dstack

dstack is a local adaptation of pstack for Codex, Claude Code, and Cursor. It keeps a closed set of pstack skills and dstack-mode playbooks close to upstream, with harness-neutral wording: one skill calls another with "Call the Skill tool with skill-name.", and helpers spawn through the active harness's native subagent tool. PSTACK_SYNC.md lists what is excluded and how weekly upstream changes are absorbed.

The project is under active development. Configuration is schema version 3; a version 2 file is not migrated, so delete or rewrite ~/.dstack/config.json and rerun setup-dstack.

Install

Skills install canonically under ~/.agents/skills. Preview, then install:

python3 install.py --dry-run
python3 install.py

Claude Code also needs the managed compatibility links under ~/.claude/skills:

python3 install.py --with-claude-links

To remove dstack, run python3 uninstall.py --skills-only to keep ~/.dstack, or python3 uninstall.py --all to remove it too. Add --dry-run to preview. Unrelated skills are preserved.

Set up

Invoke setup-dstack explicitly in each harness and repository where you use dstack. It maps the harness to codex, claude, or cursor, asks you to confirm four profiles (fast-explorer, feature-worker, bug-worker, skeptical-reviewer), each a concrete model and effort, and records this repository's transcript directory in ~/.dstack/config.json. There is no environment or command-line override for that path.

Skills read the file only through the dstack-config skill, which stops and asks you to rerun setup-dstack when anything is missing.

Then start a task with /dstack-mode, a goal, and a check that can pass or fail. /dstack-help explains the rest of the pack.