darkstar1227
UserBridge gstack-reviewed plans into execution-ready Superpowers implementation specs for AI coding workflows.
Categories
Indexed Skills (16)
autoresearch-impl
Runs a Karpathy-autoresearch-style keep-or-discard iteration loop over an already-implemented branch — propose one variant, run it against tests/benchmarks, keep it if the metric improves or discard and revert, repeat within a fixed budget — before handing off to code-review/QA. Use this after subagent-driven-development (or opencode-subagent-driven-development) finishes a task, whenever the user wants to "squeeze out more performance", "try a few implementation variants", "iterate on this implementation before review", or mentions autoresearch/Karpathy about code that already runs.
autoresearch-plan
Runs a Karpathy-autoresearch-style comparative experiment loop over candidate technical approaches (API designs, algorithms, data-flow strategies) after a plan has been reviewed but before it's handed to an implementer. Picks a winning approach against an explicit metric and records the baseline, so downstream implementation isn't built on an unvalidated guess. Use this whenever the user wants to "test approaches before committing", "spike a few options", "compare designs experimentally", mentions autoresearch/Karpathy in a planning context, or after gstack's /autoplan when the plan contains more than one viable technical direction.
benchmark-opencode-models
Deep-benchmark which OpenCode models are actually viable for opencode-bridge — ping each candidate model, then run 5 canned superpowers-style task prompts per model (feature vs bugfix, short vs detailed, plus a dedicated TDD red-to-green prompt), independently verify every result by executing the generated code (never trust OpenCode's self-reported "done"), and score each run on time/quality/completeness/autonomy/discipline/red-green-accuracy/test-call-discipline. For a fast pass/fail availability check with no scoring, use check-opencode-models instead.
check-opencode-models
Fast two-stage check for a list of OpenCode models — ping (reachability/auth) followed by one real, minimal single-shot prompt test to catch models that ping fine but are actually slow or wrong under real dispatch. Reports which models are usable right now vs. reachable-but-slow vs. reachable-but-wrong vs. unreachable. Lighter than a full benchmark. For a deep quality/discipline benchmark across multiple task shapes, use benchmark-opencode-models instead.
full-pipeline
Orchestrates the user's full end-to-end workflow in order — gstack office-hours, autoplan, autoresearch-plan, superpowers writing-plans and subagent-driven-development, autoresearch-impl, code-review, and qa — invoking each skill in sequence and handing its output forward as the next step's input. At two fixed checkpoints (after the plan is locked, and after implementation lands) it judges whether the change touches database schema and if so invokes supabase:supabase-postgres-best-practices before continuing. Use whenever the user wants to "run the whole pipeline", "go from idea to shipped", "do the full flow", or asks Claude Code to chain their office-hours → autoplan → superpowers → qa workflow automatically instead of invoking each skill by hand.
gstack-to-plan
Bridge a gstack-reviewed plan into a Superpowers writing-plans implementation plan. Reads the latest approved gstack plan, transforms it into a Superpowers-compatible spec, then invokes superpowers:writing-plans.
init-project
Detect a target project's stack (Python/uv, Node/TypeScript, Docker, Supabase, Git) and initialize or audit it against standard conventions — ruff/PEP8 or eslint/prettier, docker-compose profiles, Supabase migrations, .env hygiene, commit/branch conventions, folder layout — then write the result into the target repo's CLAUDE.md.
opencode-bridge
Delegate a coding task to OpenCode and get back a structured handoff report (done/failed/timed-out, files changed, implementation summary) — for use as an OpenCode-backed implementer step in subagent-driven-development or executing-plans. For a full plan-execution loop with review gates already wired in, use bridge:opencode-subagent-driven-development instead of calling this standalone.
opencode-subagent-driven-development
Wraps Superpowers' subagent-driven-development plan-execution loop, asking upfront whether the implementer step should be a Claude subagent (default) or OpenCode via opencode-bridge — spec-compliance and code-quality review stay identical either way.
review-pipeline-logs
Query a project's local log files and review whether a just-developed application pipeline actually ran the way it was designed to. Reads the project's design/plan doc for the expected steps, parses the relevant log file(s), then checks the run against the plan step-by-step, flags errors/exceptions, judges whether logging is detailed and leveled clearly enough to debug from, and confirms output values matched expectations. Produces a detailed report (step-by-step results, error detail, suggested fixes) — not just a pass/fail line. Trigger this whenever the user asks to "review the logs," "check if the pipeline ran correctly," "查logs" "review pipeline log對不對", wants to confirm a feature just shipped is actually working end-to-end, or pastes/points at a log file and asks whether something looks right, even if they don't say the word "skill."
send-update-email-batch
Scan a parent folder of repos and send each configured repo's accumulated update email via Resend — unattended, no confirmation, meant for /loop.
send-update-email
Send a readable, version-grouped update email via Resend for a single repo's accumulated commits since the last send — shows the rendered content and waits for confirmation before actually sending.
setup-email-updates
Create or edit the .bridge/email-config.json that send-update-email and send-update-email-batch need — recipients, last-sent tracking, and a per-repo Resend MCP connection — for a single repo or in bulk across a parent folder.
setup-env
Export the user-scope Claude Code plugins, marketplaces, and gstack installation on this machine into a portable manifest, or install from an existing manifest onto a new machine to reproduce the same Claude Code environment. Trigger when the user wants to back up their plugin setup, replicate their Claude Code environment on a new machine, or bootstrap a fresh machine with the same plugins/skills they already use elsewhere.
superpowers-pipeline
A lighter orchestrator than full-pipeline — chains only the core Superpowers loop (writing-plans, subagent-driven-development, finishing-a-development-branch) without gstack's office-hours/autoplan or the autoresearch experiment steps, for when a spec already exists or gstack-level planning isn't wanted. Still judges at two checkpoints whether the change touches database schema and invokes supabase:supabase-postgres-best-practices if so. Use when the user wants "just the superpowers flow", already has a spec/plan and wants straight to implementation, or explicitly says they don't need the gstack review layers or autoresearch for this task.
agy-second-voice
Get an independent, read-only second opinion from Antigravity CLI (`agy`, Google's coding agent) — review the current diff, adversarially challenge it, or consult it on a question/plan — as a cross-model check alongside Claude's own analysis. Use when asked to "agy review", "agy challenge", "ask agy", "second opinion", "antigravity review", or "second voice".
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.