onfiala/dots-brain

v0.3.0-alpha.2

Install and operate a personal, source-aware memory service on your own host.

Changelog

Add an official-SDK shell client for an existing MCP connection, with schema discovery, JSON input, scoped calls, and recovery of an enabled local service.
Verify an installed instance on the user-confirmed shared Work/Dot VM with local CPU embeddings. Public ingress and native web-client activation remain unverified.

Reject IPv6 HTTP issuer origins before changing configuration: SDK 1.30 does not support them. HTTPS issuers and exact localhost/127.0.0.1 development origins remain supported.
Reject issuer URLs with an empty user-information delimiter as well as real credentials.
Add regression cases and clarify the supported development origins.
Supersede the alpha.1 release candidate before publication; its Git tag is preserved.

Add optional VM-local OAuth to the existing HTTP process and SQLite database, using official MCP SDK handlers with no new runtime dependency.
Add discovery, dynamic registration, S256 PKCE, resource-bound code exchange, scoped project grants, rotating refresh tokens, and client/owner revocation.
Add deterministic owner commands to inspect, approve, deny, and revoke a specific connection without printing its tokens. Default approval excludes deletion and can narrow the client's requested scopes; all-project access remains explicit.
Persist grants across restart and invalidate pending flows and refresh tokens on OAuth disable, issuer changes, and whole-instance uninstall.
Bound registration metadata, request bodies, client count, and pending flows; reject invalid callback schemes and malformed registration JSON.
Enforce token-request resource binding and normalize the optional public-client revocation field around the SDK's current handlers. Accept case-insensitive HTTP Bearer scheme names.
Require the tested MCP SDK 1.30 or newer within the existing major version.
Verify 17 OAuth tests, including the official client's full flow and automatic refresh over live local HTTP, concurrent code reuse, scope/project boundaries, expiry, revocation, capacity recovery, and retained-data removal.
Document the agent workflow, storage boundaries, and remaining ingress/provider requirements. Public deployment and actual web-provider setup remain unverified.

Add repeatable uninstall with a read-only preview, host credential revocation, managed-process stopping, and removal of unchanged registered client entries.
Preserve memories, revisions, deletion suppression, models, unrelated client settings, and modified or unreadable configurations; report incomplete cleanup.
Block automatic restart after removal until explicit up --resume or reinstall.
Add disconnect, an integration inventory, and separate credentials for newly configured profiles. Recognize unchanged legacy entries and explicit custom paths.
Preserve user-supplied remote credentials and report required issuer revocation.
Document software removal as a separate step; add uninstall, upgrade, everyday usage, troubleshooting, and documentation-index guides in English.
Extend the setup skill to cover removal and upgrades, and validate links across root documentation, nested guides, and skills.
Verify 47 tests on Python 3.11 and 3.12 with one unchanged opt-in model test skipped; verify clean bootstrap, real Claude Code health, removal, reinstall, data retention, and removal of a disposable test checkout and package.

Add repeatable Linux background startup, crash recovery on local client reconnect, generated-bridge verification, and private credential reuse.
Add Claude Code, Cursor, Codex, and generic MCP configuration adapters with preserved settings and explicit application-activation status.
Extend the bootstrap to start the service and optionally connect a client.
Keep MCP reads responsive while SQLite writers wait, and bound read/write workers separately.
Avoid scanning the full-text index when inserting a new source record.
Drain embedding backlogs in short bounded batches instead of sleeping between every batch.
Recover the canonical port after a killed process finishes releasing its sockets.
Add bounded synthetic storage, HTTP, and multilingual embedding stress harnesses.
Add a manually triggered GitHub release workflow that verifies an existing annotated tag, builds on both supported Python versions, and uploads checked packages to a draft without overwriting published assets.
Document the repeatable release procedure.

Fix intermittent first-run failures when concurrent installers enable SQLite WAL. Journal-mode lock contention now retries within a bounded deadline, while other errors still fail immediately. The existing schema transaction remains atomic.
Test temporary lock recovery and timeout recovery without losing existing data.
Let both supported Python versions finish CI even when one matrix job fails.

Establish the product contract, English documentation, and MIT license.
Add transactional SQLite memory, provenance, explicit revisions, project-filtered full-text search, export, and deletion with reimport suppression.
Verify concurrent retry handling, restart persistence, and project isolation.
Add six MCP tools, stdio and authenticated loopback HTTP, scoped client credentials, revocation, a stdio-to-HTTP bridge, and machine-readable CLI setup.
Add pinned multilingual CPU embeddings with artifact verification, automatic background indexing, and hybrid retrieval; no paid inference API is used.
Add a deterministic bootstrap, an onboarding skill package, English operating documentation, release validation, and GitHub CI for Python 3.11 and 3.12.
Verify a live HTTP process through the stdio bridge and fresh source installation.

The HTTP service binds loopback and does not implement web OAuth or a tunnel.
Provider capture, automatic remote configuration, native ChatGPT UI, and MCP Events are not implemented. Dot VM persistence, costs, and strong credential isolation are not verified.
The embedding smoke test is not a Cortex comparison or a production benchmark.