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.
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.
Superseded by PR-Agent for Gitea reviews on 2026-10-30.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,
authenticates with OPENCODE_API_KEY from the environment only, runs
opencode-go/claude-haiku-5-5#medium (covered by the Go plan subscription,
no per-token cost), denies edit/shell/webfetch/websearch via a
global permission config written outside the checkout, and posts the
review as a PR comment only if the output ends with a VERDICT: line.
Why each piece
Piece
Why
Provider prefix opencode-go, not opencode
opencode is the pay-per-call Console gateway; the Go plan lives under its own provider id. The wrong prefix failed every run with "Insufficient account funds".
Model claude-haiku-5-5#medium
Cheap per token and cheap in the usage window, with a real #medium reasoning level. deepseek-v4.1-flash only exposes low/high/max, so its reviews always paid for the maximum effort.
The v2 installer (/v2/install)
The classic installer ships OpenCode 1.x (v1.19.3 on arm64), which cannot resolve model variants and failed with ProviderModelNotFoundError for #max.
Env-only auth, no opencode.json in the checkout
A config file in the workspace holds the API key in plaintext where the review agent can read it; a malicious PR diff could prompt-inject the agent into quoting the key in the publicly visible review comment. Environment variables are invisible to file reads.
Deny write tools via a global config outside the checkout
Headless runs cannot answer permission prompts; denials prevent both hangs and repo modification. The config file itself holds no secret, so its location leaks nothing.
Verdict line as the posting gate
opencode can exit non-zero after a fully completed review (observed: benign permission denials). A completed review always ends with a verdict line; error dumps never do. The verdict, not the exit code, decides whether a comment is posted.
--print-logs
Provider and auth errors otherwise die inside the captured output file, invisible in the job log.
Diff passed as a file, lockfile excluded
The diff as a CLI argument exceeded MAX_ARG_STRLEN (128 KiB per argument; the diff was 190 KB, mostly pnpm-lock.yaml). The agent reads .pr-review-diff.patch with its file tools.
The review prompt
The prompt names the stack (Nuxt 4 frontend, Go backends to come), lists
what the other CI jobs already catch so the model does not repeat them,
and focuses on SSR/hydration pitfalls, reactivity bugs, injection and
authz gaps, runtime-config leaks, and Go error/concurrency hygiene. It
prescribes the output format ([BLOCKER]/[WARN]/[NIT] file:line) and
derives the verdict from the findings, so the gate and the prompt agree.
Evidence
First real review on PR #5 (2026-10-18): produced five findings,
including a genuine regression in the workflow itself (a lost
chmod +x on the droast binary), verified upstream release assets and
the v2 permission schema live, and returned VERDICT: REQUEST_CHANGES.
Fresh-container replica test (node:22-bookworm, key-only auth)
reproduced the CI failure exactly, then confirmed the v2 + #max fix
before it was pushed.
Model history: deepseek-v4-pro#max was replaced by
deepseek-v4.1-flash#max after two consecutive endpoint outages, then
by claude-haiku-5-5#medium (2026-10-20) to stop paying maximum
reasoning effort on every review. The flash model had no #medium
variant to fall back to.
Not verified
Whether #medium still resolves after a future CLI upgrade — the
variant list comes from the model catalog, not a contract.
Whether the smaller model still catches the same class of findings as
the first flash review. The PR #14 findings are the baseline; re-run
the eval on the next few PRs.
Long-run behaviour of Go plan usage windows under real review
frequency, now that the model is cheaper.
Revisit when
A PR review fails again with a new error class — check the provider
prefix, the CLI version, and the usage limits first; they caused every
failure so far.
A flash-tier model gains a #medium variant and beats the current
model on the PR #14 eval set — the present pick is a cost/quality
compromise, not a permanent one.
Go Plus or pay-as-you-go balance becomes relevant.