feat(ci): replace opencode review with PR-Agent
PR Check / Security scan (Snyk) (pull_request) Skipped
PR Check / Lint + typecheck (pull_request) Canceled after 45s
PR Check / Secret scan (gitleaks) (pull_request) Canceled after 0s
PR Check / Dependency audit (pull_request) Canceled after 0s
PR Check / Dockerfile lint (droast) (pull_request) Canceled after 0s

Run PR-Agent as a pinned webhook server on the Gitea Docker network and review on PR open plus slash commands. The per-PR opencode installer and verdict-posting step are removed.

The server-side .env stays outside the repository; deploy/pr-agent/ holds the compose definition and operator notes.
This commit is contained in:
Noel Sabia committed 2026-10-30 21:00:53 +01:00
1 parent 9ecf504843
commit c626cc9c1d
11 files changed
+251 -236

No files matched your search

+5 -1
View File
@@ -3,7 +3,7 @@ type: Decision
title: AI code review setup
description: The review job runs opencode with a cheap medium-effort opencode-go model, env-only auth, a read-only permission set, and posts only reviews that end with a verdict line.
tags: [decision, ci, ai, opencode, security]
status: draft
status: deprecated
generated: { by: opencode-go/glm-5.3-flash, at: 2026-10-18T16:05:00Z }
verified: { by: human:noel, at: not yet }
sources:
@@ -15,6 +15,10 @@ sources:
title: OpenCode Go plan documentation
---
> Superseded by [PR-Agent for Gitea reviews](pr-agent.md) on 2026-10-30.
> The opencode job was removed from `pr-checker.yml`; this document remains
> as the record of the earlier setup and the failures it fixed.
# Decision
The review job installs opencode v2 from the dedicated v2 installer,
+2 -1
View File
@@ -1,7 +1,8 @@
# Decisions
* [PR pipeline layering](pr-pipeline-layering.md) - What runs pre-commit, what runs in Gitea Actions, what the AI review does, and the cost rule that decides.
* [AI review setup](ai-review-setup.md) - Plan model, env-only auth against prompt injection, read-only agent, and the verdict line as the posting gate.
* [PR-Agent for Gitea reviews](pr-agent.md) - A pinned webhook server replaces the per-PR opencode job; internal-only, scoped bot, OpenCode Go model.
* [AI review setup (opencode)](ai-review-setup.md) - Superseded 2026-10-30; kept for the env-only auth and verdict-gate reasoning.
* [Vulnerability audit policy](vulnerability-audit-policy.md) - Blocking audit, reviewed ignore-list for unfixable advisories, and the pnpm-11 settings home.
* [Release engineering](release-engineering.md) - release-please in manifest mode, conventional commits as input, message format enforced at commit time.
* [Snyk paused and secrets policy](snyk-paused-and-secrets-policy.md) - Why the job is skipped rather than deleted, and the token scopes every workflow secret needs.
+106
View File
@@ -0,0 +1,106 @@
---
type: Decision
title: PR-Agent for Gitea reviews
description: A pinned PR-Agent webhook server on server0 reviews pull requests through a scoped Gitea bot user, replacing the per-PR opencode job.
tags: [decision, ci, ai, gitea, pr-agent, security]
status: draft
generated: { by: opencode-go/deepseek-v4.1-flash, at: 2026-10-30T20:00:00Z }
verified: { by: human:noel, at: not yet }
sources:
- id: compose
resource: deploy/pr-agent/docker-compose.yml
title: PR-Agent compose service
- id: env-example
resource: deploy/pr-agent/.env.example
title: Environment variables, without secrets
- id: pr-checker
resource: .gitea/workflows/pr-checker.yml
title: PR Check workflow, opencode job removed
- id: install
resource: https://docs.pr-agent.ai/installation/gitea/
title: PR-Agent Gitea installation
- id: images
resource: https://hub.docker.com/r/pragent/pr-agent
title: Published gitea_app image tags
- id: gitea-egress
resource: https://docs.gitea.com/administration/config-cheat-sheet#webhook
title: Gitea webhook egress allowlist
---
# Decision
PR-Agent runs as `pragent/pr-agent:0.49.0-gitea_app` in a long-lived
container on `server0`, attached to the existing `gitea_gitea` Docker
network. It publishes no ports. Gitea delivers pull-request and comment
events to `http://pr-agent:3000/api/v1/gitea_webhooks`; the bot user
`pr-agent` posts the result with a scoped personal access token.
The opencode review job is removed from `pr-checker.yml`. PR-Agent runs
`/describe` and `/review` on pull request open/reopen and answers slash
commands in the conversation. Push-triggered re-review is disabled, so a
PR is reviewed once when it opens, not on every push.
The model is OpenCode Go's OpenAI-compatible endpoint:
```text
model openai/glm-5.3-flash
fallback openai/glm-5.3
api base https://opencode.ai/inference/go/openai/v1
```
`CONFIG__CUSTOM_MODEL_MAX_TOKENS=128000` supplies the context limit the
