mcp: work on several updates in parallel (workspaces, background jobs, lock wait)

[plusky]

Lets one client — or several agents — keep more than one update moving at once, without each load_template tearing the previous update's hosts down. Five commits.

1. Named workspaces

Every tool gains an optional workspace selector (default "default"). Each name resolves, via the existing SessionRegistry, to its own isolated McpSession (own loaded template, targets, per-session lock). Calls in different workspaces run concurrently, so a single stdio client can advance update B while update A's slow host op runs. stdio now uses the registry too (idle sweeper disabled). New tools: list_workspaces, close_workspace. The default workspace reproduces today's behaviour.

2. Async background jobs

Slow host commands (run, update, downgrade, prepare, install, uninstall, set_repo, reboot) take background=true: returns a job id immediately instead of holding the request open, running under the workspace lock. Poll with job_status, fetch with job_result (job_list / job_cancel too).

3. Wait for a busy refhost

Separate agents are separate processes, so a refhost lock genuinely excludes them. [lock] wait (seconds, default 0 = fail-fast) makes lock queue on a busy host — polling every [lock] wait_poll seconds until released/reaped/ours — instead of erroring. The connect-time warning is untouched; a warning is logged on wait-start/timeout so a REPL user still sees the host is busy.

4. Refhost pool selection

When refhosts.yml lists several interchangeable hosts for the same test target and [refhosts] pool_select is on, add_host connects just one free host per target instead of the whole matrix: tries candidates in turn, skips any locked by another agent, and claims (locks) the one it takes — so parallel agents draw distinct hosts. The target is the full product + version + arch + addons the update asks for, not just arch — so an update spanning all arches of SLE15-SP5 and SP7 still gets a host per (service-pack, arch); only genuine duplicates collapse to one. Searched across all locations (location ignored). Off by default.

Together: a pool gives several agents distinct hosts; lock wait makes them queue when the pool is exhausted; workspaces + background jobs keep several updates moving per agent/client.

Notes

• Workspaces inside one process share the OS pid, so the (user+pid) refhost lock treats them as one owner — they never block each other on a host; coordinate same-process workspaces so they don't drive the same host destructively at once. [lock] wait and pool selection matter across separate processes/agents.
• location is optional: with none configured/specified it defaults to the default bucket of refhosts.yml; pool selection ignores location entirely.

Tests

Full suite green (1312 passed locally). New coverage for workspaces, background jobs, lock-wait, and pool search/selection (incl. the SP5-vs-SP7 distinct-slot case). Docs: new "Working on several updates in parallel" section in Documentation/mcp.rst.

[codecov]

Codecov Report

❌ Patch coverage is 92.64706% with 30 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.85%. Comparing base (15761b8) to head (9a8dd8a).

Files with missing lines	Patch %	Lines
mtui/mcp/session.py	83.33%	13 Missing ⚠️
mtui/test_reports/testreport.py	89.28%	12 Missing ⚠️
mtui/mcp/tools.py	94.38%	5 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #216      +/-   ##
==========================================
+ Coverage   86.52%   86.85%   +0.32%     
==========================================
  Files         160      161       +1     
  Lines        8980     9369     +389     
==========================================
+ Hits         7770     8137     +367     
- Misses       1210     1232      +22     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.