Releases & upgrading

0.2.8 — September 13, 2026

This release collects the changes since 0.2.4, the previous PyPI release. The package requires Python 3.11 or newer.

The MCP SDK dependency stays on compatible 1.x releases. Fresh package installation is checked separately from the locked development environment.

Install or upgrade

pip install --upgrade coordination-memory-mcp==0.2.8

For an existing uv tool installation:

uv tool upgrade coordination-memory-mcp

Restart your MCP client after upgrading. Run comem init in each participating repository to refresh the managed protocol section and tool adapters, then review the generated changes. Keep using the same absolute COORDINATION_MEMORY_DB path so agents and the dashboard share one store.

MCP client migration

Discovery now returns compact summaries. Start with list_workspaces, inspect one workspace with get_workspace_detail, and select a team before calling get_team_board, list_pending_reviews, get_snapshot, list_contracts, or get_attention_board. These reads require one exact team_id.

get_team_board returns a flat list containing only assignment_id, title, and status. Its optional status filter accepts one exact status. With no filter, accepted, rejected, cancelled, and superseded assignments are omitted. Use get_assignment_detail to read a task's full contract.

Use get_event_detail, get_run_detail, and get_contract_detail after selecting a specific identifier. Do not expect full records in discovery responses.

Eleven event-writing tools return a compact acknowledgement without echoing the payload: append_event, submit_handoff, heartbeat_run, checkpoint_run, raise_attention, record_run_binding, request_intervention, respond_intervention, review_event, accept_event, and reject_event. claim_assignment still returns its full response, including active_run_id; actor, workspace, team, assignment, and contract mutations retain their existing responses.

See the tool reference for response shapes and filters.

Human oversight

checkpoint_run refreshes the Human Resume Brief. raise_attention records non-blocking yellow or resolved green Attention; request_intervention represents a red blocking request. The dashboard presents Human Attention before the assignment board. Updating a Brief does not change run status or create Attention. See the worked example.

Offline evaluation preparation

comem eval imports immutable baseline snapshots, builds reproducible experiment manifests, validates them, and exports JSON Schemas. These commands prepare evaluation inputs; this release does not run benchmarks.

comem eval --help
comem eval baseline --help
comem eval manifest build --help
comem eval validate --help

See the evaluation foundation and the full changelog.