Wink Pings

Tencent Open-Sources LoopForge: The Coding Agent Goes into Docker, but Approve Is Still in Your Hands

Assign work in chat, fix flaky tests in a sandbox, and you must approve before a PR. The comments first ask: what credentials does the agent hold while testing, and how many times did the flaky test run?

Tencent has open-sourced a coding agent workspace called LoopForge, with code at github.com/Tencent/LoopForge. Usage is very direct: type “fix the flaky login test and open a PR” into the chat box, and the agent will plan, fix, review, and test, with the process running inside a Docker sandbox. PRs are not merged automatically; you must approve them. It supports Codex, Claude Code, Cursor CLI, and 7 other agents, is self-hosted, and is free.

It targets a very specific gap. Most coding agent demos can change a few lines of code, but they start to fall apart at steps like flaky tests, dependency installation, opening a PR, and waiting for CI. LoopForge puts the Docker sandbox in the middle of the workflow, at least acknowledging that a real codebase is not a clean demo repo. One commenter in the replies said the Docker sandbox is the key point, and that almost every agent demo dies on the flaky test before opening a PR.

Someone in the replies linked to Orchestrator.inc and said it feels quite similar. That project is also a local workspace; it assigns each task an agent, a branch, and a worktree, and includes a board, CI failure reassignment, and a reviewer agent. The interface looks like this:

Orchestrator.inc's session board, with columns for Working, Needs you, In review, and Ready to merge

Orchestrator.inc splits the billing feature across three workers, each using a different coding agent in its own worktree

These images come from Orchestrator.inc; LoopForge's interface does not appear in the source material and is included only as a reference to a similar project. Back to LoopForge: the comment section did not stop at “Tencent open-sourced something too.” Several questions are more worth reading than the feature list.

The first category is sandbox lifecycle. One commenter asked how the Docker sandbox handles dependencies: does it pull a new image every run, or keep a warm state between tasks? Another also asked whether the sandbox starts independently per task or whether all agents share one environment. The answer directly affects two things: whether flaky tests can be reproduced reliably, and how long each fix has to wait. In real projects, dependency caching and image reuse are the dividing line in the experience, and you cannot see that in a demo.

The second category is credentials. One commenter said that before letting an agent open a PR, you need to know what credentials it holds while running tests. Fixing a flaky login test does not require push access to main, and production secrets should not be sitting in environment variables. Self-hosting solves the question of whose machine the data lands on, but it does not automatically solve least privilege. What the agent can touch before approval should be written into configuration. Trust alone is not enough.

The third category is review effectiveness. One commenter pointed out that writing the fix and reviewing the fix happen in the same box, and that this needs watching. Skipping the flaky test can also make the tests pass. Another commenter put it more precisely: if the reviewer only gets the diff, it is just grading the coding agent's own assumptions; the reviewer should be given the failure logs and a command that must turn green. Another asked directly whether the reviewer agent will block the fixer or just rubber-stamp most of the time. This concern is not abstract; “rerun until green” has already happened many times in CI.

Flaky tests themselves were also called out separately. One commenter said a flaky test going green once proves nothing, and the PR should show 50 runs before and after the fix. That is a high bar, but the direction is right. If LoopForge makes test results part of the approval page, humans can at least see how many times it ran and how the failure rate changed, rather than just looking at a green checkmark.

Finally, there is the “you must approve before merge” part. One commenter said that line is doing a lot of work, and most people will just click approve. Another reply was more direct: approval cannot become a human clicking yes on another agent's review. This design has value, but it puts responsibility back on the human. Whether the workflow is usable depends on whether the approval page provides evidence: failure logs, test counts, sandbox configuration, and what permissions the agent holds. If it only gives a diff and an Approve button, the gate will quickly become muscle memory.

Orchestrator.inc's PR timeline, where failed checks and review comments are routed back to Claude Code, and the fix is pushed again

There is also a problem that demos easily obscure: long-task state. One commenter said open-source agent workspaces are interesting, but the real test is whether they can track long-task state in a real codebase, not a demo repo. LoopForge needs to prove that states such as multiple rounds of fixes, CI reassignment, and review-comment reassignment are not lost. Self-hosting and being free are the entry point; what comes next is competing on these boring engineering details.

LoopForge has already provided a usable default workflow: assign work in chat, run in Docker, and have a human approve before a PR. This combination is more substantial than any number of agent names. What is worth watching next is its sandbox lifecycle, permission model, and approval records. If in the end it just makes a human responsible for clicking Approve, then it is not much different from today's PR workflow, except that the colleague writing the code does not drink coffee.

发布时间: 2026-11-18 22:29