fix(server): idle shells stop reloading every project once a minute - #14893
Merged
juliusmarminge merged 1 commit intoOct 2, 2026
Merged
juliusmarminge merged 1 commit into
juliusmarminge merged 1 commit into
Conversation
A shell subscriber answered every project-enrichment change by re-enriching all projects. That re-requested each root whose cache entry had expired, each resolution published another change, and every subscriber reloaded every project again. Once started (the PR sync sweep restarts it each minute), the chain kept enrichment running continuously: about two thirds of the server's trace output on a real install. Build the refresh from the identities the changes already carry, for the changed roots only, so a refresh no longer requests anything. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
vouch:trusted
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
Contributor
ApprovabilityVerdict: Approved at Macroscope's review found this PR approvable — This is a focused one-file performance fix that limits enrichment refreshes to changed projects while preserving the existing client-facing refresh shape and leaving initial/resume loading unchanged. The remaining concerns are test coverage and service layering suggestions rather than demonstrated correctness or compatibility problems. Notes:
You can add or adjust custom eligibility rules. Learn more. |
|
Navigate logical layers of code changes, visualize relationships, and explore their blast radius. 🧰 Additional context used📚 Code guidelines (1)No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository: pingdotgg/t3code/.coderabbit.yaml Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (1)
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review. 📝 WalkthroughWalkthroughShell enrichment refreshes now build metadata snapshots for projects whose workspace roots appear in the received changes. The refresh uses repository identities from those changes and retains the event-store sequence and original changes. ChangesShell enrichment refresh
Priority: ➖ Normal Estimated code review effort: 2 (Simple) | ~10 minutes Change: Bug fix Suggested reviewers: Merge Risk: ⚪ Minimal · up to The scoped refresh has no identified merge-blocking issue; it can proceed through normal checks. Architecture SummaryArchitecture risk: 🔵 Low · up to The change affects 1 system. Changed systems: Architecture concerns Review detailsSystems and components
Before / after behavior
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
juliusmarminge
added
the
macroscope-review
ChatGPT Codex Connector
macroscopeapp
Bot
dismissed
their stale review
October 2, 2026 20:56
Dismissing prior approval to re-evaluate 5865621
| ? [{ ...project, repositoryIdentity: identities.get(project.workspaceRoot) ?? null }] | ||
| : [], | ||
| ); | ||
| return shellStreamItemFromEnrichmentRefresh({ |
Contributor
There was a problem hiding this comment.
This changes backend refresh behavior but has no focused test for the changed-project selection and identity propagation. Could you add a service-level test using test layers for external dependencies that covers multiple projects and verifies only changed roots appear, including a resolved null identity? This requires a separate test file, so there is no single-hunk suggested diff.
Posted via Macroscope — Effect Service Conventions
Comment on lines
+957
to
+964
| const identities = new Map( | ||
| Array.from(changes, (change) => [ | ||
| change.workspaceRoot, | ||
| change.enrichment.repositoryIdentity, | ||
| ]), | ||
| ); | ||
| const snapshotSequence = yield* applicationEvents.latestApplicationSequence; | ||
| const changedProjects = (yield* projects.listShells()).flatMap((project) => |
Contributor
There was a problem hiding this comment.
This adds multi-step project snapshot construction and a persisted shell read directly to the WebSocket subscription. Could you move the changed-project refresh construction into the project/orchestration service that owns shell metadata, leaving this transport to subscribe and emit its result? The move spans files, so there is no self-contained suggested diff.
Posted via Macroscope — Effect Service Conventions
github-actions Bot
added a commit
to davidvanderklay/t3code-flake
that referenced
this pull request
Oct 3, 2026
## T3 Code Orchestrator V2
This nightly is the first build of T3 Code Orchestrator V2. V2 rebuilds the part of T3 Code that runs your agents: how turns start, stop, queue, and resume; how subagents and background work are tracked; how threads move between providers; and how history is saved. **Read "Before you update" if you use the mobile app or OpenCode.**
### The short version
- **Pi** and **OpenCode 2** support
- **`delegate_task`**: an agent can start child agents on any provider or model and wait for their results. This is the way to use several harnesses together.
- **Fork any thread**
- **Switch provider or model mid-thread**, with caveats (see below)
- **The T3 Code MCP**: agents can create, launch, message, wait on, read, search, and interrupt threads
- **ACP Registry**: add any registry agent (Devin, Cline, Kimi, Droid, and more)
- **Cursor** runs on the official Cursor SDK instead of the CLI
- **A real server-side queue**: queued messages survive restarts, and you can steer or queue each message
- **Subagents you can follow**: every child shows up with its model, status, and results
- **Interrupted threads continue after a restart**
- **Auto-resume when usage limits reset**
- **Scheduled tasks**
### Highlights
**Agents that manage agents**
- `delegate_task` lets an agent hand work to a child agent on any provider and model, with its own options and role. Each child runs as its own thread, keeps its own provider-native history, and reports a result. The parent can wait for children or keep working, and it wakes up when they finish, with results batched together. **This is the recommended way to combine harnesses**, for example Claude planning and Codex implementing.
