Repository navigation
Releases
TokenAtlas 1.26.2
v1.26.2 Sprint 1: fix flaky CI, same-size rewrite refresh, faster overhead; r…
TokenAtlas 1.26.1
Changed
topandreportare faster on large histories (TOKEN-189, partial): on about 700,000 requests a changedcollectrun takes about 61 s fortopand 103 s forreport, down from 75 s and 135 s. The history loads with fewer allocations, limit and budget windows use a time index instead of a full scan per window, timestamps are parsed once, and the cyclic garbage collector is paused while a command runs (peak memory about 8 % higher). The embedded report payload uses gzip level 6, so its bytes differ while the decoded data is the same (#191).collectlogs the history size (requests=) beside each step's timing (#191).
TokenAtlas 1.26.0
Added
- Optional
tokenatlas[zstd]extra (zstandard) for reading compressed Codex rollouts; thezstdbinary on PATH works as a fallback (#185).
Changed
- List-price coverage is shown beside every cost total: the headline figures, the usage and work panels, and the model-share, price-comparison and interrupted-turn insights report the share of requests without a complete USD list price and the largest unpriced models, following the report filters.
doctorcounts only requests with no list price and warns separately about price entries older than 30 days.codex-auto-reviewis listed as explicitly unpriced (#189). - Price table re-verified on 2026-11-14: current Claude list prices,
inference_geo=usmultipliers, Ultrafast tiers with their own long-context prices forgpt-6.1-solandgpt-6-astra, per-model OpenRouter endpoint sources and OpenCode Zen free models at $0 (#189).
Fixed
overheadexplains an empty result with areason_code(no_rows_stored,filter_excludes_all,rescan_found_nothing,root_missing) and says whenoverhead --refreshis needed; overhead data is still collected only by that rescan (#184).- OpenCode v2 session storage is imported alongside the legacy layout with stable keys; legacy rows win for shared message ids, v2-only messages are linked to their turn by sequence, and a stale-parser check looks at the source data (#186).
- Codex rollouts compressed to
rollout-*.jsonl.zstand the~/.codex/archived_sessionsroot are imported with the same parser and keys as plain rollouts. Reading compressed files needs the optionaltokenatlas[zstd]extra (pip install 'tokenatlas[zstd]',pipx inject tokenatlas zstandardoruv tool install 'tokenatlas[zstd]') or thezstdbinary on PATH. Without either, one diagnostic and a count of unreadable files are reported anddoctorand the coverage section show the import as partial. Rollout files with an unrecognised suffix count as unreadable too. - Concurrent commands no longer fail with
usage: database is locked(TOKEN-190). The history uses WAL mode, so long analytics commands andrefresh/collectrun side by side. Analytics commands (report,insights,topwithout--keep-text,overhead,session,doctor,quota show) are read-only and no longer migrate the schema: after an upgrade runtokenatlas refreshonce. A lock between writers exits 75 with adatabase:prefix, andcollectretries and logs the steps it skips.
TokenAtlas 1.25.0
TokenAtlas 1.25.0 adds Berget's provider-reported per-request CO₂e and IT-energy estimates.
- Capture
usage.emissionsfrom retained Pi responses and import separately captured Pi or OpenCode responses with exact request matching. - Preserve estimates across history and snapshots, with coverage-aware CLI summaries and private/shared report exports. Missing estimates remain unknown; provider estimates stay separate from TokenAtlas's token-based energy proxy.
- Add an explicit-input Python port of Berget's published calculator and methodology notes.
Live validation covered ordinary and streamed Berget chat completions for one active model. Other models and endpoints have not been verified. OpenCode currently uses the offline import path.
After upgrading, rebuild a saved report with tokenatlas open --no-refresh.
TokenAtlas 1.24.1
TokenAtlas 1.24.1 makes the subscription dashboard easier to read when saved Codex quota readings contain different plans or limit identifiers.
- Highlights the latest dated saved Codex reading; equally newest scopes remain visible rather than being arbitrarily chosen or merged.
- Keeps older and undated Codex scopes in a collapsed, keyboard-accessible section with a count, observation dates and an explicit warning when the observation time is missing.
- Explains in Swedish and English that separate readings may belong to the same account; inferred counters are labelled quota groups, not authenticated accounts.
- Preserves raw percentages, reset times, stale/expired states, Claude scope order and report privacy. No collection, history or settings changes.
After upgrading to 1.24.1, rebuild the saved dashboard with tokenatlas open --no-refresh.
