- assertUndoCovers never asks batch-check about a path with a newline (counts as
unsaved), so answers can't shift.
- The workspace-busy probe matches rows by workspace id or worktree realpath and
treats a truncated listing as busy.
- Rewind dialog opens with focus on Cancel; a disabled option reads as disabled.
- Notes: unreadable tracked deletes keep their staged blob; the turn-starting
mark is process-local by design.
- Refuse a restore that would write or delete inside a nested repository, and
verify every on-disk path it touches is in the undo tree first (git add can
skip a path without failing).
- A turn counts as busy from its snapshot until the hand-off; terminal agents
working in the workspace make restores refuse; the preview reports busy and
the dialog disables the files choice.
- W10 copy: statement description, i18n plurals, state-dependent detail line,
option directly under the description, 'The chat was rewound, but…' errors.
- Index restore keeps skip-worktree bits; large-file scan honours the deadline.
- Component fixture for the rewind prompt states.
- A restore that fails partway names its undo point; the error toast offers Undo.
- Refuse when an ignored file sits where the checkpoint needs a folder, or when
a file to overwrite is unreadable (neither could be saved first).
- Busy check runs inside the per-worktree queue and covers every chat in the
worktree; the rewind stays pending until files are back.
- add --ignore-errors exit 1 (unreadable/nested repo) no longer disables
checkpoints; untracked files over 10 MB are left out of snapshots (saved only
if a restore would overwrite them); one deadline per capture.
- Skip-worktree/assume-unchanged entries are left alone; stale temp indexes swept;
retention never prunes the checkpoint just written.
Before each turn a message opens, the host snapshots the chat's worktree as a
stash-like commit under refs/orca/checkpoints/<session>/turn/<item> (working
tree incl. untracked files, staged tree as second parent), via a private temp
index. HEAD, the branch, the user's index and the stash list are untouched.
Rewind to here now offers 'Also restore files (N changed)', on by default when
a snapshot exists. Restore first saves an undo checkpoint (also covering ignored
files it overwrites), refuses on merge conflicts or file/folder clashes before
writing anything, never moves HEAD, and restores staging only while HEAD is
unchanged. The toast offers Undo. Retention: newest 50 turn / 10 undo
checkpoints per chat, nothing older than 30 days.
Measured: capture p50 356 ms on a 40k-file repo, 637 ms on 200k files; restore
0.44 s / 0.78 s.
* feat(cloud): pdx1 default region with iad1 fallback and one golden per region; Standard is 8 vCPU / 16 GB (U17, U18)
- golden_images and environment_builds record their region (0040, backfilled iad1); rotation,
git-source builds, smoke and activation run per region; lifecycle picks the region's golden.
- New creates default to pdx1; a create with no pinned region falls back to iad1 automatically on a
capacity/provider error (VERCEL_FALLBACK_REGIONS). Until pdx1 has an image, creates use iad1.
- Environment builds boot and smoke in their own region; workspaces from a build boot there.
- One size: Standard = 8 vCPU / 16 GB (ORCA_CLOUD_SANDBOX_SPEC). Legacy 'large' requests map to
standard; 0040 rewrites existing large rows.
* merge cloud/integration (ORCA_CLOUD_SANDBOX_SPEC landed in #196)
Hosts with credential rules (model APIs, api.github.com) are answered by Vercel's egress proxy with a
certificate from its CA. Vercel's command env trusts it (NODE_USE_SYSTEM_CA, NODE_EXTRA_CA_CERTS,
SSL_CERT_FILE) but orcad starts under env -i, so Node clients in terminals used Node's bundled roots:
Gemini CLI retried forever ('This is taking a bit longer') while the same call from a command
returned at once. Found on the live proof of the new agents; NODE_USE_SYSTEM_CA=1 in the terminal
fixed it there (1 s to the provider's answer).
Staging sign-in now works end to end in dev without a dev operator key on the user's machine.
Same account ids as billing dev (which trusts staging already); never honored when SANDBOX_ENV=prod;
a staging key cannot pass as the prod issuer.
Security review: a Copilot BYOK key without a usable base URL (or with one from another scope) had
no route, so it went to the VM as itself. Every swappable name now becomes a placeholder, routed or
not, and the Copilot pair resolves like the other agent keys (shared beats personal, as a pair).
The Cursor exchange reply is parsed whatever its content type and cut down to accessToken; headers
(set-cookie) are not passed. opencode.ai is not allowlisted (anonymous share uploads).
Image: pinned opencode-ai 1.19.5, @google/gemini-cli 0.67.0, @github/copilot 1.0.108 (npm) and
cursor-agent 2026.11.05-86047df (checksummed tarball). opencode's unused linux build variant is
dropped (~205 MB). First-run screens are answered at every boot/secret refresh (folder trust for
Gemini, Cursor and Copilot; Gemini's auth picker when a Gemini key is set; OpenCode's gateway model
from ANTHROPIC_MODEL/OPENAI_MODEL); self-update is off for all four.
Egress: GEMINI_API_KEY / GOOGLE_GENERATIVE_AI_API_KEY (generativelanguage.googleapis.com,
x-goog-api-key), CURSOR_API_KEY (api2.cursor.sh, only POST /auth/exchange_user_api_key; the
refresh token is stripped so nothing long-lived lands on the VM) and COPILOT_PROVIDER_API_KEY /
BEARER_TOKEN (host of COPILOT_PROVIDER_BASE_URL, same scope only) are placeholders in the VM and
swapped by the proxy. Allowlist adds the agents' API and sign-in hosts. API/MCP accept the image's
agents; routines stay on claude/codex.