- The **T3 Code MCP** gives agents tools to create, launch, message, wait on, read, search, and interrupt threads. They can also:
- fork and merge back;
- edit and reorder queues;
- answer questions;
- rename threads, regenerate titles, link pull requests, and settle threads;
- manage scheduled tasks and projects;
- hand work off to a worktree;
- list and close previews.
Agents can't approve their own permission requests.
- Attach another thread as context: type `@` and the thread's name, or drag a thread from the sidebar into the composer.
**Fork any thread**
- Fork from any finished run, including failed, interrupted, and limited ones. Forks stay on the same provider by default and use the provider's native fork where it has one (Codex, Claude, Pi, OpenCode 2), so the full history comes along. When you're done, merge the fork's context back into its source.
**Switching providers mid-thread**
- You can change provider, account, model, or options between turns, and queue a switch without stopping the current turn. Per-provider and per-model options are remembered.
> [!WARNING]
> **Switching providers is supported, but it isn't the recommended way to use multiple harnesses.** The handoff is lossy. The new provider gets a budgeted selection of the conversation's messages, but not the previous provider's reasoning, its tool calls and results, or attachments. The agent can look up omitted messages, but the working state the first provider built up doesn't carry over. To get another harness's help, have the agent use `delegate_task` instead. Changing model or options *within* the same provider usually continues the provider's own session, so nothing is lost.
**Subagents and background work you can see**
- Native subagents from Claude, Codex, OpenCode 2, Cursor, and Pi show as child threads in a lineage view, with their model, status, progress, results, and duration. Several subagents fold into one collapsible card, and the mobile app has an agents sheet.
- When background work finishes, the agent wakes up and says which subagent, command, or monitor finished.
- **Stop works on background work too**, including after the foreground turn has finished.
- After a restart, T3 Code tells the agent which background work died, and threads that lost background work can continue.
**More providers, and deeper support**
- **Pi**, including Pi 1.0: native resume, fork, rollback, and steering; skills, extensions, and extension dialogs; T3 Code tools inside Pi; and permission modes Pi enforces itself.
- **OpenCode 2** (2.0.18+):
- native steering, forks, rollback, compaction, and plan mode;
- child sessions as subagent threads;
- T3 Code tools injected into the server;
- recovery after restarts.
T3 Code detects whether you run 1.x or 2.x. OpenCode 1.x keeps the support it had in V1, plus forking.
- **ACP Registry**: search the official registry and install an agent on the server you choose, with shared sign-in across your devices.
- **Cursor** on the official SDK: browser sign-in, model discovery, project skills and rules, images, plans, and resuming local agents.
- Many fixes for **Claude**, **Codex**, **Grok**, and **Antigravity**. Questions, plans, and todos work again, along with compaction and resume. Subagent output stays in the right thread, and approvals route correctly.
**A queue you can trust**
- The queue lives on the server and keeps its order across restarts. After a restart it waits for you to resume it. It also survives usage limits.
- Edit, reorder, remove, and steer queued messages, including their attachments. This works on mobile too.
- Choose whether a follow-up message steers the running turn or queues: **Settings → General → Follow-up behavior**. `Cmd/Ctrl+Enter` does the opposite of your default.
**Limits and recovery**
- Threads stopped by a usage limit show as **Limited**. Resume manually, **resume at reset**, or **snooze until reset**.
- With **Settings → General → Continue threads after restarts** on, interrupted threads continue after an update, crash, or reboot.
- A live context meter, plus usage that stays correct through compaction and session changes.
**Scheduled tasks**
- Create, edit, pause, resume, run, and delete automations on web and mobile, filtered by project and environment. Fixed schedules use the server's time zone.
**Thread details and the sidebar**
- Workspace, Git, scripts, linked pull requests, automations, and lineage now live in a thread-details panel. Merge controls follow the current checks.
- Sidebar status reflects what's actually running: working, waiting, limited, or failed. Settling is handled by the server and accounts for linked PRs, pins, and background work.
- **Working section (beta)**: busy threads stay out of the way until they need you.
- Configurable worktree branch names, per project if you like.
**Faster and lighter**
- Long conversations load in pages.
- Streaming updates are batched, and the app re-renders less.
- Memory use is lower, and mobile scrolls more smoothly.
### Before you update
- **Your threads come with you, but it's a one-way copy.** The first time V2 starts, it copies your existing database and imports your threads. V1 keeps its own copy, and changes made in one version don't show up in the other. See "Moving from V1 to V2" in the pinned issue.
