v0.5.1

What's Changed

Maintenance release: no new features, no breaking changes. Existing
settings and chat history are read as-is.

Changed

• Codebase reorganised into explicit layers — core/, domain/,
  application/, infrastructure/, presentation/ — with the dependency
  direction enforced by an architecture test. Domain contracts now live
  next to their consumers; the OpenAI wire format is confined to one codec;
  ChatSession is split into an accumulator, a persister and the session
  itself; every conversation action (create / fork / rename / delete) goes
  through one controller; MCP runtime state (connection, tools, prompts)
  is no longer stored in the persisted server config. No user-facing
  behaviour change beyond the fixes below.

• Stricter static analysis. strict-casts, strict-inference,
  strict-raw-types and ~50 additional lint rules; the whole tree is
  dart formatted.

• Test suite grew from 164 to 254 tests, now covering the streaming
  pipeline end to end (ChatSession, ChatSessionManager,
  ToolExecutor), providers and the main widgets.

• Outbound HTTP requests now identify the client. Every request to the
  LLM API and to MCP servers carries
  User-Agent: SpecterChat/<version> (<os>) Dart/<runtime> instead of
  Dart's default Dart/3.x (dart:io), so server-side logs can tell
  SpecterChat apart from other clients and which version is talking. A
  User-Agent set in a server's custom headers still takes precedence.

• The MCP clientInfo reports the real app version. It was pinned to
  0.1.0 regardless of the build; it now follows pubspec.yaml, like the
  About section.

Fixed

• Stop really stops. Pressing Stop while the model had already emitted a
  complete tool call used to execute that tool and start another model
  turn: the transport reported a cancelled request the same way as a
  finished one. Cancellation is now its own stream event; a stopped turn
  keeps its partial text and never runs tools or retries.
• Deleting a conversation tears down its streaming session first, so an
  in-flight stream can no longer keep writing to rows that are gone.
• Message ids are strictly monotonic. ORDER BY id is the canonical
  message order, but UUIDv7 ids minted within the same millisecond (parallel
  tool results, a fast local model) carried random low bits and could sort
  out of creation order. The generator now uses the 12-bit rand_a field
  as an intra-millisecond counter.
• A conversation created while the session cap was reached could be
  handed to the UI already disposed — the LRU pass no longer evicts the
  session it is creating.
• Conversations created within the same second now keep a stable order
  in the sidebar (updated_at has one-second precision; ties break on id).
• Per-conversation settings edits are no longer lost when switching
  conversations during the 500 ms save debounce.
• A tool call no longer fails for good once the server has dropped the MCP
  session. Streamable HTTP servers evict idle sessions (rmcp does after
  5 minutes without a request) and answer every later call carrying the old
  mcp-session-id with 404 Session not found. The client used to surface
  that as a generic Error POSTing to endpoint (HTTP 404) to the model,
  retry the dead id on its SSE stream, and stay marked connected; the only
  way out was a manual reconnect. The transport now treats that 404 for what
  the spec says it is — the session is gone — forgets the id and stops the
  reconnection loop, and McpService opens a new session and replays the
  call once, so a pause in the conversation is invisible to the model.

Full Changelog: v0.5.0...v0.5.1