pi-gsd: TypeScript port of GSD for Pi, with deterministic commands and TOON output

[fulgidus]

Howdy, I'm new to GSD but I really like the nimble approach.

Since I'm moving away from opencode to pi, and its many flexibilities, I started with a 1:1 port that evolved into something more.
I write this issue mainly because I think some of these approaches would benefit the main project.

I'd like to share with you pi-gsd (npm, repo).

It's fully compatible with GSDv1.30.0 files (didn't like the idea of not being able to go back to opencode if I so desired), doesn't collide with it, so you can have both installed globally in your system and has a few nice perks that I value a lot; namely:

• Typescript + Zod runtime typings enforcement
• Deterministic templating and parsing
  • This enables some neat tricks, like LLM-free gsd-progress, gsd-next, etc... for instant responses and cost savings
• TOON instead of JSON for LLM-centric gsd-tools calls (~50% fewer tokens)

Here is a full comparison of what I've done (or in the process thereof) until now:

Feature	gsd v1.30	pi-gsd	Details
.planning/ data format	✔️	✔️	100% compatible - projects are portable across tools
Workstreams	✔️	✔️	Full workstream isolation
4 model profiles	✔️	✔️	quality / balanced / budget / inherit
18 subagents	✔️	✔️	Identical agent definitions
57 GSD skills	✔️	✔️	All commands available via pi prompt dispatcher (replaces skill system)
Different skills paths for pi	✔️	⚡	All 57 skills moved to .pi/gsd/ to enable advanced pi-gsd-tools integration
pi harness (.pi/)	❌	✔️	New - GSD installs into pi's config dir
Background hooks (pi)	❌	✔️	TypeScript extension (gsd-hooks.ts) installed via postinstall
Pi session history ingestion	❌	✔️	/gsd-profile-user reads pi JSONL sessions from ~/.pi/agent/sessions/
/gsd-setup-pi onboarding	❌	✔️	Setup skill for bun install where postinstall is skipped (default untrusted behavior)
gsd-tools → pi-gsd-tools CLI	✔️	⚡	Same commands basic signatures as original (gsd-tools) but enhanced
[-o|--output] [toon|json] output	❌	⚡	Token-efficient toon renderer output (or json, if LLM absolutely needs it...)
[-p|--pick] {JSONPath} extraction	❌	⚡	Field extraction from CLI output
TypeScript source	❌	⚡	Full TS port of gsd-tools (~9k lines)
Compile-time type safety	❌	⚠️	Zod schemas defined; in progress (~25 anys remaining)
Runtime validation (Zod)	❌	⚠️	Zod schemas for all .planning/ types; wired to config.json + validate health in progress
Smarter --repair	❌	📃	Schema-driven repair planned
Instant commands (no LLM cost)	❌	✔️	/gsd-progress, /gsd-stats, /gsd-health, /gsd-help, /gsd-next — zero LLM, editor pivot
/gsd-next auto-advance	❌	✔️	Deterministic phase routing, pre-fills editor with the correct next command
Prompt-dispatch for all skills	❌	✔️	54 pi prompt templates — clean autocomplete, arg hints, direct workflow dispatch

Legend: ✔️ done · ⚡ enhanced · ⚠️ in progress · 📃 planned · ❌ not available

Cheers, Alessio

[trek-e]

Cool project — the deterministic command execution approach is interesting. The structured output format (TOON) is a pragmatic choice for Pi's constraints.

If you find specific patterns from pi-gsd that you think would benefit upstream GSD, the best path is to open individual Enhancement issues scoped to specific improvements rather than a broad integration proposal. That way each idea can be evaluated on its own merits.

Thanks for sharing.