[Hackathon] Support Agent Skills standard for install/setup

[powxenv]

Discord Username / User ID

1169538417365876806

What does this improvement do?

This proposes replacing the CLI's bespoke, per-agent skill-install formats with the Agent Skills open standard. Today the CLI hand-wires 8 targets (claude, cursor, cline, antigravity, codex, kiro, windsurf, copilot) — each with an own-file vs managed-section mode, a per-agent wrap() for custom frontmatter, and special cases like a managed section in codex's AGENTS.md and a size-capped compact body for windsurf. The proposal: install one canonical .agents/skills/<skill>/SKILL.md per the standard, and broaden support to ~70 agents. Shared agents (Cursor, Gemini CLI, Copilot, Codex, Cline, Amp, OpenCode…) read .agents/skills/ directly — one install serves many — and the rest get a symlink back to that single source of truth.

Details / implementation notes

Why agentskills: it's an open standard already adopted by the agents users run, and its client guide calls .agents/skills/ the "widely-adopted convention for cross-client skill sharing," auto-discovered by every compliant client. Adopting it does three things at once — it drives broader adoption of the CLI (works with 72+ agents out of the box), makes the install code easier to maintain (one uniform code path instead of per-agent special cases), and lets the CLI stop handling non-standard rules per agent (the wrap()/mode/compact-body/codex-section machinery all goes away in favor of one standard SKILL.md).

What changes:

• Canonical .agents/skills/<skill>/SKILL.md with standard name/description frontmatter.
• A universal flag per target: shared agents need no symlink; others get one symlink back to the canonical file (no drift between agents).
• agent status checks the canonical file and the symlink landing, with clear status codes.
• --target accepts standard registry ids plus legacy short aliases (claude, kiro, copilot) so old scripts keep working; defaults to claude-code in CI.
• Backup + overwrite semantics for forced reinstall on both canonical and symlinked files.

Supported agents

Universal (18) — read .agents/skills/ directly: Amp, Antigravity, Antigravity CLI, Cline, Codex, Cursor, Deep Agents, Dexto, Firebender, Gemini CLI, GitHub Copilot (copilot), Kimi Code CLI, Loaf, OpenCode, Replit, Warp, Zed, PromptScript.

Symlinked (54): AiderDesk, AstrBot, Autohand Code CLI, Augment, IBM Bob, Claude Code (claude), OpenClaw, CodeArts Agent, CodeBuddy, Codemaker, Code Studio, Command Code, Continue, Cortex Code, Crush, Devin for Terminal, Droid, Eve, ForgeCode, Goose, Hermes Agent, inference.sh, Jazz, Junie, iFlow CLI, Kilo Code, Kiro CLI (kiro), Kode, Lingma, MCPJam, Mistral Vibe, Moxby, Mux, OpenHands, Ona, Pi, Qoder, Qoder CN, Qwen Code, Reasonix, Roo Code, Rovo Dev, Tabnine CLI, Terramind, Tinycloud, Trae, Trae CN, Windsurf, ZCode, Zencoder, Zenflow, Neovate, Pochi, AdaL.

Verified by: https://github.com/vercel-labs/skills#supported-agents

Confirmations

• I have searched existing issues and this is not a duplicate.
• I have provided my Discord identity above for reward coordination.

[powxenv]

i'm open to discussion since it's probably breaking changes ^_^, the draft pr is at #271 so it's clear what i mean and it can be ready to review right away if this feature is wanted. my goal is so all agents can directly use testsprite.

[powxenv]

/assign

[testsprite-hob]

Assigned to @powxenv. Thanks for taking this on. (2 more slots available.)

Your open assigned issues (1/3):
(none)

[zeshi-du]

Maintainer position on the direction — recording it here so #271 has something concrete to aim at:

Direction: accepted. The Agent Skills standard is real, vendor-neutral, and increasingly what agents actually read; converging our 8 hand-wired per-agent formats onto .agents/skills/ + symlinks is the right long-term shape, and it deletes exactly the machinery (managed-section rewriting, compact-body truncation workarounds) that has cost us the most maintenance.

Pace: not yet. This is a breaking change to a shipped subsystem, so it lands after the next release train, behind these gates:

1. Per-agent verification, not aggregate claims. For every target reclassified as "reads .agents/skills/ natively" or "symlink into <dir>", the PR needs a citation to that agent's current docs. Windsurf is the one we're most skeptical of (our compact-body workaround exists because Cascade truncated long rule files; "reads a skills dir" needs proof, or Windsurf stays on its bespoke path until it does).
2. Migration story for existing installs. agent install --force re-running is not enough — users who hand-edited outside the old managed sections must not lose content silently. Detect-and-back-up (which feat(agent)!: align skill install with the Agent Skills standard #271 already does well for overwrites) needs to extend to the transition, with a printed summary of what moved where.
3. CONTRIBUTING must change in the same PR — the standing per-target policy (proposer maintains the target, documented mechanism per editor) doesn't survive a 72-target registry; replace it with a policy about the universal/symlink classification instead.
4. CI must actually run — the triage gate has been cleared and runs approved on feat(agent)!: align skill install with the Agent Skills standard #271, so push a commit (or re-request) and let the 2k-test suite speak instead of the checklist.

Practical effect on in-flight work: #236 (Gemini bespoke target) is paused pending this — if #271 lands, Gemini is covered natively and #236 closes; its author has been told.

@powxenv: take it off draft when 1–3 are addressed. Genuinely good direction — thanks for pushing it with a working implementation instead of a proposal doc.

[powxenv]

thanks for confirming!

Regarding Windsurf, I've checked and they've actually rebranded to Devin Desktop. It now supports agent skills right out of the box and doesn't need a compact body anymore—unlike rules, which are capped at 12,000 characters per file. Also, Gemini was replaced by Antigravity back on June 18, 2026 (official blog post).

I've updated the README with the latest docs for each agent, and I've manually verified they're all correct. If anyone wants to chip in and add a new agent, I've updated CONTRIBUTING.md to keep things aligned with the changes.