v0.10 — Vexa goes cloud-native

The biggest refactor ever. v0.9 was a bot that joined meetings. v0.10 is a cloud-native agent platform where meeting bots are one workload type among many. From single-container all-in-one → first-class Kubernetes, scalable horizontally, with a general-purpose agent runtime on top.

723 commits · 1014 files changed · +142k / −44k lines.

The new architecture

Two-layer model: runtime + applications

v0.10 introduces a clean two-layer split that didn't exist before:

                 ┌──────────────┐  ┌────────────┐
  applications   │  meeting-api │  │ agent-api  │   … future workloads
                 └──────┬───────┘  └─────┬──────┘
                        │                │
                 ┌──────▼────────────────▼──────┐
  orchestration  │          runtime-api          │
                 │  process | docker | kubernetes │
                 └────────────────────────────────┘

runtime-api is the new lower-level primitive. Pluggable orchestrator backends: process (local dev), docker (compose / single-VM), kubernetes (LKE / any K8s). One scheduler, many workloads. Both meeting bots AND AI agents run on top of it.

meeting-api (application) — meeting bot orchestration, transcription aggregation, webhooks, recording pipeline. Asks runtime-api to spawn bot containers; owns meeting state.

agent-api (application, net-new) — general-purpose agent platform. Runs Claude Code / Codex / any AI agent sandboxed in containers, scaled by runtime-api. Workspace persistence, container isolation. Use cases span real-time meeting support, post-meeting AI processing, and anything else you'd want a containerized agent to do — not limited to meeting intelligence. This is the foundation for Vexa as an agent platform.

Browser sessions on demand

New browser-session profile on runtime-api: on-demand Chromium containers with Playwright-connected CDP (Chrome DevTools Protocol). Any agent can request a browser, drive it programmatically, and the runtime tears it down when done. Powers interactive bots (authenticated login flows, web automation from agents) that couldn't exist before.

WhisperLive retired

The separate whisperlive service is gone. Real-time audio buffering moved into vexa-bot directly. One less hop, one less service to deploy, simpler pod topology on K8s. The transcription backend (remote Fireworks / local faster-whisper) is now selected via TRANSCRIPTION_SERVICE_URL from inside the bot container.

Kubernetes / Helm — first class

Previously helm was "experimental". v0.10 makes it the production-grade path:

deploy/helm/charts/vexa/ — full chart, 30+ templates, per-service deployments, StatefulSets for postgres/minio, PVCs, RBAC, ingress, PDB, priority class.
Production-tuned values.yaml — explicit requests + limits (cpu + memory) on every service; global hardened securityContext; Redis memory cap + eviction; DB pool sized; PodDisruptionBudget templates (off by default, toggle per service).
imagePullPolicy: Always, imagePullSecrets, external-postgres opt-in (postgres.enabled: false + credentialsSecretName).
Upgrade path from existing dev clusters: helm upgrade vexa deploy/helm/charts/vexa -f values-staging.yaml.

Dashboard in-tree

services/dashboard/ — Next.js web UI brought into the monorepo. Magic-link login, meeting CRUD, bot control, live WebSocket updates, webhook config, recording playback, transcript search.

More new services

calendar-service — Google / Outlook calendar integrations.
telegram-bot — Telegram interface to Vexa.
vexa-agent — agent runtime.

New top-level structure

deploy/ — {compose, lite, helm} (replaces old docker/* tree + root docker-compose.yml + alembic.ini).
features/<name>/ — prose README + machine-readable dods.yaml per feature (14 features tracked today).
packages/ — OSS npm packages: vexa-client, vexa-cli, transcript-rendering.
libs/ — shared Python libs.
tests3/ — release protocol (see below).
CLAUDE.md, VERSION — AI-operator entry point + version pin.

Release protocol (new primitive)

Because v0.10 is this large, the release process got an overhaul too. The release cycle is now a strict stage state machine — every make release-* target and every AI skill guards on the current stage; illegal transitions hard-fail:

idle → groom → plan → develop → provision → deploy → validate ⇄ triage → human → ship → teardown

Five primitives documented in tests3/README.md:

Scope (releases/<id>/scope.yaml) — per-release contract declaring issues + proves[] bindings into the registry.
README-as-contract — every features/<name>/README.md references a dods.yaml sidecar (machine-readable DoDs).
Registry (tests3/registry.yaml) — one source of truth for checks (consolidated from two prior files).
Fresh-infra lifecycle — every cycle: provision from zero → validate → teardown. No shared staging that decays.
Stage state machine — tests3/.current-stage + .state/stage-log.ndjson + lib/stage.py; AI skills load per-stage prompts from tests3/stages/*.md.

Cycles shipped under the new protocol:

260417-webhooks-dbpool — webhook gateway injection + meeting-api DB pool fix.
260418-webhooks — non-meeting.completed status webhooks actually fire when opted-in.
260419-helm — helm chart production-tuning + LKE cluster correctly sized for bot workload.

Webhooks

Gateway injects validated X-User-Webhook-* into meeting.data on POST /bots; client-supplied headers stripped first (spoof-resistant).
Fixed: meeting.completed double-fire on status path.
Fixed: stop_requested gate no longer silences non-terminal status webhooks.
bot.failed, meeting.started, meeting.status_change now actually deliver when opted-in via webhook_events (prior release's green gate was false-positive).
Dashboard /webhooks page shows all event types.

Reliability

Meeting-api DB pool exhaustion fixed (leaked AsyncSession transactions); pool sized 20/20/10s.
idle_in_transaction_session_timeout: 60s guard on postgres.
Reconciliation scheduler known-bug (kills prod bots on container-based deployments) disabled at runtime, pending topology-aware rewrite next cycle.

Breaking / operator-visible

Layout: docker/* → deploy/*. Compose users: make -C deploy/compose all. Lite: make -C deploy/lite all. Helm: helm install vexa ./deploy/helm/charts/vexa.
Services gone: bot-manager (split into meeting-api + runtime-api), transcription-collector (absorbed into meeting-api), whisperlive (folded into vexa-bot).
API surface: public routes unchanged on meeting-api. runtime-api is internal-only. /transcripts/* now served by meeting-api.
Env: .env layout moved; see deploy/env-example.

Upgrading

# Full stack via Docker Compose
make -C deploy/compose all

# Single container trial (Vexa Lite)
make -C deploy/lite all

# Production Kubernetes (Linode LKE / any K8s cluster)
helm install vexa deploy/helm/charts/vexa \
  -f deploy/helm/charts/vexa/values-staging.yaml

DockerHub :latest now points to 0.10.0-260419-1140.

Known gaps (scoped for next cycle)

HELM_VALUES_RESOURCES_SET static check is grep-single-pattern; tighten to per-service verification.
release-build doesn't propagate fresh image tag into .env automatically (stale-image risk on helm upgrade).
aggregate.py reads all .state-<mode>/reports/ regardless of scope.deployments.modes — stale out-of-scope reports can contaminate gate math on scope-narrowed cycles.
values-staging.yaml retains hosted-environment identifiers; candidate for redaction or move to -example.yaml.
Reconciliation scheduler is topology-unaware; rewrite to dispatch per orchestrator backend.

Credits

Shipped via the new release protocol. Full audit trail under tests3/releases/260417-webhooks-dbpool/, tests3/releases/260418-webhooks/, tests3/releases/260419-helm/ — each cycle's scope.yaml, plan-approval.yaml, triage-log.md, code-review.md, human-approval.yaml.

Vexa-ai/vexa v0.10 v0.10 — Vexa goes cloud-native · agent platform · K8s first-class on GitHub