agents-library

Aligns code with the project's established architecture without changing behavior. Use when asked to fix architectural inconsistencies — misplaced logic, layer-boundary violations, configuration read directly from the environment, inline definitions that belong in a dedicated location, duplicated shared helpers, or inconsistent response/return shapes. Makes one focused, same-pattern change per run and opens a reviewable PR.

Removes dead code without changing live behavior. Use to clean up unused imports and variables, commented-out blocks, unreachable branches, orphaned files, stale TODO/FIXME comments, or dead parameters. Confirms nothing references the code (including dynamic dispatch) before removing it, then opens a reviewable PR.

Fills documentation gaps without changing code. Use to add or update doc comments on public functions/classes, module overviews, API/contract docs, or stale project docs (README/architecture notes) after a component changes. Matches the project's existing documentation style and opens a docs-only PR.

Micro-refactors that improve clarity without changing behavior. Use to extract duplicated logic into a helper, replace magic numbers/strings with named constants, flatten deep nesting with early returns, rename vague identifiers, or simplify redundant boolean logic. Keeps diffs small and reviewable and opens a PR; does not fix bugs or change behavior.

Light security-hygiene fixes without changing business logic. Use to add a missing auth guard to a protected endpoint, stop internal error details or stack traces leaking to clients, move a hardcoded secret/config value into central config, add missing input validation, or remove sensitive data from logs. Handles hygiene gaps (not vulnerability research) and opens a reviewable PR.

Triage the project's open pull requests and batch the trivial ones onto a branch you name. Lists every open PR, fans out one review sub-agent per PR that reads the diff and judges it across four lenses — size/scope, change type, mergeability/CI, and an actual correctness read — then recommends include or skip with reasoning. Consolidates the verdicts into one ranked list, lets you pick which to take, and locally git-merges the chosen PRs into your target branch (no push, nothing closed on GitHub), reporting conflicts and a final summary. Use this whenever the user wants to sweep open PRs, batch or bulk merge pull requests, "clean up the PR queue", collect the easy/trivial/safe PRs, merge the low-risk ones together, or assemble several PRs onto one branch — even if they don't say the word "skill". Requires the `gh` CLI; works on the local repo, never pushes or closes PRs on its own.

Explain how a codebase is shaped — the read-to-explain counterpart to review-pr. Orients on the project, then fans out read-only explorer sub-agents to map it across three lenses (layering & entry points incl. config/auth/error handling, data & persistence, conventions & build), consolidates their findings into one skimmable orientation brief with file:line references, and optionally writes that brief to a doc. Supports three scopes: whole repository (default), a path/glob for one subsystem, or a feature/flow trace that follows a single execution path end to end. Use this whenever the user wants to understand an unfamiliar codebase, "explain how this works", get oriented, onboard onto a repo or subsystem, map the architecture, write an ARCHITECTURE.md, or trace how a feature/endpoint flows — even if they don't say the word "skill". Read-only by default; never modifies code and writes a doc only on explicit confirmation. No `gh` or remote required.

Review the work-in-progress on the current branch before it is finalized — the reviewer you start in a fresh agent after implementing a change. Orients on the project (AGENTS.md, README, conventions), computes the local branch diff, fans out specialized review sub-agents across nine quality domains (correctness, architecture, dead code, docs, refactor, testing, UX polish, security, conventions), consolidates and ranks their findings, presents them for you to pick from, and then either applies a chosen subset or runs an autonomous improve-until-converged loop. Use this whenever the user wants to review a PR, review a branch, review their changes before merging, "check what I just built", clean up a diff, or asks for a pre-merge / pre-PR review — even if they don't say the word "skill". Works on the local diff before a GitHub PR exists; no `gh` or remote required.

Survey a target — the whole repository, a path/glob, or the current branch diff — for behavior-preserving simplification opportunities, then optionally apply them. Orients on the project (AGENTS.md, README, conventions), builds and shards a scan surface, fans out parallel sub-agents across four simplification lenses (redundancy & dead code, complexity & structure, clarity & idiom, docs), consolidates and ranks their findings, presents them for you to pick from, and then either applies a chosen subset or runs an autonomous improve-until-converged loop — every applied fix gated on the project's lint and tests and behavior-preserving by construction. Use this whenever the user wants to simplify the codebase, find code or docs that can be simplified, reduce complexity, declutter or tidy a repo, find duplication / dead code / over-engineering, or survey the whole project (not just a diff) for cleanup opportunities — even if they don't say the word "skill". Works on the local branch; no `gh` or remote required.

Triage the project's open GitHub issues into a ranked action plan. Lists the untriaged open issues, fans out one assessment sub-agent per issue that reads the issue *and the actual codebase* and judges it across five lenses — validity (does the bug still exist in the code?), completeness, classification, duplicate detection (search-then-confirm), and a code-grounded effort estimate — then recommends an action with evidence. Consolidates the verdicts into duplicate clusters and one ranked decision table (easy wins, duplicates, needs-info, larger work), lets you pick which actions to take, and then — only after your confirmation — applies labels, posts needs-more-info comments, closes confirmed duplicates via `gh`, and fixes the easy wins you approve: one fix sub-agent per issue, each in its own git worktree, each opening its own reviewable pull request that references the issue. Use this whenever the user wants to triage issues, clean up or groom the issue tracker/backlog, find duplicate issues, find and fix easy wins or good first issues, label or categorize open issues, or figure out what's worth fixing first — even if they don't say the word "skill". Requires the `gh` CLI; never writes to GitHub without explicit per-action confirmation, and never commits to the default branch — every fix lands as its own PR.