AGENTS.md · diff
git:20260906.da4694e to git:20260909.89c48da
19 added, 0 removed. Audit A to A.
# AGENTS.md
OpenWork is a free, open-source desktop app (macOS, Windows, Linux) for doing
work with AI agents on your own files — an open-source alternative to Claude
Cowork and Codex, built on OpenCode, running any model from 50+ providers.
Desktop mode keeps files local; cloud is optional. Three surfaces live in this
repo:
- **Desktop app** (`apps/`, `packages/`) — local-first agent workspace: chat on
files, skills, browser automation, scheduled automations, Anthropic-compatible
plugins.
- **OpenWork MCP gateway** (`ee/apps/den-api`) — one URL
(`api.openworklabs.com/mcp/agent`) that brings org-assigned skills, plugins,
and connections (Google Workspace, Microsoft 365, MCPs) into Codex, Claude
Code, Cursor, or any MCP client via `search_capabilities` /
`execute_capability`.
- **OpenWork Den** (`ee/apps/den-*`) — the org control plane: provision
inference, manage teams and access, set desktop policies, publish skills and
plugins through marketplaces.
The app consumes OpenWork server surfaces (self-hosted or hosted) rather than
inventing parallel behavior. Anything OpenCode can do is available in OpenWork,
even before a dedicated UI exists.
## Confidentiality (hard rule — this repo is public)
Never let a branch name, commit, PR text, comment, fixture, or evidence identify
a customer, prospect, partner, or outside person; use internal ticket IDs, and
escalate any leak instead of rewriting history.
## Verification (every change)
- Proof is a journey spec in `evals/specs/**` or `scenarios/<name>/e2e.test.ts` (`*.e2e.test.ts` drives the app/Den; `*.test.ts` hits a server or gateway boundary). Unit tests are not proof and never live in journey directories; the boundary ratchet rejects new ones that import product source, read the repo, or spawn test runners.
- Default is zero new test files: run the journey spec that covers the change; extend it for a gap; new file only for a new user journey.
- Run `pnpm evals:e2e <slug>` or `pnpm evals:pr specs/<name>.test.ts`; report the printed placement and verdict.
- `Passed` requires an observable assertion for every claim; skips are never passed.
- Docs/comments, types-only, and inert agent config may skip runtime proof — say so.
- Skill chain: `write-a-spec` → `run-tests` → `diagnose-a-red-run` when red → `publish-evidence`.
+ - After journey checks, one owner runs the final local Warden review; see
+ `.warden/README.md`. Record `git HEAD` and `origin/dev` before and after it,
+ use bare `pnpm warden:check` against the clean committed branch, and report
+ expected versus actually reviewed skills. Unstaged and untracked files are
+ not reviewed. Do not require a commit unless the user authorized one; for
+ uncommitted work, `--staged` is diagnostic only and its limits must be stated.
+ - Only completed expected applicable skills, verified scope, and no blockers
+ means reviewed with no blockers. Exit 0 with no files or no matching triggers
+ is `Not reviewed` or `Not applicable`, with scope explicit, never clear.
+ Missing credentials, errors, cancellation, or partial coverage are
+ `Incomplete`. After the last edit, rebase,
+ base update, policy change, or model change, rerun the full final review with
+ all applicable skills. `--skill <name>` and `--staged` are scoped diagnostics,
+ not final clearance. Do not use `--fix` or run Warden in every parallel
+ subagent.
+ - Finish with finding IDs, classification, evidence, `Clear when`, disposition,
+ exact command/exit code/scope/run reference, and the recorded refs. A local
+ result never authorizes GitHub approval. A suggested repro is not an executed
+ test.
## Pull requests
- Do not default to draft PRs. A request to create or make a PR means a
ready-for-review PR once the required proof is published. Use a draft only
when the requester explicitly asks for one or the current verdict is
`Incomplete` or `Failed`, and state exactly what proof is missing.
- Run tests and report commands + results. A runtime-observable change is not
done until its test evidence is visible on the PR. If validation cannot run,
say why and give exact repro steps.
## Local headless web (agents)
- `pnpm world up dev-headless --detach` launches an isolated browser UI +
local `openwork-server` without Electron as a detached script world.
`pnpm dev:headless-web` remains a compatibility alias with its prior foreground
default (`--detach` still works). Read
`tmp/dev-headless-web.json` for the owner-only runtime manifest.
It does not use `~/.config/openwork/server.json`, and its engine keeps its own
sessions database at `tmp/dev-headless-opencode.db` instead of the desktop
app's `~/.local/share/opencode/opencode.db`. Stop a running script with
`pnpm world down dev-headless`; pass script options after `--`, for example
`pnpm world up dev-headless --detach -- --replace --keep-tokens`. Cloud sign-in
is copy/paste handoff (Den cannot redirect grants to localhost): Account → Sign
in → copy OpenWork link on Den → Paste sign-in code in Settings.
## Coding
- pnpm only, never npm/yarn. TypeScript: never `any`, typecasts, or `as` unless
100% necessary or instructed.
- Prefer Tailwind, React, shadcn/ui (Base UI), TanStack Query, Zustand, Zod,
Drizzle, Better-Auth. Reuse `@/components`; end users are non-technical.
- Smallest possible diff, then make it smaller. Propose the simpler solution. No
fallback expressions when types or control flow already guarantee a value.
- If asked to do too much at once, stop and say so.
## Hands-on PR previews
For a preview the user can test inside Codex, use
[preview-my-work](.opencode/skills/preview-my-work/SKILL.md): named Daytona
worlds for Den or real Electron through noVNC, with isolated scenarios,
frontend updates, reset and teardown.
## Reusable scenarios
Start at [scenarios/README.md](scenarios/README.md) for workflows, videos, and documentation screenshots. Keep each scenario together; reuse the existing world, testkit, behavior, and DocShot packages. Author ordinary TypeScript functions and React/Remotion components.