chore(workspace): what the blank-chat session taught this workspace

.claude/hooks/glob-guard.test.mjs

@@ -26,10 +26,17 @@ writeFileSync(path.join(CWD, 'notes.md'), '');
 process.on('exit', () => rmSync(CWD, { recursive: true, force: true }));
 
 /** @returns {{blocked: boolean, message: string}} */
+// WHY an empty fake /proc: the hook refuses to signal Destin's LIVE app, looked up
+// by pid in /proc. Against the real one, a literal like `kill -9 4321 5678` failed
+// the day pid 4321 was the live app's zygote (2026-09-27). The live-app cases below
+// build their own fake /proc (runProc); every other case sees no processes at all.
+const NO_PROC = mkdtempSync(path.join(tmpdir(), 'glob-guard-noproc-'));
+process.on('exit', () => rmSync(NO_PROC, { recursive: true, force: true }));
+
 function run(command, tool_name = 'Bash') {
   const r = spawnSync('python3', [HOOK], {
     input: JSON.stringify({ tool_name, tool_input: { command }, cwd: CWD }),
-    encoding: 'utf8',
+    encoding: 'utf8', env: { ...process.env, GLOB_GUARD_PROC: NO_PROC },
   });
   return { blocked: r.status === 2, message: (r.stderr || '').trim() };
 }

ROADMAP.md

@@ -74,9 +74,9 @@ Target: `v1.3.1`
 | [dev-workspace](docs/roadmap/dev-workspace.md) — building the app, not the app | 101 | 32 | 4 | 9 |
 | [native-harness](docs/roadmap/native-harness.md) — the app's own agent doing work | 57 | 10 | 5 | 25 |
 | [user-interface](docs/roadmap/user-interface.md) — shared primitives, chrome, layout, copy | 40 | 16 | 2 | 6 |
-| [remote-access](docs/roadmap/remote-access.md) — reaching the app from another device | 29 | 6 | 1 | 4 |
+| [remote-access](docs/roadmap/remote-access.md) — reaching the app from another device | 30 | 7 | 1 | 4 |
 | [sync](docs/roadmap/sync.md) — moving your stuff between devices | 26 | 10 | 8 | 5 |
-| [chat-data](docs/roadmap/chat-data.md) — everything kept about a chat | 22 | 7 | 2 | 2 |
+| [chat-data](docs/roadmap/chat-data.md) — everything kept about a chat | 23 | 8 | 2 | 2 |
 | [other-features](docs/roadmap/other-features.md) — real features too small for their own area | 22 | 5 | 6 | 5 |
 | [marketplace](docs/roadmap/marketplace.md) — finding, installing and rating plugins and themes | 20 | 7 | 0 | 5 |
 | [files](docs/roadmap/files.md) — documents the user opens, edits or organises | 17 | 4 | 0 | 9 |

docs/active/investigations/2026-09-10-android-parity-audit.md

@@ -403,6 +403,17 @@ dev-workspace).
 
 ### Found after the consolidation
 
+- **Swiping the app away and reopening it leaves every running conversation's chat blank**
+  (2026-09-27). Swiping from recents closes the screen but the background service keeps the
+  sessions running (`SessionService.onTaskRemoved` does nothing, on purpose). The reopened
+  screen asks each session for its first history page, and Android has no `transcript:page`
+  handler (the shim refuses it quietly) while its transcript watcher only pushes new lines — so
+  the chat shows "Start a conversation" while the terminal and Claude still have everything,
+  until close + resume. Read from the code during the desktop blank-chat fix (youcoded#582),
+  not reproduced on a handset. The desktop cure (read a rebuilt screen's first page to the end)
+  needs the Kotlin tail reader the paging item already asks for.
+  `android` `needs-verify` `checked 2026-09-27`
+
 - **Dogfood APKs count as a separate person in our own numbers** (2026-09-13). Android's device
   id is scoped to the app-signing key, and `android-test-build.yml` signs with a throwaway debug
   key — so a dogfood build installed beside the real app on one phone reports a DIFFERENT device

docs/roadmap/chat-data.md

@@ -3,6 +3,11 @@ Filing test: everything kept about a chat — transcript, title, tags, notes, se
 resume state. Not here: the model is running right now (native-harness); the files a chat
 produced and the panel that shows them (files).
 
