Skip to content

gstack-lessons.md

Latest commit

 

History

History
43 lines (29 loc) · 2.46 KB

gstack-lessons.md

File metadata and controls

43 lines (29 loc) · 2.46 KB

Lessons From gstack

This note captures what codex-workflows should learn from /Users/vsumner/src/github.com/garrytan/gstack without turning into a gstack clone.

Core Lesson

gstack's useful pattern is not "many slash commands." It is a layered system:

  • fat skills encode judgment, phase order, and role behavior;
  • deterministic tooling handles stateful or noisy execution;
  • resolvers and generation keep skill docs aligned with source metadata;
  • local learnings compound without bloating the always-loaded prompt.

That matches this repo's direction: Codex-native workflow skills plus small deterministic CLIs only where existing primitives are too noisy.

What To Copy

  • Host-specific adaptation. Codex skills should be generated or written for Codex's actual plugin, skill, memory, and CLI surfaces. Do not make Codex ingest Claude-specific skill internals by default.
  • A bias toward committed, checkable skill artifacts. If plugin skills grow enough to drift, introduce a small template/check pipeline before the drift becomes manual review work.
  • Command metadata as a source of truth. When this repo adds a real CLI, its command definitions should feed examples, docs, and validation checks.
  • Cheap validation first. Parse examples, check paths and commands, and validate generated artifacts before adding expensive agent evals.
  • Append-only learnings for non-obvious operational facts. Search them on demand, include confidence/source metadata, and refuse to log obvious facts.

What Not To Copy

  • Do not port gstack wholesale.
  • Do not create a new CLI that only wraps git, gh, or Codex app/plugin primitives.
  • Do not recreate ~/.codex/skills pollution with broad global installs.
  • Do not add template generation before there is enough plugin-skill drift to justify it.
  • Do not let preambles become a new giant global prompt.

Next Bet

The next useful primitive is probably a Codex-native learning loop, not an orchestration CLI.

A first version should be deliberately small:

  1. Search local learnings only when a skill asks for prior context.
  2. Record learnings as append-only local JSONL with type, key, insight, confidence, source, and relevant files.
  3. Log only non-obvious findings that would save a future session time.
  4. Keep memory-extension consolidation separate from local searchable learnings.

This would complement codex-threads: threads are evidence from past sessions; learnings are curated operational conclusions from that evidence.