AI coding agents: what they do in DevFlow

The general idea

A coding agent pairs a large language model with tools: it can read files, search a codebase, run shell commands and write changes. Instead of answering one prompt, it loops through a goal, some reasoning, tool calls and a synthesis until the work is done. The output is a diff, not a chat reply.

How DevFlow uses them

DevFlow does not rely on one general-purpose assistant. It runs a chain of specialised roles: a planner, a task creator, an implementer, a verifier and a summarizer, plus helpers such as a diagnostician, a build doctor and a conflict resolver. The Go orchestrator passes the structured output of one step to the next as its prompt.

Each one gets its own tool list. The task creator and the summarizer may only read; the implementer may edit files and run commands; the verifier may read and run tests. A reviewing step therefore cannot quietly rewrite the code it reviews.

Which model runs where

Every invocation goes through the OpenCode CLI by default, against the model the workspace selects. Resolution goes from a per-task override, to a workspace override for that role, to a tier (fast, pro or max), to the workspace default. The shipped default is openrouter/minimax/minimax-m3, and each workspace brings its own provider key.

Limits on a single run

A run that never finishes is a cost problem. Each invocation has a hard wall-clock cap of 90 minutes and an idle watchdog that stops a process silent for 45 minutes.

A subprocess never inherits the server environment either. It gets an allowlisted set of variables, so database passwords and OAuth secrets stay out of reach of a prompt-injected command.