+- [ ] A streamed reply was split mid-sentence across two chat bubbles ("…for your review. Now" |
+      "the invites.") in one practice-app run while the computer was busy; seen once in ~30
+      loaded runs (2026-09-27), not yet seen in the real app
+      `chat` `desktop` `needs-verify` `checked 2026-09-27` `needs-repro`
+
 - [ ] The Projects page's conversation list shows only Claude Code conversations and carries no
       tags, note or last-used model, so its cards look sparser than the same conversations in the
       Resume browser, and assistant conversations in that folder never appear there (seen while

docs/roadmap/dev-workspace.md

@@ -60,16 +60,15 @@ seen-on is always n/a here.
       bigger budget
       `n/a` `confirmed` `checked 2026-09-16`
 
-- [ ] Three small tooling papercuts from the 2026-09-18 premium-motion session, each costing a
-      whole cycle: (1) `run-dev.sh --stop` lists the dev window's own `claude` children with no
+- [ ] Two small tooling papercuts from the 2026-09-18 premium-motion session, each costing a
+      whole cycle (a third, line budgets only reported at the end of verify, was fixed
+      2026-09-27: verify checks them first): (1) `run-dev.sh --stop` lists the dev window's own `claude` children with no
       sign they are its children — have it say so from the process's ancestry, instead of a doc
       asking sessions to remember; (2) a fresh session worktree has no `scratch/perf-lab/assets`,
       so `fixture.mjs --ensure-assets` re-downloads ~490 MB and a slow link times the whole run
       out after the 3-minute build and the quiet-machine wait — copy or hardlink from the shared
-      checkout's `scratch/perf-lab/assets` when it is there; (3) `verify.sh` found six separate
-      line-budget overruns of one to thirteen lines, each only at the END of a full run — run
-      `tests/line-budgets.test.ts` first, it takes under a second
-      `n/a` `confirmed` `checked 2026-09-18`
+      checkout's `scratch/perf-lab/assets` when it is there
+      `n/a` `confirmed` `checked 2026-09-27`
 
 - [ ] The perf lab's new "blank on arrival" count has never been shown to FAIL. It was added
       2026-09-18 (`scenario-workload.mjs` → `switchBlank`) because the painted clock counts entry

docs/roadmap/remote-access.md

@@ -1,6 +1,13 @@
 # remote-access — reaching the app from another device
 Filing test: reaching the app from another device — the protocol, the browser client.
 
+- [ ] A phone connected to the computer can lose the messages it is showing: when the computer
+      sends its copy of the conversations and its copy of one is empty (it had not loaded that
+      conversation yet, or its own chat went blank), that empty copy replaces the phone's. Read
+      from the code during the 2026-09-27 blank-chat fix; left out when Destin scoped that work
+      to desktop
+      `desktop` `needs-verify` `checked 2026-09-27`
+
 - [ ] With a phone connected, the computer sends it the terminal output and chat events of every
       open session, not only the one the phone is showing — more sessions running means a busier
       computer and phone. Found by the 2026-09-23 performance review; not built because remote

docs/roadmap/user-interface.md

@@ -120,8 +120,12 @@ layout, copy. Not here: one screen only — that screen's area, with the surface
       `all` `confirmed` `checked 2026-09-11` `v1.3.1` → docs/active/investigations/2026-09-10-error-inventory/README.md
 
 - [ ] Chat panel vanished from a live session (beta.16, 2026-08-27) — no messages, and new
-      sessions showed no "Start a conversation" text; Destin said ignore for now
-      `chat` `desktop` `needs-verify` `checked 2026-08-27` `needs-repro` → docs/archive/investigations/2026-08-27-terminal-black-glyphs-mipmap-driver.md
+      sessions showed no "Start a conversation" text; Destin said ignore for now. 2026-09-27: a
+      friend on Windows and Destin on Linux saw messages disappear MID-conversation behind
+      "Start a conversation" while Claude and the terminal still had them. Four causes fixed
+      (youcoded#582); what rebuilds the chat mid-conversation in the built app is still unknown —
+      next step is a tripwire that logs what emptied an open chat, or the friend's desktop.log
+      `chat` `desktop` `needs-verify` `checked 2026-09-27` `needs-repro` → docs/archive/investigations/2026-08-27-terminal-black-glyphs-mipmap-driver.md
 
