agent-kit

Resume and execute implementation work from a plan-tracking GitHub issue that carries issue-hosted source, plan, and execution-state snapshots, preserving progress on the issue when local docs/plans files may be cleaned up.

Use when the user wants to open or continue a GitHub issue as the durable timeline for a discovered problem, investigation, blocker, implementation handoff, or unresolved follow-up loop.

Finalize and close plan-tracking GitHub issues after issue-backed execution is complete and the user or project policy has approved closure.

Open or preview a provider-aware GitHub or GitLab tracking issue for a docs/plans plan bundle, bootstrapping the source doc and plan first when the user has a converged feature discussion but no plan file yet.

Deliver a dispatch-ready plan by sprint: split plan tasks into PR lanes, dispatch subagent work, enforce acceptance gates, and advance through final integration without main-agent implementation.

Close out an existing plan-issue runtime where main-agent owns orchestration and review only, subagents own implementation task lanes, and close gates require approval plus merged PRs.

Main-agent PR review workflow that enforces explicit PR comment links, routes follow-up back to subagent-owned task lanes, mirrors decisions to the issue timeline, and controls merge/close outcomes.

Deliver an existing plan-tracking GitHub issue end to end by resuming issue-backed execution, requiring specialist review and repair for every PR, merging completed PRs, updating issue-hosted execution state, and closing through tracking-issue-closeout.

Subagent workflow for assigned task-lane implementation, blocker clarification handoffs, draft PR creation, and review-response updates linked back to the owning issue.

Deliver GitLab merge requests end to end with one workflow: preflight, create via create-gitlab-mr, wait/fix pipeline until green, then close via close-gitlab-mr.

Create a standard execution-ready implementation plan under docs/plans/<slug>/ with phases/sprints, atomic tasks, dependencies, validation, and Read First source context. Use when the user asks to plan implementation or split work. For required complexity scoring, sprint scorecards, PR grouping, or subagent-reviewed execution modeling, use create-dispatch-plan.

Open a draft GitHub dispatch lane PR for a plan-issue implementation lane using the assigned dispatch record, PLAN_BRANCH base branch, and canonical dispatch lane PR body schema.

Deliver GitHub pull requests end to end with one provider-specific workflow: preflight, create via create-github-pr, wait/fix checks until green, then close via close-github-pr.

Merge and close GitLab merge requests with glab after provider-specific pipeline gating, draft readiness, merge, and local cleanup.

Merge and close GitHub pull requests with gh after provider-specific checks gating, draft readiness, merge, and local cleanup. Use with `kind=feature` or `kind=bug`.

Capture redacted static HTTP evidence bundles through the nils-cli web-evidence command.

Create, read, validate, and clear edit-scope locks through the nils-cli agent-scope-lock command.

Record failing-test evidence, waivers, and final validation through the nils-cli test-first-evidence command.

Generate a repo-local project implementation retrospective by consuming the released nils-cli `repo-retro` command; use when the user asks for project retrospectives, weekly repo reviews, implementation trend snapshots, or local engineering review agendas without remote API enrichment.

Validate SVG inputs and convert svg/png/jpg/jpeg/webp inputs to png/webp/jpg

Decompose risky or broad diffs into read-only specialist review passes and merge normalized findings.

Manage curated HEURISTIC_SYSTEM inbox and operation-record case folders through the nils-cli heuristic-inbox primitive with judgment-first workflow guidance.

Autostage the full in-scope working tree delta and commit it using Semantic Commit format when the active agent owns the complete change set and the user should not manually stage files.

Commit a staged or deliberately selected change set using Semantic Commit format.

Prune outdated docs/plans coordination markdown and reconcile plan-related docs safely for a target project.

Record skill invocation intent, linked evidence, validation, outcome, and failures through the nils-cli skill-usage command.

Aggregate read-only AI and technology trend signals from Polymarket, Hacker News, GitHub, arXiv, Hugging Face, official sources, and news search into source-grounded Markdown or JSON digests. Use directly for source-level radar scans, machine-consumable output, source tuning, or as the lower-level data engine for daily-brief.

Prepare a source-grounded daily information brief for recurring AI/technology or user-selected topics. Use as the default daily entrypoint when the user asks for today's news, this week's signals, a morning brief, "what should I know", or a personalized synthesis; orchestrates topic-radar JSON output with stable preferences and does not implement its own source fetchers.

Record review findings and final validation through the nils-cli review-evidence command.

Canonical browser-facing QA workflow for report-only, acceptance-gate, and fix-loop checks using Browser, Chrome, Playwright, web-evidence, browser-session, and workflow evidence primitives.

Fully automated GitHub Actions CI fixer: inspect failing checks, apply fixes, semantic-commit-autostage + push, then watch CI and repeat until green.

Execute project release workflows by following a repo-provided release guide when present; otherwise use a changelog-driven fallback flow with helper scripts.

Compatibility entrypoint for legacy web-qa requests; route new browser-backed QA work to browser-qa while preserving web-evidence and browser-session examples.

Resume and execute plan-driven implementation work with a durable execution-state ledger.

Convert completed requirements, design, feasibility, or customer-facing discussion into an implementation-readiness source document.

Convert review findings, risks, lessons learned, or "fix later" notes into a repo-local improvement source document. Use when the user asks to preserve a review as docs, create an improvement/backlog source record, or write a docs-based follow-up artifact before planning or handoff. Do not use for ordinary implementation plans, discussion-to-implementation handoffs, or copy-ready next-session prompts.

Audit and remove obsolete durable implementation artifacts after execution is complete.

Execute a markdown implementation plan by spawning parallel subagents for unblocked tasks, then integrate results and validate. Triggers on explicit "/execute-plan-parallel <plan.md> [optional selector]".

Create a dispatch-ready execution-modeling plan under docs/plans/<slug>/ with per-task complexity, required sprint metadata/scorecards, PR grouping, split-prs validation, and subagent review. Use when the plan must be ready for sprint dispatch, PR lanes, or subagent execution. Do not use as the primary artifact when the user only needs a durable review finding or improvement backlog; route that to review-to-improvement-doc first.

Initialize baseline agent-docs files safely for new repositories using dry-run-first workflow.