Kodiai is an installable GitHub App that provides AI-powered PR auto-reviews and conversational code assistance via @kodiai mentions. It replaces the current approach of forking anthropics/claude-code-action and maintaining workflow YAML files in every repo — instead, repos just install the app and optionally drop a .kodiai.yml config file.
When a PR is opened or @kodiai is mentioned, the bot responds with accurate, actionable code feedback — inline review comments with suggestion blocks, or contextual answers to questions — without requiring any workflow setup in the target repo.
- ✓ Webhook server receives GitHub events and verifies signatures — v0.1
- ✓ GitHub App authenticates via JWT and mints installation tokens — v0.1
- ✓ Event router classifies webhooks and dispatches to handlers — v0.1
- ✓ Per-repo
.kodiai.ymlconfig loaded with sensible defaults (zero-config works) — v0.1 - ✓ PR auto-review on open/ready + manual re-request (
review_requested) — v0.1 - ✓ Inline review comments with suggestion blocks — v0.1
- ✓ Fork PR support works natively — v0.1
- ✓
@kodiaimention handling across issue/PR/review surfaces — v0.1 - ✓ Tracking comments show progress and update on completion/error — v0.1
- ✓ Content sanitization + TOCTOU protections — v0.1
- ✓ Bot ignores its own comments (no infinite loops) — v0.1
- ✓ Job queue with per-installation concurrency limits + ephemeral workspaces — v0.1
- ✓ Review UX improvements (eyes reaction,
<details>wrapping, conditional summaries) — v0.1 - ✓ Production deployment to Azure Container Apps (Docker + probes + secrets) — v0.1
- ✓ Review-request reliability hardening (delivery correlation, runbook, output idempotency) — v0.1
- Code modification via @mention (branch creation, commit, push) with guardrails
- Write-mode reliability polish (clearer failures, safer retries, reduced false positives)
- Expand operator evidence capture (delivery correlation, queue/failure metrics, grep-ready runbooks)
- Direct SDK agent loop for non-Claude LLMs — Phase 2+ after v1 is stable
- Bedrock / Vertex / API key auth backends — OAuth only for v1
- Public GitHub Marketplace listing — small group of known users for now
- Real-time streaming UI or dashboard — GitHub comments are the interface
- CI/CD pipeline automation — deployment is manual or separate
- GitHub App: Registered (App ID 2822869, slug
kodiai) - Production: Deployed on Azure Container Apps
- FQDN:
ca-kodiai.agreeableisland-d347f806.eastus.azurecontainerapps.io - Webhook:
https://ca-kodiai.agreeableisland-d347f806.eastus.azurecontainerapps.io/webhooks/github
- FQDN:
- Test repo:
kodiai/xbmc(public fork) used to validate PR review + mention flows - Core stack: Bun + Hono, Octokit, Agent SDK (
query()), in-process MCP servers, in-process queue (p-queue) - Execution model: clone workspace -> build prompt -> invoke Claude Code -> publish outputs via MCP tools
v0.1 ships an installable GitHub App that:
- Automatically reviews PRs (inline comments + suggestions; silent approvals for clean PRs)
- Responds to
@kodiaimentions across GitHub comment surfaces - Is deployable and running in production with observability and runbooks
- Runtime: Bun — used for both dev and production
- Framework: Hono for the HTTP server
- Auth: Claude Max OAuth token only for v1
- Deployment: Azure Container Apps (needs provisioning)
- Audience: Private use — small group of known users, not public marketplace
- Compute: Self-hosted (our Azure infra), not GitHub Actions runners
| Decision | Rationale | Outcome |
|---|---|---|
| Bun + Hono over Node + Express | Bun is fast, has native TypeScript support, Hono is lightweight and runs anywhere | ✓ Good |
| Claude Code CLI over direct API calls | Gets full Claude Code toolchain for free — file editing, MCP, tool use | ✓ Good |
| In-process p-queue over external queue | Simpler to start; can migrate to durable queue later if needed | ✓ Good |
| Shallow clone per job | Avoids large repo downloads; 50 commits gives enough diff context | ✓ Good |
.kodiai.yml config over env vars |
Per-repo customization without touching the app server | ✓ Good |
Last updated: 2026-02-09 after v0.1 milestone