[codex] rebuild ClayCode feature rollout

.codex/artifacts/qa/codex-auth-refresh-and-send.md

@@ -0,0 +1,126 @@
+# Codex Auth Refresh And Send QA
+
+Date: 2026-04-07
+Repo: `/Users/canal/.codex/worktrees/1cce/t3code`
+Desktop state: `~/.t3/userdata`
+Build under test: local desktop rebuild from this worktree
+
+## Scope
+
+Verify the reported auth-refresh/send failure:
+
+1. The visible provider refresh control should complete instead of spinning forever.
+2. A resend on a previously errored Codex thread should recover instead of inheriting the stale `Quota exceeded` session state.
+3. A brand-new Codex thread should still send and receive a response normally.
+
+## Inventory
+
+- Visible settings refresh on the running desktop app
+- Existing errored thread recovery via the live orchestration websocket
+- Fresh thread send via the live orchestration websocket
+
+## Environment
+
+- Desktop app launched with `bun run start:desktop:main-state`
+- Live ports from the patched launch:
+  - `http://127.0.0.1:61902`
+  - `http://127.0.0.1:61903`
+- Codex provider reported `Authenticated · ChatGPT Pro Subscription`
+
+## Results
+
+### 1. Visible refresh control
+
+Status: Passed
+
+Steps:
+
+1. Opened `http://127.0.0.1:61903/settings/general`
+2. Clicked `Refresh provider status`
+3. Waited for the checked timestamp to update
+
+Evidence:
+
+- Page text updated to `Checked just now`
+- Provider row still showed `Codex v0.118.0`
+- Auth label still showed `Authenticated · ChatGPT Pro Subscription`
+- Refresh button was re-enabled after completion
+
+### 2. Existing errored thread recovery
+
+Status: Passed
+
+Target thread:
+
+- Thread id: `a7172a15-7002-4de4-8942-221f1ce58f9c`
+- Existing title: `test123`
+- Previous persisted state before retry:
+  - session status: `error`
+  - last error: `Quota exceeded. Check your plan and billing details.`
+
+Action:
+
+- Dispatched a new `thread.turn.start` against that same thread with:
+  - user text: `Reply with exactly OK. QA_AUTH_RETRY_1775590365845`
+
+Observed outcome:
+
+- Provider log showed a fresh session restart:
+  - `session/connecting`
+  - `session/threadOpenRequested` with `Attempting to resume thread 019d6909-55e7-7e00-adb6-ef516eea229c.`
+  - `session/threadOpenResolved`
+- Persisted session row moved to:
+  - status: `ready`
+  - last_error: `NULL`
+  - updated_at: `2026-04-07T19:32:56.517Z`
+- Persisted messages now include:
+  - user: `Reply with exactly OK. QA_AUTH_RETRY_1775590365845`
+  - assistant: `OK`
+
+Interpretation:
+
+- This is the exact stale-session recovery path we needed.
+- The old errored thread no longer stays poisoned after a resend.
+
+### 3. Fresh thread send
+
+Status: Passed
+
+Target thread:
+
+- Project id: `2dd348e7-e576-411c-98ba-dd161318420a` (`t3code`)
+- New thread id: `c244983d-5fb8-44dc-9941-0f63d57429d4`
+- Title: `QA_AUTH_FRESH_1775590432926`
+
+Action:
+
+1. Created a new thread in `t3code`
+2. Sent:
+   - `Reply with exactly OK. QA_AUTH_FRESH_1775590432926`
+
+Observed outcome:
+
+- Provider log showed a fresh Codex thread start
+- Persisted session row ended in:
+  - status: `ready`
+  - last_error: `NULL`
+  - updated_at: `2026-04-07T19:34:02.480Z`
+- Persisted messages now include:
+  - user: `Reply with exactly OK. QA_AUTH_FRESH_1775590432926`
+  - assistant: `OK`
+
+## Ship Readiness
+
+Passed:
+
+- Visible provider refresh control
+- Existing errored-thread resend recovery
+- Fresh thread send/response
+
+Not directly verified:
+
+- The exact same click path inside the Electron sidebar/composer chrome, because this desktop shell was still rendering the empty `No projects yet` sidebar state in automation even while the underlying state DB and websocket server were healthy.
+
+Residual risk:
+
+- There may still be a separate Electron/sidebar state hydration issue, but the auth refresh path and the Codex send/recovery path both behaved correctly against the patched backend/runtime.

