Index Shared Agent Skills

.agents/AGENTS.md

@@ -0,0 +1,51 @@
+# Global Agent Instructions
+
+## Git Sign-off
+
+When creating commits with DCO sign-off (`git commit -s`), always use:
+
+```
+Signed-off-by: Prasanth Baskar <bupdprasanth@gmail.com>
+```
+
+Never hardcode or guess the sign-off email. Always use `bupdprasanth@gmail.com`.
+
+## Harbor Worktree Push Rules
+
+See `~/.agents/harbor-worktree-rules.md` for full details. Summary:
+
+**CRITICAL: Wrong remote = private code leaked to public repos.**
+
+Harbor (`~/code/OSS/harbor/`) is a bare repo with worktrees named `{remote}-{description}`. The prefix before the first `-` IS the push target remote. Derive: `basename "$PWD" | cut -d'-' -f1`
+
+- `8gcr-*` → `8gcr` (**PRIVATE**) | `next-*` → `next` (**PRIVATE**) | `bupd-*` → `bupd` (fork) | `upstream-*` → `upstream` (**PUBLIC**) | `glab-*` → `glab` (**PRIVATE**)
+
+1. ONLY push to the remote matching the directory prefix.
+2. NEVER push `8gcr` or `next` code to `upstream`.
+3. When unsure, ask. Confirm remote + branch before force-push.
+
+## Neovim PR Review
+
+When the user provides a GitHub PR URL or PR number and wants local Neovim review, use the `neovim-pr-review` skill or run:
+
+```bash
+python ~/.agents/skills/neovim-pr-review/scripts/prepare_pr_review.py <github-pr-url>
+```
+
+This workflow requires a clean worktree, creates a dedicated review branch, and leaves the PR changes unstaged with `git reset --mixed` so Neovim highlights the changed files. Do not use `git reset --hard` for this workflow.
+
+For a raw PR patch, prefer GitHub CLI auth:
+
+```bash
+gh pr diff <number-or-url> -R <owner/repo> --patch > pr.patch
+```
+
+For public repositories, `<pr-url>.patch` and `<pr-url>.diff` also work.
+
+## Shared Agent Skills
+
+Shared skills live under `~/.agents/skills/<skill-name>/SKILL.md`. Each skill uses YAML frontmatter with `name` and `description`; optional references, scripts, and assets live beside that `SKILL.md`.
+
+Agents that support skills should index `~/.agents/skills/*/SKILL.md` at session start and autoload a skill when the user's request matches its description. Load only the matching `SKILL.md` first, then load bundled files referenced by that skill as needed.
+
+For Codex, OpenCode, Claude-compatible agents, or custom tools without native skill indexing, use `~/.agents/skills/README.md` and `~/.agents/skills/registry.json` as discovery aids. Do not load every skill body by default.

.agents/README.md

@@ -0,0 +1,10 @@
+# Shared Agent Rules
+
+Canonical rules for all AI agents. See `~/dotfiles/AGENT-CONFIG.md` for the full management plan.
+
+- `AGENTS.md` — Global agent instructions (source of truth, symlinked by `.claude/CLAUDE.md` and `.codex/AGENTS.md`)
+- `harbor-worktree-rules.md` — Harbor multi-remote push safety
+- `git-signoff.md` — DCO sign-off identity
+- `skills/` — Shared skills with `SKILL.md` frontmatter for agent autoloading
+- `skills/README.md` — Human-readable skill discovery index
+- `skills/registry.json` — Machine-readable skill discovery manifest

.agents/git-signoff.md

@@ -0,0 +1,5 @@
+# Git Sign-off
+
+Use `git commit -s` with: `Signed-off-by: Prasanth Baskar <bupdprasanth@gmail.com>`
+
+Never hardcode or guess the sign-off email. Always use `bupdprasanth@gmail.com`.

.agents/harbor-worktree-rules.md

@@ -0,0 +1,23 @@
+# Harbor Worktree Push Rules
+
+**CRITICAL: Wrong remote = private code leaked to public repos.**
+
+Harbor (`~/code/OSS/harbor/`) is a bare Git repo with worktrees named `{remote}-{description}`.
+The prefix before the first `-` in the directory name IS the Git remote to push to.
+
+Derive remote: `basename "$PWD" | cut -d'-' -f1`
+
+## Remotes
+
+- `8gcr-*` → `8gcr` (container-registry/8gcr — **PRIVATE**)
+- `next-*` → `next` (container-registry/harbor-next — **PRIVATE**)
+- `bupd-*` → `bupd` (bupd/harbor — personal fork)
+- `upstream-*` → `upstream` (goharbor/harbor — **PUBLIC**)
+- `glab-*` → `glab` (8gears/container-registry/harbor — **PRIVATE**)
+
+## Rules
+
+1. ONLY push to the remote matching the directory prefix.
+2. NEVER push `8gcr` or `next` code to `upstream` — that is public OSS.
+3. When unsure, ask before pushing.
+4. Confirm both remote AND branch before any force-push.

.agents/skills/README.md

@@ -0,0 +1,31 @@
+# Shared Skills
+
+This directory is the shared skill root for OpenCode, Codex, Claude-compatible agents, and custom agent tooling.
+
+## Discovery
+
+- Canonical source: `~/.agents/skills/<skill-name>/SKILL.md`
+- Metadata: YAML frontmatter `name` and `description`
+- Autoload rule: match the user's task to a skill description, then load only that skill's `SKILL.md`
+- Bundled resources: load referenced files from the same skill directory only when needed
+- Machine-readable manifest: `~/.agents/skills/registry.json`
+
+## Matt Pocock Skills
+
+Installed from `mattpocock/skills`:
+
+- `obsidian-vault` - Search, create, and organize Obsidian notes.
+- `edit-article` - Restructure and tighten article drafts.
+- `caveman` - Ultra-compressed communication mode.
+- `handoff` - Produce handoff notes for another agent.
+- `write-a-skill` - Create new agent skills.
+- `diagnose` - Disciplined debugging and regression diagnosis.
+- `grill-with-docs` - Stress-test plans against docs, domain language, and ADRs.
+- `triage` - Triage issues through project workflow states.
+- `improve-codebase-architecture` - Find architecture and refactoring opportunities.
+- `setup-matt-pocock-skills` - Configure repos for the engineering skill set.
+- `tdd` - Use red-green-refactor development.
+- `to-issues` - Break plans into implementation issues.
+- `to-prd` - Turn conversation context into a PRD.
+- `zoom-out` - Explain code at a higher abstraction level.
+- `prototype` - Build throwaway prototypes for logic or UI questions.