0.5.1

Patch release, 2026-10-23. Correctness only: no new features, no pricing-catalog
change, no peer-range change, no new dependency.

What changed

0.5.0 shipped a large UI release. This patch is the result of a full review
sweep over the plugin's data path, pricing engine and client state, run as
small, individually tested fixes. 95 commits since v0.5.0; every confirmed
defect carries a regression test under tests/regression/.

Aggregation and time boundaries

• A session whose activity spans midnight no longer loses tokens in either day
  window, and UTC slots are attributed to the local day that contains them.
  The previous slot loop keyed days by slot * 30min, so a local evening in a
  non-UTC zone could land on the wrong date.
• updatedAt is now the maximum of the real timestamps instead of
  max(...) || null, so a session whose only timestamp is 0 keeps it.
• An invalid or out-of-range timestamp no longer aborts the whole export or the
  whole aggregate; it degrades to an empty field.
• One unreadable session no longer fails the entire aggregate: the session is
  reported as unavailable with a warning and the rest of the data still loads.

Pricing

• DeepSeek legacy peak rates now exclude Beijing weekends, matching the
  time-of-use rules. Previously a Saturday morning was billed at peak.
• contextOver512k: false is honoured instead of being treated as "unknown".
• providerId comparisons are case-insensitive, so DeepSeek and deepseek
  no longer resolve to different rules (and no longer report unsupported).
• A historical USD total converts at the FX row for its own usage day instead of
  the newest row; the day is carried through merges and the wire payload.
• Calendar validation rejects impossible dates (2026-02-30) and an empty
  catalog is rejected instead of validating successfully.
• tierMultipliers with a falsy value can no longer bypass validation, and a
  multiplier of 1 is no longer shown as an extra pricing condition.
• A summary mixing free rows with unpriced rows no longer reports free.

State, cache and versions

• Route rows stay ordered after a cache hit; previously an updated row was
  spliced into a stale position.
• The pricing store keeps the previous settings when a settings read fails,
  restores the pinned catalog version exactly, and returns a copy of its history
  instead of the internal array.
• Projection cache rows are rejected when their identity (id, cwd,
  createdAt) or lifecycle does not match the session header.
• Warning objects no longer leak their internal count field to the RPC payload.

Null and prototype safety

• Lookups keyed by user data (__proto__, constructor) use own-property checks
  in the project filter, balance chips, route cache and trend series.
• null projects, sessions, windows and balance entries are skipped instead of
  throwing, and an empty-string balance is treated as "not reported" rather than
  "invalid".

Client behaviour

• The balance alert bell appears again (it was suppressed by a truthiness bug),
  and the alert input no longer keeps a stale value after an external change.
• An expired resetsAt no longer reads as "resets soon".
• The session table's "sessions" column sorts by a real value; the model share
  ring, heatmap legend and average line match what is actually drawn.
• The day-mode button restores the day you left instead of writing null, so
  switching Overview → All → Day no longer resets to the newest day.
• Closing the pricing drawer returns focus to the trigger, and Tab stays inside
  the modal drawer.
• CSV export uses the session-level authoritative cost when the host provides
  one, so a session no longer shows two different amounts in the summary column
  and in its detail rows.

Kept as-is on purpose

One review finding was rejected during release review because the proposed
change contradicted an existing, documented contract:

• The subagent group header keeps showing the project-wide subagent count. A
  per-page count makes the same header change while paging and breaks the
  browser fixture baseline.

Compatibility

Unchanged from 0.5.0: same peer ranges, same engines.dsh
(>=0.1.0-rc.6 <0.3.0), same signed catalog. The only wire change is one
optional pricing field on a cost total; older payloads remain valid and the
schema still rejects unknown fields.

Harness	Verification
0.2.1-rc.2	isolated real-module contract: 6/6 checks pass on Node 22.22.0
earlier rows	unchanged, see docs/compatibility.md

Verification

• Full test suite: 652 tests / 152 files pass in the default timezone and
  TZ=UTC (up from 505 at 0.5.0; 125 regression tests).
• Browser fixture: 50 check groups / 80 screenshots, 0 console errors,
  0 runtime exceptions, 0 failed requests at 1440/390/320 px, including the
  preview-before-save hint added in this release.
• Real Harness contract on 0.2.1-rc.2: TYPERT manifest, JSONL persistence,
  projection init/restore/checkpoint with fork exclusion, aggregate wire schema
  and live snapshotEvents agreement.
• Signed catalog 2026101824: 900 rules, verified against the pinned public key.
• Fresh build matches committed lib/*; npm pack preview contains the same ten
  files as 0.5.0 (282 kB).

Upgrade

# Web
dsh plugin --profile web add -w @rongyi7/dsh-stats@0.5.1

# Desktop app: use the CLI bundled with the app
"/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" plugin \
  --profile desktop add -w @rongyi7/dsh-stats@0.5.1

Restart the host after replacing the plugin. The signed catalog updates
independently; do not reinstall merely to receive new prices.

