Adds the session_checkpoints section to the control-plane DB substrate
architecture doc: schema version and row identity, the column groups, the
public API surface, and the hard rules (redaction at write, reconcile
rather than blind restore, unknown live state is not a mismatch, drain
fails closed on an incomplete checkpoint, no transcript storage).
Satisfies acceptance criterion 1 of #660 ("Schema documented and
versioned") for the implementation added in 7e18dcc.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
8.6 KiB
Control-plane DB substrate (#613)
Status: Implemented (SQLite single-writer MVP)
ADR: mcp-allocator-control-plane-observability-adr.md
Module: control_plane_db.py
Architecture statement
DB coordinates, Gitea records, Sentry/GlitchTip observe; the bridge is the only path that turns observations into Gitea work.
What this ships
| Capability | Notes |
|---|---|
| Schema | sessions, work_items, leases, assignments, terminal_locks, events, incident_links, session_checkpoints |
| Atomic assign+lease | ControlPlaneDB.assign_and_lease — one BEGIN IMMEDIATE transaction |
| Mutation gate | require_valid_assignment — live lease + allowed action + non-terminal work + non-stale head |
| Heartbeat / release / expire | Lease lifecycle helpers |
| Terminal-lock index | Routing signal for #600 (terminal path first) |
incident_links |
Provider-neutral link model for #612 — not assignable work; scope keys NULL-safe |
session_checkpoints |
Durable session resume state for #660 — redacted at write, reconciled (never restored) on boot |
Hard rules (enforced in code)
- Assignable
work_items.kind∈ {issue,pr} only — never raw Sentry/GlitchTip incidents. - Two concurrent sessions cannot both receive an active assignment on the same open work item (second gets
wait). - Merged/closed work items return
no_safe_workat assign time, andrequire_valid_assignmentfails closed if the work item becomes terminal later. - Assignments pin
expected_head_sha; mutations fail closed if the work item head drifts. - SQLite path is the single-writer MVP (
GITEA_CONTROL_PLANE_DB, default under~/.cache/gitea-tools/control-plane/). Multi-session multi-host production requires Postgres or a single allocator daemon (ADR §6). incident_linksoptional scope fields are stored as empty strings (never NULL) so UNIQUE is canonical across minimal upserts.- Legacy
incident_linksmigration collapses NULL-scope duplicates only when Gitea targets and all meaningful observation metadata agree (fingerprint, status, event_count, permalink, timestamps, linked PRs, etc.). Conflicting metadata fails closed — no silent discard.
Dependency chain
#613 control-plane DB (this) → #600 allocator API → #612 incident bridge
- #600 must call this substrate (not file locks / comment-only leases alone) for completion.
- #612 must write
incident_linkshere and create Gitea issues; the allocator assigns those issues, not raw incidents.
Allocator API (#600)
Module: allocator_service.py · MCP tool: gitea_allocate_next_work
- Workers call
gitea_allocate_next_work(apply=false|true)instead of self-selecting work. apply=falsereturns a dry-run selection (outcome=preview) with skip reasons.apply=truereserves viaControlPlaneDB.assign_and_lease(atomic assignment+lease).- Coordination source is always the control-plane DB — not file locks or comment-only leases.
- Routing follows ADR §5.3 (REQUEST_CHANGES → author, terminal path first, foreign lease → wait).
- Raw monitoring incidents are never candidates.
Lease lifecycle API (#601)
Module: lease_lifecycle.py · MCP tools: gitea_*_workflow_lease(s)
Active control-plane leases are first-class workflow state:
| Tool | Purpose |
|---|---|
gitea_list_workflow_leases |
List active (or all) leases for a repo |
gitea_inspect_workflow_lease |
Freshness + safe_next_action for one lease id |
gitea_adopt_workflow_lease |
Owner-resume or sanctioned reclaim with provenance |
gitea_release_workflow_lease |
Explicit owner release (audited) |
gitea_expire_workflow_leases |
Deterministic expire of past-expires_at leases |
gitea_abandon_workflow_lease |
Abandon with required proof |
gitea_reclaim_expired_workflow_lease |
Expire + re-assign with provenance |
Hard rules
- Control-plane DB is the coordination authority — file locks / comment-only leases are not authoritative alone.
