Test coverage begins — and catches a real reconnect bug on day one

Why Care?

For the entire life of augment-it, agents have written and rewritten this codebase with zero automated tests guarding it — verification was typechecks, builds, hand-run prove-scripts, and the operator walking the surface. That debt came due this month: humain-vc's corpora failing to load, corpus creations that never landed, a rail that renders "empty" indistinguishably from "broken." We couldn't tell a bug from an empty workspace because nothing asserted the difference.

This is the first automated test coverage in the repo. And it earned its keep immediately: the very first property test we wrote failed on its first run — as the registry spec explicitly predicted it would — and the failure was not a test artifact. It was a real reconnect bug in the client transport that plausibly underlies the zombie-session and "invokes never reach the workspace" symptoms the operator has been hitting. The test found it; the fix closes it; the test now guards it.

The bug the test caught

The client transport (packages/workspace/src/transport.ts) reconnects to the workspace-service after a dropped socket. But it only scheduled that reconnect from the WebSocket close handler. It turns out that when a connection is refused — the exact situation during a workspace-service restart or redeploy, when the browser's reconnect attempt hits a port with nothing listening — Node's native WebSocket (and, per spec, browsers too) fires only error, never close.

So the reconnect chain died on the first failed attempt. A surface that lost its socket during a backend restart would never reconnect on its own; it stayed wedged until a full page reload. Invokes fired into that window hung to their 120-second deadline and then blamed the server — precisely the "the workspace did not reply" mystery. This is a strong candidate for one of the mechanisms behind [[Search-And-Add-Invokes-Never-Reach-The-Workspace]] and the zombie-session class.

The fix: an error on a socket that never opened now runs the same reconnect tail the close handler does, guarded by a per-attempt flag so a browser that fires both error and close still schedules exactly one reconnect. Established-socket drops, auth-death (4401/4403), and deliberate close() all keep their existing behavior untouched.

Visible proof it works: the property test's whole suite dropped from 2.1s to 0.9s after the fix, because the invoke now reconnects promptly instead of hanging to its deadline.

What landed

The harness

Vitest is wired as the repo's test runner (unit + service tiers; Playwright for the future E2E tier per the plan). pnpm test runs the repo's turbo test task; each package runs vitest run. No new runtime dependencies — the only additions are the test library itself. In particular, the transport tests needed a WebSocket server to connect to, and rather than pull the ws package back in (deliberately removed from this repo), the harness includes a small hand-rolled RFC-6455 endpoint over node:http — zero dependencies, mirroring the real workspace-service invoke/claim contract.

Group C — transport resilience (4 tests, packages/workspace)

The registry's ✓-phrases, now green: no invoke silently vanishes across every socket-churn scenario (the one that caught the bug); a pre-open invoke is delivered exactly once after open; close 4401 fails all pending work immediately with "session expired"; and after auth-death the transport tries one silent refresh-reconnect, then waits — no storm. A timing seam was added to transport.ts so minutes-scale deadlines and backoff are assertable in milliseconds (production defaults unchanged).

Group D — workspace registry (3 tests, services/workspace)

The default_domain_type resolution that suspect 1 hangs on — humain-vc's thesis vs the strategy fallback — plus workspace.json org_id winning over the WORKSPACE_ORG_MAP env fallback, and a .env-less workspace still listing with has_env: false. Each test builds a throwaway clients/ root on disk and loads the module fresh, because the env map parses once at import.

Group H — didi's corpora slab (2 tests, services/workspace)

The "Existing corpora" slab didi reads before resolving a corpus name: it lists every domain in the workspace regardless of type (the load-bearing difference from the curator rail's typed query), and it degrades to an empty slab — never a failed turn — when the resolver hiccups. Capability dispatch and NATS are mocked, so the contract is pinned without a live bus.

Where this sits

Nine tests, three of the registry's ten MECE groups. Groups A (identity contract, ExUnit in id-didi-sh), B (session tenancy), E (canonical CRUD), F (corpus files), G (curator state), I (end-to-end), and J (the DB↔disk↔UI alignment audit) remain proposed in the registry, written as the ✓-phrases we want to see turn green next. The registry spec ([[Corpora-Builder-Harmony-Test-Registry]]) tracks every one, Proposed or Implemented, with its purpose and location.

See also

• [[Corpora-Builder-Harmony-Test-Registry]] — the living test registry

• [[Test-Coverage-Harness-And-Regression-Floor]] — the harness plan

• [[Troubleshooting-Workspace-DB-State-Alignment]] — the live bug hunt this serves

• [[Search-And-Add-Invokes-Never-Reach-The-Workspace]] — the open issue this fix likely closes

• [[Session-Expiry-Turns-The-App-Into-A-Zombie]] — the zombie class the reconnect bug feeds