Codex Skills Library

Codex Skills Library turns repeatable agent behavior into source-controlled folders rather than loose prompts or chat memory. The project keeps each skill portable with trigger metadata, instructions, references, scripts, examples, assets, and agent metadata, then uses generated catalogs to make the suite inspectable. The important decision is operational: workflows should be reusable and reviewable, but still checked inside the target repository where local instructions, constraints, and verification commands matter.

The repository separates skill membership, lifecycle metadata, generated catalog output, plugin backups, and public docs-site pages. Catalog scripts detect drift before regeneration, while the maintainer decides whether a local skill belongs in the curated suite or should stay outside the public catalog. That distinction matters because the repo is public-learning-oriented but still preserves provenance-sensitive skill folders, third-party notices, local plugin backups, and unresolved license boundaries.

This intake records the repo locally as a Hyphenomenon project without mutating the production knowledge database. The captured state includes a local docs-site review, source-backed hydration artifacts, and an explicit drift finding: an in-progress hyphenomenon-chat-intake skill folder exists outside skills/SUITE_SKILLS.txt. That is useful proof of the real operating model because the project depends on human/operator review before deciding whether local skill work becomes curated public catalog content.

The September 9 public-presence review found SKILL.md, name and description in all 94 inspected skill folders, then separated structural validity from portable execution. The recommended next proof was a clean installation and representative task with declared dependencies and personal assumptions removed. Context Layer was the proposed first release candidate; RPS production remained dependent on the operator’s Photoshop, printers and integrations. A portable release was recommended, not verified.

A September 8 Shopify Headless reconciliation provides a concrete Context Layer example: preserve local evidence, integrate upstream, resolve six docs conflicts, renumber the V45 pilot to 1188, refresh lifecycle references and validate the resulting checkpoint. Its source reports 26 upstream commits and 20 local files reconciled at 57f30e43. The generalizable result is a preservation and handoff method, not permission to discard dirty files or reproduce the old branch state.

The public-presence audit also produced a searchable decision review with signed-out evidence and saved-state readback. It kept recommendations separate from operator-authorized visibility changes and did not confuse private builder-profile settings with a public URL.

Queue latest Faire order exposed execution failures despite an existing duplicate-and-resize procedure: incomplete cloud discovery was treated as absence, source images/local files were used instead of the approved cloud master, a clipping warning was bypassed, and queue success was overstated. The September 2 source reports a focused rps-print-order skill patch: Shopify-derived orientation including Faire orders; exact current Creative Cloud folder checks; canvas-first duplication for missing sizes; positive absence proof for new art; saved .psdc identity; visual preview stop conditions; and paused queues unless resumption is authorized. Current skill-source inspection confirms the shared submission procedure's orientation, warning and filename checks. Print new order adds a September 11 accounting lesson: canceled jobs may report printed sheets, reprints must be separated, and missing history on one Mac cannot establish unprinted work across hosts. The reusable improvement is explicit evidence and operator review at each state boundary, with physical output verified separately.