fastedd27
UserWorking agent-orchestration skills from The Spire Library. Build pipeline plugin: design-swarm → run-list → conductor with probe-verified execution, a swappable config layer, and a two-question setup helper.
Categories
Indexed Skills (6)
model-eval
Deep security evaluation of an AI model artifact — a public HuggingFace model (especially custom-code / trust_remote_code models) OR a model folder already on disk. Runs a deterministic collector, authors a dual-pass behavioral claim and checks it against the code's actual call graph, traces generation-to-sink paths, and produces a plain-English report (decision on top, evidence below) with an audit-opinion verdict. Use when asked to evaluate, audit, vet, or decide whether to adopt a model that ships custom code, or after model-scorecard returns tier C or D. EXECUTES NOTHING from the artifact and NEVER issues a clearance; a code-execution artifact seen static-only receives a "disclaimer of opinion".
model-scorecard
Zero-friction safety gut-check for a public HuggingFace model repo. Uses only curl + jq against the public HF API (no tokens, no download) to detect weight format (safetensors vs pickle-family), custom code (trust_remote_code), and assign a coarse (format, loader) risk tier with a plain-language card. Use when asked for a quick "is this model safe to pull?" check, before adopting a HuggingFace model, or as fast triage ahead of the deeper model-eval. NEVER issues a clearance.
conductor
(v2026-08-05.1) Execute a run-list document — pipeline stage 3 of design-swarm → run-list → conductor. The conductor reads the run-list tolerantly (no compiler, no sidecar), front-loads ALL operator input into one decide-once brief, then runs the rows: tiered workers dispatched per row, every row verified by a mechanical probe before it ticks, checkpoints kept in the run-list itself, and anything canonical/physical/unknown parked to a plain-English return board. Trigger on: 'execute the run-list', 'run the run-list', 'resume the run', 'conductor session on <run-list', or any ask to actually run or resume a staged run-list. Do NOT trigger for planning a build (that is run-list) or designing one (that is design-swarm).
design-swarm
(v2026-08-06.1) Run a design swarm: a divergence-first DESIGN pass that fans independent lens agents over a hard problem, competes approaches under different axioms, kills the weak ones at a critique gate, and stops at a written design the operator ratified. Never builds. Trigger on: 'design swarm', 'swarm this design', 'run a design pass on', 'brainstorm swarm', 'swarm the design process', or any ask to design/architect something non-trivial where the operator wants competing options rather than one answer. Do NOT trigger for build/implementation work (that is run-list territory) or quick single-answer questions.
run-list
(v2026-08-05.1) Plan and drive a multi-session build with a single durable run-list (a "build scope checklist" / "build sequence" / "session plan"). Use this WHENEVER a build is big enough to span several dependent work sessions and/or more than one repo or surface — the moment you'd otherwise start firing off sessions ad hoc. Trigger on: "run-list", "build sequence", "session plan", "map out the sessions", "break this build into sessions", "how should we sequence this", "plan the build", "build scope checklist", or any point where the user is about to kick off a large build across work sessions and needs the sessions ordered, prompted, and tracked. Also reach for it mid-build to log verified completions, capture ops/deploy state, and keep the doc posting-ready. Don't wait for the exact word "run-list" — if the work is clearly a multi-step build that will outlive one sitting, this is the tool.
setup
(v2026-08-06.1) First-run onboarding helper for the spire-pipeline config layer: replays the config discovery order, and either reports on an existing house-config.md (diff mode — read-only) or interviews the operator with two plain questions and writes a minimal starter config. Trigger on: 'setup', 'set up the pipeline', 'configure the pipeline', 'first-run config', 'house-config help', 'help me make a house config' — and 'check my config' / 'audit my config' for diff mode. Explicit invocation ONLY: no other skill calls this, and nothing in any skill's config-load step detects a missing config and routes here. Do NOT trigger for running the pipeline itself (design-swarm / run-list / conductor).
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.