Skip to content

Experience-Driven Black-Box Cost Control for Coding Agents

1.9 KB
Raw Blame

Tool pruning

How tool pruning (see the main README) decides what to drop, per agent.

  • Evidence pruning (Claude Code, Copilot CLI; stage.ToolingStage). After each session, a loaded tool's unused streak grows or resets. In the experiments a tool is proposed once its streak reaches evidence sessions. The plugin asks for more (plugin/fold.py): unused over toolprune_evidence prompts across at least toolprune_min_sessions sessions, or over toolprune_sessions sessions, and three times as long for NotebookEdit. A protected work-surface tool (tools.PROTECTED_TOOLS) or one ever seen in use is never proposed. In the experiments only, a per-task judge (tools.judge_tools_for_task) can also drop tools the task text makes irrelevant, its cost charged to the ledger; the plugin has no task text and never calls a model to prune.
  • Capability pruning (Codex; codex.CodexUntoolProfile). Codex's unit is a capability package (collaboration, goals, image/media, ...) switched off in its config. A package is demoted only after enough loaded-but-unused sessions, at most two per session, and a single use protects it for good. The plugin proposes a demoted package only once it also meets the prompt-and-session evidence above (three times as long for image/media).

This package decides what to drop. Writing the decision into an agent's own settings file is the plugin's job (tokeneconomizer/plugin/denylist.py): by itself under toolprune=auto (the default), once the repository's sessions have been idle toolprune_quiet_minutes, or on toolprune apply under propose; either way it takes effect from the next session.

Module
stage.py ToolingStage: evidence across sessions, the proposal, the Codex mask
codex.py CodexUntoolProfile: Codex's evidence-gated capability controller
tools.py the protected and inapplicable tool sets; the experiments' per-task tool judge