pstack for OpenCode

pstack is Lauren Tan's set of agent skills for rigorous engineering work: verification-first bug fixing, disciplined delegation, multi-model review panels, and 24 one-rule principles. It was built for Cursor. This repo ports it to OpenCode: the same skills, patched so the subagents, model roles, transcripts, and loops all work in OpenCode v2.

One command installs 72 skills, 5 subagents, and 72 slash commands.

Requirements: OpenCode v2, git, python3, node.

git clone https://github.com/devalade/pstack-opencode.git
cd pstack-opencode
./install.sh

Then:

1. Start a new OpenCode session (or opencode2 service restart if anything looks stale).
2. Run /setup-pstack to map your models onto the poteto-agent and poteto-judge subagents and pick a budget.
3. Start a task with /poteto-mode build <feature>. verify your changes and show me proof.

Targets and upstream can be overridden:

Skills (72). All 54 pstack skills (including the 24 principle-* leaves and the 23 poteto-mode playbooks) plus the 18 cursor-team-kit skills pstack references: deslop, control-ui, control-cli, fix-ci, loop-on-ci, review-and-ship, thermo-nuclear-code-quality-review, and friends. Almost all are disable-model-invocation: true, matching pstack's design: they load when you invoke them, not on the model's whim.

Subagents (5). poteto-agent (code delegates), poteto-judge (read-only review, judgment, and explanation seats), comment-sicko (comment reviewer for /no-comments), ci-watcher (PR check monitoring), thermo-nuclear-code-quality-review (harsh maintainability audit).

Commands (72). Every skill is a slash command: /poteto-mode, /how, /why, /interrogate, /unslop, /tdd, /teach, /create-verification-skill, /principle-prove-it-works, and so on. Skills also load by @skill-id mention or plain request.

/poteto-mode fix this regression. repro first, then root-cause it and prove the fix with runtime evidence.

/poteto-mode i'm going to bed. land the stack even if ci flakes. i want everything merged by morning.

/how do we cancel runs? do we have an n+1 when we look up every run to cancel?

/interrogate review this pr.

New here? Type /poteto-help with a question.

/setup-pstack is rewritten for OpenCode: it detects models with opencode2 models and writes model: lines into the agent files. Panels get their model diversity from poteto-agent vs poteto-judge (add a second writer agent file for multi-model writing races).

1. Fetches cursor/plugins (sparse: pstack + cursor-team-kit) at UPSTREAM_REF.
2. Copies all 72 skills into PSTACK_SKILLS_DIR, replacing existing copies of the same names.
3. Vendors the pstack guide into poteto-help/references/guide/ (with an OpenCode note).
4. Applies the OpenCode rewrites from files/ and patches/ (the patch scripts fail loudly if upstream text moved).
5. Copies the 5 subagents into OPENCODE_AGENTS_DIR.
6. Generates one slash command per skill into OPENCODE_COMMANDS_DIR.

Re-running install.sh is idempotent at the same ref: skills are restored from a fresh clone before patching. Point UPSTREAM_REF at a newer commit to upgrade; if upstream text moved, the patch scripts name the exact string and fail loudly rather than half-applying.

• pstack's automations/benny (Cursor automations) and cursor-team-kit's .mdc rules (Cursor always-apply rules).
• /make-bot-ui is installed as-is, but it is inherently Grok Bot specific.

The skills are Lauren Tan's pstack; the companion skills are cursor-team-kit by Eric Zakariasson and the Cursor team. Both are MIT licensed and vendored here with patches. This is an unofficial community port, not affiliated with Cursor or Lauren Tan.

The port tooling in this repo is MIT. See LICENSE and UPSTREAM.md.