Skip to content

SESSION_STATUS.md

Latest commit

 

History

History
119 lines (95 loc) · 5.45 KB

SESSION_STATUS.md

File metadata and controls

119 lines (95 loc) · 5.45 KB

Session Status

Last updated: 2026-03-18 (v0.9.11)

What's Working

Infrastructure

  • Postgres data plane — migrated from SQLite+PVC to shared Postgres (db-custom-1-3840, 200 max connections)
  • GCS blob storage — segments stored inline in Postgres + GCS (preplanai-kailab-blobs), range reads for fast file access. Inline is safety net; GCS is best-effort.
  • Zero-downtime deploys — 2 replicas with RollingUpdate (maxUnavailable=0), health probes
  • GitLab CI/CD — push to gitlab.com:rite-day/kaylayerhq/kai-server triggers test → build → deploy pipeline
  • Cache headers — no-cache on HTML (prevents stale JS chunk 404s after deploy), immutable on hashed assets

CLI (v0.9.11)

  • kai capture -m "message" — attach message to snapshot, shown as CI run headline
  • kai push — sends git HEAD commit message via X-Kailab-Message header
  • kai push --force — skips negotiate, re-sends all objects (data recovery)
  • kai review open/approve/close/comment — full review lifecycle
  • kai fetch --review <id> — syncs review + comments from server to local
  • kai ci runs/run/logs/cancel/rerun — remote CI management
  • kai ws create/checkout/list — workspace management
  • kai mcp serve — MCP server for AI coding assistants (12 tools)

Web UI (kaicontext.com)

  • File view — IDE-style split panel (tree + content), file search/filter, type-specific icons, language breakdown bar, keyboard navigation, breadcrumb, doc tabs (README/CONTRIBUTING/LICENSE/SECURITY)
  • CI runs — commit messages as headlines, SSE live updates, auto-scroll logs, 30min default timeouts, pod GC for stale jobs
  • Reviews — create/view/approve/merge/abandon, inline line commenting, semantic + line diff toggle, threaded comments, relative timestamps
  • Header — logo mark, refined spacing, soft shadow, desaturated avatar
  • History — batched changeset fetches (5 concurrent, max 20)
  • README links — SPA navigation for internal links

Email Notifications (Postmark)

  • CI pipeline results → snapshot author
  • Review comments → review author
  • Review state changes (approved, merged, abandoned, changes requested)
  • @mention notifications
  • Org invitation/removal

Review System (end-to-end)

  • CLI: kai review open → creates review targeting changeset
  • CLI: kai push → sends review + changeset to server
  • Web: view review, see semantic diff, add inline comments
  • CLI: kai fetch --review → syncs comments back to local
  • Web: approve/merge → updates snap.main to changeset head
  • Email: notifications on all state changes

Known Issues

Edge Accumulation (partially fixed)

  • Root cause: SQLite treats NULL as unique in PRIMARY KEY, so INSERT OR IGNORE on (src, type, dst, at) with at=NULL never ignored duplicates
  • Fix applied: Changed PK to (src, type, dst), auto-migration deduplicates on first run, DISTINCT on queries
  • Status: HAS_FILE edges stable at 197. DEFINES_IN/IMPORTS edges still accumulate on re-capture because AnalyzeSymbols creates new edges each time. The skipAnalysis optimization helps when snapshot ID is unchanged.
  • Files: kai-cli/internal/graph/graph.go, kai-cli/schema/0001_init.sql, kai-cli/internal/snapshot/snapshot.go

Capture Performance

  • Full capture (197 files): ~4 seconds (first run), ~7 seconds (subsequent due to symbol edge accumulation)
  • skipAnalysis triggers when snapshot ID is byte-identical to snap.latest (no file changes)
  • Incremental per-file analysis not yet implemented — all files re-parsed even when only one changed

Push Performance

  • Sends ALL edges every push (~20K), not just new ones
  • Edge negotiate/dedup not implemented on the push protocol
  • Segments accumulate in Postgres (408MB for kai repo across 28 segments)

CI

  • Stale job pods can exhaust namespace quota — mitigated by GC (deletes pods >30min old every 5min)
  • Runner image built separately from main pipeline (Dockerfile.runner)

Architecture

kai-cli (local)
  ├── .kai/db.sqlite     — local graph (nodes, edges, refs)
  ├── .kai/objects/       — content-addressed file blobs
  └── .kai/message        — capture message (consumed by push)

kai push → kailab (data plane)
  ├── Postgres            — segments, objects, refs, edges, ref_history
  ├── GCS                 — segment blobs (range reads)
  └── notifyPushCI →

kailab-control (control plane)
  ├── Postgres (shared)   — users, orgs, repos, workflows, CI runs
  ├── SvelteKit frontend  — embedded in Go binary
  ├── Postmark            — email notifications
  └── kailab-runner (CI)  — K8s pods per job, 3 runner replicas

Deployment

kai-server (GitLab)

git push  # triggers: test → build → deploy

kai CLI (GitHub)

# Build and install locally
cd kai-cli && go build -o kai-darwin-arm64 ./cmd/kai
cp kai-darwin-arm64 ~/go/bin/kai

# Push to kai (captures + pushes semantic graph)
kai capture -m "message" && kai push

Manual operations

# Check CI pods
kubectl get pods -n kailab-ci

# Clean stale CI pods
kubectl delete pod -n kailab-ci -l app=kailab-ci --field-selector=status.phase!=Running

# Check Postgres connections
kubectl exec -n kailab deployment/kailab -- sh -c 'apk add --no-cache postgresql-client > /dev/null 2>&1; PGPASSWORD="..." psql "postgres://..." -c "SELECT count(*) FROM pg_stat_activity;"'

# GCS bucket
gsutil ls -lh gs://preplanai-kailab-blobs/