[EPIC] Port canonical skills to every supported agent format

## Goal Make `lost-rob0t/skills` a platform-neutral Agent Skills repository instead of an OpenCode-only tree. The durable skill source must live once, under a canonical `skills/<skill-id>/` tree. Client-specific layouts are deployment adapters/views, not separately edited copies. ## Source-of-truth policy - Canonical skill body, metadata, scripts, references, and assets: `skills/<skill-id>/...` - Never maintain divergent copies of the same skill for OpenCode, Claude Code, Codex/`.agents`, Agent Zero, or other compatible clients. - `$HOME/skills` is the canonical local checkout for editing this repository. - Raw imported backups, target data, credentials, client names, hostnames, addresses, scan results, and other private operational state must never be committed. - Imported pentesting/security skills keep reusable techniques, tooling, sequencing, validation, and support code, but target-specific state is stripped before commit. ## Initial supported adapters 1. OpenCode - Home Manager `programs.opencode.skills` 2. Agent Skills generic / `.agents/skills` 3. Claude-compatible `.claude/skills` 4. Agent Zero - export/copy layout compatible with `usr/skills` 5. Additional clients - add only after verifying their current documented skill discovery/format rather than inventing adapters ## Shipped First portability slice merged in #12 (`adbf9714c83237e752d00b16a690ff27e91be30b`). - [x] Define canonical multi-agent source and private-import policy in `docs/formats.md`. - [x] Move every currently exported OpenCode skill under canonical `skills/`. - [x] Remove the obsolete OpenCode-only source tree. - [x] Expose generic `lib.skills` / `lib.skillNames` from the flake. - [x] Keep backwards-compatible `lib.opencodeSkills` / `lib.opencodeSkillNames` aliases. - [x] Add OpenCode, `.agents`, Claude and Agent Zero adapter metadata; add Home Manager modules where the install root is well-defined. - [x] Update README, AGENTS.md and CI for the canonical catalog. - [x] CI verifies flake metadata, canonical names/packages and adapter roots. ## Remaining work - [ ] Import the user's existing Agent Zero skill corpus after local-only sanitization and deduplication. - [ ] Preserve scripts/references/assets when importing skills. - [ ] Dedupe by semantic purpose and content rather than blindly keeping every historical copy. - [ ] Verify additional client adapters against current upstream documentation before adding them. - [ ] Update dotfiles after the import batch with one skills flake input/lock bump. ## Backup blocker The first supplied Agent Zero backup artifact was not a valid ZIP; the backup process returned a timestamp compatibility error instead of archive bytes. Do not persist raw backup contents or private target data while retrying/importing. ## Acceptance - A skill has one durable source tree. - Current OpenCode deployment continues to work from the flake. - At least OpenCode, `.agents`, Claude-compatible, and Agent Zero deployment/export paths are documented and mechanically derivable from the canonical tree. - CI rejects malformed or missing `SKILL.md` files and broken exported catalogs. - No raw target/private data from imported operational backups is committed. - Imported historical skills are generalized, deduplicated, and represented in the canonical catalog before this epic closes.

lost-rob0t commented

2026-08-22 05:32:44 +00:00

(Migrated from github.com)

Goal

Make lost-rob0t/skills a platform-neutral Agent Skills repository instead of an OpenCode-only tree.

The durable skill source must live once, under a canonical skills/<skill-id>/ tree. Client-specific layouts are deployment adapters/views, not separately edited copies.

Source-of-truth policy

Canonical skill body, metadata, scripts, references, and assets: skills/<skill-id>/...
Never maintain divergent copies of the same skill for OpenCode, Claude Code, Codex/.agents, Agent Zero, or other compatible clients.
$HOME/skills is the canonical local checkout for editing this repository.
Raw imported backups, target data, credentials, client names, hostnames, addresses, scan results, and other private operational state must never be committed.
Imported pentesting/security skills keep reusable techniques, tooling, sequencing, validation, and support code, but target-specific state is stripped before commit.

Initial supported adapters

OpenCode
- Home Manager programs.opencode.skills
Agent Skills generic / .agents/skills
Claude-compatible .claude/skills
Agent Zero
- export/copy layout compatible with usr/skills
Additional clients
- add only after verifying their current documented skill discovery/format rather than inventing adapters

Shipped

First portability slice merged in #12 (adbf9714c83237e752d00b16a690ff27e91be30b).

Define canonical multi-agent source and private-import policy in docs/formats.md.
Move every currently exported OpenCode skill under canonical skills/.
Remove the obsolete OpenCode-only source tree.
Expose generic lib.skills / lib.skillNames from the flake.
Keep backwards-compatible lib.opencodeSkills / lib.opencodeSkillNames aliases.
Add OpenCode, .agents, Claude and Agent Zero adapter metadata; add Home Manager modules where the install root is well-defined.
Update README, AGENTS.md and CI for the canonical catalog.
CI verifies flake metadata, canonical names/packages and adapter roots.

Remaining work

Import the user's existing Agent Zero skill corpus after local-only sanitization and deduplication.
Preserve scripts/references/assets when importing skills.
Dedupe by semantic purpose and content rather than blindly keeping every historical copy.
Verify additional client adapters against current upstream documentation before adding them.
Update dotfiles after the import batch with one skills flake input/lock bump.

Backup blocker

The first supplied Agent Zero backup artifact was not a valid ZIP; the backup process returned a timestamp compatibility error instead of archive bytes. Do not persist raw backup contents or private target data while retrying/importing.

Acceptance

A skill has one durable source tree.
Current OpenCode deployment continues to work from the flake.
At least OpenCode, .agents, Claude-compatible, and Agent Zero deployment/export paths are documented and mechanically derivable from the canonical tree.
CI rejects malformed or missing SKILL.md files and broken exported catalogs.
No raw target/private data from imported operational backups is committed.
Imported historical skills are generalized, deduplicated, and represented in the canonical catalog before this epic closes.

2026-08-22 06:50:30 +00:00

Agent Zero follow-up

The shipped #12 slice provides Agent Zero deployment-path compatibility (usr/skills) and an install-root-aware Home Manager helper. It does not yet provide a generated Agent Zero-native metadata view.

Track that separately in #14. Current upstream Agent Zero parses tags and triggers/trigger_patterns and weights triggers strongly during skill search, so #14 will add native discovery metadata without forking canonical skill bodies or abandoning the repo's terse keyword descriptions.

This distinction should remain explicit in docs: target root != client-specific format transformation.

## Agent Zero follow-up The shipped #12 slice provides Agent Zero **deployment-path compatibility** (`usr/skills`) and an install-root-aware Home Manager helper. It does not yet provide a generated Agent Zero-native metadata view. Track that separately in #14. Current upstream Agent Zero parses `tags` and `triggers`/`trigger_patterns` and weights triggers strongly during skill search, so #14 will add native discovery metadata without forking canonical skill bodies or abandoning the repo's terse keyword descriptions. This distinction should remain explicit in docs: target root != client-specific format transformation.

nsaspy referenced this issue

2026-08-27 19:57:33 +00:00

feat(skills): make catalog platform-neutral #12

Rows
Columns

[EPIC] Port canonical skills to every supported agent format #11