Persistence, Multi-agent Symposia/Teams, and Dashboard Improvements

v0.16.4

Changes since v0.16.3:

Agents and CLI

• Expanded the LAMMPS agent with user-supplied potentials, input templates and
  structure data, GPU launch support, potential-selection-only mode, automatic
  result summarization, improved repair history, and structured progress events
  (#277).
• Added richer default CLI logging with MIME-aware rendering for source code,
  command output, diffs, plans, and tool/agent progress. Fixed artifact
  rendering for named agents (#278, #288).
• Fixed PlanningAgent state and message accumulation, and simplified its
  review/revision flow to avoid invalid model message sequences (#285, #286).
• Added --config support to ursa mcp-server (#275).

Dashboard

• Added secure API-key management using the operating-system credential store,
  while retaining environment-variable and keyless configuration options
  (#275).
• Added dashboard launching, monitoring, continuation, and cancellation of agent
  teams and symposia (#280).
• New sessions now require an explicit persistent or temporary workspace instead
  of receiving a hidden default workspace; legacy session workspaces remain
  recoverable (#282).
• Refined dashboard layout, controls, settings, and model-provider handling
  (#275, #276).

Persistence

• Unnamed CLI sessions are now ephemeral and no longer create checkpoint
  databases. Named sessions remain persistent, with a migration warning for
  legacy workspace checkpoints (#279, #287).
• Persistent-agent SQLite databases now prune obsolete checkpoint history after
  successful runs, substantially reducing long-running storage growth (#286).

Documentation and maintenance

• Reorganized and expanded the agent, logging, API, environment, persistence,
  installation, and tutorial documentation (#278).
• Added versioned documentation publishing and updated MCP server guidance
  (#281).

Changes before that are described in the general documentation.

Hotfixes for a couple bugs

There was a structured output bug for specific endpoints that emerged when the previous output bug was fixed. Missed testing on the last release.

Also two other important fixes:

• Context summarization had a bug that could inflate the number of tokens sent to the summarizing invoke. This could hit context limits and crash runs. It is now fixed by proper handling of the messages object in that function.
• The dashboard had a couple places where model kwargs, api_key and base_url, were being handled in a manner that didnt generalize to multiple endpoints well. This is now fixed.