.codex/artifacts/qa/codex-import-durable-thread.md

@@ -0,0 +1,53 @@
+# Codex Import Durable Thread QA
+
+Date: 2026-04-17
+Branch: `codex/rebuild-feature-rollout`
+
+## Environment
+
+- Branch-backed local dev server in Chrome via the Computer Use plugin
+- Target project: `t3-qa-codex-import-project`
+- Full verification gate rerun on this checkpoint:
+  - `bun fmt`
+  - `bun lint`
+  - `bun typecheck`
+  - `bun run test`
+  - `bun run build`
+  - `bun run build:desktop`
+
+## Targeted automated coverage
+
+- `cd apps/web && bun x vitest run --config vitest.browser.config.ts src/components/ChatView.browser.tsx -t "imports a Codex transcript into a durable thread from the global shortcut"`
+- `cd apps/server && bun x vitest run src/codexImport/Layers/CodexImport.test.ts`
+
+## Manual QA
+
+### Scenario 1: Import a local Codex session into a durable ClayCode thread
+
+1. Opened the branch-backed app in Chrome on a fresh project-backed thread context.
+2. Triggered `Cmd+Shift+I`.
+3. Verified the `Import from Codex` dialog opened and loaded live local Codex sessions.
+4. Selected a real local Codex session.
+5. Chose the target project `t3-qa-codex-import-project`.
+6. Confirmed the import.
+7. Verified the app navigated to a durable thread route (`/$environmentId/$threadId`) instead of a `/draft/...` route.
+8. Verified the imported transcript content rendered in the message timeline.
+9. Verified the thread activity showed the import provenance for the Codex session.
+
+Result: pass
+
+### Scenario 2: Reopen the import dialog after import
+
+1. Reopened `Import from Codex` on the imported thread.
+2. Verified the imported session row showed the `Imported` pill in the list.
+3. Verified the preview pane showed `Import status: Already imported`.
+4. Verified the primary action label changed to `Open imported thread`.
+5. Confirmed the action and verified the app stayed on the already-imported durable thread rather than creating a duplicate thread.
+
+Result: pass
+
+## Observations
+
+- The rebuild now matches the intended durable-history model: importing creates a real local thread instead of just prefilling a draft.
+- Repeat imports are idempotent at the UX level: the UI clearly marks the existing imported thread and reopens it instead of duplicating content.
+- The earlier stale React Query state bug is fixed. After importing, the list row, preview pane, and action button all update consistently without requiring a full dialog refresh.

.codex/artifacts/qa/deep-search-and-project-search.md

@@ -0,0 +1,56 @@
+# Deep Search + Project Search QA
+
+Date: 2026-04-17
+Branch: `codex/rebuild-feature-rollout`
+
+## Terminal gate
+
+- `bun fmt`
+- `bun lint`
+- `export PATH="$HOME/.nvm/versions/node/v24.13.1/bin:$PATH" && bun typecheck`
+- `export PATH="$HOME/.nvm/versions/node/v24.13.1/bin:$PATH" && bun run test`
+- `export PATH="$HOME/.nvm/versions/node/v24.13.1/bin:$PATH" && bun run build`
+- `export PATH="$HOME/.nvm/versions/node/v24.13.1/bin:$PATH" && bun run build:desktop`
+
+Result:
+
+- passed
+- `bun lint` still reports the same 8 pre-existing warnings and 0 errors
+
+## Targeted automated coverage
+
+- `cd apps/web && bun run test -- keybindings.test.ts globalThreadSearch.test.ts projectFolderSearch.test.ts quickThreadSearch.test.ts`
+- `cd apps/web && bun run test:browser -- src/components/GlobalThreadSearchDialog.browser.tsx src/components/ProjectFolderSearchDialog.browser.tsx src/components/QuickThreadSearchDialog.browser.tsx`
+- `cd apps/web && bun run test:browser -- src/components/GlobalThreadSearchDialog.browser.tsx src/components/ProjectFolderSearchDialog.browser.tsx src/components/QuickThreadSearchDialog.browser.tsx src/components/ChatView.browser.tsx -t "global shortcut"`
+
+Result:
+
+- passed
+
+## Computer Use QA
+
+Environment:
+
+- Chrome against local dev server at `http://localhost:5737`
+
+Checks:
+
+1. Opened the command palette from the live app.
+2. Confirmed `Search all threads` was present as a top-level action.
+3. Opened `Search All Threads`.
+4. Searched for `new` and verified the current `New thread` result appeared with title highlighting.
+5. Closed the dialog and reopened the command palette.
+6. Confirmed `Search project folders` was present as a top-level action.
+7. Opened `Search Project Folders`.
+8. Verified the live sidebar project appeared as a selectable result.
+9. Selected the result and confirmed the app navigated to a fresh draft-thread route (`/draft/...`).
+
+Result:
+
+- passed for both dialog flows and the project-selection navigation path
+
+Notes:
+
+- Direct shortcut injection through Computer Use remained browser-sensitive, so the authoritative live verification came from opening the dialogs through the app surface rather than relying only on raw modifier chords.
+- The live command palette showed older shortcut hints for `threads.searchAll` / `projects.search`. That was not a checked-in code regression: this machine has saved overrides in `~/.t3/userdata/keybindings.json` and `~/.t3/dev/keybindings.json` mapping those commands to older shortcuts. The checked-in defaults, docs, and tests now expect `Cmd/Ctrl+Alt+F` and `Cmd/Ctrl+Alt+P`.
+- Queue-row `Alt+Up/Down` parity was already covered in `apps/web/src/components/QueuedFollowUpsPanel.browser.tsx`; a fresh manual live pass for that interaction was blocked here because the local Codex provider on this dev server timed out before a runnable queued-follow-up state could be created.

