espensev
UserPortable multi-agent skill packages for Codex, Claude, and Gemini, plus worktree tooling for parallel agent workflows.
Categories
Indexed Skills (23)
delegate
Decide whether a narrow, well-scoped sub-task should go to a LOCAL Ollama model vs stay with the controller, and route it if so. Grounded in the ollama-telemetry MCP delegation tools with a static-guidance fallback. Use when offloading a bounded transform (summarize/classify/extract/rewrite/review/draft) on material the controller already fetched.
diagnosing-bugs
Use when the user wants the cause or fix for a reproducible bug, runtime error, flaky behavior, or performance regression. Builds a red-capable feedback loop before changing code. Do not use merely to run or classify a known test suite (use qa), review a diff (use review), or audit runtime efficiency broadly (use deep-audit).
docs-sync
Detect documentation drift, stale claims, path mismatches, and merge conflict markers. Use when checking README, docs, status, or architecture consistency.
handoff
Use when the session opens by pointing back at earlier work instead of describing new work: resume, continue, take over, follow up, pick this up, read the handoff, or a path to a handoff/state/plan file. Reconstructs the prior state from the memory store and the repo, verifies every checkable claim in it, and states what to do next before touching anything. Do not use for a fresh task with no prior state, or to perform the resumed work itself - it routes, it does not build.
manager
Use when the user explicitly wants a multi-agent campaign executed or managed: parallel worktrees, agent launches, dependency-aware progress, merges, verification, or campaign status. Do not use for plan-only design (use $planner), bounded codebase research, or one tight local change.
planner
Use when the user wants a structured multi-agent campaign designed before execution: work decomposition, agent tasks, dependencies, gates, or a phased refactor/migration. Do not use to execute an approved campaign (use manager) or for one bounded local change.
qa
Use when the user wants tests run, coverage checked, failures classified, a configured app smoke-tested, or regression tests added. Do not use for findings-first diff review (use review), a hard bug needing a new reproduction-and-fix loop (use diagnosing-bugs), or a broad runtime-efficiency audit (use deep-audit).
review
Use when the user asks for a findings-first review or asks you to review and advise — on a branch, pull request, worktree, staged diff, or fixed-point diff, or on a written artifact such as a plan, design doc, handoff note, spec, or proposal — against repo standards, specs, and regression risk. Do not use to implement a fix, run a test campaign (use qa), diagnose one bug (use diagnosing-bugs), or audit runtime efficiency (use deep-audit).
ship
Stage, commit, and optionally push validated work. Handles multi-file campaigns, commit grouping, message drafting, and exclusion of temp/sensitive files. Use when the user asks to commit, package, land, push, or prepare validated changes for delivery.
skill-authoring
Create or revise Agent Skills with concise discovery metadata, progressive disclosure, and portable support files. Use when adding a new SKILL.md, changing skill frontmatter, splitting long instructions into references, or preparing Codex/Claude skill packages.
smart-test
Map changed files to the minimal useful test subset instead of running the full suite. Use for fast feedback before full QA, commit, merge, or build validation.
telemetry-live-ops
MACHINE-LOCAL ops skill (not a portable skill): start, verify, and inspect the live ollama-telemetry deployment spanning observer_hub on MAINDESK and host_agent on snd-host. Retarget it with the OLLAMA_TELEMETRY_* env overrides. Use when tasks mention snd-host, observer_hub, MachineTelemetryHostStack, live telemetry deployment, remote telemetry health, or machine-wide telemetry verification.
usage-stats
Use when reviewing token/cost usage, budgets, rate-limit forecasts, session tool or timeline activity, agent/campaign efficiency, or execution closeouts. Prefers measured telemetry with an explicitly estimated fallback. Do not use to operate a live telemetry deployment.
browser-control
Use when a task needs a live browser: navigating pages, reading DOM or page text, screenshots, form fills, console/network inspection, or web app testing. Routes ALL browser automation through the DevHome Opera Developer CDP endpoints (devbrowser, ports 9000/9001) via chrome-devtools-mcp or direct CDP. Never use harness-bundled browser tooling (browser/chrome plugins, browser-harness, playwright) - it is disabled on purpose.
deep-audit
Use when the user asks for an evidence-backed, multi-pass runtime-efficiency or scalability audit across real execution paths. Covers hot-path reconstruction, CPU, allocation/retention, I/O, concurrency, queues, retries, scheduling, and lifecycle. Do not use for diff review, one known regression with a requested fix, tests, security assessment, bounded feasibility research, or parallel remediation.
devhome-lifecycle
Check, synchronize, or diagnose the machine-local DevHome Codex safety and Remember lifecycle hooks plus the shared Claude/Codex Handoff Relay on enrolled snd-desk. Use when asked about hook drift, lifecycle plugin status, automatic handoffs, or updating either DevHome hook projection.
parallel-agents-light
Use when the user explicitly asks Codex for lightweight parallel agents, subagents, independent reviewers, or faster split execution without a full manager campaign. Routes among local work, bounded sidecars, split review, and split implementation; do not use for a single tight file set or a durable managed campaign.
repo-conventions
Follow the target repository's established conventions before applying fallback guidance for API design, backend or frontend structure, Playwright E2E tests, current research, or a bounded local work cycle. Use when implementing or reviewing one of those areas; do not use when a more specific installed skill owns the task.
verification-loop
Run a repo-appropriate build, lint, typecheck, test, and diff review sequence before handoff. Use after significant code changes, refactors, or before declaring a branch ready for review.
documentation-lookup
Fetch current library, framework, and API documentation before answering library-specific questions. Use when a request depends on up-to-date behavior, setup, migrations, configuration, or code examples for a named library or framework.
exa-search
Use Exa as an optional research accelerator for current web, code, company, or people search. Use when Exa MCP is available and the task benefits from higher-quality search or page extraction than the default web workflow alone.
loop-master
Backward-compatible alias for multi-round orchestration. Routes to repo-conventions for the immediate bounded step and to planner/manager for durable multi-agent campaigns. Use when older references invoke loop-master; do not implement separate orchestration logic here.
mcp-server-patterns
Build and maintain MCP servers with current SDK semantics, schema validation, and clear transport choices. Use when implementing tools, resources, prompts, stdio servers, or HTTP-based MCP services.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.