Known limitations

• The sweep is not exhaustive: open findings that would need a product decision
  rather than a defect fix are left for a later release.
• Live balances, unpriced-price flows and partial/stale banners remain
  fixture-only on this release; no new real-account run was performed.
• The browser fixture runs headless Chrome; it cannot cover OS-level focus
  behaviour in the desktop shell.

0.5.0

Pre-1.0 feature release, 2026-10-21.

Highlights

• DSH 0.2.1-rc.2 support: peer ranges, isolated Harness contract matrix and
  compatibility docs now include 0.2.1-rc.2. The contract smoke accepts both
  plugin export shapes used by 0.2.0 (default) and 0.2.1 (apply object).
• Official new-model onboarding: the 30-minute pricing workflow collects
  OpenAI, Anthropic and Google official pages. A new text model with an
  unambiguous standard price is appended as an estimated rule carrying a
  [provisional] note and auto-signed. Existing model price changes,
  retired/limited models, non-text models and ambiguous identities remain
  review candidates.
• Longest-known model fallback: route suffixes such as
  global:...-sg can resolve through the longest known alias instead of
  becoming 待计价; estimated pricing is never presented as exact.
• Latest catalog coverage: GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon
  rules, plus provisional official coverage for current Anthropic and Google
  models.
• Project overview rework (0.5.0): a cost-first KPI row, Token-structure
  legend, active-span strip, a seven-column project table with a sticky header,
  a filter menu with hidden-state persistence, and expandable project/session
  details. The panel has a fixed height and a single body scroll for both the
  overview and the timeline. The header, status banner and pricing drawer stay
  in place. Esc closes the top layer (filter, menu, drawer) before the panel.
• Pricing settings drawer (0.5.0): the old long pricing page is replaced by
  a 560px three-section drawer (full panel width on narrow screens): Models,
  Custom rules, and Updates & versions. Models use 25-group pages with
  used/all filtering, search, flat/tiered summaries, and complete rule detail;
  custom rules use a list/editor split with conditions, two-layer drafts,
  pending removal/undo, preview invalidation, and a fixed action footer.
  Updates & versions separates the current catalog, check/sync times, the
  auto-update draft, and 5-at-a-time catalog snapshots. Restoring a snapshot
  pins it and pauses automatic updates while preserving custom rules; unpinning
  requires a previewed save and does not trigger a refresh. Model history rates
  remain distinct from whole-catalog snapshots. See
  docs/pricing-settings-0.5.0.md.
• Timeline rework (0.5.0): the day view now uses project lanes above
  half-hour cost/token bars, a keyboard-reachable half-hour detail popover and
  expandable session rows. The all view is a calendar matrix, 30 days per page,
  with intensity/project coloring, idle rows, week gaps, a time-of-day profile
  and daily bars. The timeline has no internal viewport; the panel keeps its
  single body scroll.
• Balance page rework (0.5.0): all connected accounts now appear in one
  urgency-ordered list instead of a provider dropdown. A summary strip totals
  available balance per currency, highlights the tightest quota and counts the
  accounts that need attention. Rows show the balance breakdown or remaining
  quota bars, can link to the provider console, and support a per-account alert
  line stored in the browser only. First load uses a stats/providers skeleton;
  a single failed account stays in its own row while a whole-query failure moves
  to the header and a retry banner.
• Responsive detail: 320/390/640 layouts keep session time, turns/steps,
  LLM and tool time, cache hit, input/output and model in a non-truncated meta
  line; wide rows keep every column unclipped.
• Trends rework (0.5.0): the trends tab now offers 7/30/90-day ranges and a
  cost/token switch, with a summary strip, daily bars, a fixed 26-week heatmap
  and a model ring plus sortable table. Clicking a bar or a heat cell opens that
  day in the overview, and the page follows the project filter.
• Repository entry: the panel title (“DSH 用量” / “DSH Usage”) is now a link
  to https://github.com/rongyishuaige7/dsh-stats, opened in a new window with
  noopener noreferrer. It is keyboard reachable, keeps the dialog's accessible
  name, and is not reachable while the pricing drawer is open.
• README and bilingual guide: the README now covers desktop and Web
  installation side by side, and the detailed install, pricing, account,
  data-status and export material moved to docs/usage-guide.md and
  docs/usage-guide.en.md. README.zh.md stays a pointer page.