.codex/artifacts/qa/electron-desktop-rebuild.md

@@ -0,0 +1,65 @@
+# Electron Desktop QA
+
+Date: 2026-04-17
+Branch: `codex/rebuild-feature-rollout`
+App: `ClayCode (Alpha)` built desktop bundle
+State dir: `/tmp/t3-electron-qa-state`
+
+## Build + launch
+
+- Passed `bun run build:desktop`
+- Passed `bun run test:desktop-smoke`
+- Launched the built Electron app via `apps/desktop` `bun run start`
+
+## QA inventory
+
+- Desktop app launches from the built Electron bundle
+- Project onboarding works in Electron
+- Draft-thread creation works in Electron
+- Snippet picker opens and inserts a snippet into the composer
+- Quick thread search opens from the keyboard shortcut and navigates to a thread
+- Desktop Connections settings render the Tailnet access row and network-access confirmation dialog
+- Identify any Electron-specific blockers for provider-backed turn flows
+
+## Passed
+
+- Added the repo as a project from the desktop `Add project` flow using `/Users/canal/.codex/worktrees/28c4/t3code`
+- Verified the app created a fresh draft thread route on project open
+- Switched the provider selector from unavailable Codex to `Claude Sonnet 4.6`
+- Opened the snippet picker from the composer and confirmed built-in snippets render in the dialog
+- Filtered/selected the built-in `Write Tests` snippet and verified it inserted into the composer
+- Opened Quick Thread Search with `Cmd+Shift+F`
+- Verified the Quick Thread Search dialog rendered with the expected search/help copy
+- Searched for the existing thread and confirmed navigation back into that thread route
+- Created another new draft thread from the sidebar and verified the route changed to a new `/draft/<id>`
+- Opened `Settings -> Connections`
+- Verified the `Tailnet access` row rendered with the live Tailnet hostname and IP
+- Toggled `Network access` and verified the confirmation dialog opened with `Cancel` and `Restart and enable`
+- Cancelled the dialog successfully and confirmed the settings page returned to its prior state
+
+## Hotkeys
+
+- `Cmd+Shift+F`: passed. Opened Quick Thread Search and navigated back into the existing thread route from a fresh draft thread.
+- `Cmd+Shift+S`: passed. Opened the snippet picker in Electron and exposed the built-in snippet list; I also verified snippet insertion into the composer in the same desktop run.
+- `Cmd+[` / `Cmd+]`: attempted against real draft/thread route history, but no route change was observed in the later Electron session. Because Computer Use key injection became inconsistent for command shortcuts in that session, I am treating this result as inconclusive rather than claiming a definite product regression.
+- `Tab` while a turn is running: passed. I typed `Then list the exact Queue + Steer functions and where each one is called.` during an active `GPT-5.4` run, pressed `Tab`, and verified the queued panel appeared with `1 queued follow-up`, `Ready to dispatch in order.`, and the expected queued row actions.
+- `Enter` while a turn is running: passed. I then typed `Also explain how Steer with Enter differs from Queue with Tab.` during the same live run, pressed `Enter`, and verified it posted immediately as a new live user turn instead of going back into the queue.
+- `Shift+Tab` composer mode toggle: not separately re-verified in Electron once shortcut delivery became inconsistent; this still needs a clean follow-up pass if we want explicit desktop-only evidence for every shortcut.
+
+## Blocked / not fully verified
+
+- Codex provider-backed turn execution is blocked in this Electron run because the desktop app reports `Codex CLI is not authenticated. Run \`codex login\` and try again.`
+- Claude-backed turn execution is also blocked for full happy-path validation because the attempted message failed immediately with `Credit balance is too low`
+- I was able to re-verify live-response behavior in Electron with `GPT-5.4`, including:
+  - successful response rendering from a live provider-backed turn
+  - queueing a follow-up during an active response with `Tab`
+  - steering immediately during an active response with `Enter`
+  - queued follow-up auto-dispatch after the active turn settled
+- I did not re-verify queued follow-up save-from-queue behavior in this Electron pass.
+
+## Notes
+
+- The Electron shell, routing, project onboarding, settings surfaces, snippet picker, quick thread search, and draft-thread navigation all behaved correctly in the built desktop app
+- Live provider-backed turns are workable in Electron through `GPT-5.4`, even though the Codex and Claude providers remained blocked in this environment for separate auth/quota reasons
+- In the successful queue/steer pass, the queued panel disappeared after dispatch and both follow-up user turns were visible in-thread while the assistant continued working, which is the expected user-facing behavior
+- Command-shortcut delivery through Computer Use became inconsistent later in the session, so the sidebar-history hotkeys need one cleaner follow-up check before I would mark them definitively passed or failed in Electron

