tokenuse 0.0.12, named Credit Ledger, catches the pricing books up with the mid-2026 model wave, teaches the Copilot integration GitHub's new AI-credit billing, restores Copilot CLI session capture from its new SQLite stores, and adds an Agent Setup audit page.

Highlights

• Pricing books now cover Claude Fable 5, Claude Mythos 5, Claude Opus 4.8 (including its fast mode), and Claude Sonnet 5, so Claude 5-family usage no longer falls back to Sonnet 4.6 rates.
• Claude Sonnet 5's introductory pricing steps up to standard rates automatically on September 1, 2026 through a dated override row.
• Copilot pricing tracks GitHub's AI-credit model lineup: GPT-5.6 Luna/Sol/Terra, Claude Opus 4.8, Claude Sonnet 5, Claude Fable 5, Gemini 3.5 Flash, MAI-Code-1-Flash, and Kimi K2.7 Code.
• Copilot limit gauges understand GitHub's June 1, 2026 switch to AI-credit billing: post-switch quota snapshots read as "AI Credits" with the remaining balance in credit units.
• Copilot ingestion reads the CLI's SQLite stores, so sessions from CLI builds that stopped writing events.jsonl around May 2026 show up again, with real token totals where data.db records them.
• Added an Agent Setup page to the TUI and desktop app with a local-only, redacted audit snapshot of agent home folders, archive usage, project coverage, tool knowledge files, MCP declarations, and recent context hygiene.

Pricing

The LiteLLM import filter now includes the claude-fable, claude-mythos, and grok families, and the curated Anthropic rows add Claude Fable 5, Claude Mythos 5, Claude Opus 4.8, and Claude Sonnet 5. Fast mode is modelled per row: Opus 4.8 fast at 2x base pricing, and Opus 4.7 fast at its 6x rate (deprecated upstream, with removal scheduled for July 24, 2026).

Claude Sonnet 5 ships with introductory pricing that Anthropic raises on September 1, 2026. The upstream book carries the introductory rate while a curated override with effective_from: "2026-09-01" takes over from that date, so no manual follow-up is needed when the step-up lands.

Copilot's tool-scoped rows follow GitHub's post-switch pricing page, which now uses per-provider tables with tiered Default and Long-context rows; row matching selects the Default tier. Models GitHub retired from the page (GPT-4.1, GPT-5.2, GPT-5.2-Codex, Grok Code Fast 1, Goldeneye) keep their last-known prices for historical usage. The claude-opus alias now points at Opus 4.8 and openai-auto at GPT-5 mini, Copilot's included model.

Copilot AI Credits

GitHub moved every Copilot plan to usage-based AI-credit billing (1 credit = $0.01) on June 1, 2026 while keeping the legacy premium_interactions key in its usage payload. The limits parser now detects the credit era through the new token_based_billing field or the observation date, labels post-switch gauges "AI Credits" with the remaining balance in credit units, and parses the newer quota_reset_date_utc reset field. Older sidecars keep the legacy "Premium Interactions" label, and the pacing insight now speaks in AI-credit terms.

Copilot CLI Capture

The Copilot CLI stopped writing per-session events.jsonl files around May 2026, which made newer sessions invisible. tokenuse now discovers and parses the CLI's central SQLite stores. Turns in ~/.copilot/session-store.db are estimated with the same chars / 4 rule as VS Code transcripts and take their project from the session's working directory or repository. Sessions in ~/.copilot/data.db carry authoritative token totals and real model ids; each becomes one aggregate call that the archive refreshes in place as live sessions grow.

Both stores run in WAL mode, so parsing works on a private copy of the database and its -wal/-shm sidecars, and the adapter's source fingerprint folds in -wal metadata so archive syncs notice un-checkpointed turns.

VS Code transcripts now fall back to the session's declared model when tool-call id inference finds no known prefix, instead of pricing those sessions as the generic copilot-auto bucket. Inference still wins when a known prefix appears, because declared models have been observed to disagree with the backend that actually served the session.

Notes

• Local usage analysis remains local-only. Network calls are limited to explicit Config-page downloads, opt-in quota sync actions, local CLI advice generation selected by the user, or maintainer refresh and release paths.
• Archived calls keep their import-time cost by design, so usage imported while the Claude 5 family was missing from the books stays at fallback pricing unless corrected manually. Refreshed pricing applies to newly imported calls.
• The release tag should be v0.0.12 and must match the root Cargo package, desktop Cargo package, desktop package.json, and Tauri config versions.