Repository navigation
feat(settings): ship Closura fleet defaults for linking and provider updates - #73
Conversation
4365a54 to
2b48708
Compare
|
Independent review: the one bot round for this PR. Reviewer: a separate Claude Opus 5.5 subagent. It read range Result: no blockers. Linked environments on by default doesn't weaken linking auth: the setting only adds refusals, and approval still needs a pairing code, or a session holding What came out of it:
Head after fixes: |
…updates Closura machines link to each other and pin their provider CLIs, but every new install started with Linked environments off, so links and link cards were refused, and with stock provider update checks on. Fleet machines had both set by hand. Make them built-in defaults: Linked environments on, provider update checks off. A machine that already stored a value keeps it. Two tests that relied on the old defaults now set them explicitly. Model: claude-opus-5-5 Task: implementation Agent: Builder Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d tests These tests flip Linked environments to show that a change reloads tools; with the setting on by default they now flip it off. Model: claude-opus-5-5 Task: fix Agent: Builder Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
They also gate the model catalog download, so say where to turn them on. Model: claude-opus-5-5 Task: docs Agent: Builder Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
f78fcc3 to
f3a4bb2
Compare
|
Build artifacts from the merge commit
macOS: blocked, not built.
Install notes for the fleet rollout:
|
Problem
Kameron, 2026-11-18 about 05:20 ET: "didnt connect mac air. its settimgs require pairing code which needs to be off for all t3 machines. our t3code fork should be opinioated and have batteries included such as our own 'default' settings and auto cloudflare joining (if can be done very simply, otherwise not worth), etc."
Every new Closura T3 install started with Linked environments off. While it is off, the machine refuses links, link cards and calls over a link, so a fresh machine can't join the fleet. Installs also started with stock provider update checks on. The fleet had been fixing both by hand on each machine. ASUS and R720 both carry
enableProviderUpdateChecks: false, and closura-agent-config'st3-linkscomponent turns Linked environments on.Change
Built-in defaults in
packages/contracts/src/settings.ts:linkedEnvironmentsEnabledenableProviderUpdateChecksHow existing machines move: the server drops any value equal to the default when it writes settings. So a machine where someone turned Linked environments off, or left update checks on, under stock defaults has no stored key and takes the Closura default on upgrade. Only a value written by hand that differs from the new default (
linkedEnvironmentsEnabled: false,enableProviderUpdateChecks: true) survives. The reverse holds too: after the first write on this build, the hand-setenableProviderUpdateChecks: falseon ASUS and R720 is dropped as redundant, so a machine rolled back to stock T3 returns to stock defaults (links off, checks on).The cost of checks off: the same switch gates the T3 model catalog download, Antigravity release checks and background installation update lookups. A fresh Closura machine won't see new models until a T3 release bundles them or someone turns the checks on in Settings → Maintenance. ASUS and R720 already run this way.
docs/user/updating.mdnow says so.User docs that told people to turn Linked environments on now say it is on by default. Tests that relied on the old defaults now set them explicitly: the McpHttpServer test for "while the beta is off", the ModelManifest fetch helper, and the two ProviderSessionManager reload tests, which now switch the setting off to exercise a change. The agent-context budget isn't affected by this change: it measures the same with either default, and #72 already refreshed it on the base branch.
Checked, and already what Closura wants, so no change:
requireConnectDeviceApprovalalready defaults to off. Linked environments on doesn't loosen authentication: the setting only adds refusals, and approving a link still needs a pairing code or a session holdingaccess:writeplus a CSRF token. Linking a machine this app is connected to mints the code automatically when the session holdsaccess:write, which every standard and T3 Connect session does. A typed code is only needed for a bare address. Linking still goes through the existing OAuth code exchange.Founderiskameron | jeff.newWorktreesStartFromOrigin: already defaults totrue.scripts/closura/build-*.sh) embed no update feed, so the packaged app never auto-updates to stock. The server never updates on its own.Deliberately left out
balanced. The brief said "turn on background activity".balancedalready runs background refresh, with git fetch every 30s and a provider refresh every 5 minutes, and both ASUS and R720 run it today.performancewould also keep background work going for unfocused clients. But it means a git fetch every 15s and a provider refresh every minute on every machine, including the 8 GB laptop. Changing the profile alone also leaves the legacy interval aliases inconsistent, and two tests caught that. My recommendation is to keepbalanced. It's a small follow-up if Kameron wantsperformance.t3-linksworkstation component (closura-agent-config PR fix(shell): bound PATH repair fallbacks pingdotgg/t3code#831), with the token kept out of git and logs. Re-implementing it as a first-run step would mean shipping or fetching tunnel credentials from the app. The rollout should runt3-linksinstead.idleCompactionare specific to each machine, or are unverified as fleet-wide. No secrets were added.t3when someone asks. Closing that path cleanly touches the CLI, self-update and the client gating. That's worth its own PR if wanted.Verification (ASUS, Node 26.13.1; ASUS load average was about 45)
apps/server/src/providerandpackages/provider-*, plusws,serverSettingsandProviderSessionManager: 171 files.@t3tools/contractstypecheck: 0 errors.vp fmt --checkand lint are clean on the changed files.--no-verify. The pre-commit hook's format and lint steps pass, but itsknip:checkfails on the base branch: 4 files left unused after the upstream workflows were removed (.github/scripts/mirror-play-apk.cjs,apps/web/scripts/verify-built-boot.ts,scripts/resolve-nightly-release.ts,scripts/smoke-cli-archive.ts). This diff doesn't touch them, and the fork's CI doesn't run knip.47007e11f4, which includes feat(inbox): agent-launched threads go to Agents unless the agent picks an inbox #72: budget, ProviderSessionManager, McpHttpServer, ModelManifest, contracts settings and shared settings tests all pass (6 files, 209 tests).git range-diffshows the only change is dropping the budget edit.Model: claude-opus-5-5
Task: implementation
🤖 Generated with Claude Code