model table does not know; PR-Agent's default `config.max_model_tokens`
still caps the review budget at 32000.
# Why
* One container replaces a per-PR CLI install, binary cache and prompt
wrapper. The review path is then the same code for every PR.
* PR-Agent owns the Gitea webhook and comment plumbing, including the slash
commands (`/review`, `/improve`, `/ask`) the repository previously had no
way to trigger.
* The webhook endpoint stays on the Docker network. Nothing new is exposed
to the internet, and comment links still use the public Gitea URL.
* The image tag is pinned. A new PR-Agent release is a deliberate change,
not a rollout underneath the repository.
# Gitea configuration
* Bot user `pr-agent` is a collaborator with write permission on
`CV_Website/CV_Site_Nuxt`. The token scopes are `read:repository`,
`write:issue` and `read:user`. Gitea has no comment-only collaborator
role, so write is the narrowest role that can post comments and labels.
* `[webhook] ALLOWED_HOST_LIST = private:3000` is required: Gitea's default
lax egress policy allows public targets but refuses private ones, and the
webhook target is a Docker-network address. The API is reached at
`http://gitea:3000`; `GITEA__WEB_URL=https://git.noelsabia.dev` keeps
published links on the public host.
* The webhook sends `push`, `issues`, `pull_request` and `issue_comment`
events. Only the last two change behaviour today; `push` is kept because
PR-Agent may handle push commands in a later configuration.
* Secrets live only in `/home/noel/self-hosted/pr-agent/.env` (mode 600) on
`server0`. The repository contains the compose file and an example env
file, never the token or webhook secret.
# Not verified
* The OpenCode Go API key was not present when this was committed, so no
model call and no real review have run. The first PR after the key is set
is the verification.
* Whether `glm-5.3-flash` handles PR-Agent's structured review output well
enough. `glm-5.3` is configured as the fallback.
* Whether two gunicorn workers and the 1 GB memory limit hold under
concurrent reviews. The host has roughly 2.5 GB available with Gitea and
its runner idle.
* Whether the default review prompt is specific enough for this repository.
A `.pr_agent.toml` with project instructions is deliberately absent until
the first real review shows what it misses.
# Revisit when
* The first review fails, returns unstructured output, or misses the kind
of finding the opencode prompt used to catch.
* A newer `gitea_app` image tag is worth taking; bump the tag and run
`docker compose up -d`.
* Review volume makes the container's memory limit tight, or Gitea adds a
role that can comment without push access.
+11 -6
View File
@@ -26,8 +26,8 @@ Every pull request is checked by three layers, ordered by cost:
in one job (one install, cached pnpm store), gitleaks over full
history, pnpm audit (blocking), droast Dockerfile lint, Snyk
(currently disabled).
3. **AI code review** (per PR push): opencode with a plan-covered model,
verdict-based, posted as a PR comment.
3. **AI code review** (on PR open, or on a slash command): the PR-Agent
webhook server, posted as a PR comment.
The split rule: anything that finishes in seconds and auto-fixes goes in
pre-commit; anything slow, network-bound or quota-bound goes in CI; anything
@@ -45,15 +45,16 @@ already passed the deterministic gates.
# Caching
Two caches, both served by the runner's own cache server (no external
service, entries stay on the home server):
One cache, served by the runner's own cache server (no external service,
entries stay on the home server):
* **pnpm store**, keyed on the store layout version (v11) and a `sha256sum`
of `pnpm-lock.yaml`. Installs hardlink from the store instead of
downloading tarballs again. The key is built in a shell step because
`hashFiles()` evaluates empty in this act runner.
* **opencode binary**, keyed on the pinned CLI version (`~/.opencode/bin`).
Small next to the model call, but it removes a network fetch per review.
The opencode binary cache is gone with the opencode job. PR-Agent is a
long-running container on `server0`, so a review no longer downloads a CLI.
Lint and typecheck also share one job: same image, same lockfile, same
install. Splitting them duplicated the install and made both compete for
@@ -71,6 +72,10 @@ exposure in the review setup — each fixed where it was cheapest.
# Failure modes deliberately designed against
The opencode-specific rows are historical since the review moved to
[PR-Agent](pr-agent.md); they are kept because the failure classes still
apply to any review implementation.
| Failure seen | Guard now in place |
|---|---|
| Error logs posted as review comments | Posting gate is the output contract (a `VERDICT:` line), not the exit code |