Repository navigation
v2.6.12
简体中文: 查看本版本的中文更新日志
2.6.12 - 2026-10-09
Added
- Opening an OpenCode session fills the modal's chat timeline and inspector drawer, which until now said no transcript was stored.
_opencode_rich_session_detailreads OpenCode's database read-only and returns per-message metadata -- role, timestamp, model, character and token counts, and whether the message carries reasoning or tool calls -- plus per-tool metadata for name, status and whether it had arguments. Metadata only: no prompt, response, reasoning or tool output text is read or sent to the browser. It picks the v2session_messagehistory or the legacymessagetable the same way the session list does, so the two surfaces cannot disagree about which store is authoritative, and it readspartsonly for messages with no inline content. The read runs when a session is opened and never during a refresh; if it fails, the request logs a warning and still returns the session and its turns with empty message and tool lists rather than failing the modal. (#210, thanks @blasphemy) - The Sessions tab's client filters carry the same tool logos and names as the rest of the dashboard instead of text-only chips, and a name collapses to its logo when the row runs out of room rather than pushing the row wide. The measurement that decides when to collapse is kept stable so it does not flap between states as the row re-renders, and the Mimo mark stays readable in dark mode. (#209)
docs/development/ARCHITECTURE.mdtraces the path a usage event takes -- local client logs through the parsers into the SQLite usage store, through pricing andcompute.py, and out to the API, the TUI and the statusline integrations -- with a section on the quota subsystem.docs/reference/QUOTA.mdgains a sequence diagram of the background poll and its reset-boundary samples. (#188, thanks @Dhruv1401)docs/reference/CONFIG.mddocuments 23TOKDASH_*variables that lived only in the code, across four new sections: background warming, sessions and parsing, quota polling, and TUI and update checks. The quota entries are one line each and defer the design toQUOTA.md.TOKDASH_COMPUTE_CONCURRENCYnow gives its real default,max(2, min(8, cpus // 2)), and says that excess requests wait for a slot rather than failing outright, andTOKDASH_CLAUDE_PROFILESis described as quota-only, since usage and session parsing always scan~/.claude*/projects. (#186, thanks @Dhruv1401)- All six READMEs carry a CI status badge linking to the workflow runs and the current PyPI version linking to the project page. (#185, thanks @Dhruv1401)
- New pull requests start from a template with sections for the description, related issues, category, testing and a checklist, the checklist restating the rule in
RELEASING.mdthat a change toREADME.mdupdates all five translations in the same PR.pytestmoved intorequirements-dev.txtso the contributing guide installs it from there rather than separately. (#184, thanks @Dhruv1401) /openapi.jsondescribes itself: a description, the package version, contact and license, and six tags (Pricing, Usage, Quota, Sessions, System, Updater). All 27 API routes carry a tag and most carry a one-line description taken from their docstring; the dashboard page, manifest and service worker stay untagged. (#193, thanks @Dhruv1401)- Four models from the 2026-10-08 provider scan:
gpt-6.1-sol-pro,claude-haiku-5.5,mistral-large-4-0andgemini-nano-banana-2.1, taking the pricing DB to 2.0.34 and 364 models.switchyard(nvidia) was the scan's fifth auto-approved addition and is left out on purpose -- OpenRouter publishes that router with prompt/completion-1as a variable-pricing sentinel, which scales to -$1,000,000 per MTok and would price every request through it negative. The scan's 95 price changes are review-required and stay in the updater's diff report for a human pass. (#208)
Changed
- The dashboard no longer serves Swagger UI or ReDoc.
/docsand/redocloadedswagger-ui-dist@5andredoc@2fromcdn.jsdelivr.netpinned to a major version with no integrity check, and since the vendored assets in 2.6.10 they were the only third-party scripts Tokdash ran. A script served on the dashboard's own origin can fetch/api/csrf-tokenexactly as the dashboard does, so a compromised CDN package opened on localhost would have picked up the write token along with read access to sessions; opening either page also sent the visitor's IP and the dashboard's address to jsDelivr, and both rendered blank offline./openapi.jsonstays -- plain JSON, no third-party code -- so OpenAPI viewers and client generators keep working, andAPI.mdremains the human-readable reference. (#201) - Freebuff has its real mark. What shipped was a placeholder: a rounded badge with a plain sans
Fin a green freebuff.com does not use. It is the site's own glyph now, redrawn from the two polygons its header publishes, in the#22c55ethat site repeats across its landing styles, and being monochrome it rides the dark-mode inversion the wayzed.svgdoes. Muse Code has no product mark of its own -- its docs site wears the Meta symbol -- so it keeps its letter tile, but the tile's accent and its README pill had been two different hues, one of them purple; both are#0064E0now, the blue that gradient uses, with a comment on each side tying them together.SUPPORTED_CLIENTS.mdwas missing the Muse Code pill all six READMEs carry, andmake_agent_pills.pyrecords thatpills/devin.pnghas no source mark to rebuild from. (#207) - The contributing guide installs
requirements-dev.txtthe way CI does instead of pytest alone, so the tests that needhttpx(FastAPI's test client) orbuild(the packaging tests) no longer fail for someone following it, and it gained Windows activation commands. Running from source now setsTOKDASH_DATA_DIR=output/dev-data, matchingCONFIG.md, because a checkout ahead of the installed Tokdash migrates~/.tokdashto a schema the installed version then refuses to read. (#200)
Fixed
- The six Overview sparklines draw recorded usage instead of a drawing. Each carried a fixed decorative path -- the same rising curve for every user, every date range and every server -- and now derives its week curve from the raw per-server Usage and Stats payloads the dashboard already fetches, so no curve costs an extra request. The daily fold sums its numerators rather than averaging daily percentages, which is the difference between a 1/1 day beside a 0/9 day reading 10% and reading 50%; the Top Model curve tracks the model the card actually named rather than whichever model won each day; a date outside the Stats response's own coverage is a gap rather than a measured zero, since a payload computed yesterday cannot have measured today; and a curve is dropped instead of carried over when its range, server set or in-flight generation is not the one the headline belongs to. Where a curve cannot be drawn honestly the card says why, in all six locales -- no hourly breakdown in what the page already loads, daily values that disagree with the current total, today still in progress -- rather than showing a shape. One regression case runs against a response captured from a running build, because a guard written against a field the API has never sent passes every invented fixture and then clears the real payload. (#205)
- Devin CLI reports usage from a live
sessions.db. A populated store keeps each call's tokens atmetadata.metrics, the model atmetadata.generation_modeland the row id atmessage_id, and the parser read none of them, so a real store reported no usage and logged the containerless warning. It now checksmetadata.metrics,metadata.usageand both performance-metrics spellings after the documented locations, takes the model from the node's metadata, and keys an entry bymessage_idwhen the node has no id.SUPPORTED_CLIENTS.mdrecords the verified layout in place of the "not yet verified" note. (#206, thanks @MurmuredVelpro) - On a screen wider than about 1936px the content column sat flush against the sidebar with all the spare width on the right. It lost its
mx-autoin the design rebase in #129; restoring it centres the header and every tab view within the main area, leaves layouts at 1920px and below unchanged, and keeps the Report view on its own narrower centred width. (#211, thanks @674019130) - Three CLI errors say what to do next. A
tokdash db verifythat still finds the stored database disagreeing with the live parse after its retries carries aguidancefield pointing attokdash db resync; an unknowntokdash dbaction lists the valid ones, which is reachable becausedbandquotashare one positional choice list sotokdash db pollgets past argparse; and a bad--portcatches only theValueErrorthatint()raises and words its failure the same way its sibling does. (#182, thanks @Dhruv1401)