opencode-pstack

pstack for OpenCode v2

pstack is an engineering style for agents: playbooks for the work, principles that settle tradeoffs, and verification against the real artifact. This repository is a native OpenCode v2 port of cursor/plugins pstack (v0.15.29, upstream 190d287). It is expressed in OpenCode's own primitives: skills, agents, commands, and config. There is no OpenCode fork, no vendored runtime, and no plugin shim; the files are the port.

What you get

• 56 skills under skills/pstack/<name>/SKILL.md, with their playbooks, references, and scripts.
• 23 subagents under agents/pstack/<role>.md, every one mode: subagent with its own model: line.
• 56 commands, one per skill, under commands/pstack/<name>.md, typed as /pstack/<name>.
• pstack-docs/, the installed guide and the OpenCode adapter.

Install

git clone https://github.com/bitgineer/opencode-pstack opencode-pstack
cd opencode-pstack

node scripts/install.mjs --dry-run   # print the plan, write nothing
node scripts/install.mjs             # -> ~/.config/opencode
node scripts/install.mjs --project   # -> ./.opencode in the current directory
node scripts/install.mjs --check     # exit 0 only if the install matches this tree

Add --dest <dir> to target an explicit config directory. Add --project or --dest to --check as well, so the check reads the same destination the install wrote.

Installing through an agent? docs/AGENT-INSTALL.md is the same procedure written as a runbook: preconditions, the destination decision, the exact output each step should print, and the failure table. It also covers the two traps in this install — a v1 binary shadowing v2 on PATH, and running verify-port.mjs without the destination flag, which makes it inspect a directory with no install and still exit 0.

The installer owns four paths and replaces each whole: skills/pstack, agents/pstack, commands/pstack, and pstack-docs. Nothing else in the config directory is touched, and nothing this port does not own is removed. A re-run after an update converges, so a skill deleted upstream does not survive in the installed copy.

If the OpenCode server was already running, restart it so the new skills, agents, and commands are registered:

opencode service restart

Use

/pstack/setup-pstack    # optional: pick a model for each role
/pstack/poteto-mode     # start work that needs rigor
/pstack/poteto-help     # which skill fits this request
/pstack/<name>          # any skill directly, e.g. /pstack/unslop

A skill's ID is its directory name, so skills/pstack/unslop/ is the skill unslop. The model loads a skill with the skill tool by that bare ID; there is no @ prefix. Slash commands are the path you type, and each one loads its skill and passes the rest of the line through.

poteto-mode sets disable-model-invocation: true, so it never appears in the model's skill list and starts only when you type /pstack/poteto-mode. Every other skill is advertised by ID, which is what lets the model choose unslop, how, why, tdd, or a principle-* skill on its own. The gate fails if any other skill hides.

How it maps to OpenCode

What is different from Cursor pstack

• Role models are agent files, not a rules file. ~/.cursor/rules/pstack-models.mdc becomes the model: line of each agents/pstack/<role>.md, written by /pstack/setup-pstack from opencode models. The agent is the unit of model choice, so the old reasoning-level guidance ("raise the effort a step, drop -fast") has no target and is gone.
• No todo tool. OpenCode v2 has none, so the plan lives at the top of the reply and is updated as each step lands.
• Cursor-only features skip with a substitute. Automations, Bugbot-as-a-service, cloud agents, Custom Modes, /loop, and cursor-team-kit do not exist here. A step that needs one writes Skipped in OpenCode: <step>. <reason>. and continues. Review threads become ordinary GitHub review threads, and cloud agents become worktrees.
• opencode run does not expand slash commands. The TUI sends POST /api/session/<id>/command, and the server writes the command's expanded text into the prompt. opencode run "/pstack/bro ..." passes the string through as ordinary text, so use it for prose prompts and the TUI or the command endpoint for /pstack/<name>.

The full mapping, including every tool name and skip line, is pstack-docs/harness/opencode.md in the installed copy.

Verification

node scripts/verify-port.mjs is the static gate. It checks, and exits non-zero on any failure:

Both gates are reproducible from a clean clone:

node scripts/install.mjs --dest /tmp/oc
node scripts/verify-port.mjs --dest /tmp/oc          # 10 checks passed, 0 failed
node scripts/install.mjs --check --dest /tmp/oc      # check: install matches the port

docs/AGENT-INSTALL.md is the same procedure as a runbook: preconditions, the destination decision, the exact output each gate prints, the live smoke checks (debug agents, a skill load by bare ID, and /pstack/<name> through the command endpoint), the failure table, and a report template. Every claim on this page can be reproduced with a v2 opencode binary and the commands above.

Re-porting

Upstream changes are re-ported and the artifacts regenerated with the pipeline in the private opencode-pstack-maintainers repository. This repository ships only the ported result.

Provenance

Upstream pstack is by Lauren Tan (poteto), MIT licensed, at cursor/plugins/pstack, tag v0.15.29, commit 190d287. This port keeps the MIT license and keeps the original skill text wherever OpenCode has an equivalent; the edits are names, locations, and the passages that describe Cursor-only mechanisms. pstack-docs/ is the upstream guide, adapted to the OpenCode adapter in pstack-docs/harness/opencode.md.

License

MIT. See LICENSE. Copyright (c) 2026 Lauren Tan.