.codex/artifacts/qa/final-electron-qa-2026-05-02.md

@@ -0,0 +1,55 @@
+# Final Electron QA - 2026-05-02
+
+Branch: `codex/rebuild-feature-rollout`
+
+Workspace: `/Users/canal/.codex/worktrees/28c4/t3code`
+
+Electron QA state:
+
+- App launched from branch with `T3CODE_HOME=/tmp/t3code-electron-qa-home-final-2`.
+- QA project: `/tmp/qa-project-claycode-replay`.
+- App title and sidebar masthead rendered as `ClayCode (Alpha)` / `ClayCode ALPHA`.
+- Electron rebuild completed before the final manual pass, so the visible app used current branch assets.
+
+Manual QA performed with Computer Use:
+
+- Added `/tmp/qa-project-claycode-replay` from the desktop Add project flow in a clean profile.
+- Sent a live GPT-5.4 thread seed and verified a rendered model response.
+- Verified `cmd+k`, `cmd+shift+s`, `cmd+shift+k`, `cmd+shift+f`, `cmd+alt+f`, and `cmd+alt+p` in Electron after the physical-key fallback fix for Option-modified macOS keys.
+- Verified sidebar traversal hotkeys in Electron: `cmd+shift+]`, `cmd+shift+[`, `alt+Down`, `alt+Up`, `alt+shift+Down`, and `alt+shift+Up`.
+- Verified Settings rebrand text and Tailscale/network-access confirmation copy in Electron.
+- Verified Queue + Steer in Electron with a controlled `sleep 60` run:
+  - A running turn showed the stop button and active working state.
+  - The composer displayed `Steer` and `Queue` actions while the turn was active.
+  - Clicking `Queue` created the `1 queued follow-up` panel with `Alt+Up/Down` and `Alt+Shift+Up/Down` row guidance.
+  - The queued follow-up auto-dispatched after the running turn settled and rendered `queued follow-up processed`.
+  - Clicking `Steer` inserted a live steer message during the active turn; the model acknowledged it before the queued follow-up ran.
+- Verified controlled command execution path during Queue + Steer: model ran `sleep 20` and `sleep 60`, then rendered the requested completions.
+
+Issues found and fixed during this QA pass:
+
+- Sidebar masthead still showed legacy `T3 Code`; fixed by rendering `APP_BASE_NAME` in `Sidebar`.
+- macOS Option-modified direct hotkeys such as `cmd+alt+f` and `cmd+alt+p` did not resolve reliably; fixed by adding `event.code` letter aliases in keybinding resolution.
+- Disconnect/reconnect toast copy still hardcoded `T3 Server`; fixed to use `ClayCode Server` via `APP_BASE_NAME`.
+- Electron launcher attempted the renamed launcher path during local start; fixed by only using the renamed launcher when `T3CODE_USE_RENAMED_ELECTRON_LAUNCHER=1`.
+
+Additional gap QA performed after the final Electron session:
+
+- Verified GitHub PR pill rendering in Electron with a local project on branch `codex/fake-pr-qa`, a fake GitHub remote, a local upstream tracking ref, and a temporary fake `gh` shim that returned open PR `#4242`. The header action changed to `View PR`, and the sidebar thread row exposed `#4242 PR open: QA fake PR pill`.
+- Verified disconnect state in Electron by repeatedly killing the embedded server process. The visible toast and accessibility tree both showed `Disconnected from ClayCode Server`, retry countdown text, and a `Retry now` button.
+- Found and fixed a reconnect-success edge case: the success toast could be skipped if WebSocket recovery passed through an intermediate `connecting` state before `connected`. The recovered toast now keys off a remembered prior disconnect timestamp instead of only the immediate previous UI state.
+- Rebuilt and relaunched Electron from the patched branch, repeated the server-kill recovery test, and visually verified `Reconnected to ClayCode Server` with the disconnect/reconnect timestamps.
+
+Residual risks / not fully re-exercised manually:
+
+- GitHub PR pills were verified with a controlled local `gh` shim rather than a live GitHub API response, so the UI path is covered but live credential/network behavior remains dependent on the real `gh` environment.
+
+Automated gates to run after this artifact:
+
+- `bun fmt`
+- `bun lint`
+- `bun typecheck`
+- `bun run test`
+- `bun run build`
+- `bun run build:desktop`
+- `bun run test:desktop-smoke`