SPB Git

Repository Public

INTEGRATION-TODO.md 3.4 KB

# Integration notes — what is still open

State at commit 45f0d03, 2026-11-18 (late evening). This file used to be the integration checkpoint; everything that was resolved has been folded into the reference documents (ARCHITECTURE.md, SECURITY.md, OPERATIONS.md, API.md, ACCEPTANCE.md). Only what is really still open remains here. Delete the file when it is empty.

Closed since the checkpoint (verified in the code): nested labeled values read by the API and the alert engine; agent overhead keys; pki.ca fingerprint key; UTC timestamps; relay upstream taken from MCC_INGEST_SIGNED_LISTEN; ops plane, web and deployment merged; Linux agent, install scripts and the signed transport under launchd exercised by the production rollout.

# Code

  1. TestJobRunsToSuccess fails fixed — (services/internal/jobs/runner_test.go:215): it still expects a target object in the job.run wire parameters, removed by commit 365d278. make ci fails at test until the assertion is updated.
  2. SSE payloads differ from the contract: node.state ({node_id, ts, values, availability}), action.output (batch of lines[]), audit.appended ({seq}), node.updated from ingest (reason instead of change), service.changed with state: "gone", resync ({}), plus the undeclared hello frame and the never-emitted model.changed. The web client reads the contract shapes for node.state and action.output. Details: API.md §5.3 and §8.
  3. Error codes differ from the contract (session_expired, webauthn_failed, refresh_reuse_detected, gone are never emitted); the web tests for webauthn_failed. API.md §8, row 7.
  4. Settings without effect: retention.raw_days, retention.1m_days, retention.1h_days (ingest reads MCC_RETAIN_* at start-up) and alerts.default_* (rule creation uses hard-coded defaults). retention.action_events_days is read by housekeeping but cannot be set through the API.
  5. Rate limits differ from the contract (one shared public bucket per IP instead of per-operation limits).
  6. Version stamping: make build stamps internal/api only; telemetry-ingest and action-orchestrator report 0.1.0-dev / dev.
  7. model.server.start / model.server.stop are catalogued as available but no agent implements them.
  8. go test -race has not been run on the ingest and agent packages.
  9. No automated test of control-api during a database outage; nothing tests an interrupted deployment.

# Not yet verified on real targets

  • Any authenticated flow in production: no owner account exists yet (the owner must open the one-time bootstrap link with his own passkey).
  • Real platform authenticators; real Safari and Firefox; real passkeys in the native apps.
  • service.start / stop / restart, cache.purge, agent.restart, job templates on a real pm2 / launchd (every agent runs the default read-only policy).
  • APNs (no key exists).
  • Restore on another node; key rotation; rollback of a release; moving the control plane.
  • Certificate renewal in production (first certificates issued on 2026-11-18, 30 days).

# Documentation follow-ups

  • Regenerate METRICS-CATALOG.md with the Linux column now that BHS64 reports.
  • DESIGN-BRIEF.md §1 still lists the development ports and a LAN-facing ingest listener; production uses 8800–8805 with ingest on loopback (ARCHITECTURE.md §2).
  • docs/APPLE-APPS.md is written separately.
Files main
Menu

Theme