v1.0.0 - Stable minimal kernel

v1.0.0 — stable minimal kernel

The planned local-first, single-agent kernel is complete. This release focuses on
compatibility and release evidence, not a larger capability surface.

Highlights

• Stable documented 1.x host API and data contracts; direct package-root imports
  for settings, runtime, audited read tools, events, and model/context extension types.
• Atomic SQLite upgrades from supported legacy schemas without data loss; newer
  databases are refused before modification.
• Strict configuration and evaluation input validation, preventing silent typos.
• One active Web task at a time; overlapping requests return HTTP 409.
• Versioned evaluation reports; missing/all-zero usage stays unknown.
• Fresh wheel/source installs checked on Linux and Windows, with CLI, embedding,
  history/audit, receipts, and packaged Web smoke tests.

Validation

• 91 automated tests passed locally, plus lint, format, and dependency checks.
• The six CI jobs passed on the release commit: minimum/current dependency tests,
  Linux/Windows fresh installs, and optional MCP integration.
• Ollama qwen3:8b: workspace reads 3/3 and safety boundaries 2/2.
• Configured compatible API / openai/gpt-4o-mini: workspace reads 3/3 and safety
  boundaries 2/2; the route did not provide measured token usage.
• These are ten compatibility smoke cases, not a statistical benchmark or a cost estimate.

Upgrade and install

Python 3.11+ is required. Download the wheel and install it in a virtual environment:

python -m pip install minimal_local_agent-1.0.0-py3-none-any.whl
minimal-agent --version

The source archive is also provided. This release is not published to PyPI.
Before upgrading, stop active runs and back up SQLite using its backup API.
Correct v0.9 configuration remains valid; fix unknown/misspelled fields that were
previously silently ignored. Serialize tasks sharing a database/workspace.

See compatibility and migration notes,
validation records,
reviewed smoke results,
and the release CI run.

Scope and privacy

Five built-in tools remain. No shell, browser, hidden sub-agents, background
orchestration, or automatic context compression was added. The explicit
context_reducer seam remains available for trusted hosts and disabled by default.
Actual model addresses, credentials, personal defaults, and full local reports
are excluded from Git and distributions. Custom Python/MCP code is trusted host
code, not a sandbox. See the threat boundary in SECURITY.md.

Development was split into nine reviewable, independently verifiable commits.
v1.x prioritizes maintenance and small host-owned adapters.

v0.9.0 - Context budget and deterministic evaluation

Adds a per-request serialized-message byte budget with explicit rejection and an opt-in host reducer seam. Full conversation history remains in SQLite even when the model view is reduced; automatic summarization is not enabled. Evaluation cases now have isolated workspaces and deterministic file assertions, with task-family success and per-success latency/token reports. Includes the focused v1.0 roadmap and privacy-safe documentation. Local tests: 66 passed; real local qwen3:8b evaluations: built-in 3/3 and adversarial 2/2; GitHub CI: all 5 jobs passed. This remains an alpha release.

v0.8.0 - Explicit Python read tools

Adds explicit, policy-visible Python read-only tool registration for CLI and local Web hosts, with audit receipts and a runnable example. Simplifies the extension boundary by removing raw toolset injection. Includes documentation and validation updates. Python extensions are trusted code, not sandboxed. Local connection settings remain unpublished.

v0.6.0 - Local Web console and compatible model endpoints

Adds a loopback-only Web console, opt-in OpenAI Chat Completions-compatible model endpoints, safer local configuration, session usability improvements, and cross-platform/package regression checks. Ollama remains the public default. Real service addresses, API keys, and personal model defaults stay in ignored local settings. See CHANGELOG.md and README for details.

v0.5.0 - Auditable Agent Kernel

Full Changelog: v0.2.0...v0.5.0

v0.2.0 - Verifiable local agent

Adds mechanical read-only mode, diff-first write approval, audit inspection and JSON export, deterministic local-model evaluation, simpler small-model-friendly tool schemas, bilingual comparison documentation, and Windows SQLite cleanup fixes. Both qwen3:4b and qwen3:8b passed the 3/3 local integration checks.

v0.1.0 - Initial release

Initial minimal local-agent loop with workspace-confined tools, confirmed writes, SQLite sessions and audit events, tests, and CI.