- **The mobile app must be updated too.** V2 servers only accept V2 clients, and the current App Store and Play Store releases are V1. Join the beta for a compatible app:
- **Android:**
1. Join https://groups.google.com/g/t3-code-v2-beta.
2. With the same Google account, open https://play.google.com/apps/testing/com.t3tools.t3code.
- **iOS:** TestFlight. https://testflight.apple.com/join/XgaxaRtd
- **OpenCode users:** V2 is built for OpenCode 2.0.18 and newer, and 1.x keeps limited support. "Limited support" means OpenCode 1.x works in V2 much as it did in V1 (steering, rollback, subagents, approvals, plan mode, compaction, T3 Code tools), and gains forking. The one thing it loses is that an approval or question still waiting when T3 Code restarts is not shown again; answer or stop it before updating. New OpenCode features go to OpenCode 2 only. OpenCode 2 is a separate package (`@opencode/cli`), and it converts OpenCode's shared database the first time it runs. Don't run 1.x and 2.x side by side on one machine.
- **Minimum provider versions:** Codex 0.159+, Claude Code 2.1.280+, Grok 1.0.13+, OpenCode 2.0.18+, Pi 0.80.5+ (1.0 recommended). **Settings → Providers** tells you if a version is too old.
Read this issue for more info: pingdotgg/t3code#14871
### Changed or removed
- The desktop app uses a new browser profile, so you may need to sign in again to websites opened inside the app.
- The separate agents panel is gone. Subagents appear in the lineage view and inline in the thread.
- Workspace, Git, and script controls moved into thread details.
- The token-by-token streaming setting is gone.
- Grok no longer offers Auto-accept edits. Those threads run in Supervised mode.
- Antigravity receives images directly. Other attachments are passed as file paths.
- Subagent threads are read-only. Message the parent thread instead.
We'll be polishing this over the next few days, so expect frequent updates. Please report anything that feels off.
## What's Changed
* fix(mobile): opening a thread no longer crashes Android release builds by @bompus in pingdotgg/t3code#14850
* fix(server): stop a second server from resending Claude turns by @cestercian in pingdotgg/t3code#13295
* docs(triage): leave protected PRs open by @juliusmarminge in pingdotgg/t3code#14852
* feat(orchestrator): introduce new orchestrator by @juliusmarminge in pingdotgg/t3code#2829
* fix(server): Claude MCP tools show their names and icons by @scratchyone in pingdotgg/t3code#14860
* fix(server): released builds keep the OpenCode policy for their own version by @juliusmarminge in pingdotgg/t3code#14863
* fix(server): idle shells stop reloading every project once a minute by @SunkenInTime in pingdotgg/t3code#14893
* fix(server): Claude model changes no longer kill running background agents by @juliusmarminge in pingdotgg/t3code#14726
* fix(server): API-key Codex installs no longer warn on every start by @yordis in pingdotgg/t3code#14903
* fix(server): steers keep the newest next-turn selection by @juliusmarminge in pingdotgg/t3code#14725
* fix(settings): Pi and ACP Registry no longer show an Early Access badge by @juliusmarminge in pingdotgg/t3code#14915
* fix(server): Codex auth tokens are an omitted key, never undefined by @yordis in pingdotgg/t3code#14908
* chore(lint): forbid declaring tests inside for loops by @juliusmarminge in pingdotgg/t3code#14921
* fix(desktop): markdown no longer loads invisible in new dark-mode browser tabs by @shivamhwp in pingdotgg/t3code#14727
* test(server): the fake Claude CLI exits with its parent by @yordis in pingdotgg/t3code#14926
* fix(mobile): production OTAs no longer reach the previous major's store binaries by @juliusmarminge in pingdotgg/t3code#14944
**Full Changelog**: pingdotgg/t3code@v0.0.45-nightly.20261002.2595...v0.0.46-nightly.20261003.2610
Upstream release: https://github.com/pingdotgg/t3code/releases/tag/v0.0.46-nightly.20261003.2610
github-actions Bot
added a commit
to omarcresp/t3code-flake
that referenced
this pull request
Oct 3, 2026
## T3 Code Orchestrator V2
This nightly is the first build of T3 Code Orchestrator V2. V2 rebuilds the part of T3 Code that runs your agents: how turns start, stop, queue, and resume; how subagents and background work are tracked; how threads move between providers; and how history is saved. **Read "Before you update" if you use the mobile app or OpenCode.**
### The short version
- **Pi** and **OpenCode 2** support
- **`delegate_task`**: an agent can start child agents on any provider or model and wait for their results. This is the way to use several harnesses together.
- **Fork any thread**
- **Switch provider or model mid-thread**, with caveats (see below)
- **The T3 Code MCP**: agents can create, launch, message, wait on, read, search, and interrupt threads
- **ACP Registry**: add any registry agent (Devin, Cline, Kimi, Droid, and more)
- **Cursor** runs on the official Cursor SDK instead of the CLI
- **A real server-side queue**: queued messages survive restarts, and you can steer or queue each message
- **Subagents you can follow**: every child shows up with its model, status, and results
- **Interrupted threads continue after a restart**
- **Auto-resume when usage limits reset**
- **Scheduled tasks**
### Highlights
**Agents that manage agents**
- `delegate_task` lets an agent hand work to a child agent on any provider and model, with its own options and role. Each child runs as its own thread, keeps its own provider-native history, and reports a result. The parent can wait for children or keep working, and it wakes up when they finish, with results batched together. **This is the recommended way to combine harnesses**, for example Claude planning and Codex implementing.
- The **T3 Code MCP** gives agents tools to create, launch, message, wait on, read, search, and interrupt threads. They can also:
- fork and merge back;
- edit and reorder queues;
- answer questions;
- rename threads, regenerate titles, link pull requests, and settle threads;
- manage scheduled tasks and projects;
- hand work off to a worktree;
- list and close previews.
Agents can't approve their own permission requests.
- Attach another thread as context: type `@` and the thread's name, or drag a thread from the sidebar into the composer.
**Fork any thread**
- Fork from any finished run, including failed, interrupted, and limited ones. Forks stay on the same provider by default and use the provider's native fork where it has one (Codex, Claude, Pi, OpenCode 2), so the full history comes along. When you're done, merge the fork's context back into its source.
**Switching providers mid-thread**
- You can change provider, account, model, or options between turns, and queue a switch without stopping the current turn. Per-provider and per-model options are remembered.
> [!WARNING]
> **Switching providers is supported, but it isn't the recommended way to use multiple harnesses.** The handoff is lossy. The new provider gets a budgeted selection of the conversation's messages, but not the previous provider's reasoning, its tool calls and results, or attachments. The agent can look up omitted messages, but the working state the first provider built up doesn't carry over. To get another harness's help, have the agent use `delegate_task` instead. Changing model or options *within* the same provider usually continues the provider's own session, so nothing is lost.
**Subagents and background work you can see**
- Native subagents from Claude, Codex, OpenCode 2, Cursor, and Pi show as child threads in a lineage view, with their model, status, progress, results, and duration. Several subagents fold into one collapsible card, and the mobile app has an agents sheet.
- When background work finishes, the agent wakes up and says which subagent, command, or monitor finished.
- **Stop works on background work too**, including after the foreground turn has finished.
- After a restart, T3 Code tells the agent which background work died, and threads that lost background work can continue.
**More providers, and deeper support**
- **Pi**, including Pi 1.0: native resume, fork, rollback, and steering; skills, extensions, and extension dialogs; T3 Code tools inside Pi; and permission modes Pi enforces itself.
- **OpenCode 2** (2.0.18+):
- native steering, forks, rollback, compaction, and plan mode;
- child sessions as subagent threads;
- T3 Code tools injected into the server;
- recovery after restarts.
T3 Code detects whether you run 1.x or 2.x. OpenCode 1.x keeps the support it had in V1, plus forking.
- **ACP Registry**: search the official registry and install an agent on the server you choose, with shared sign-in across your devices.
- **Cursor** on the official SDK: browser sign-in, model discovery, project skills and rules, images, plans, and resuming local agents.
- Many fixes for **Claude**, **Codex**, **Grok**, and **Antigravity**. Questions, plans, and todos work again, along with compaction and resume. Subagent output stays in the right thread, and approvals route correctly.
**A queue you can trust**
- The queue lives on the server and keeps its order across restarts. After a restart it waits for you to resume it. It also survives usage limits.
- Edit, reorder, remove, and steer queued messages, including their attachments. This works on mobile too.
- Choose whether a follow-up message steers the running turn or queues: **Settings → General → Follow-up behavior**. `Cmd/Ctrl+Enter` does the opposite of your default.
**Limits and recovery**
- Threads stopped by a usage limit show as **Limited**. Resume manually, **resume at reset**, or **snooze until reset**.
- With **Settings → General → Continue threads after restarts** on, interrupted threads continue after an update, crash, or reboot.
- A live context meter, plus usage that stays correct through compaction and session changes.
**Scheduled tasks**
- Create, edit, pause, resume, run, and delete automations on web and mobile, filtered by project and environment. Fixed schedules use the server's time zone.
**Thread details and the sidebar**
- Workspace, Git, scripts, linked pull requests, automations, and lineage now live in a thread-details panel. Merge controls follow the current checks.
- Sidebar status reflects what's actually running: working, waiting, limited, or failed. Settling is handled by the server and accounts for linked PRs, pins, and background work.
- **Working section (beta)**: busy threads stay out of the way until they need you.
- Configurable worktree branch names, per project if you like.
**Faster and lighter**
- Long conversations load in pages.
- Streaming updates are batched, and the app re-renders less.
- Memory use is lower, and mobile scrolls more smoothly.
### Before you update
- **Your threads come with you, but it's a one-way copy.** The first time V2 starts, it copies your existing database and imports your threads. V1 keeps its own copy, and changes made in one version don't show up in the other. See "Moving from V1 to V2" in the pinned issue.
- **The mobile app must be updated too.** V2 servers only accept V2 clients, and the current App Store and Play Store releases are V1. Join the beta for a compatible app:
- **Android:**
1. Join https://groups.google.com/g/t3-code-v2-beta.
2. With the same Google account, open https://play.google.com/apps/testing/com.t3tools.t3code.
- **iOS:** TestFlight. https://testflight.apple.com/join/XgaxaRtd
- **OpenCode users:** V2 is built for OpenCode 2.0.18 and newer, and 1.x keeps limited support. "Limited support" means OpenCode 1.x works in V2 much as it did in V1 (steering, rollback, subagents, approvals, plan mode, compaction, T3 Code tools), and gains forking. The one thing it loses is that an approval or question still waiting when T3 Code restarts is not shown again; answer or stop it before updating. New OpenCode features go to OpenCode 2 only. OpenCode 2 is a separate package (`@opencode/cli`), and it converts OpenCode's shared database the first time it runs. Don't run 1.x and 2.x side by side on one machine.
- **Minimum provider versions:** Codex 0.159+, Claude Code 2.1.280+, Grok 1.0.13+, OpenCode 2.0.18+, Pi 0.80.5+ (1.0 recommended). **Settings → Providers** tells you if a version is too old.
Read this issue for more info: pingdotgg/t3code#14871
### Changed or removed
- The desktop app uses a new browser profile, so you may need to sign in again to websites opened inside the app.
- The separate agents panel is gone. Subagents appear in the lineage view and inline in the thread.
- Workspace, Git, and script controls moved into thread details.
- The token-by-token streaming setting is gone.
- Grok no longer offers Auto-accept edits. Those threads run in Supervised mode.
- Antigravity receives images directly. Other attachments are passed as file paths.
- Subagent threads are read-only. Message the parent thread instead.
We'll be polishing this over the next few days, so expect frequent updates. Please report anything that feels off.
## What's Changed
* fix(mobile): opening a thread no longer crashes Android release builds by @bompus in pingdotgg/t3code#14850
* fix(server): stop a second server from resending Claude turns by @cestercian in pingdotgg/t3code#13295
* docs(triage): leave protected PRs open by @juliusmarminge in pingdotgg/t3code#14852
* feat(orchestrator): introduce new orchestrator by @juliusmarminge in pingdotgg/t3code#2829
* fix(server): Claude MCP tools show their names and icons by @scratchyone in pingdotgg/t3code#14860
* fix(server): released builds keep the OpenCode policy for their own version by @juliusmarminge in pingdotgg/t3code#14863
* fix(server): idle shells stop reloading every project once a minute by @SunkenInTime in pingdotgg/t3code#14893
* fix(server): Claude model changes no longer kill running background agents by @juliusmarminge in pingdotgg/t3code#14726
* fix(server): API-key Codex installs no longer warn on every start by @yordis in pingdotgg/t3code#14903
* fix(server): steers keep the newest next-turn selection by @juliusmarminge in pingdotgg/t3code#14725
* fix(settings): Pi and ACP Registry no longer show an Early Access badge by @juliusmarminge in pingdotgg/t3code#14915
* fix(server): Codex auth tokens are an omitted key, never undefined by @yordis in pingdotgg/t3code#14908
* chore(lint): forbid declaring tests inside for loops by @juliusmarminge in pingdotgg/t3code#14921
* fix(desktop): markdown no longer loads invisible in new dark-mode browser tabs by @shivamhwp in pingdotgg/t3code#14727
* test(server): the fake Claude CLI exits with its parent by @yordis in pingdotgg/t3code#14926
* fix(mobile): production OTAs no longer reach the previous major's store binaries by @juliusmarminge in pingdotgg/t3code#14944
**Full Changelog**: pingdotgg/t3code@v0.0.45-nightly.20261002.2595...v0.0.46-nightly.20261003.2610
Upstream release: https://github.com/pingdotgg/t3code/releases/tag/v0.0.46-nightly.20261003.2610
IanWorley
added a commit
to IanWorley/t3code
that referenced
this pull request
Oct 3, 2026
* fix(web): stop clipping the bottoms of diff file names (pingdotgg#14375) * fix(web): name the step that registers a mobile client (pingdotgg#10963) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: shivam <91240327+shivamhwp@users.noreply.github.com> * test(desktop): Keep WSL busy-runtime fixtures visible when sh is bash (pingdotgg#14351) * chore: bump vite-plus to 1.0 (pingdotgg#14462) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(web): show double bolts for Codex Ultrafast (pingdotgg#14479) * docs: define contribution triage policy (pingdotgg#14480) * docs: use explicit contribution triage exemptions (pingdotgg#14485) * fix(server): Grok CLIs older than 1.0.13 are marked broken (pingdotgg#14486) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(clients): hide disconnected environments when adding projects (pingdotgg#14490) * feat: start threads without a project (pingdotgg#13612) Co-authored-by: maria-rcks <maria@kuuro.net> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): Claude /compact no longer ends early and leaves the thread busy (pingdotgg#14497) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(web): Dark+ and Light+ themes import instead of colliding with built-in ids (pingdotgg#14499) * fix(web): make composer suggestions usable with screen readers (pingdotgg#10154) Co-authored-by: Khai Shern, Toh <khaishern8235@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * perf(release): build and publish npm platform packages concurrently (pingdotgg#14028) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * perf(ci): run PR checks side by side and balance server shards (pingdotgg#14025) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * perf(release): trim Windows packaging setup (pingdotgg#14037) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * perf(release): start Windows builds without waiting for the Linux job (pingdotgg#14027) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * perf(release): build Vercel deployments early and go live after publish (pingdotgg#14029) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): Claude subagents with their own model no longer show the parent's model (pingdotgg#14540) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * feat(web): restart the agent session from cmd+k to load new skills and plugins (pingdotgg#14542) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(web): hotkey settings record plain keys and Tab (pingdotgg#14548) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(release): Windows CLI smoke test no longer fails on temp dir cleanup (pingdotgg#14553) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * feat: start a new project from just a name (pingdotgg#14527) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(desktop): agent clicks in the browser no longer pop Save dialogs (pingdotgg#14573) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * chore(mobile): upgrade to Expo SDK 58 and React Native 0.88 RC (pingdotgg#12045) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * feat(mobile): render the Android subscription widget with expo-widgets (pingdotgg#12046) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * feat(mobile): degrade the agent Live Activity once its content goes stale (pingdotgg#12047) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * feat(notifications): stack agent alerts by thread on both platforms (pingdotgg#12048) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * chore(mobile): drive dev-menu suppression from the dev-client launch URL (pingdotgg#12049) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * fix(mobile): read display scale and width from the view's scene, not UIScreen.main (pingdotgg#12050) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * refactor(mobile): adopt Expo Modules 2.0 for function-only native members (pingdotgg#12051) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * feat(mobile): suppress only the on-screen thread's alert on Android (pingdotgg#12052) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * feat(marketing): replace the screenshot crop with a designed social card (pingdotgg#13487) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * docs: update user count in AGENTS.md (pingdotgg#11413) * test(web): remove duplicate sidebar ordering tests (pingdotgg#14558) Co-authored-by: t3-code[bot] <269035359+t3-code[bot]@users.noreply.github.com> * Update marketing stats for GitHub stars and users * fix: cloned projects show their favicon instead of a monogram (pingdotgg#14635) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * docs: server features are services, and handlers stay thin (pingdotgg#14613) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(ci): pin eas-cli so mobile PR previews deploy again (pingdotgg#14659) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): background GitHub polling uses ~74% fewer calls with batched GraphQL (pingdotgg#14673) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(web): show the Docker icon on dockerfile code blocks (pingdotgg#14675) * feat(web): update providers on every machine with one click (pingdotgg#14678) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * feat(web): beta Working section hides busy threads until they need you (pingdotgg#13926) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(web): PR checks status collapses to its icon instead of wrapping (pingdotgg#14700) * fix: remove outdated restart setting update advice (pingdotgg#14710) Co-authored-by: t3-code[bot] <269035359+t3-code[bot]@users.noreply.github.com> Co-authored-by: maria <254055478+maria-rcks@users.noreply.github.com> * fix(web): keep the terminal toggle clear of the last header action (pingdotgg#13426) * fix(web): group composer undo the way the Lexical composer did (pingdotgg#14674) * fix(desktop): markdown pages no longer render invisible in the dark-mode browser (pingdotgg#14601) Co-authored-by: PR Batch Tester <agent@local.test> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(mobile): one No project row at the top of the project picker (pingdotgg#14532) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Remove unused isCloudDebugEnabled and isTerminalDebugEnabled (pingdotgg#14367) * fix(mobile): update Expo 58 to restore widgets and Live Activities (pingdotgg#14734) * fix(mobile): project picker cards match the settings card color (pingdotgg#14756) * chore(mobile): bump app version to 1.4.0 Co-authored-by: codex <codex@users.noreply.github.com> * fix(markdown): keep Windows paths intact in link and image destinations (pingdotgg#12615) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(mobile): opening a thread no longer crashes Android release builds (pingdotgg#14850) * chore(release): prepare v0.0.45 * fix(server): stop a second server from resending Claude turns (pingdotgg#13295) * docs(triage): leave protected PRs open (pingdotgg#14852) * feat(orchestrator): introduce new orchestrator (pingdotgg#2829) Co-authored-by: maria-rcks <maria@kuuro.net> Co-authored-by: Bilal Bakr <62337003+Bil0000@users.noreply.github.com> Co-authored-by: shivam <91240327+shivamhwp@users.noreply.github.com> Co-authored-by: Vitalii Yehorov <vitalyiegorov@gmail.com> Co-authored-by: Jake Leventhal <jakeleventhal@me.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Alex Southwell <saphid@gmail.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Nicholas Wasmiller <derped@mineperial.com> Co-authored-by: PB <poilmb@gmail.com> Co-authored-by: Exotic <118054752+extoci@users.noreply.github.com> Co-authored-by: Yash Singh <saiansh2525@gmail.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: Theo Browne <me@t3.gg> Co-authored-by: Gabriel De Andrade <30420087+gabrielelpidio@users.noreply.github.com> Co-authored-by: Dara Adedeji <76637177+SunkenInTime@users.noreply.github.com> Co-authored-by: scratchyone <scratchywon@gmail.com> Co-authored-by: Dominic Roy <dominic@sdko.org> Co-authored-by: chukfinley <chuk@chuk.dev> Co-authored-by: Primož Ajdišek <bigpod@bigpod.si> Co-authored-by: benthecarman <benthecarman@live.com> Co-authored-by: NaveDanan <nave0712@gmail.com> Co-authored-by: aaditagrawal <103925638+aaditagrawal@users.noreply.github.com> Co-authored-by: Aditya Garud <153842990+yashranaway@users.noreply.github.com> Co-authored-by: Nick Anisimov <n.anisimov.23@gmail.com> Co-authored-by: MacKinley Smith <smithmackinley@gmail.com> Co-authored-by: Yordis Prieto <yordis.prieto@gmail.com> Co-authored-by: t3-code[bot] <269035359+t3-code[bot]@users.noreply.github.com> Co-authored-by: AKolenda <akole779@mtroyal.ca> Co-authored-by: Guillermo Casanova <75276669+Gigioxx@users.noreply.github.com> Co-authored-by: Otavio Salvador <otavio@ossystems.com.br> Co-authored-by: Shirish Pothi <183252392+shirishpothi@users.noreply.github.com> Co-authored-by: Ishaan Kothari <ishaanko.mail@gmail.com> Co-authored-by: Bob Fowler <bob@rjf.ca> Co-authored-by: Anton Bezdenezhnykh <gamer392@yandex.ru> Co-authored-by: ValeraZSD <48602572+ValeraZSD@users.noreply.github.com> Co-authored-by: Ephraim <ephraim39hr14m@gmail.com> Co-authored-by: Ryan Ilano <ryanilano@users.noreply.github.com> Co-authored-by: Alex <me@pixp.cc> Co-authored-by: maco <gosarmarcel7@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Tristan Knight <admin@snappeh.com> Co-authored-by: PR Batch Tester <agent@local.test> Co-authored-by: oliver <97427849+flamboh@users.noreply.github.com> Co-authored-by: kamkm <99585688+Kamkmgamer@users.noreply.github.com> Signed-off-by: Yordis Prieto <yordis.prieto@gmail.com> * chore(mobile): bump app version to 2.0.0 Co-authored-by: codex <codex@users.noreply.github.com> * fix(server): Claude MCP tools show their names and icons (pingdotgg#14860) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): released builds keep the OpenCode policy for their own version (pingdotgg#14863) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): idle shells stop reloading every project once a minute (pingdotgg#14893) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): Claude model changes no longer kill running background agents (pingdotgg#14726) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): API-key Codex installs no longer warn on every start (pingdotgg#14903) Signed-off-by: Yordis Prieto <yordis.prieto@gmail.com> * fix(server): steers keep the newest next-turn selection (pingdotgg#14725) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(settings): Pi and ACP Registry no longer show an Early Access badge (pingdotgg#14915) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(server): Codex auth tokens are an omitted key, never undefined (pingdotgg#14908) Signed-off-by: Yordis Prieto <yordis.prieto@gmail.com> * chore(lint): forbid declaring tests inside for loops (pingdotgg#14921) Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(desktop): markdown no longer loads invisible in new dark-mode browser tabs (pingdotgg#14727) Co-authored-by: PR Batch Tester <agent@local.test> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * refactor(pi): remove legacy ACP connection * fix(providers): preserve Kiro and VibeProxy behavior --------- Signed-off-by: Yordis Prieto <yordis.prieto@gmail.com> Co-authored-by: shivam <91240327+shivamhwp@users.noreply.github.com> Co-authored-by: Seth <sethmr21@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Mike Olson <mwolson@member.fsf.org> Co-authored-by: Julius Marminge <julius0216@outlook.com> Co-authored-by: Theo Browne <me@t3.gg> Co-authored-by: maria-rcks <maria@kuuro.net> Co-authored-by: oliver <97427849+flamboh@users.noreply.github.com> Co-authored-by: Andrew Johnson <andrew@johnson5.net> Co-authored-by: Khai Shern, Toh <khaishern8235@gmail.com> Co-authored-by: kamkm <99585688+Kamkmgamer@users.noreply.github.com> Co-authored-by: t3-code[bot] <269035359+t3-code[bot]@users.noreply.github.com> Co-authored-by: Chris Deeming <chris@xenforo.com> Co-authored-by: maria <254055478+maria-rcks@users.noreply.github.com> Co-authored-by: Igor Makowski <56691628+Mnigos@users.noreply.github.com> Co-authored-by: PR Batch Tester <agent@local.test> Co-authored-by: Kriday Dave <technocratix902@gmail.com> Co-authored-by: codex <codex@users.noreply.github.com> Co-authored-by: Dara Adedeji <76637177+SunkenInTime@users.noreply.github.com> Co-authored-by: Aaron Queen <bompus@users.noreply.github.com> Co-authored-by: Cestercian <yashafaid@gmail.com> Co-authored-by: Bilal Bakr <62337003+Bil0000@users.noreply.github.com> Co-authored-by: Vitalii Yehorov <vitalyiegorov@gmail.com> Co-authored-by: Jake Leventhal <jakeleventhal@me.com> Co-authored-by: Alex Southwell <saphid@gmail.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Nicholas Wasmiller <derped@mineperial.com> Co-authored-by: PB <poilmb@gmail.com> Co-authored-by: Exotic <118054752+extoci@users.noreply.github.com> Co-authored-by: Yash Singh <saiansh2525@gmail.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: Gabriel De Andrade <30420087+gabrielelpidio@users.noreply.github.com> Co-authored-by: scratchyone <scratchywon@gmail.com> Co-authored-by: Dominic Roy <dominic@sdko.org> Co-authored-by: chukfinley <chuk@chuk.dev> Co-authored-by: Primož Ajdišek <bigpod@bigpod.si> Co-authored-by: benthecarman <benthecarman@live.com> Co-authored-by: NaveDanan <nave0712@gmail.com> Co-authored-by: aaditagrawal <103925638+aaditagrawal@users.noreply.github.com> Co-authored-by: Aditya Garud <153842990+yashranaway@users.noreply.github.com> Co-authored-by: Nick Anisimov <n.anisimov.23@gmail.com> Co-authored-by: MacKinley Smith <smithmackinley@gmail.com> Co-authored-by: Yordis Prieto <yordis.prieto@gmail.com> Co-authored-by: AKolenda <akole779@mtroyal.ca> Co-authored-by: Guillermo Casanova <75276669+Gigioxx@users.noreply.github.com> Co-authored-by: Otavio Salvador <otavio@ossystems.com.br> Co-authored-by: Shirish Pothi <183252392+shirishpothi@users.noreply.github.com> Co-authored-by: Ishaan Kothari <ishaanko.mail@gmail.com> Co-authored-by: Bob Fowler <bob@rjf.ca> Co-authored-by: Anton Bezdenezhnykh <gamer392@yandex.ru> Co-authored-by: ValeraZSD <48602572+ValeraZSD@users.noreply.github.com> Co-authored-by: Ephraim <ephraim39hr14m@gmail.com> Co-authored-by: Ryan Ilano <ryanilano@users.noreply.github.com> Co-authored-by: Alex <me@pixp.cc> Co-authored-by: maco <gosarmarcel7@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Tristan Knight <admin@snappeh.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Replaces #14707. GitHub closed it when the V2 branch it targeted was merged into main (#2829) and deleted. Same change, now based on main.
Problem
With a client connected, the server reloads every project's metadata over and over even when nothing changed. On my real install (37 projects), these refreshes were 64% of all server trace output. The trace files rotate at 10 MB, so that noise left only about 28 minutes of history, and real failures get pushed out. An Oct 1 incident where three runs stopped recording progress had already rotated away by the time I looked.
Cause
Each project's repository identity is cached for one minute. When a cached entry is refreshed, the enrichment service publishes a change, even if nothing changed. Each shell subscriber answers a change by reloading metadata for all projects (
enrichmentRefreshes→loadProjectMetadataSnapshot→getAvailablefor each project). That reload re-requests every entry that has since expired, and each of those resolves and publishes again. Something that reads project info each minute keeps the chain going: here it's the PR sync sweep, for threads with a linked PR.On my install the call chains show it: about 44k of ~45k
getAvailablecalls came fromloadProjectMetadataSnapshotunder the shell subscription, and the rest fromPullRequestSyncReactor.sweep.Change
enrichmentRefreshesinws.tsbuilds the refresh from the identities the changes already carry, for the changed projects only, and no longer callsgetAvailablefor every project. Each refresh frame has the same contents as before: only the changed projects, with their new identity. The one difference is that a refresh no longer re-requests other projects' metadata. Initial loads and resumes are unchanged.One file, about 30 lines. There's no new state and no TTL change.
Scope and approval
This is a small, focused fix for an obvious performance bug. It touches one server stream, with no contract or client change.
Verification
Live run on a dev build (Windows 11). I measured on the V2 branch just before it merged; the
ws.tsrefresh code is identical on main. One web client is connected, and the dev home has 13 projects: the worktree, 6 git repos with a remote, and 6 plain folders. One thread has a linked PR, so the PR sweep runs each minute. Each window is 5.8 minutes, measured from the server trace with a client connected. For "before" I swappedws.tsback to the base version.getAvailablecalls per minuteThe PR sweep kept running during the "after" window.
Without the linked PR, a fresh dev home shows no churn at all: the first load settles and nothing restarts the chain. That's why it took a PR sweep to reproduce.
Tests:
vp test run src/orchestration-v2/ShellStream.test.ts src/project/ProjectEnrichmentService.test.ts: 27 passed (rerun on main).Typecheck:
vp exec tsc --noEmit -p .inapps/server: no errors (rerun on main).Lint:
vp lint apps/server/src/ws.tsreports only three existing unused-identifier warnings, on lines I didn't change.Not checked: multiple clients and remote connections; per-client cost scales down the same way. There's no new automated test for the cascade, because the shell subscription has no test harness and building one would be larger than this fix.
Diagnosed, implemented and measured by Claude Opus 5.5 in Claude Code, running inside T3 Code.
🤖 Generated with Claude Code