TokenAtlas 1.24.0
TokenAtlas 1.24.0 separates saved usage by subscription plan and preserves custom Claude Code statuslines while collecting quota readings.
- Filter report usage by subscription plan, with plan columns for requests and quota windows.
- Use
top --plan teamfor Business usage and see plan attribution and partitioned costs in top/insights output. - Keep unknown plans and incomplete costs explicit. Repeated-counter evidence may distinguish separate same-plan quota windows; inferred numbers do not identify or link accounts across resets or window lengths.
- Wrap an existing Claude Code statusline with
statusline --wrap -- COMMAND...: quota is recorded first, input and both output streams pass through unchanged, and the command's exit status is preserved. - Generate setup and undo snippets for the effective settings file; metadata is preserved and files are not edited. Select the original shell with
--wrap-shell(required on Windows; POSIX sh-compatible commands default to/bin/sh). - Bound wrapped commands and cleanup waits with a configurable timeout. POSIX cleanup covers the isolated process group, Windows cleanup covers the job; POSIX descendants that deliberately leave the group and wrapper SIGKILL remain limitations.
- Apply the same dated session plan attribution to requests, quota-only events, limit hits and budget calculations; reprice older Codex calibrations under this rule without rewriting saved readings. Retain original quota percentages and reset times.
- Keep plan filters reflected in hit counts and efficiency export metadata.
- Reduce redundant history annotation in plan cost facts (31% lower traced peak memory in a 400,000-record benchmark).
After publication, upgrade explicitly with tokenatlas upgrade --version 1.24.0, then rebuild saved history with tokenatlas open --no-refresh.
Quota percentages and reset times come from saved provider readings. Plan names and inferred quota-window labels are not authenticated account identities.
TokenAtlas 1.23.0
TokenAtlas 1.23.0 adds a focused subscription dashboard and clearer prompt previews.
- See the latest saved used and remaining percentages for five-hour and seven-day limits.
- See each reset time and its countdown, with reading age and explicit stale, missing and expired states.
- Switch between the subscription view and the full offline dashboard in Swedish or English.
- Read prompt excerpts more easily in the dashboard.
Upgrade explicitly with tokenatlas upgrade --version 1.23.0. Rebuild saved history with tokenatlas open --no-refresh.
The dashboard uses saved provider readings. It does not estimate percentages from token counts or costs, or claim separate account identities from plan names.
TokenAtlas 1.22.0
TokenAtlas 1.22.0 adds tokenatlas upgrade (#159).
- Check the latest compatible stable version explicitly with
upgrade --check. - Upgrade a verified pipx, uv tool, or dedicated venv installation; pin a version for deliberate rollback.
- Confirm interactive changes or supply
--yesin automation. - Keep a verified prior-environment backup and coordinate upgrades with scheduled collection.
- Verify the installed package and launcher, preserving usage history, configuration and schedules.
Rebuild an existing dashboard with tokenatlas open --no-refresh after upgrading. Older TokenAtlas versions need their original package manager for the initial update.
TokenAtlas 1.21.0
TokenAtlas 1.21.0 adds reproducible token-efficiency analysis to the offline dashboard and insights --json.
- Inspect concentration across work turns, large-context input, delegation volume and changes between equal time windows.
- Follow formulas, missing-data coverage and bounded pseudonymous contributors; automatic review remains separate from subagent work.
- Export compact evidence for your agent without prompt text, session identities or project paths.
- Calculate an explicitly selected hypothetical input reduction without claiming predicted savings or changing your workflow automatically.
- Use the Swedish or English dashboard, including mobile layout, filters and exact chart zoom.
Existing JSON fields remain compatible; the new token_efficiency extension is versioned. No LLM or network connection is required for the analysis.
Closes #157.
TokenAtlas 1.20.0
TokenAtlas 1.20.0 adds a token-focused analytics overview to the offline report.
- Browse 7- and 30-day windows with daily stacked token history and absolute/share legends.
- Compare token classes, models, agent roles, surfaces, harnesses and effort levels.
- Inspect tokens per request, tokens per linked turn, cache reuse and adjacent-period changes.
- Expand largest sessions and weekly breakdowns; inspect request and observed-turn timelines.
- See activity from retained turn context with explicit collection limits.
- Use the view in English or Swedish, including narrow screens and shared reports.
Unknown measurements remain unknown. Reasoning is included in output, and token efficiency is not presented as proof of answer quality. Local logs do not provide a full ChatGPT, plugin or skill activity stream.
No collector schema migration or new runtime dependency is required.