vibepod-agents

vibepod-agents builds and publishes the VibePod agent images using a root compose.yml.

Proxy image publishing is intentionally not part of this repository anymore.

Repository Layout

compose.yml
docker/
  claude/
  gemini/
  opencode/
  devstral/
  auggie/
  copilot/
  codex/
  pi/
  agy/
  tau/
  jcode/
.github/workflows/
  build.yml
  auto-release.yml
automation/
  agent-versions.json
scripts/
  bootstrap-wiki.mjs
  check-agent-updates.mjs
  apply-agent-updates.mjs
  smoke-agent-images.mjs
  lib/
    catalog.mjs

Image Targets

Target	Docker Hub	Version Source
claude	vibepod/claude	github:anthropics/claude-code
gemini	vibepod/gemini	npm:@google/gemini-cli
opencode	vibepod/opencode	npm:opencode-ai
devstral	vibepod/devstral	github:mistralai/mistral-vibe
auggie	vibepod/auggie	npm:@augmentcode/auggie
copilot	vibepod/copilot	npm:@github/copilot
codex	vibepod/codex	npm:@openai/codex
pi	vibepod/pi	npm:@earendil-works/pi-coding-agent
agy	vibepod/agy	github:google-antigravity/antigravity-cli
tau	vibepod/tau	pypi:tau-ai
jcode	vibepod/jcode	github:1jehuang/jcode

Defaults:

• AGENT_NAMESPACE=vibepod
• IMAGE_TAG=latest
• CLAUDE_CODE_CHANNEL=latest

Local Usage

cd /workspace/vibepod-agents
cp .env.example .env

docker compose version
docker compose config --services

Build and push all targets:

IMAGE_TAG=0.3.0 docker compose build
IMAGE_TAG=0.3.0 docker compose push

Build and push selected targets:

IMAGE_TAG=0.3.0 docker compose build claude gemini codex pi agy
IMAGE_TAG=0.3.0 docker compose push claude gemini codex pi agy

Smoke Tests

Run smoke tests against already-published images:

node scripts/smoke-agent-images.mjs --pull

Build local images and smoke test them:

IMAGE_TAG=ci-smoke node scripts/smoke-agent-images.mjs --build

Run one or more selected targets:

node scripts/smoke-agent-images.mjs --targets claude,codex --pull
IMAGE_TAG=ci-smoke node scripts/smoke-agent-images.mjs --targets pi,agy --build

Each target runs a simple argument-forwarding command through the image and the catalog version_command. If Docker is not installed or the daemon is not available, the script prints a SKIP message and exits successfully.

Claude image note:

• docker/claude follows the official Claude Code setup via the native installer (https://claude.ai/install.sh).
• Use CLAUDE_CODE_CHANNEL=stable (or another supported channel) when building if needed.

CI Pipelines

• build.yml: validates every Docker context on PRs and pushes to main
  • multi-arch builds: linux/amd64, linux/arm64
  • devstral and agy are intentionally linux/amd64 only
  • each matrix target also builds a local image and runs the smoke test script
• auto-release.yml: checks upstream CLI versions every 6 hours (and on manual dispatch)
  • publishes to Docker Hub namespace vibepod (https://hub.docker.com/u/vibepod)
  • uses agents.json as the source of truth for image definitions and build configuration
  • uses the repository wiki as the source-of-truth release state store
  • canonical state file in wiki repo: automation/agent-versions.json
  • generated tracking page in wiki repo: Container-Versions.md
  • tracking page URL: https://github.com/VibePod/vibepod-agents/wiki/Container-Versions
  • CI is update-only (fails if wiki state/page are missing)
  • creates a new image only when tracked upstream version changes
  • manual dispatch supports force_recreate_latest=true to rebuild unchanged agents and publish a new dated tag plus latest
  • after fixing a stale image whose tracked version is already current, manually dispatch with force_recreate_latest=true so the corrected build configuration is used
  • publishes <YYYY.MM.N> and latest for changed containers only
  • updates wiki state after a successful publish

Versioning Rules

• Automated image tags follow YYYY.MM.N (UTC).
• N starts at 1 for each container at the first release in a new month.
• N increments by 1 when that container's wrapped CLI version changes again in the same month.
• Every automated publish also updates the latest tag.

Manual Updates And Catalog

agents.json is authoritative for image definitions and build configuration. The wiki state file automation/agent-versions.json is authoritative for tracked agent versions and release history. Auto-release merges those sources by target before planning builds.

• One-time local bootstrap of wiki state from Docker Hub:

git clone https://github.com/VibePod/vibepod-agents.wiki.git wiki
node scripts/bootstrap-wiki.mjs --wiki wiki --all-tags

• fetches all tags from Docker Hub, pulls each image, and runs --version to resolve agent versions
• --all-tags inspects every historical tag (not just the latest); omit it to only inspect the most recent tag per agent
• --skip-inspect skips Docker image inspection entirely (agent versions recorded as "unknown")
• Commit and push the initialized wiki state manually:

cd wiki
git add automation/agent-versions.json Container-Versions.md
git commit -m "chore: initialize agent version state"
git push

• For fully automated containers (automation.enabled=true), the workflow checks upstream sources (npm packages, PyPI projects and GitHub releases) and publishes only on version change.
• Current GitHub-release tracked containers:
  • claude -> anthropics/claude-code
  • devstral -> mistralai/mistral-vibe
  • agy -> google-antigravity/antigravity-cli
  • jcode -> 1jehuang/jcode
• Current PyPI tracked containers:
  • tau -> tau-ai
• If you publish manually for Dockerfile/image-only changes, update tracked.agent_version, tracked.image_tag, and release_history in the wiki catalog so automation stays in sync.
• After manual state changes, re-run bootstrap to regenerate the tracking page (in wiki clone):

node scripts/bootstrap-wiki.mjs --wiki wiki --skip-inspect