add mini-swe-agent workspace adapter#540
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
zomux
left a comment
There was a problem hiding this comment.
Reviewed in depth (adapter pattern + security). Faithful clone of the amp.js pattern for the mini-swe-agent Python CLI. Spawn-safe (prompt as last argv element, shell:false on POSIX, injection-tested), secrets via child env only + redaction, correct Windows/node resolution (mini is a pip console-script — uses whichBinary + shouldUseShellForBinary, sidesteps the cline CI bug), proper process-group stop/teardown + idle watchdog, and CI-safe tests (stubs _miniBin + redirects sessions dir, so no dependence on a real binary or ~/.openagents portable node). --yolo auto-exec is consistent with the Goose/Copilot precedent. Non-blocking nits: idle-timeout surfaces 'exited with code null' instead of a timeout message; consider pipx install (PEP 668); one dead test param. LGTM.
What
Adds mini-SWE-agent as a Workspace coding-agent adapter.
It follows the existing Cline/Copilot pattern: configured by the Launcher, managed by the daemon, and executed in the user's project directory.
This is not a SWE-bench integration. No dataset browsing, Docker harness, evaluation, worktrees, or
predictions.jsonsupport is included.Key behavior
miniprocess per Workspace message.--yolo --exit-immediatelyin the task workspace.shell: false, argv arrays, and ignored stdin.MSWEA_CONFIGURED=trueonly for the child process to bypass mini's first-run wizard.MSWEA_MINI_CONFIG_PATH.Files
MiniSweAgentAdaptermini-swe-agentin the adapter map and registryTesting
A successful model-backed file modification was not tested because the environment does not have a valid model API key.
Out of scope
CORE_AGENTSbuild-registry.jsissues