Adds `.core/workspace/db.duckdb` — the permanent record of dispatch
cycles described in RFC §15.5. Stats rows persist BEFORE workspace
directories are deleted so "what happened in the last 50 dispatches"
queries survive cleanup and sync drain.
- `workspace_stats.go` — lazy go-store handle for the parent stats DB,
build/record/filter/list helpers, report payload projection
- `commit.go` — writes a stats row as part of the completion pipeline so
every committed dispatch carries forward into the permanent record
- `commands_workspace.go` — `workspace/clean` captures stats before
deleting, new `workspace/stats` command + `agentic.workspace.stats`
action answer the spec's "query on the parent" use case
Co-Authored-By: Virgil <virgil@lethean.io>
Introduce an optional go-store persistence layer for the three state
groups described in RFC §15.3 — queue, concurrency, registry — plus
runtime_state and dispatch_history used by the sync pipeline.
- statestore.go lazily opens `.core/db.duckdb` via go-store when
available; nil-safe helpers return cleanly so in-memory/file-based
fallbacks survive when the store cannot open (graceful degradation,
RFC §15.6)
- prep.go tracks the store reference on the subsystem and closes it on
shutdown; hydrateWorkspaces now consults the registry group before
the filesystem scan so ghost agents are marked failed across
restarts, and TrackWorkspace mirrors updates back into the cache
- runtime_state.go persists backoff + fail-count snapshots into the
go-store runtime group so dispatch backoff survives restarts even
when the JSON file rotates
- commit.go writes the completed dispatch record into dispatch_history
for RFC §16.3 sync push to drain without rescanning workspaces
- statestore_test.go covers lazy-once init, restore/delete round trip,
ghost-agent failure marking, and runtime-state replay across
subsystem instances
Co-Authored-By: Virgil <virgil@lethean.io>
All RegisterTools and internal register*Tool methods updated from
*mcp.Server to *coremcp.Service. Tool registration calls updated to
use svc.Server() for SDK AddTool calls. Monitor subsystem updated
to store *coremcp.Service and access Server() for Sessions/ResourceUpdated.
Tests updated to create coremcp.Service via New() instead of raw SDK server.
Co-Authored-By: Virgil <virgil@lethean.io>
