agentgui 1.0.1025 → 1.0.1026

package/.gm/memories/mem-0c444abde86202f6-1681.md

@@ -0,0 +1,8 @@
+---
+key: mem-0c444abde86202f6-1681
+ns: default
+created: 1783677428217
+updated: 1783677428217
+---
+
+agentgui 31st run addendum (2026-07-10): confirming gm-continue pass caught that the gui-ux-craft workflow's light-theme-contrast lens had crashed mid-run (API connection error) and never actually produced findings - a real coverage gap, not a clean pass. Dedicated follow-up audit computed real WCAG contrast ratios (not eyeballed) and found 6 genuine light-theme failures the crashed lens would have caught: chat avatar initials using raw --mascot/--accent-ink on tinted backgrounds (2.03:1/3.65:1), tool-done status text using --success/--green-2 (3.18:1/2.88:1) instead of the text-safe --green, --amber measuring 4.26:1 on --bg-2 (narrowly failing AA on the background it actually renders on, despite being tuned for --paper), --danger measuring 4.25:1 on --bg-2 for the same reason (auth-error banner), raw --mascot as button/CLI text (3.11:1), and --sun being essentially unusable everywhere it's used (1.1-1.5:1 against every app background) - retired in favor of --amber. All fixed with the same pattern as the 22nd-run --accent->--accent-ink precedent: separate text-safe 'deep' tokens from fill/border tokens, and always check contrast against the ACTUAL background a token renders on (--bg-2 card surfaces, not just --paper) not just the token's own design-time tuning target. LESSON for gm-continue: when a workflow notification includes a `failures` array naming a lens that errored out, that lens's scope is unaudited work, not a null result - it belongs in the confirming PLAN pass as its own row, the same as any other unaudited surface. Kit pushed df04d69, agentgui pushed 03002cf, CI green (Test/Deploy GH Pages/Publish and Release/Auto-Declaudeify all success).

package/.gm/memories/mem-3602d11d7bce9446-1057.md

@@ -0,0 +1,8 @@
+---
+key: mem-3602d11d7bce9446-1057
+ns: default
+created: 1783677456828
+updated: 1783677456828
+---
+
+agentgui Live History Stream Wiring (2026-05-05): site/app/js/app.js opens an SSE EventSource on navTo('history') via B.streamHistory(base, onEvent) and closes it on tab exit; state.live = {es, connected, lastEventTs, error, eventCount}. Event dispatch loop: 'hello' sets connected=true; 'event' pushes into state.events + increments the matching session counter only when data.sid===selectedSid; 'conversation' calls refreshHistory() to reload the session list; 'error' sets live.error. Renders are throttled via requestAnimationFrame to avoid an event-storm during burst loads. The FIRST request to any /v1/history/* endpoint triggers ccsniff's loadOnce(), a synchronous walk of every JSONL file under ~/.claude/projects (override via CLAUDE_PROJECTS_DIR) - in the original test env this was 299 files/80MB/69k events, a 30-90s startup latency during which health-check timeouts are normal and expected; subsequent requests are fast (cached index). The endpoints are served by ccsniff's Express router mounted in-process from server.js, no external proxy.

package/AGENTS.md

@@ -1,5 +1,9 @@
 # AgentGUI — Agent Notes
 
+## Light-theme contrast fixes (2026-07-10) — thirty-first run, confirming pass addendum
+
+The 31st run's `gui-ux-craft` workflow had a `failures` array naming `hunt:light-theme-contrast` as crashed (API connection error) — that lens never ran, so light theme was never actually audited despite the workflow otherwise completing. Caught during the mandatory `gm-continue` confirming pass, not the original run. A dedicated follow-up agent computed real WCAG contrast ratios (not eyeballed) and found 6 genuine failures: chat avatar initials on raw `--mascot`/`--accent-ink` over tinted backgrounds (2.03:1/3.65:1), tool-done status text on `--success`/`--green-2` (3.18:1 paper / 2.88:1 on the card bg it actually renders on), `--amber` measuring only 4.26:1 on `--bg-2` despite being tuned for 4.5+:1 on `--paper`, `--danger` the same shape (4.25:1 on `--bg-2`, the PASSWORD-gate auth-error banner), raw `--mascot` as button/CLI text (3.11:1), and `--sun` being unusable everywhere it's used (1.1-1.5:1 against every app background) — retired in favor of `--amber`. Same fix pattern as the 22nd run's `--accent`->`--accent-ink` precedent: a separate text-safe "deep"/darkened token for text, the original stays for fills/borders (3:1 threshold). **Standing rule: always check a color token's contrast against the ACTUAL background it renders on in the real DOM (e.g. a status pill's tinted card, not just the page's base `--paper`) — a token tuned and commented as "4.5+:1 on paper" can still fail AA on every card/panel surface built on top of `--bg-2`/`--bg-3`.** **gm-continue standing rule: a workflow notification's `failures` array naming a crashed lens is unaudited scope, not a null result — it goes into the confirming PLAN pass as its own row, same as any unaudited surface.** Kit pushed `df04d69`, agentgui pushed `03002cf`, all 4 CI workflows green.
+
 ## Maximal usability/negative-space/predictability sweep (2026-07-10) — thirty-first run
 