- Active foreign leases cannot be stolen; inspect returns
wait_foreign_active. - Abandon requires
(dead_process OR missing_worktree)andno_live_mutation_risk; foreign abandon also needsoperator_authorizedor full dead+missing+no_open_pr proof. - Adopt provenance always records
adopted_from_session_id,adopted_by_session_id, work identity, optional head SHA, and worktree path. - Reviewer→merger comment handoff (
gitea_adopt_merger_pr_lease) is unchanged and remains the Gitea-thread durability path.
python3 -m pytest tests/test_lease_lifecycle.py tests/test_control_plane_db.py tests/test_allocator_service.py tests/test_merger_lease_adoption.py -q
Incident bridge (#612)
Module: incident_bridge.py · MCP tools: gitea_observability_*
- Bridge converts Sentry/GlitchTip observations → normal Gitea issues +
incident_links. - Phase-1:
gitea_observability_reconcile_incident(apply=false|true)with JSON observation. - Dry-run performs no Gitea mutation and no DB write.
- Apply reuses existing links or creates one Gitea issue; never invents
work_items.kind=incident. - Config:
GITEA_OBSERVABILITY_PROJECTS_JSONorGITEA_OBSERVABILITY_PROJECTS_FILE. - Provider tokens never appear in issue bodies, links, or tool results.
- Allocator sees bridge work only after a Gitea issue exists.
Session checkpoints (#660)
Module: control_plane_db.py · Table: session_checkpoints · Parent #655 · Soft-depends #659 · Vision #652 · Roadmap #653
Workflow state that lived only in MCP process memory or chat did not survive a restart, so session identity, stage, lease ownership, and the next valid action had to be reconstructed by hand. The session_checkpoints table makes that state durable.
Schema version
Rows carry checkpoint_schema_version (currently 5, tracking the module-level SCHEMA_VERSION), so a reader can tell which field set a record was written under. The row identity is
UNIQUE (remote, org, repo, session_id, work_kind, work_number) — one current-state row per session per work unit, upserted rather than appended. A session-level checkpoint that is not bound to an issue or PR uses the work_number = 0 sentinel with an empty work_kind.
| Group | Columns |
|---|---|
| Identity | session_id, provider_identity, role, remote/org/repo |
| Work unit | work_kind (issue/pr/empty), work_number, worktree_path, branch, head_sha |
| Ownership | capabilities (JSON), lease_id, assignment_id |
| Progress | workflow_stage, last_completed_action, current_operation, pending_mutation (JSON) |
| Recovery | evidence (JSON), blocker, next_valid_action, recovery_instructions |
| Bookkeeping | checkpoint_schema_version, status, created_at, updated_at |
API
| Method | Purpose |
|---|---|
write_session_checkpoint |
Upsert the current-state row; redacts, and optionally enforces drain completeness |
get_session_checkpoint |
Read one checkpoint by session + work unit |
list_session_checkpoints |
List checkpoints by session or by work unit |
reconcile_session_checkpoint |
Pure diagnosis of a stored checkpoint against live head/lease state |
checkpoint_completeness |
Pure — the drain-required fields missing from a record |
Hard rules
- Redaction at write. Free-text columns are redaction-filtered and JSON columns are redacted recursively before storage, so a pasted token or credential URL cannot land in a checkpoint. Secrets are never stored (#660 AC4).
- Reconcile, never blind-restore.
reconcile_session_checkpointreturns a diagnosis (stale,head_mismatch,lease_mismatch,reconcile_action) and the caller decides. A stored head that no longer matches live Git, or a lease that is gone or reassigned, marks the checkpoint stale (#660 AC3). - Unknown live state is not a mismatch. A
Nonelive input means "not checked" and never flags staleness on its own. - Drain fails closed. A write with
require_complete=Truerefuses when any ofsession_id,role,workflow_stage,next_valid_action,recovery_instructionsis missing or blank, so drain cannot complete on a checkpoint that could not resume. - No transcript storage. Checkpoints hold resume state, not conversation history.
Non-goals (intentionally deferred)
- Full unsupervised watchdog auto-filing (prefer explicit reconcile first)
- Assuming GlitchTip writeback API equals Sentry without verification
- Gitea comment/label mirror writers for every assignment (optional later)
Tests
python3 -m pytest tests/test_control_plane_db.py -q