@madsoftwaredev/opencode-lint-feedback

opencode plugin add @madsoftwaredev/opencode-lint-feedback

lint-feedback appends project-local ESLint diagnostics to successful native write, edit, and patch results. It runs checks only, after native formatting; it does not fix files, install packages, supply rules, or run repository-wide lint or typechecks. Use the existing /lint command for safe corrections and re-checks.

OpenCode V2 does not run language servers or produce LSP diagnostics. Automatic formatting is a separate feature and does not prove a file is lint-clean. The global formatter settings use project-aware built-ins (including JavaScript/ TypeScript and Dart) with the existing bundled RuboCop override retained.

Requirements and selection

• Node on the native shell's PATH, project-installed ESLint 9 or 10, and a flat eslint.config.{js,mjs,cjs,ts,mts,cts} file. TS configs/parsers still need their project's normal dependencies. There are no production dependencies to install for this plugin itself.
• Checks cover changed .js, .jsx, .mjs, .cjs, .ts, .tsx, .mts, and .cts files. Native structured results identify additions, modifications, and move destinations. Deleted files, failed edits, unrelated tools, and filenames appearing only in patch bodies or result text never trigger a check. Unsupported native result shapes skip rather than guessing.
• The root is the current session directory, not the service cwd or the plugin instance's directory. Discovery walks upward from each edited file, stopping at that root. Nested packages use their nearest flat config and installed ESLint; hoisted dependencies inside the root work. Run a monorepo session from its root if nested packages need shared configuration/dependencies above their directory.
• Only regular files are linted. Session-root aliases and internal symlinks work; edited files, configs, or ESLint installs resolving outside the root are skipped. No global ESLint, npx, package downloads, arbitrary repository globs, or fallback rules are used. Flat config is explicit and the ESLint cwd is its directory, so config patterns keep their usual base without searching above the root.

Unsupported projects are skipped, not passed. These include legacy .eslintrc configurations, ESLint versions outside 9/10, ESLINT_USE_FLAT_CONFIG=false, Yarn PnP without node_modules, dependencies symlinked outside the root, and remote workspaces. Use /lint with the project's own command in those cases. Custom config names, repository scripts with special flags, Biome diagnostics, and full TypeScript/compiler checks also remain explicit verification tasks.

Permissions, environment, and limits

The V2 tool.transform wrapper first awaits the original edit executor. It then invokes the registered native shell tool executor, passing the original session, agent, message, call ID, and progress context unchanged. It does not use ctx.shell.create, spawn a process, or load project JavaScript in the hook. Before calling shell it only canonicalizes the session directory and approved edited file paths; all config/package inspection happens in check.mjs under shell authorization. Missing shell support has no direct-process fallback.

The authorized command is node /path/to/check.mjs /session/root /changed/file .... Normal shell rules apply to that command; permissions are never widened or auto-approved. Denial or environment failures leave the successful edit intact and append an incomplete-check message. The native shell executor performs external_directory and shell assertions using this original context before starting the command. Keep this boundary covered when upgrading OpenCode.

Each changed directory uses a separate native shell invocation, with that directory as workdir, so the existing direnv shell.create.before integration applies the correct inherited/nested environment. The helper and ESLint inherit that environment unchanged. Project config and plugins are executable code, just as with manual lint; native shell approval is not a filesystem sandbox for their imports or side effects. This plugin confines its own discovery, not arbitrary code explicitly loaded by a project's config.

Per edit: at most 32 files, 8 directories, and a 30-second scheduling budget. Each shell command has a maximum 10-second timeout; helper ESLint subprocesses share 8 seconds and have a 64 KiB stdout/stderr capture limit. Helper reports are capped at 12,000 characters and appended feedback at 16,000 characters. Remaining files, timeouts, truncation, ignored-file warnings, config failures, and command failures are reported, never converted into a clean result. Permission prompts and existing environment-hook preparation (direnv has its own timeout) are outside the lint process budget. Concurrent edits can change files before a check reads them; feedback describes the file on disk, not an immutable edit snapshot.

Reload

After changing plugin code, reload to rebuild every loaded location. This cancels pending permission requests and forms; running sessions continue with fresh services at their next step boundary. Use it only when those global side effects are acceptable:

opencode api post /api/location/reload
opencode api get /api/plugin

Look for lint-feedback. A service restart is not required. Formatting changes are global; the check only runs where the project already provides supported ESLint configuration.

Local tests and real ESLint fixture

From the repository root:

bun test ./packages/lint-feedback/index.test.js ./packages/lint-feedback/check.test.js

Tests exercise project/file selection, returned output, bounded failures, literal arguments, and the permission-preserving executor boundary with deterministic native-shell doubles. One test executes a disposable local CLI to cover the real subprocess/argv/environment path. Native permission enforcement itself must also be checked in a live V2 session; the doubles do not prove the host's policy engine.

fixtures/project is a deliberately broken, real ESLint project, isolated from this repository's rules. For a disposable manual fixture using cached npm packages only (no downloads), run from this repository:

FIXTURE=$(mktemp -d "${TMPDIR%/}/opencode/lint-real-XXXXXX")
cp -R packages/lint-feedback/fixtures/project/. "$FIXTURE/"
npm install --prefix "$FIXTURE" --offline --ignore-scripts --no-audit --no-fund
node packages/lint-feedback/check.mjs "$FIXTURE" "$FIXTURE/broken.js"

If the pinned packages are not cached, stop rather than falling back to an online install. An existing ESLint 9/10 project's node_modules can instead be copied into the disposable fixture (not symlinked outside it). Expected helper output: ESLint found lint errors and missingValue ... no-undef. The source must remain unchanged. Changing the source to export const answer = 42; should report No ESLint diagnostics.

For the native path, use a session located at the fixture and actually patch broken.js to export const answer=anotherMissingValue (not a shell write). The project's Prettier should format the file, while the successful patch result should also contain ESLint's no-undef diagnostic. A shell-denied agent should retain the edit result with an incomplete check and never execute the lint helper. The fixture is manual verification, not an E2E suite or a global lint policy.

License

MIT