p-wegner
@p-wegner
Public Skills
drive-new-project
by p-wegner
Drive a freshly-created project hands-off to a finished multi-ticket epic — the playbook for "use the board to build a new app / implement a 10+ ticket epic" tickets. Encodes the completion CONTRACT (keep the meta-ticket In Progress until the epic is N/N Done, then drive the meta itself to Done — not parked in Review), a preflight prerequisites check, fan-out epic seeding, autodrive enablement, and a REQUIRED resident watch that recovers stalls instead of abandoning them. Use when a ticket asks you to register/scaffold a new project and take it to completion via the board.
agentic-kanban
by p-wegner
Operate the agentic-kanban board — a kanban board where each card is an AI coding session in its own git worktree. Use when asked to work a ticket, drive a project hands-off, review or merge agent work, inspect what agents are doing, or wire the board into another tool. Covers the MCP tools, CLI, board views, statuses, and the review/merge gate.
sentinel
by p-wegner
The board-orchestrator WATCH role. Check the out-of-process board-monitor loop (the "Conductor") once and report a concise one-line status — loop health, board state, profile≠mock, and the current Strategy-Bullseye targets/weights it's steering by; alert with detail and recover ONLY when something needs attention. Use for "check the board monitor", "is the Conductor healthy?", or as a recurring watch scheduled with ScheduleWakeup. Distinct from the board-monitor skill (that's the system-health checklist a monitor CYCLE runs; this is what the human-side supervisor checks between cycles).
dev-server
by p-wegner
Start/stop/health-check the dev server safely — worktree ports, process-signature kills, never kill-all-node
board-monitor
by p-wegner
System-level board health — conflict scan, server/frontend health check. Board operations (relaunch/merge/nudge/auto-start) are handled by the app's built-in monitor (Settings → Workflow → Board Monitoring). Run Sections 0-3 every cycle; only do Section 4 if the app monitor is OFF.
set-provider-default
by p-wegner
Set the board's default agent provider+profile (and optional model) in ONE place, atomically, across every consumer that can drift — the Strategy Bullseye, the Settings provider/claude_profile mirror, and the provider-scoped default_model_<provider> — then verify they agree. Use whenever the user says "set the provider default", "switch the board to claude/codex/copilot", "use profile X", "make anth the default", or after a provider stall where the sources drifted out of sync.
direct-master
by p-wegner
Change master in the main checkout without the board's workspace machinery — from a quick doc fix to implementing full tickets in-session. Covers choosing between four execution modes (main agent direct, subagents on shared master, subagents in worktrees, board workspace) by time criticality, traceability, and context budget; grouping adjacent tickets (including ones pulled forward from Backlog) so the expensive gates run once per group instead of once per ticket — reading/writing the board's coupled_with edges and using ticket-group workspaces (#661, memberIssueIds) where they fit; plus the commit discipline (aggressive pathspec commits, tree never left dirty) that keeps auto-merge unblocked.
backlog-refill
by p-wegner
When the board backlog is running low, generate an aggressive batch of new well-scoped tickets — with the feature/bugfix mix set by the monitor's REFILL_FOCUS (in objective.md) — so the board always has enough queued to keep the monitor's agent target busy. Use when the monitor sees the backlog dipping below its BACKLOG_FLOOR.
app-runner-discover
by p-wegner
Discover how to build/run this project's app(s), verify a real start against a health endpoint, and write .app-runner/apps.json + .app-runner/apps.md for the App Runner dashboard.
release
by p-wegner
Pre-release verification + gated GitHub + npm release for agentic-kanban. Runs all checks, summarizes changes, drafts release notes, then PAUSES for explicit user confirmation before pushing the git tag, the GitHub release, and the npm publish.
board-navigator
by p-wegner
Comprehensive guide for agents to interact with the kanban board using MCP tools
unstuck
by p-wegner
Drive a stuck kanban issue all the way to master. Diagnose why the agent stopped, answer pending questions, run the agent to commit, then merge via the kanban endpoint and verify master actually advanced.
skill-author
by p-wegner
Scaffold, validate, and install new OpenCode agent skills — handles directory creation, YAML frontmatter, naming rules, tool/permission wiring, and placement in the correct discovery paths. Use when asked to "create a skill", "write a skill", "scaffold a SKILL.md", "make a new agent skill", "add a skill to this repo", or any request to author reusable agent behavior.
