# Verify

> Run an isolated Synara instance and verify UI behavior with real provider sessions.

- Skill: `emanuele-web04/verify` (Agent Skill)
- Install (CLI): `npx skillmds@latest add emanuele-web04/verify`
- Raw SKILL.md: https://api.skillmd.com/api/skills/emanuele-web04/verify/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: emanuele-web04 (https://skillmd.com/u/emanuele-web04)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/emanuele-web04/verify

---


# Verify: run Synara locally for runtime verification

How to launch an isolated Synara instance (server + web) to observe UI changes, without touching `~/.synara` or the default dev ports.

## Launch

```bash
# 1. Server (from the directory you want as the workspace/project cwd):
SYNARA_HOME=<scratch>/synara-home \
SYNARA_PORT=3899 SYNARA_MODE=web SYNARA_NO_BROWSER=1 \
VITE_DEV_SERVER_URL=http://localhost:5899 \
bun <repo>/apps/server/src/index.ts &

# 2. Web (vite dev):
cd <repo>/apps/web && PORT=5899 VITE_WS_URL=ws://localhost:3899 bun run dev &
```

Then open http://localhost:5899/.

## Gotchas

- Preflight the ports before launching: run `lsof -nP -iTCP:<port> -sTCP:LISTEN` on both the server and web ports (check IPv4 and IPv6 listeners — a `:::port` entry collides even when `127.0.0.1` looks free). If you use `scripts/dev-runner.ts dev`, read its dry-run output for the real port map first.
- `SYNARA_AUTH_TOKEN` is inherited from the launching shell: a server started with it set requires auth the web client does not have, which produces a healthy-but-disconnected UI. Unset it in the isolated test process only — never strip it from production policy.
- `VITE_DEV_SERVER_URL` on the **server** is required — without it the WS handshake from the vite origin is rejected with 403 (see `apps/server/src/trustedOrigins.ts`).
- `VITE_WS_URL` on the **web** side tells the app where the WS server lives (`apps/web/src/wsTransport.ts`).
- Default ports are 3773 (server) / 5733 (web), with no automatic per-checkout offset. The dev runner uses an explicit `SYNARA_PORT_OFFSET` or derives an offset from `SYNARA_DEV_INSTANCE` when supplied — pick explicit distinct ports and confirm the dry-run output to avoid colliding with a running dev instance.
- Add a disposable git workspace through the sidebar's **Projects → Add project**
  button and enter its absolute folder path. If New project in the composer
  picker does not open the dialog, use this sidebar entry point instead.
- To see diffs: select a git workspace with uncommitted changes, then click the **+N −N** toggle in the top-right chat header — it opens the DiffPanel (working-tree diff). No project/thread needed.
- Server tests: don't run the suite from a checkout under `/private/tmp` — `localImageRoute.test.ts` fails there (its "outside the workspace" fixture lands in an allowed temp root). It passes from a normal checkout and on CI.

## Playwright driving

Chrome extension may be unavailable; `playwright` is a devDependency of `apps/web` — import it by absolute path from `apps/web/node_modules/playwright/index.mjs` in a scratch script.

## Real provider feature verification

The Mind- and Kanban-specific bullets below describe the feature branches
under test (`agent/mind-feature-rebased`, `agent/kanban-v2-clean`), not
mainline UI — run them while checking out or merging those branches.

- Use an installed, authorized provider/model, selected explicitly in the UI.
  Wait for model discovery to resolve before sending; a new composer can briefly
  display loading placeholders or default to a different model than expected.
- A real UI-initiated turn exercises the browser WebSocket transport and provider
  gateway. Keep mechanical tool coverage separate from behavioral feature tests.
- For continual-learning memory, naturally teach a project preference without
  naming tools. Check its appearance in Mind, then ask a relevant question in a
  **new thread in the same project**. Expand the worked/tool rows to distinguish
  retrieval from conversation recollection or a plausible generic answer.
- If a natural targeted query misses a saved fact, retain that failure before
  trying a broad question about remembered facts. A successful broad recall does
  not retroactively prove spontaneous recall on the original task.
- To verify Mind auto-refresh, mount Mind before the agent saves, then leave it
  untouched for at least one polling interval. Navigating, refreshing, or editing
  a memory during the observation window is not independent polling evidence.
- Project chips require multiple memory-bearing projects. Use distinct facts in
  two disposable projects and verify each filter and each project profile.
- Kanban's overview and per-project board are separate views. Click the project
  heading for the four-column Attention board. For keyboard draft reorder, focus
  the inner card button before Alt+ArrowUp/Down.
- An idle-draft move-to-Done refusal requires a draft **without a goal**. Setting
  a goal can start provider work and invalidate that precondition. Read the card
  immediately before the probe, and retain the exact runtime error.
- Do not manufacture aged/stale or high-volume cap evidence by mutating a live
  database. Use authorized fixtures or report those coverage gaps explicitly.

## Devin Secrets Needed

None for an already available no-auth provider/model in an isolated instance.
Otherwise obtain the chosen provider's required credential through the approved
secret mechanism; never put secret values in this skill or test artifacts.