Forecost — AI skill for Claude Code

forecost

**The independent economic receipt for AI-agent runs.** Forecost is designed to record a content-minimized causal account of a run, keep measured usage separate from the different ways it can be valued, and show where local runtimes, gateways, pricing tables, and imported provider evidence agree or disagree.

Local-first. No signup. No hosted Forecost service.

[License: MIT](LICENSE) · Local branch test matrix: Python 3.10–3.13


*Conceptual architecture illustration—not a UI screenshot, security boundary, or claim that every adapter supplies every source shown.*

The public repository's `main` branch and badges still describe the legacy product. This unreleased 0.3 checkout exists only on the local `codex/forecost-product-core` branch at the audited baseline. Do not use a public clone, PyPI, or a marketplace command as though it installs this code.

**Historical packaged-product audit (2026-08-09).** The source declares 0.3.0, but the changelog remains **Unreleased**. Graph receipts, the local evidence journal, deterministic Run Lab, offline reconciliation, and single-host resource envelopes are experimental local features. Claude Code and LiteLLM integrations are best-effort; OpenAI Agents and LangGraph are offline mapping contracts only. Live provider HTTP ingestion, complete graph coverage, distributed enforcement, verified task correctness, and future task-cost forecasting are not claimed.

**Red-team hold (2026-08-13; still in force).** The current worktree repairs the code-level defects found in trace identity, timing, receipt serialization, evidence scoring, import authority, canonical-ledger durability, owned-state privacy, CI failure mode, and plugin launch trust. Specifically, schema v10 introduced trace-scoped `span_key`; receipt schema v2 requires explicit tim