cost-measurementlisted

# Cost Measurement **`gaia-research/skill-cost` is the canonical basis for every cost measure in the Gaia ecosystem.** Do not hand-roll token accounting, do not read harness JSONL yourself, and do not quote a price from memory. If a number is going into a report, it comes from here. Repo: <https://github.com/gaia-research/skill-cost> · registry id `gaia-research/skill-cost` --- ## Why this and not self-reported numbers `skill-cost` reads the **session logs the harness actually wrote** and prices every turn against [BerriAI/litellm](https://github.com/BerriAI/litellm)'s canonical `model_prices_and_context_window.json`, refreshed weekly. That means: - Numbers are **persisted usage**, not a model's guess about its own consumption. - Cache reads and cache writes are counted separately — which matters enormously, since cache reads routinely dominate total tokens and are priced very differently. - Every harness is priced the same way, so cross-harness arms are comparable. Asking a model how many tokens it used produces confabulation. Reading the log does not. --- ## Supported harnesses | Harness | Session storage | |---|---| | pi | `~/.pi/agent/sessions/**/*.jsonl` | | Claude Code | `~/.claude/projects/**/*.jsonl` | | OpenAI Codex | `~/.codex/sessions/**/*.jsonl` | | opencode | `~/.local/share/opencode/**/*.jsonl` | | Hermes Agent | `$HERMES_HOME/state.db` (SQLite, read-only; usage tables only, never transcripts) | **Known gap: Grok is not supported.** If a Grok arm ne