Claude Skills by sblattj

Use when comparing two models, providers, or ferry lanes on speed, cost, or behavior before switching a lane — e.g. "is X faster than Y", "compare GLM vs Gemini", "what is this lane costing", "benchmark the flash lane". Covers the thinking-token traps that make naive benchmarks lie (reasoning eats max_tokens, mandatory-reasoning surfaces, ignored thinking budgets, chunk-burst decode rates), the streaming harness at scripts/bench-models.py, real-traffic A/B from the observability stack, and th...

Use when adding a fallback hop to a lane in llm-ferry's route proxy (~/.config/ferry/litellm.yaml) so it fails over to an independent model on error — the worker lanes' single Luna hop (flash/super-flash) is the live, worked example; wiring a chain onto the heavy driver lane (legacy orch/orchestrator) is the "if you decide it needs one" case, against current house policy. Covers strict-failover wiring (separate model_name, router_settings.fallbacks, chain order), the independent-capacity rule...

Use when adding a worker lane or more worker keys to llm-ferry's load-balanced pool in ~/.config/ferry/litellm.yaml. Covers pool-vs-fallback, the per-project quota rule and the Google ToS line you must NOT cross, and lane naming/aliasing. For the heavy driver lane's strict failover chain instead, use add-fallback-orchestrator; the local GPU lanes take neither.

Use when calling the task tool from a LOCAL lane (opencode-local, or any ferry local-* model) to delegate work to subagents. Prevents malformed task calls (hallucinated task_id, missing description, nested delegation) that silently stall local agent sessions; covers the exact three-field call recipe and retry-once rule.

Use when a goal is running or requested: the prompt carries a <goal_continuation> or <goal_plan> block, the sidebar shows a Goal panel, the turn says 'Start working toward this goal now', the user types /goal, pastes a handoff, asks for unattended multi-step work or to 'keep going until done', or asks how a goal is doing. Covers setting and decomposing goals, the goal_* tools, CEV evidence, the [goal:evidence]/[goal:complete]/[goal:blocked] markers, budgets, and the sidebar.