code-review-thorough
by p-wegner
In-depth AI code review using a more capable model — catches subtle bugs and architecture issues
code-review
by p-wegner
Default AI code review prompt — customize per project to change review behavior
board-tuning-lab
by p-wegner
R&D harness for tuning the kanban BOARD (its agents, monitor, and workflow) along a chosen DIMENSION by building an increasingly complex disposable fixture project, driving it through the board, measuring where the workflow breaks, filing the gaps as dev-board tickets, and fixing them end-to-end. Dimension-parameterized — docker/multi-repo, observability, token-efficiency, ticket-sizing, and more added over time. Use for "run the board-tuning lab", "tune the board for <dimension>", "dogfood/stress the board's <X> support", "build a fixture and find gaps in <X>", "exercise the board on docker / multi-repo / observability / token efficiency / ticket sizing", or "improve how the board handles <X>".
count-code-lines
by p-wegner
Count productive code lines per package using scc, scoped to actual source directories
roast
by p-wegner
Deliver a good-natured, affectionate roast of the board's current state — the Butler's /roast party trick.
publish
by p-wegner
Publish a new version of the agentic-kanban npm package — build, bump, pack check, publish, commit, push
learning-step
by p-wegner
Analyze agent session data (kanban sessions, JSONL transcripts, or inline context), detect friction points, and produce targeted improvements to documentation, hooks, skills, or code.
cleanup
by p-wegner
"Clean up stale agentic-kanban resources including git worktrees, Claude Code sessions, and E2E test artifacts."
coverage-intelligence
by p-wegner
Build a multidimensional functional-coverage model for a software system and rank the highest-ROI verification gaps. Binds the observable-behaviour model (from behavior-discovery) against requirements (PRD/ADRs/user stories/acceptance criteria), the existing test suites (unit/integration/e2e), and historical signals (git churn, bug history) to compute — per behaviour — whether it is covered / partial / uncovered / undocumented-implemented / documented-missing, across many orthogonal coverage dimensions (capability, requirement, workflow, permission, navigation, API, error, boundary, state-transition, config, accessibility, regression, risk). Produces a coverage matrix, gap report, and an ROI-ranked test backlog. NOT line coverage. Use when the user asks "what's our real test coverage", "what behaviour is unverified", "where are the coverage gaps", "what tests give the most value", "map requirements to tests", "build a coverage model", or as phase 2 of an end-to-end verification effort.
e2e-test-author
by p-wegner
Generate the highest-ROI end-to-end tests from a prioritized coverage-gap backlog, then subject them to an adversarial review that tries to REFUTE them before they land. Consumes the verification model (_priorities.md / _coverage.json from coverage-intelligence), authors complete user-workflow tests (not isolated clicks) using the target project's own anti-flake conventions, makes each test DECLARE the coverage dimensions it contributes, runs independent adversarial reviewer subagents (overfitted assertions, flaky interactions, false positives, missed edge cases), fixes what they find, and writes the new coverage back into the model so re-running the pipeline improves. Use when the user asks to "write the missing tests", "generate high-value e2e tests", "author tests for the coverage gaps", "close the top coverage gaps", or as phase 3 of an end-to-end verification effort. For a single ad-hoc test in agentic-kanban use the project e2e-author skill instead; this skill is the gap-driven, reviewed, batch author.
behavior-discovery
by p-wegner
Build a high-confidence model of a software system's EXTERNALLY OBSERVABLE behaviour — capabilities, actors, permissions, entry points, workflows, state transitions and error states — by combining many evidence sources (domain docs, OpenAPI/GraphQL schemas, source, routing, running-UI exploration, existing tests, git history) and degrading gracefully when some are absent. Infers business semantics; never paraphrases code. Marks unknowns explicitly instead of guessing. Output is the shared _behavior-model.json that the coverage-intelligence and e2e-test-author skills consume. Use when the user asks to "model what the app does", "discover observable behaviour", "build a verification/behaviour model", "explore the running app and map its capabilities", or as phase 1 of an end-to-end verification effort.