review-fixer — token-minimized PR review-thread handling for OpenCode

review-fixer triages and resolves pull-request review threads from any reviewer and emits compact, TOKEN-MINIMIZED text so an LLM spends few tokens per call. It lists unresolved threads on a PR and, once you've fixed the code, posts each reply and resolves the thread — in pure TypeScript. No Go binary, no download.

It ships as:

• an OpenCode custom tool (review-fixer) — pure TS, registered by the plugin,
• a plugin (opencode-review-fixer) that registers the tool, and
• a skill (SKILL.md) that's auto-installed into your opencode config dir on postinstall (best-effort; copyable by hand if needed).

Layout

src/index.ts        plugin entry: registers the review-fixer custom tool
src/actions.ts      top-level wiring (token/repo/PR resolve → list/apply/verify)
src/github.ts       minimal GitHub GraphQL + REST client
src/resolve.ts      token / repo / PR discovery (flag → env → gh / git)
src/report.ts       pure rendering + filtering (byte-exact output contracts)
src/apply.ts        reply+resolve orchestration over a batch of items
src/deps.ts         side-effect seam (env / fetch / subprocess) for tests
install-skill.mjs   best-effort postinstall: copies SKILL.md into ~/.config
testdata/*.golden   committed golden text used by tests (source of truth)

Install

Enable the plugin in your opencode.json by adding it to the plugin array, then reference the published npm package:

{
  "plugin": {
    "opencode-review-fixer": "npm:opencode-review-fixer"
  }
}

bun install   # or npm install / pnpm install

The bundled SKILL.md is auto-copied to $XDG_CONFIG_HOME/opencode/skills/review-fixer/SKILL.md (default ~/.config/opencode/skills/review-fixer/SKILL.md) by postinstall. The copy is best-effort — if it fails (permissions, read-only mount, etc.) install still succeeds and a warning is printed to stderr. You can always copy SKILL.md by hand from the package if needed.

review-fixer tool args

| arg | type | actions | | -------- | ----------------------------------- | ------------------- | | action | enum (list/apply/verify) | (required) all | | pr | int | list, apply, verify | | repo | string (owner/name) | list, apply, verify | | author | string (substring filter on login) | list, apply, verify | | items | array of { threadId, body } | apply |

pr, repo, and author are auto-resolved from gh / git when omitted (gh auth token, gh repo view, current branch → gh pr list). author omitted means all reviewers.

Two-call workflow

1. { "action": "list" }                                  → see unresolved threads
2. { "action": "apply", "items": [...] }                 → reply + resolve all at once
3. { "action": "verify" }    (only if remaining > 0)     → confirm what's left

Example:

{
  "action": "apply",
  "items": [
    { "threadId": "PRRT_kwDOExample", "body": "Fixed: extracted the helper and added a nil check." }
  ]
}

Resolve a thread only after you've actually addressed (or rebutted) it.

Develop & test

bun install
cd plugins/review-fixer
bun test                  # all unit tests
bun run typecheck         # tsc --noEmit
GOLDEN_UPDATE=1 bun test  # REGENERATE testdata/*.golden from current output
bun test --coverage       # tests with coverage (target: 100%)

The test suite is table-driven with committed golden files for every text-producing function (renderList, renderVerify, renderApply, applyLine, end-to-end runAction). Run GOLDEN_UPDATE=1 bun test whenever you intentionally change a byte of output, then review the diff in testdata/*.golden.

Release

Published to npm via release-please on tag opencode-review-fixer-v* (Conventional Commits drive the version). No GoReleaser, no binaries — npm publish is the whole story.