ptp-run-at-model

Runs ptp command work in a subagent or Codex shell-out at a deterministic model and effort.

Updated Jun 8, 2026
One-click install
npx skills add https://github.com/AlmogMaayan/ptp --skill ptp-run-at-model-almogmaayan
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: ptp-run-at-model
Source: https://github.com/AlmogMaayan/ptp/tree/main/skills/ptp-run-at-model
Command: npx skills add https://github.com/AlmogMaayan/ptp --skill ptp-run-at-model-almogmaayan

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Claude Code sessions run work at whatever model and effort the session happens to be set to, and no tool can change the session model in place. This Skill solves that by running a ptp command's real work in a sub-context with a model override, so each pipeline step executes at a deterministic model and effort rather than the session default. ## Core Features & Use Cases - Deterministic model dispatch: Accepts a <model>.<effort> target (or an effort.md line-1 reference) from the calling command and runs the work at exactly that setting. - Dual main-agent branches: Runs work either as a foreground Claude subagent with an injected effort directive (main=claude) or as a write-capable codex exec shell-out (main=codex), resolved via layered config. - Ordered contract: Enforces outer abort-preconditions, branch-guard handling, telemetry ledger bracketing, and a verbatim result relay of completed, refused, or needs-human-action terminal states. - Use Case: When /ptp:plan needs to run at opus.high but the session is on Sonnet, the command passes the target to this Skill, which spawns one foreground subagent at Opus with the high-effort directive and relays the plan result back. ## Quick Start Reference this Skill from a ptp command by passing a target like opus.high along with a description of the work the resolved main agent should perform.

Frequently Asked Questions about ptp-run-at-model

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run a Claude Code command at a specific model and effort?▼

Pass a `<model>.<effort>` target such as opus.high or sonnet.medium to the dispatch skill, which spawns one foreground subagent at that model with the effort injected as a prompt directive. The session blocks until the subagent returns and relays its result.

How does the model override token work in ptp commands?▼

Commands may support an opt-in `model:<model>.<effort>` token embedded in the argument text, overriding the default target for that single invocation only. Invalid or duplicate tokens cause the command to refuse before any branch guard or subagent spawn.

Can ptp run its main work through Codex instead of Claude?▼

Yes, when layered config resolves main=codex, the skill shells out to a write-capable `codex exec` via Bash instead of spawning a Claude subagent. Model and reasoning effort are then sourced from tier-based config or the codex-model override token.

What happens if effort.md is missing or unparseable?▼

When a caller passes the read-from-effort.md instruction and the file is missing or line 1 is not a parseable model.effort pair, the skill defaults to opus.high and notes the defaulting. The run proceeds rather than aborting.

Why does the branch guard run in the outer session instead of the subagent?▼

Cutting a branch uses the ptp-branch-prep Workflow, which neither a subagent nor a codex exec shell-out can launch due to one-level Agent nesting. The guard therefore runs in the main session before the spawn, and the subagent is told its own guard check will be a no-op.

When should a step be downgraded to a cheaper model?▼

Only steps that are fully specified, leave no design judgment, and have a single verifiable correct outcome qualify as mechanical and may use cheaper tiers like haiku. Planning, decomposition, and review steps are judgment-carrying and must never be downgraded to save latency or cost.