Coding agents in a local AI loop
Windshift's runner gives coding agents a controlled path from assigned work to a reviewable pull request on infrastructure you manage.
A coding-agent demo can begin and end in chat. A production workflow cannot. Teams must decide which repository the agent can touch, where it runs, how it gets credentials, which model it uses, and where people review the result.
Windshift uses an assigned work item as the starting point. The request, comments, linked pages, owner, status, and repository context are already there.
The work item starts the run
Assign the agent a bounded bug investigation, documentation update, cleanup task, or implementation. Windshift queues the run, and the runner checks out the repository, starts a sandboxed workspace, and streams its log to the item. If the agent changes code, Windshift opens a draft pull request.
Not every run needs a commit. An investigation note, blocker report, or review summary may be the right result. Whatever the outcome, it stays attached to the request that produced it.
Start every run from a clean checkout
The Windshift server manages the queue and run state. A runner host executes the job in a new agent container with a per-run checkout.
The container receives the repository and tools configured for that binding. Broad credentials stay outside it, and old state does not carry into the next task. In a small self-hosted setup, Windshift and the runner can share an internal Docker network. Windshift handles model and Git access, so the agent container does not need general internet access for a normal run.
Use cheaper models for routine work
Background tasks do not all need the most capable model. A lower-cost model may be enough to:
Each agent binding selects its own LLM connection. Teams can reserve stronger models for uncertain changes and use cheaper ones for routine maintenance without changing the surrounding workflow.
Teach the agent how the repository works
Repositories rarely contain every working agreement. An agent may need to know which tests to run, how migrations are written, when documentation must change, and how large a pull request should be.
Windshift stores these instructions as managed Markdown skills. Admins attach the relevant skills to an agent binding, giving each run the team's conventions without one large repository-wide prompt.
Read the run log before the diff
Every run records its status and event log. The item detail view shows commands, tool results, failures, and the final summary.
If the run produces commits, Windshift opens a draft pull request and links it to the item. In one example, it opened realigned/windshift-core#14 from `agent-runs/run-45` into `main`.