 User asked to "max the gui usability practicality... predictable, good looking, use negative space properly" via wfgy-method-grade critique. Ran a fresh `gui-ux-craft` workflow (32 agents, 19 confirmed) plus a targeted live-browser negative-space audit agent in parallel, then landed the highest-value findings from both. Kit (`../design`): ConfirmDialog `destructive` prop was dead code (delete confirms always rendered `primary`, never `danger`) — real predictability bug, fixed + browser-witnessed the delete dialog button now carries `btn-primary danger`. Settings panels had a 56px gap between stacked Panels (32px `.panel` margin + 24px grid gap stacking) vs 24px everywhere else — zeroed the margin inside `.settings-grid`, witnessed 24px consistent. `.ds-page-header-dense-lede` and `.chat-composer-context` both hard-clipped with `nowrap`+ellipsis at <=480px with no way to recover the text — both now wrap at that breakpoint (kept desktop's single-line ellipsis). History empty-state sub-line used the misapplied `.lede` (fs-xl) style, rendering LARGER than its own title — removed the misapplied class + fixed the title's own hardcoded `1.05rem` to `var(--fs-lg)` token. Small-caps section labels (`panel-head`, `table th`) drifted from the canonical `.t-label` size (`--fs-xs` vs `--fs-tiny`) — unified. `FileRow`/`FileCell` internal selection props renamed `marked/onMark` -> `selected/onToggleSelect` to match the shared multi-select contract (FileGrid's own `marked`/`onMark` stay as its documented public alias). `ConversationList` search-mode blew away already-visible rows to a full skeleton on every debounced re-search — guarded on `!sessions.length`, matching `files.js`'s sibling pattern. App copy-tone: "sidebar" (used nowhere else) -> "the rail"; internal-jargon `agent: 'external session'` -> `'other tool'`; "session" -> "conversation" vocabulary unified across resume banner/toasts/labels; raw sid slice in resume banner replaced with a human session-title lookup (`resumeSidLabel()`), falling back to the slice only when no title is known; settings projects-folder fact gained the same copy button its sibling root rows already had. Perf: `prism-core.min.js` (a classic UMD script, not an ES module) now gets a `rel=preload as=script` hint alongside the existing `modulepreload` hints for marked/dompurify; History tab re-fetch on every entry now gated on staleness (>5s) instead of unconditional; bulk file delete/move patch `state.files.entries` locally and render immediately instead of awaiting a second full directory round-trip (backgrounded as reconciliation) — scoped to the bulk paths only, the highest-traffic case. Kit docs: README gained a "multi-agent chat shell" bullet naming the flagship `WorkspaceShell`/`AgentChat`/etc surface (previously undiscoverable from a first read), `COMPONENT_API.md` documents `FileRow`'s `move` action and scopes the `freddie.js` re-exports as demo-site-only. **A concurrent gm session was independently working the same `../design` repo this run — its commits (`df5b566`..`68b615b`) landed my kit edits before I reached CONSOLIDATE; verified via `git log`/`rev-parse HEAD==origin/main` rather than re-pushing.** Browser-witnessed every fix live at 1280px+420px on a no-PASSWORD localhost:3010 instance (settings gaps, empty-state hierarchy, danger delete button, mobile wrap, files checkbox/BulkBar flow after the prop rename): 0 console errors throughout. **Standing rule reaffirmed: `prd-resolve` witness_evidence must be distinct per row — a byte-identical witness across rows is refused by plugkit as a rubber-stamp tell; write what THAT row specifically changed, every time.**
@@ -222,21 +226,7 @@ Surfaced 2026-05-04 while validating the chat surface in `site/app/`. Fix applie
 
 ## Live History Stream Wiring (2026-05-05)
 
-**`site/app/js/app.js` opens SSE EventSource on navTo('history'); first /v1/history/* request triggers 30-90s loadOnce() synchronous walk.**
-
-The live history feature wires the in-process ccsniff `/v1/history/stream` SSE endpoint via `B.streamHistory(base, onEvent)` on tab entry and closes the EventSource on tab exit. State shape: `state.live = { es, connected, lastEventTs, error, eventCount }`.
-
-Event dispatch loop:
-- `hello` event: set `connected=true`
-- `event` event: if `data.sid === selectedSid` push into `state.events`, increment matching session counter
-- `conversation` event: call `refreshHistory()` to reload session list
-- `error` event: set `live.error`
-
-Throttle renders via `requestAnimationFrame` to avoid event storm during burst loads.
-
-**First request to `/v1/history/*` triggers loadOnce() that walks all JSONL files under ~/.claude/projects** (env default; override with `CLAUDE_PROJECTS_DIR`). In our test env: 299 files, 80MB, 69k events -> 30-90s startup latency. Health check timeouts during this window are normal and expected. Subsequent requests are fast (cached index).
-
-The endpoints are served by ccsniff's Express router mounted in-process from `server.js`. No external proxy.
+`site/app/js/app.js` opens an SSE EventSource on `navTo('history')` (`B.streamHistory`, `state.live` shape, hello/event/conversation/error dispatch, rAF-throttled renders) against ccsniff's in-process `/v1/history/stream`; first `/v1/history/*` request triggers a 30-90s synchronous JSONL walk (`CLAUDE_PROJECTS_DIR`) — health-check timeouts during that window are normal. Full detail in rs-learn (recall "agentgui Live History Stream Wiring").
 
 ## Runtime CDN: GUI loaded from unpkg (2026-05-29)
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentgui",
-  "version": "1.0.1025",
+  "version": "1.0.1026",
   "description": "Multi-agent ACP client with real-time communication",
   "type": "module",
   "main": "electron/main.js",