Skip to content

zoltanmaric/coppersushi / .agents/skills/model-selection

model-selection

Verified

Selects models for authorized delegation and independent reviews using available models, task difficulty and recorded quota usage. Use before launching a sub-agent, selecting a reviewer or escalating a model.

Model selection

Honor explicit model choices and the current harness's delegation permissions. This skill chooses how to staff authorized work; it does not authorize extra agents or change the current conversation's model.

Discover available models and usage tools before claiming either is unavailable. For Codex, call get_usage_limits. After each successful check, pass the response's JSON payload (not its tool envelope) on stdin to python .agents/skills/model-selection/scripts/quota.py record, using the selected local Python environment. Transfer JSON as data, not interpolated shell code. The recorder uses only the standard library on macOS/Linux.

The recorder stores account/bucket identifiers, observation times, usage percentages and reset windows in ~/.local/state/agent-quota/usage.csv, shared across repositories. It omits other response fields and uses a file lock for concurrent writers. Never commit live responses or the log. Use --log for isolated test data. If access is denied, report the recording limit and use the live reading; do not relocate private history into the repository or claim it was saved.

Choose the lightest suitable available model:

WorkPreferred family, when available
Clear, bounded inspection or routine workTerra
Ordinary implementation and investigationSol
Difficult reasoning or architecture that needs stronger reasoningAstra

Use equivalent models when the harness exposes different names. Prefer Astra for a first full independent review when the scope warrants it and quota allows it. When quota is tight, use a lighter model even for the first full review. For follow-up reviews of specific fixes, default to the lightest suitable model. Match reasoning effort to difficulty. Do not launch duplicate attempts by default.

Use all applicable quota windows: remaining capacity, time until reset and task scope. Compare observed consumption rates with the recorder's sustainable rate (remaining percentage points divided by hours until reset). Its early/late rates describe observations, not a forecast or per-model cost. Sparse or brief coverage does not establish a full-window trend. Do not infer spare capacity in one bucket from another. Respect live availability flags and retry delays even if a percentage suggests room.

If quota data is unavailable, keep concurrency modest and use task difficulty to choose. If recording fails, use live headroom without claiming a historical trend. Recheck during long tasks and before expanding concurrency or escalating models; do not poll just to fill the log. The recorder saves material changes and an hourly observation when checked, while each summary includes the live reading.

Apply the choice at launch

Before each launch, choose the effective model and concurrency from the task and the latest quota evidence. When remaining capacity is tight and recent consumption is materially above the sustainable rate, default to one delegated agent at a time. Treat a brief spike as a reason for caution.

If continuing at the current pace is likely to exhaust quota before reset, tell the user before launching. Give the remaining quota, reset time, the lighter model you will use and any work it may leave unresolved. Proceed with that model on work it can handle. Identify any work that still needs stronger reasoning. Honor explicit user model choices and live availability limits.

Inherit the parent model only when it matches the chosen model. Otherwise pass the chosen model explicitly in the spawn call. For example, when an audit calls for Sol and the parent uses Astra, set model: "gpt-6.2-sol". If an override requires a fresh context, give the agent a self-contained brief. Keep a short record of the task, effective model, concurrency and quota reason.

Before escalating, distinguish missing context or tool failure from a capability gap. Give concrete feedback and escalate only when the remaining task needs it. The lead integrates and verifies delegated results. Give independent reviewers a fresh context and a self-contained brief, even when they inherit the parent model. In Codex, set fork_turns: "none" at launch.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall