Attention implementation. Unreleased commits on main are not part of this Developer Preview until a new 1.0.0-rcN release is published.

edge-scaffold​

The current Edge Scaffold pins live on Current Versions. Preview capability summary: iOS app template generation from Edge Studio export; ScaffoldConfig-based customization; four-tier model delivery (Cache → Bundle → ODR → HuggingFace).