Agent skills for turning an issue or a work description into planned, locally committed code, opening pull requests with a fixed-shape description, and reviewing a GitHub pull request from a local clone.
These are plain SKILL.md directories — no build step, no runtime, no package
manager. They work with any agent that reads the SKILL.md convention.
| Skill | What it does | Trigger |
|---|---|---|
local-triage-work |
Turns a GitHub issue link or a text description of a bug, feature, or tech debt into an agreed plan and local commits: interviews the user, designs a strategy, presents the plan for review, then on approval cuts a feature branch from the updated base branch and commits each step. Never pushes or opens a PR. | You paste an issue link or describe the work and ask to triage, plan, or implement it |
local-open-pr |
Opens the PR for the current branch with a fixed shape: one sentence for the problem target with a Closes line for each issue it resolves, one for how it is solved, then a Mermaid data-flow diagram of the Dart, Kotlin, and Python components the diff touches. |
"open a PR", "raise a pull request", "put this up for review" |
local-code-review |
Reviews a GitHub PR from a clone sitting on the PR's updated base branch: short PR summary, diff-health checklist, integration with the local code, tests and regressions, a green/yellow/orange/red diagnostic, then per-finding triage that publishes the accepted findings to GitHub. Read-only: never checks out or executes PR code. | You paste a github.com/.../pull/... URL and ask for a review |
The three are stages of one pipeline, not alternatives: local-triage-work starts
from an issue or a description and ends at local commits on a feature branch,
local-open-pr starts once those commits exist and owns only the PR, and
local-code-review starts from a reviewer's clone and owns only the review.
Global (available in every project):
npx skills add CodandoTV/agent-skills -g --skill local-triage-work --skill local-code-review --skill local-open-prPreview without installing (lists what the repository contains):
npx skills add CodandoTV/agent-skills -lProject-scoped instead of global (commit-able, shared with everyone who clones that repository):
npx skills add CodandoTV/agent-skills --skill local-triage-work --skill local-code-review --skill local-open-prThe installer symlinks into the detected agent directories by default. From a
local checkout where the source may move, add --copy.
Manual install — copy the folders into any skill root your agent watches:
cp -R skills/local-triage-work skills/local-code-review skills/local-open-pr ~/.agents/skills/Update later with:
npx skills update local-triage-work local-code-review local-open-prVersions are git tags. The installer has no semantic versioning of its own — a version is whatever ref you ask it to clone. Latest release: v2.1.0.
Installing without a ref follows main, so corrections arrive with
npx skills update (recommended):
npx skills add CodandoTV/agent-skills -g --skill local-triage-work --skill local-code-review --skill local-open-prPinning a release holds the skill still:
npx skills add CodandoTV/agent-skills#v2.1.0 -g --skill local-triage-work --skill local-code-review --skill local-open-prThat difference is not cosmetic. npx skills update re-resolves the ref recorded
in the lock file, so a pinned install is frozen on that tag and never moves — to
upgrade it you re-run add with the new tag. Pin by tag, not by commit SHA:
the installer clones with git clone --branch, which rejects a SHA.
One version covers the whole repository because the three skills are stages of one pipeline that ship together. What counts as major, minor, and patch is written down at the top of CHANGELOG.md.
On a machine that installs these skills there are two copies — this checkout and
the one the agent reads in ~/.agents/skills. Pick one writable side, or they
drift silently.
Recommended: link the installed copies to the checkout (once).
make linkThis replaces ~/.agents/skills/<name> with a symlink into skills/<name> here,
so editing the installed copy is editing the repository: nothing to sync, and
make drift stays quiet by construction. The skill provider follows symlinked
entries, so discovery is unaffected. Old copies are moved to
.backup-<timestamp>-<name>, never deleted. make link refuses when a copy has
diverged, because linking would silently pick a side.
If you edit the installed copy instead (another machine, or before linking):
make sync-dry # what would change, including deletions
make sync # installed copy wins; validates, then prints git statusIf you edit here and want the agent to use it right away:
make deploy # checkout wins; the agent re-reads the body on its next loadThen publish, the same way in every case:
# 1. add "## [x.y.z] - YYYY-MM-DD" with its entries to CHANGELOG.md
# 2. commit
make release VERSION=x.y.zmake release validates, refuses a dirty tree or a missing CHANGELOG section,
tags vX.Y.Z, pushes, and publishes notes taken from that section. To see the
plan without touching GitHub: bash tools/release.sh x.y.z --dry-run.
Do not run npx skills add (or update) for these three names once linked: the
installer materializes its own copy under ~/.agents/skills, replacing the link,
and its lock entry then disagrees with your edits.
All three:
git, with a clone of the repository the work belongs to.
local-triage-work additionally uses an authenticated GitHub CLI (gh auth login,
repo scope) only when the input is a GitHub issue link, and runs the target
repository's own test and typecheck commands.
local-code-review additionally uses an authenticated gh (repo scope), jq,
and ripgrep (rg) or the agent's own search tool, and reads the local code around
the diff, so the clone should be on an up-to-date base branch.
local-open-pr uses an authenticated gh (repo scope) to resolve issues and create
the PR, and draws its blast-radius diagram as Mermaid, so the component listing is
limited to Dart, Kotlin, and Python files in the diff.
skills/
local-code-review/
SKILL.md
references/
diff-health-checklist.md
integration-checklist.md
tests-and-regressions-checklist.md
scripts/
changed-lines.sh # maps added lines to inline-comment anchors
local-open-pr/
SKILL.md
local-triage-work/
SKILL.md
references/
interview-checklist.md
plan-template.md
tools/
validate-skills.py # frontmatter, catalog cap, cited paths
drift.sh # installed copy vs checkout
sync-from-global.sh # installed copy -> checkout
deploy.sh # checkout -> installed copy
link.sh # replace installed copies with symlinks
release.sh # validate, tag, push, publish
Makefile # make validate | drift | sync | deploy | link | release
.github/workflows/
validate-skills.yml # runs the validator plus a discovery smoke test
references/ and scripts/ are loaded on demand by the skill body, so keep them
beside the SKILL.md that mentions them.
Run the validator before pushing — CI runs the same thing and will fail the PR:
pip install pyyaml
python3 tools/validate-skills.pyIt rejects a name that does not match its directory, a description over the
500-character catalog cap (the tail with the trigger phrases is what gets
truncated), a body that cites a references/ or scripts/ path that does not
exist, and camelCase invocation flags. Adding a behavioural change means adding a
CHANGELOG.md entry and tagging the release.