pandastack

pandastack

Personal context-aware AI operator OS — one substrate, four runtimes, no vendor lock-in.

I built pandastack to run my own work across multiple AI CLIs without dotdir sprawl. Skills are version-controlled markdown. Personas are replaceable. Context recipes ship as TOML. Same content runs across Claude Code, Codex CLI, Gemini CLI, and Hermes; per-CLI shims handle syntax differences. No data-layer vendor lock-in.

The stack is 26 skills focused on dev, writing, and knowledge workflows, tiered into 23 core (markdown-only, fresh-clone runnable) and 3 ext (publicly installable CLI). Anchored on a personal Obsidian vault as SSOT.

v2.2 philosophy: pandastack ships verbs. The brain (gbrain or your own knowledge store) keeps state. Lifecycle discipline is your job, not the package's. v2.2 dropped the flows/ directory (7 spec files) and moved all personal-tier cadence skills (brief-morning, evening-distill, bird, curate-feeds) to the pandastack-private overlay. The public package is now self-contained: clone + install gives you everything in the manifest.

Stability scope (read this first):

v2 is personal-substrate stable: API, schema, and skill content are stable for the author's daily use. v2.2.0 (2026-05-09) is a scope-tightening release — 38 → 26 skills, 7 → 0 flows. Fresh-clone Core install runs without author hand-holding; verified-user-install count is still 0 because the v2.2 surface has not been validated by external A-class users yet.

What this means for you:

• If you are the author or a fork-and-learn power user, v2.2 is stable for daily use.
• If you are a fresh A-class user (Obsidian + Coding Agent power user willing to bring your own vault and CLIs), bash scripts/bootstrap.sh reports what runs now and what install steps remain. Core (23 skills) should run on a clean clone.
• If you use Logseq / Roam / Notion instead of Obsidian, skills will reference Panda's vault conventions (Blog/_daily/, Inbox/ship-log/, etc.) — these are prompt defaults, not hard-coded interfaces. You'd adapt them per session or by editing skill text. There's no built-in adapter layer; whether that matters depends on your tolerance for hand-tuning conventions.

Who this is for:

• Multi-CLI users who want the same skills across Claude Code, Codex CLI, and Hermes
• Vault-centric operators building on Obsidian (Tier=personal skills additionally need private CLIs: gog, bird, pdctx)
• Personal-OS builders who want substrate-first architecture instead of dotdir sprawl
• v1 dogfood reality: 1 user (the author). Verification window for v1.3.0+ structural fix opens now.

Quick start

git clone https://github.com/panda850819/pandastack.git
bash pandastack/scripts/bootstrap.sh --claude    # or --codex

bootstrap.sh reports:

• substrate state (~/.agents/AGENTS.md only)
• 23 core skills runnable on this clone with no external CLI
• 3 extension skills with the exact brew install / npm install -g to enable each
• Personal-tier skills are no longer in this manifest — they live in the pandastack-private overlay (private CLIs: gog, bird, feed-server)

After install:

1. /pandastack:init once inside your project
2. /office-hours — bring a fuzzy idea, walk out with a written brief
3. /sprint — 1-2h focused execution, ends in SHIPPED / PAUSED / FAILED
4. /ship knowledge <path> on a finished note in your vault
5. Stop there. You'll know if pandastack fits how you work.

Tier model (v2.2): Skills are tiered in plugins/pandastack/manifest.toml. Core = markdown-only, runs on a fresh clone. Ext = needs a public CLI install. Personal-tier skills (private CLI dependencies) live in the pandastack-private overlay, not in this manifest. capability-probe only ABORTs when substrate is missing, not when ext CLIs are absent.

Lifecycle map

Which skill do I run when? This is the 30-second answer. pandastack v2.2 has 3 documented compositions; everything else is ad-hoc skill chaining you decide on.

                  pandastack lifecycles (v2.2)
                  ────────────────────────────

  dev        DEFINE ──▶ PLAN ──▶ GATE ──▶ BUILD ──▶ VERIFY ──▶ REVIEW ──▶ SHIP
             office-h   /plan   careful  build     qa         review     ship
             or grill                                          (codex
                                                              cross-check)

             Express path: /sprint chains DEFINE → SHIP internally (1-2h cap).

  writing    CAPTURE ──▶ STRUCTURE ──▶ DRAFT ──▶ SHIP
             direct      write          write     ship write
                         (slop check)             (Close + Extract + Backflow)