 - [ ] Switching sessions in terminal view redraws the letters of EVERY open terminal each time
       (the black-glyph safety net in `TerminalView.tsx`). Measured 2026-09-10 with the new perf-lab

docs/wrap-ups.md

@@ -782,3 +782,14 @@ recurred — the repetition is the data.
 - Line budgets raised on App.tsx, SettingsPanel, SyncPanel, StatusBar, SessionStrip. RECURRENCE of the budget friction already in `dev-workspace.md` → no new entry
 - Destin: "the simplest possible version … fewest error cases … least thinking from the model", and journeys run automatically because "a check nobody runs goes stale" (his call when asked) → applied: spec decisions + the verify.sh comment
 - deleted/merged: run-review.sh, cdp-ports.sh, probe-ports.sh, montage.sh, make-gallery.py and their 3 tests + CI step deleted; 82 click plans archived (site-gallery kept); 16 per-shot port/offset mechanisms replaced by one engine; ui-probe and the two drag probes deliberately kept (reasons in the spec)
+
+## 2026-09-27 — chat goes blank over a live conversation; native "unreadable" read as "empty" (youcoded#582, session/chat-vanish)
+- Line budgets tripped six times (App.tsx, types, ipc-handlers, native-session-host, remote-server, chat-reducer), each only at the END of verify, plus two line-budgets.json merge conflicts. RECURRENCE (4th session) of the 2026-09-18 papercut → applied: `scripts/verify.sh` runs the budget test first (~2 s) and prints the overruns before the parallel checks start; shown red with a one-line probe; roadmap papercut (3) closed in `dev-workspace.md`  [~3 extra verify runs]
+- Static reading could not find what empties a live chat in the built app; the renderer has no way to write to desktop.log, so the trigger stays unknown → roadmap: the existing `user-interface.md` "Chat panel vanished" item updated with the four fixed causes and the tripwire as next step
+- Two independent reviews of my own work each found a real gap (duplicate tool card from an overlapping page; resume's sync header read made the new retry inert) → applied in #582; no new rule: the global rules already ask for unintended-consequence review, and it worked
+- A pre-existing journey flake (resume-conversation, ~1 in 6 under load, on master too): step 2's "Resume" matched a chatsearch card's own Resume button when the menu opened slowly → applied: the journey waits for `chat/sessions` (8/8 under 32-core load)
+- I read a STALE verify log (verify keeps logs only for failed runs) and briefly believed a green --full run had failed → dropped: verify prints its own log path on failure; my mistake, caught by re-running the suite
+- The glob-guard hook blocked a cp-backup/restore toggle for a red-check; a reversible sed toggle worked → no change: the hook did its job
+- Destin: "ignore android" (scope), "investigate more thoroughly and fix… check our own work" → no new rule (global.md covers both); Android and phone findings filed (android audit appendix, `remote-access.md`), a split reply bubble filed in `chat-data.md`
+- The workspace hook suite failed 1/430 at close-out, on master too: `glob-guard.test.mjs` asserted `kill -9 4321 5678` is allowed, and pid 4321 was Destin's live app that day, so the live-app guard (correctly) blocked it → applied: every non-live-app case runs against an empty fake /proc (`GLOB_GUARD_PROC`); shown red against the real /proc
+- deleted/merged: App.tsx's first-page closure moved to `state/first-page-loader.ts` (App.tsx −27 lines, budget lowered to match)

scripts/verify.sh

@@ -267,6 +267,22 @@ fi
 # caught before release. Checkouts that predate tsgo fall back to tsc.
 TSC=tsc
 [[ -x "$DESKTOP/node_modules/.bin/tsgo" ]] && TSC=tsgo
+
+# Line budgets FIRST, synchronously (~2 s), printed before the parallel checks start.
+# WHY: an over-budget file was the most repeated verify failure in docs/wrap-ups.md
+# (four sessions; six overruns in one, 2026-09-27), and it only surfaced when every
+# check had finished, minutes later. Reported now, it is fixed while the rest run.
+# Still counted as a failure at the end; the related-tests run repeats it harmlessly.
+BUDGETS_FAILED=0
+if [[ -f "$DESKTOP/tests/line-budgets.test.ts" ]]; then
+  if ! ( cd "$DESKTOP" && npx vitest run tests/line-budgets.test.ts ) >"$LOGDIR/budgets.log" 2>&1; then
+    BUDGETS_FAILED=1
+    echo "FAIL  line budgets (checked first — the other checks still run)"
+    grep -E 'lines, budget' "$LOGDIR/budgets.log" | grep -vE '^\s*[-+]' | sed 's/^ */      /' | sort -u
+    echo ""
+  fi
+fi
+
 start types "types ($TSC --noEmit)" npx "$TSC" --noEmit -p tsconfig.json
 # The test tree is its own TS project (different module resolution, allowJs for
 # the .mjs orchestrator). Separate check so a failure names which tree broke.
@@ -326,6 +342,7 @@ fi
 
 FAILED=0
 FAILED_KEYS=()
+if [[ $BUDGETS_FAILED -eq 1 ]]; then FAILED=1; FAILED_KEYS+=("budgets"); fi
 for key in types testtypes tests knip lint design invariants screens journeys; do
   [[ -n "${PID[$key]:-}" ]] || continue
   wait "${PID[$key]}"; rc=$?