No user visibility into state — needs a State-Inspector

The symptom

There is no way, from inside the running app, to see what the app currently believes. When something looks wrong — a stale card, a flow that mounts the wrong remote, a search rail with no launch context — the only inspection tools are DevTools spelunking (per-remote, and federation scrubs cross-origin errors) and reading localStorage by hand.

Where state actually lives here (the inspector’s table of contents)

1. Per-remote Svelte 5 runes singletons — deliberately NOT shared across federation (no shared block): activeFlow + layout in the shell, searchContext in search-and-add, and a separate @augment-it/workspace singleton instance in EVERY remote, each converging via WS broadcasts rather than shared memory. When they desync, nothing shows which remote believes what.
2. localStorage keys — the cross-remount contract: augment-it:active-flow, augment-it:session-token, augment-it:active-record-set, augment-it:search-request, augment-it:org-workbench:active-org, per-page sort prefs, per-set idx bookmarks, mapping caches. Undocumented as a set; no registry.
3. WS-session state — what workspace-service currently holds for this connection (active client, didi identity, pinned flag).
4. Server-side stores — row-store record sets/rows, response-store, prompt-store; visible only through whichever remote happens to render them.
5. The canonical layer — SurrealDB; inspectable only via scripts/MCP.

Directions (jotted, not decided)

• A State-Inspector surface — dev-mode panel or its own small remote: dumps (1) this remote’s singleton snapshots, (2) all augment-it:* localStorage keys parsed, (3) the current WS session frame, (4) recent cross-remote CustomEvents (a ring buffer listener on augment-it:*). Read-only first; editing state comes much later, if ever.
• A localStorage key registry — even just a doc/table with owner + shape per key would halve the spelunking.
• Event tap — the augment-it:* window-event namespace is the app’s nervous system and completely invisible; a 50-line listener that logs the last N events with payloads would have explained most historical “why didn’t it refresh” mysteries.
• Composes with [[Live-Not-Live-Indicator-Tooling-And-Cross-Service-Error-Surfacing]] — liveness answers “is it up,” the inspector answers “what does it think”; they may share a surface.

Flagged up the tree

This is not an augment-it quirk — every ai-labs app has the same hole with different state substrates. The universal framing lives at ai-labs/context-v/issues/State-Inspector-As-A-Universal-Need-Across-Apps.md; this file stays the augment-it-specific instance (the five-place inventory above is this app’s).

Open questions

• Dev-mode-only, or shipped-but-tucked-away (the didi chat could gate a /state verb)?
• One inspector in the shell that asks each remote to self-report (a augment-it:state-report request/response event pair), vs per-remote panels?
• Does the WS/event ring buffer belong in @augment-it/workspace so every remote gets it for free?