Decisions Session

Decisions: Session Facts

This page holds two facts about what a session costs. Read it before you change how a slice reports cost or how setup sets the prompt cache.

Why a Call Reports Its Own usage

Choice: A hook that calls $.model.complete takes the cost and tokens from the usage of that result. It never takes them from total_cost_usd of the session. The advise tool of 0.2.0 took its tokens this way. The tool is removed in 0.3.0.

Reason: In all 4 measured runs (Claude Code 2.1.302), total_cost_usd did not count the model call of a hook.

Evidence: ModelCompleteResult.usage in the mod API types. The runs are on Evidence.

Why Setup Does Not Set the Prompt Cache TTL

Choice: The setup snippet has no promptCacheTtl and no subagentPromptCacheTtl (D18). Setup removes a promptCacheTtl of 5m that an earlier setup wrote.

Reason: Claude Code picks the TTL itself when the keys are unset. The main conversation gets 1 hour on a Claude subscription within its usage limits, and 5 minutes on an API key, Bedrock, Vertex, or Foundry. Everything outside the main conversation gets 5 minutes, unless ENABLE_PROMPT_CACHING_1H=1. A fixed value in the snippet is wrong for one of these accounts.

Evidence: Both names are in the settings schema of the installed Claude Code 2.1.305, in schemas/claude-code-settings.json and schemas/claude-code-managed-settings.json. just schema extracts the schema from the bundle. The env vars CLAUDE_CODE_PROMPT_CACHE_TTL and CLAUDE_CODE_SUBAGENT_PROMPT_CACHE_TTL take precedence over the keys.

Related Pages

• Decisions Engine
• Evidence
• Decisions