Skip to content

Develop - #81

Merged
fseiffarth merged 31 commits into
mainfrom
develop
Nov 16, 2026
Merged

fseiffarth merged 31 commits into
mainfrom
develop

Conversation

@fseiffarth

Copy link
Copy Markdown
Owner

No description provided.

fseiffarth and others added 30 commits November 16, 2026 12:18
The crate compiles twice for cargo test, as the library and as its test
harness, and cargo runs the two at once: about 7 GB and 10 GB of compiler
memory side by side on a runner with 7 GB. The macOS backend test step spent
38 minutes on five minutes of compiler work. test-macos now sets
CI_CARGO_SPLIT_BUILD, and ci-failed-first.sh then builds the harness first,
alone; the full build finds it fresh. Dependencies still build in parallel.

CI-Only: macos

Co-Authored-By: Claude Fable 5.2 <[email protected]>
…er the chrome

The ring is the nearest visible control's own outline and the connector is
drawn in the card's box beneath the sticky header and bottom bar, so neither
is redrawn a frame late over the chrome; layout changes re-measure once per
frame.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…weep

ssh probes and connects are bounded (15 s / 30 s) and a timed-out run reads
as unreachable without retrying other credentials. The sweep runs six probes
at a time, skips busy/usage logins on hosts it found stale, and a late
connect answer never overwrites a newer lamp action. A working login clears
the stale mark; disconnect leaves job teardown to the machine.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…edge

2.1.325 keeps the pane's last column free when it wraps a long option, so
wrapsLabel counts that column as no room. A notice too long for the pane
takes the dialog edge's whole row down to one stroke; LABELLED_DIALOG_EDGE
reads that as the edge so the /model sheet keeps its title. prompt.py takes
extra widths to hold the wrap.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
….0.112 as verified

Verified versions and the newer-version notice move up; the update checklist
gets each re-check, and copilot/chat_send.py (new) sends prompts with the
Chat's and the phone's exact writes into a real Copilot.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Antigravity tabs resumed with --continue, the folder's newest conversation,
so two tabs of one folder reopened the same chat. A PreInvocation hook
(services::antigravity_session, registered under its own name in each
home's .gemini/config/hooks.json, gated on 1.5.2+) now records the tab's
own conversation in live_sessions/<uid>, skipping an `agy -p` run from the
tab's shell tool (ANTIGRAVITY_AGENT) and subagent calls. With a launch
note and the hook's proof, google_chat::resolve relaunches the recorded
conversation, starts a never-prompted tab fresh, and the Reader and the
Resume menu's This chat mark follow the same choice; a tab from before the
hook keeps today's rule.

Probed offline on 1.5.2-1.5.9 (scripts/verify-cli/antigravity/hooks.py,
stub_gemini.py now forces tool calls). Not run live.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
A UserPromptSubmit hook in each home's .config/muse/settings.json records
the session each prompt of the tab's own Muse goes into, in the tab's
live-session record. Muse hands hooks an allow-listed environment, so the
script reads the tab's variables from the parent Muse (/proc, ps -E without
it); a nested muse exec is refused by MUSE_SESSION_ID (1.4.15) or a Muse
above it (1.3.0). Registration merges only schema_version 1 settings,
atomically, for a probed Muse >= 1.3.0; nothing is registered on Windows.

resolve follows the plan's table: a launch note per spawn makes the record
authoritative, a noted tab with no record starts fresh once the hook's
proof exists, and resume --last is kept only where a prompted session
exists (Muse skips unprompted ones and exits). The Reader reads a noted
tab's record first.

Probed offline on 1.3.0 and 1.4.15 with scripts/verify-cli/muse/resume.py.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
… table corrected

OpenCode and Vibe already resume each tab's own conversation; the README
table and the guide said folder's latest. Adds the plan, the manual tests
and the hook survey for the follow-up run.

Co-Authored-By: Claude Fable 5.2 <[email protected]>
The single watch slot meant the last watch_dir won and any tree's unwatch_dir
ended the watch for every other tree. Box sections, popouts and Files tabs
mount several trees at once, so each folder now keeps its own watcher, released
by path; the oldest is dropped past a cap of 64.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…mtime

A same-sized doc written within one clock tick matched both, so a replaced file
read as unchanged. Take the volume serial and NTFS file id from the open handle.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…ut a mouseleave

The fullscreen header sliding away, the pointer leaving the window or the strip
scrolling a tab off left the portaled card hanging. useHoverAnchorGuard closes
it on the next pointer move elsewhere, pointer-out or window blur; status-bar
and tab-strip cards (main and popout) use it.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
HeaderBar draws its run from APP_ICONS alone, so alerts, Mobile, VPN, Machines,
resources and the dev-build chip vanished from the bar once dragged out of the
status cluster. Move them into APP_ICONS; StatusCluster keeps only the members
that need its own state. A test pins that every draggable id has a node.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…d installs missing ones

Connected project hosts are probed for the agent binaries (probe_remote_agent_binaries,
never dialing a disconnected host) and the menu lists a CLI only where it is
installed. A missing CLI installs in a visible terminal from a submenu
(install_agent_project_command).

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Uses the Agents view's default order, reusing the Needs-you activity poll; a
hand-arranged order still wins and agentless projects keep the host's order.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…rified key

Unknown CLIs no longer get a guessed Shift+Tab or /model. Mode and model
controls are gated per family on phone and desktop Reader; the Open Code and
agy spellings resolve to their families and OpenCode's slash catalog lists the
commands its mini interface accepts.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The Reader treated it as the Clear key and replaced the location picker.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…ardening

Side panel: one collapsible section per member, any number open; header ⋯
folds the others or opens all; search covers every open member. Box sheet
applies changes at once; box pill and list rows get a ⋯ menu; ✕ on a member
removes it with an Undo toast; a member pill has Close project in a box.

Inside a box activeId stays on the project the box was entered from, so side
panel, tab jumps, the steering ring and header search compare against the tabs
scope. Box writers edit under the file lock, folder is only authored by
ensure_box_folder, link manifests and generated agent docs are treated as
untrusted, imported project ids that imitate a box scope are not kept, and
dissolving a box closes its tabs and drops its sessions, prompts and schedules.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
…hooks

Antigravity's hook now also runs on Stop and SessionStart: PreInvocation
writes working, Stop (any end: no tool call left, a failed model call) done,
and a new conversation's SessionStart injects the send hint as a system
message. The event is the script's argument, since 1.5.2 and 1.5.3 put no
hookEventName in the payload. Probed offline on 1.5.2-1.5.9.

Muse's hook now also runs on SessionStart (the send hint as
additionalContext), Stop and, from 1.4.15, Interrupt (async; done) and
SessionEnd (idle); UserPromptSubmit writes working. A subagent's events are
dropped by session id, and the nested-run rules hold for every event.
Probed offline on 1.3.0 and 1.4.15 with the echo provider; approval,
notification and API-error events were not reachable and stay unregistered.

Registration migrates the entries the resume work wrote. New probes:
hooks.py turns, muse/turns.py; stub_gemini gains STUB:FAIL, STUB:SLOW and
marks. Not run live.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Copilot gets a turn hook of its own (copilot_session/turn.sh, its own
.copilot/hooks file, gated on the probed 1.0.112): working on a prompt and a
finished tool, decision on a shown dialog's notification, done on agentStop
and a failed model call, idle at the session's end. Only the session the tab
is pinned to counts, and a copilot -p from the tab's shell tool
(COPILOT_AGENT_SESSION_ID) is refused; the hint's session record now
refuses it too.

The OpenCode plugin writes the same record from the prompt and the bus
(V1 event hook, V2 event subscription), for the recorded session only. V2
gives the shell tool a per-session copy of the environment, so the plugin's
claim never reached a nested `opencode run`, which took the tab's session
record over: the claim now goes into every shell (shell.env, V2
create.before).

Probed offline (scripts/verify-cli/{copilot,opencode}/hooks.py) on Copilot
1.0.112 and OpenCode 2.2.7 / 2.3.0; OpenCode 1.x not run. Never live.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ooks

Offline probes of the real binaries (scripts/verify-cli/*/turn_hooks.py,
no login, stub models) showed the gaps the turn record had:

- Claude 2.1.325 ends a turn the API refused with StopFailure and no Stop,
  and reports a failed tool with PostToolUseFailure instead of PostToolUse,
  so a working or decision mark had nothing to end it. Both are registered
  (done / working, nothing for an interrupted tool), as are Elicitation
  (decision, at the dialog; its notice is 6 s late) and ElicitationResult.
- Codex 0.169.0 fires PermissionRequest as its approval dialog opens: now
  registered (trust label permission_request, hash matching one Codex
  wrote), so the wait no longer depends on reading the screen. A home that
  trusted the older events gets Codex's "Hooks need review" screen once;
  skipped, the new hook stays inert and the screen reading remains.
- Vibe 3.1.2's TUI fires all its hook types. The session is recorded at
  pre_agent (each prompt) and post_agent, with working / done, post_tool,
  the approval_requested notice and session_end for the rest. A vibe -p run
  from the tab's bash tool carries VIBE_SESSION and is refused; before, its
  post_agent took the tab's record over.

Both hook twins implement the same rule; the POSIX one runs in the tests.
Adds stub_anthropic.py and namespaced Responses tools to stub_openai.py.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…CLI's hooks

The design doc's cross-CLI paragraph, the guide, nine manual tests and the
follow-up run's outcome in the plan.

Co-Authored-By: Claude Fable 5.2 <[email protected]>
A home that trusted the session hooks and skipped Codex's "Hooks need review"
for PermissionRequest reported Enabled, with no way to fix it from the app.
The reported state now counts that hook where it is registered; the rollout
binder's gate stays on SessionStart alone. Also corrects a comment in the
activity store that said Codex has no approval hook.

Co-Authored-By: Claude Fable 5.2 <[email protected]>
… connection

Consent records the SSH endpoint/root (CompletionProjectPolicy.remote); a local
consent never authorizes SSH. The pool keeps the spec each connection was opened
with, and remote buffers sync into an empty local workspace.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
@fseiffarth
fseiffarth merged commit d237af8 into main Nov 16, 2026
26 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant