T3 Code fork: shipping my own fixes before upstream
Case study: my T3 Code fork merges upstream with real merge commits and ships my own fixes first, like a steer delay cut from 238.9 s to 3.1 s.
The problem: my complaints arrive faster than upstream ships
I use T3 Code to run Claude Code and Codex. It is also someone else's project, with about 3,500 open issues and pull requests. When something bothers only me, I can file an issue and wait, or I can fix it and run my own build. The upstream README says it plainly: "If you don't like something, fork it."
One problem pushed me to act. When I sent a message while Claude was mid-turn, running a long command in the foreground, the message sat unread until the command ended. In real runs that took 238.9 and 244.0 seconds. T3 Code already sends a steer message as "now", but Claude Code only lets a running web fetch step aside for it. A running shell command holds it. Upstream already moves subagents out of the way. It does not do that for commands.
The second problem showed up the same day. I run Codex through my own CLI proxy, and in that setup Codex lists only the models built into the installed Codex CLI, 0.169.0 on my machine. GPT-6.2 Sol and GPT-6.2 Astra ship in Codex 0.169.1, which Homebrew had not published yet. The app said "up to date" and hid two models my proxy already served.
How the T3 Code fork stays mergeable
T3 Code is a monorepo with an Electron desktop app, a web app, an Expo mobile app and a server that runs the agents. I created the fork in late October 2026, and for about three weeks it never merged upstream. The first merge landed on November 17, 2026.
Since then amaraa-mode follows upstream main with real merge commits, not a rebase, so git knows what is already merged and each sync carries only the new commits. scripts/upstream-sync.sh refuses to run on a dirty working tree, fetches both remotes, and prints Up to date when upstream has nothing new. Otherwise it creates a sync/upstream-<date> branch, starts the merge and lists the conflicts.
The conflict rule is written down in docs/upstream-sync.md: keep every fork feature, unless upstream now ships the same thing. Then I take upstream's version and delete mine.
After the merge, the sync has to pass vp run check:changed, vp run i18n:check and vp run fork:check, plus a separate typecheck in apps/server, because a type error there can surface as unknown in files far from the cause. fork:check runs the tests for each fork-only feature and reports one line per feature.
A sync pull request always lands on the red merge lane, because it changes contracts, migrations and dependencies at once. When every other CI job passes and one review finds nothing that must be fixed, I merge it with an admin override. That review step matters to me: code is cheap now and review is the bottleneck.
Fixing the steer delay: move the command to the background
PR #57 changes what happens when a person types while Claude runs a foreground command. The Claude adapter first moves that command to the background, the same as pressing Ctrl+B. Claude reads the message at once and answers it. The command keeps running, and its result comes back later as a background-task notice.
This applies only to messages a person types in the web or mobile app. Messages from an agent, a schedule or a notice are left alone.
I timed it with real runs on Claude Haiku 5.5 against a throwaway data folder, from send to first reply:
| Case | Before | After |
|---|---|---|
| Turn stuck in a 4-minute foreground command | 238.9 s | 3.1 s |
| Turn waiting on a background command | 3.1 s | 3.2 s |
| Turn just restarted by a background notice, then a 4-minute command | 244.0 s | 2.7 s |
| Two prompts in a row | same as row 1 | 4.5 s, then 3.3 s, answered in order |
| Codex (GPT-6-Luna) running a 4-minute command | 1.1 s | no change needed |
Codex already answered in about a second, so I left it alone. One case stays slow: a foreground MCP tool call, such as a long wait on another thread, still holds the message, because Claude Code cannot move those to the background. For that case the composer's Stop button ends the turn at once.
The whole fix is 137 added lines across 8 files.
New models without waiting for a release
PR #58 handles the Codex proxy problem. For Codex, the model list comes from the installed CLI, not from the manifest T3 Code ships. An open upstream pull request that adds the models to the manifest would not have helped my setup.
Now a Codex instance with its own model provider also asks its proxy's /v1/models endpoint and offers newer models in a line the CLI already lists (Sol, Astra and so on), copying settings from the closest model in that line. A line the CLI has never listed stays hidden. Once the CLI lists a model itself, the fork adds nothing, so no model shows up twice.
The fork also keeps its own model file, model-manifest.fork.json. The app reads it from amaraa-mode every hour beside upstream's list. Upstream entries win unless the fork file is newer, and a bundled copy covers offline use. vp run models:sync fills the file from open upstream pull requests that add models, so I commit what it changes and skip the release.
I checked it on a throwaway data folder with only the proxy Codex setup copied over. Both models appeared on Codex 0.169.0, and a real turn on each completed.
The same PR changed the desktop update check from every 4 hours to every 30 minutes, plus a check when the Mac wakes. The fork releases several times a day, and timers stop while a Mac sleeps.
A Code map sized by what changed
PR #56 rebuilt the Code map tab. The old map sized boxes by how big the repository is, so untouched folders and lockfiles drew the eye and a one-line edit looked like a rewrite.
Now the tab shows only the files the thread changed, and box size equals lines added plus removed. New files are green, changed ones amber, moved ones blue and deleted ones red. Generated files and lockfiles fold behind a button. Risk flags mark database changes, package changes, deleted files, env or CI files, and missing tests.
That PR touched 13 files, with 1,402 lines added and 295 removed. I also checked it on a 305-file thread, which rendered as 9 boxes.
Using the agent to fix the agent
Two of the three PRs name their build setup. PR #58 says "Model and harness: Claude Opus 5.5 in Claude Code." PR #56 says "Opus 5.5 · Claude Code in T3 Code." I run Claude Code through T3 Code to patch T3 Code, and each fix that works improves the tool I use to write the next one.
Lessons
Time the bug. "It feels slow" became 238.9 seconds in one run and 244.0 in another. The after numbers, 3.1 and 2.7 seconds, are how I know the fix works.
Write the sync routine down before the fork drifts. Mine went about three weeks without merging upstream. With a script, a written conflict rule and a per-feature check, the second sync landed within a day of the first.
Some fixes can't wait for someone else's roadmap. The Codex proxy problem only exists in a setup like mine, and the upstream change would not have fixed it. With a fork I fixed it myself and shipped it without waiting for a release.
FAQ
What is T3 Code?
T3 Code calls itself "the open-source control plane for coding agents." It is a desktop, web and mobile app from pingdotgg that runs Claude Code, Codex, Cursor, OpenCode, Grok Build, Antigravity and other agents on your machine. It is MIT licensed.
Why keep a T3 Code fork when you could file issues?
Some problems are specific to one setup, like running Codex through a CLI proxy. A fork lets me fix those myself and still take everything upstream ships.
How do you keep a fork from drifting away from upstream?
Merge upstream often with real merge commits, write down a conflict rule, and test the fork's own features after every merge. My rule: keep the fork's version unless upstream now ships the same thing, then take upstream's and delete mine.
Sources
- pingdotgg/t3code on GitHub: README and repository stats on November 18, 2026
- Amartuvshins0404/t3code on GitHub: PR #56, PR #57, PR #58, docs/upstream-sync.md, scripts/upstream-sync.sh
Cover image: GitHub social preview of pingdotgg/t3code
Related articles
sysone: type-safe REST API and React hooks from one model
Case study: sysone gives you a type-safe REST API and React hooks from one model, with row-level access rules, a 2.6 KB client and support for five databases.
OpenCode 2.0: background service, Code Mode and subagents
OpenCode 2.0 adds one background service for every client, Code Mode for MCP tools, and background subagents. What changed, what breaks, and how to upgrade.
Code Has Become Cheap: AI Review Is the New Bottleneck
AI has made writing code cheaper, but identifying high-quality code, reviewing it, and delivering it reliably are becoming the new engineering bottleneck.