Trends page (0.5.0)

• Ranges: the last 7, 30 or 90 days, always ending today. The range drives
  the summary strip, the daily bars and the model section; the heatmap is fixed
  at 26 weeks and says so in its own heading.
• Metrics: cost or tokens. Switching changes the bars, the heatmap shading,
  the model shares and the ring.
• 26-week heatmap: one cell per day, shaded by quartiles of active days, with
  active days, current streak and longest streak beside it.
• Day jump: clicking a bar or a heat cell opens that day in the overview and
  moves focus to the overview tab.
• Project filter: hidden projects are excluded here exactly as in the
  overview and timeline, and the hidden state is shared across the three pages.
• Current streak: it now means a run that reaches today or yesterday, so a
  streak that ended last month no longer reads as current. The previous
  "current streak" counted back from the most recent active day.

Compatibility

Harness	Verification
0.1.3-alpha.2 through 0.2.0-rc.2	previously verified; see docs/compatibility.md
0.2.1-rc.2	isolated real-module contract: 6/6 checks pass on Node 22.22.0; real Web runs at 8b047ce (R1/R16 and UI checks) and at 9611c25 (shell, four tabs, pricing drawer, 0 errors)

0.2.1-rc.2 is supported by peer ranges. The desktop app currently bundles
this Harness release; clean installation should no longer report peer mismatches.

Verification

• Full test suite: 505 tests pass in the default timezone and TZ=UTC.
• Browser fixture: 50 check groups / 80 screenshots, 0 console errors,
  0 runtime exceptions, 0 failed requests at 1440/390/320 px. The fixture covers
  the pricing drawer model list/detail, custom drafts, preview races, pinned and
  unpinned updates, seven signed historical snapshots, empty history, restore
  success/failure, long identities/forms, short windows, and the previous
  6b–6e regressions.
• Pricing collector: the OpenAI Standard section now carries an 11-column table
  (output moved to the second column, cache writes split into 30m and 12h) next
  to the 4-column table for older models. The adapter maps columns by header
  name, keeps the 30m write as the cache-write rate, and leaves the 12h write
  unmodelled. Before this fix every row was skipped and the run failed closed.
• Real Harness contracts: matrix includes 0.1.3-alpha.2, 0.1.5-rc.1,
  0.1.5-rc.2, 0.1.5-rc.3, 0.1.6-alpha.1, 0.1.7-rc.2, 0.2.0-rc.2,
  and 0.2.1-rc.2.
• Real Web (0.2.1-rc.2, plugin built at 9611c25): npm run smoke:web passed
  with the panel, four tabs, exact status and the pricing drawer (catalog
  2026101817, 876 rules); 0 console errors, 0 runtime exceptions, 0 failed
  requests. The temporary home had no credentials or sessions, so the balance
  and overview zero states were exercised, not live data. See
  docs/compatibility.md.
• Desktop real device (DeepSeek Harness on macOS, bundled Harness 0.2.1-rc.2;
  plugin tarball built at 9611c25, whose lib/* is identical to this release):
  the maintainer reviewed every page and reported no blocking issue; the panel
  title opened the repository in the system browser; a custom rule was added,
  previewed and saved, then removed and saved again.
• Signed catalog 2026101817: 876 rules, verified against the pinned public key.
• Fresh build matches committed lib/*; npm pack preview contains the four
  runtime bundles, license and READMEs.
• Pricing auto-publish workflow remains separate from this UI slice.

Upgrade

# Web
dsh plugin --profile web add -w @rongyi7/dsh-stats@0.5.0

# Desktop app: use the CLI bundled with the app
"/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" plugin \
  --profile desktop add -w @rongyi7/dsh-stats@0.5.0

Both commands install the published 0.5.0. To try a local build first, install the
rongyi7-dsh-stats-0.5.0.tgz tarball with the same command shape. The desktop profile is managed
by the app, so a separately installed npm dsh refuses it.

Restart the host after replacing the plugin. The signed catalog updates
independently; do not reinstall merely to receive new prices.

Known limitations

• Provisional rules cover standard rates only. Fast/Priority/Batch/Flex,
  1-hour cache writes and special tiers require review or remain unsupported.
  OpenAI's 12-hour cache write is not modelled; the 30-minute write is used.
• Provisional rules are not promoted automatically. A maintainer reviews the
  [provisional] rule, changes confidence to exact, removes the marker and
  signs through sign_reviewed_catalog.
• Existing model price changes intentionally require review.
• Meta open-weight models do not have a single first-party token price; they
  follow the actual host/OpenRouter route.
• Real Web verification at 9611c25 covered the shell, four tabs, exact status
  and the pricing drawer. Live balances, non-zero statistics, unpriced-price
  flows and partial/stale banners remain fixture-only. The R1 generation Esc and
  R16 approval Esc checks date from the 8b047ce run and were not repeated.
• The fixed-height panel aligns day, all and timeline bottoms at 1440 px;
  the timeline no longer has an internal viewport. Narrow layout and
  non-truncation are asserted separately at 320/390/640 px.
• Desktop acceptance did not itemize Chinese/English, light/dark, a narrow
  window or opening a pending model directly; those states have fixture
  scree...