- local.core-workflow: composer local worktree (Blank Terminal), edit in its terminal, diff in
Source Control and the diff tab, Stage All, commit
- acc.download: Explorer Download… (20 MiB file) and Download as ZIP… (folder) with progress and
success toasts and sha256 checks; only the native Save panel's answer is stubbed in main via the
inspector (W9D_MAIN_INSPECT)
- acc.remote-desktop: workspace menu Open remote desktop -> orcaDesktop connected, noVNC canvas
sampled, a red xterm launched on DISPLAY=:1 shows up; guest screenshots over CDP
- local rows without a host badge count as local
Three startup changes, each generic:
- orcad serve starts the profile store's writer thread before its preflight (~40 ms of thread and
module startup), and the store adopts it. The thread waits for its initialization as its first
message; an unclaimed thread is unref'd and discarded after 30 s, and a dead or mismatched one
is replaced by a fresh spawn as before.
- Before the ~100 ms of synchronous module imports, orcad lets the terminal daemon's launch reach
its fork (bounded at 50 ms), so the daemon boots alongside the imports instead of after them.
- The journal's storage-row schemas are built on first use. They cost ~15 ms to construct and
nothing reads journal rows before ready (V8 coverage at ready: journalRowDeclaresMore count 0).
Measured on 4-vCPU VMs from the current golden, n=20 interleaved readyMs:
2.9 GHz host p50 617 -> 572, p95 662 -> 646; 2.5 GHz host p50 882 -> 834, p95 969 -> 905.
W10 wake-visibility P1s. stderr '- …' lines (NDJSON phase events with
--progress json; --json defaults to none; stdout keeps only the result):
- wake: 'Waking <ws>...', one 'Still waking (Ns)…' line at 10 s,
'<name> is awake (took Ns).', or, when the API answers still waking
after its wait, '<name> is still waking. Check with: …' (exit 0, status
waking kept in the result)
- sleep: 'Putting <ws> to sleep...' / '<name> is asleep (took Ns).'
- worktree create --on cloud: 'Still starting (Ns)…' at 10 s and
'Workspace <id> is running (took Ns).'
Lifecycle handler moved to cloud-lifecycle.ts; shared helper cloud-progress.ts.
* feat(w8): reassure after 10 s of waking (files are safe; a queued prompt sends once awake), under the steps with the timer still running
* fix(w8): wake-slow note: iOS-only announce, prompt sentence only while waking, whole-run reassurance from 20 s, fresh per run
A user's own orcad started from a server pane must not share or overwrite the server's
marker, so the variable joins ORCAD_SERVER_LAUNCH_ENV_KEYS. The launcher may name a directory
that does not exist yet; create it (0700) before writing the marker.
The dispatcher now takes either a method list or a loader. The RPC server's default is a loader
that imports the method registry on the first call; orcad preloads it right after its ready line,
so a reconnecting client normally finds it loaded. Concurrent first calls share one load; a failed
import answers those calls with an RPC error and the next call retries.
The legacy relay CLI imported the registry statically, which kept it on orcad's startup path; it
now loads it with its first relayed command (its dispatcher moves to its own file for that).
Measured on a 4-vCPU VM from the current golden, n=20 interleaved: readyMs p50 679 -> 621,
p95 716 -> 665.
The desktop's runtime record kept saying 'running' after a sleep from another
client until lost-contact tripped (minutes), so an open or predictive wake
returned at once without resuming, and the row read awake.
- Wake path: a cloud record still 'running' asks the control plane; when the
sandbox is suspended the record is marked asleep and the resume runs.
- Status reads (window refresh, predictive preflight, the 1 min beat) mark a
running record asleep when the control plane says sleeping, so the row
reads Asleep. Only when the record is unchanged since it was read.
- Quiet (predictive) wakes report whether a resume actually ran; one that
found the VM up is 'skipped': no unopened hold, no backoff, no short idle
window in main.
- A skipped wake (preflight no, or nothing to resume) gives the 30 s
hover/focus gap back.
The wake phase "orcad ready" (W9) runs from orcad's spawn to the control plane
having the pairing offer. After orca-sandbox-start exits, the control plane
made a second Vercel round trip (files API, ~120 ms from here) to read the
offer, because command stdout is kept as logs.
startOrcad now passes a fresh 32-byte key per call in env (env is not kept).
The start script adds "sealedOffer": "<iv>.<base64 AES-256-CTR(offer line)>"
to the pointer line it already prints, so the logs only hold ciphertext and
the control plane opens the offer locally. The key is unset first thing in the
script and in the prelude, so nothing the script starts inherits it. If the
offer is missing or does not open (older image, no openssl, wrong key), the
control plane reads the file as before. Older Workers send no key, and
current images then print the plain pointer.
The ready poll in the start script drops from 50 ms to 10 ms (~20 ms mean),
keeping the 60 s minimum timeout.
Verified on a VM from snap_aQmW, driving the start script like the control
plane: all four starts, including an idempotent second start, opened without
the files API, and the stdout never held the pairing code.