v0.3.5

• New compact button in the composer, next to handoff. It runs the built-in
  /compact through ctx.remote.commands.execute(sessionId, '/compact', []),
  so the user gets exactly the harness command — the same durable
  command/run / command/done records and the same result node as typing it —
  not a private reimplementation and not a second compaction path. The host half
  needs no change: the command is already registered in DSH. remote.commands
  is a separate Cordis service (the remote service only associates it), so the
  button is mounted by a child ctx.inject(['remote', 'remote.commands'])
  plugin; the chip and handoff still load when that service is absent.
• The button is honest about what it is doing. It disables itself and shows
  compacting… while the request is in flight, confirms success in the UI
  language, and on a failure shows the host's own reason (busy, cancelled,
  changed history, no useful summary) in the tooltip. A host without the command
  or without the command channel says so instead of pretending it worked.
• This is the one place where tokens can be spent, and only on a click: the
  plugin still makes no model calls of its own, but /compact asks the harness
  summarizer to write the summary. All three READMEs now say this next to the
  "no tokens" promise instead of leaving it as a silent exception.
• The smoke suite clicks the real component against a stub Remote: it pins the
  /compact line, the session id, the empty attachment list, the busy state,
  the localized success, the "nothing to compact" answer, a refused remote
  call and a handler error result (132 checks).

v0.3.4

• DSH 0.2.1-rc.3 support declared — that is the newest release (published
  2026-10-23). No host- or client-facing code change was needed: every contract
  the plugin touches is unchanged between 0.2.1-rc.1 and 0.2.1-rc.3 — the
  session/event payload, the tokenUsage / contextPressure projections, the
  session query surface, credentials.resolve and both balance shapes,
  webServer.register, the model catalog (llm.resolveModelInfo), the two
  client slots (conversation.session.header.utilities,
  conversation.input.right) and uiWorkspace.startSession. The compaction
  formula in dsh-compaction-basic is byte-for-byte identical to rc.1
  (min(thresholdRatio * window, window - reserve - headroom), headroom 65536),
  so the threshold and the bands stay right.
• What rc.2 and rc.3 did change does not reach the plugin: V4.1-Flash-1017 and
  DeepSeek Search/Fetch extend the model and tool set (the window still comes
  from the catalog), the new "More" plugin menus add slots but do not move the
  two the plugin uses, and the API-key model-list sync is the same account path
  the plugin already reads.
• Verified live on rc.3, not just on paper: the status route returns the token
  projections and the catalog window, the handoff route returns a document, the
  balance arrives from the stored API key, both client slots exist in the
  client contract, and dsh --profile web --dump-config keeps
  id: context-governor enabled.
• The compatibility tables in all three READMEs name 0.2.1-rc.3, and the smoke
  suite now pins the declared range (^0.1.7-rc.2 || ^0.2.0-rc.1) instead of
  only checking that some range exists, so an accidental narrowing fails before
  a release.

v0.3.3

• The handoff document now covers the whole session instead of its tail. It used
  to take the last user message as the task and the last assistant message as
  the state, so a long session read as "what happened at the end" with no path
  to it. The host now walks the model surface in order: the first task, then
  every user request with the last assistant answer to it (the session arc),
  then the current state and the touched files.
• Injected harness messages are not requests. AGENTS.md, the runtime-context
  snapshot, the skill catalog and the time snapshot arrive as user/message
  events whose source.kind is not user; before, they became arc steps and the
  answer to the previous turn stuck to them. Now only real user messages — or
  messages from an older host without a source — start a step.
• Several messages of one turn fold into one step: a user who sends text and then
  an image is one request, and the image is marked [изображение] / [image] /
  [图片]. Touched files are reduced to one form (relative to cwd), so a file no
  longer appears twice under two different paths.
• A very long session keeps its beginning and its end (24 steps) and says how
  many middle requests were skipped, instead of dumping everything.
• The smoke suite covers the arc: first task versus last, the middle, numbering,
  en/zh localization, injected-message filtering, image steps, same-turn merging
  and the long-session gap marker.

v0.3.2

• The balance is no longer short by the granted wallet on the platform-account
  path. The account service returns topped-up wallets in value and granted
  ones in bonusWallets; the plugin summed only value and reported it as the
  total, so a $29.53 grant next to a $19.26 top-up read "$19.26, granted
  $29.53, topped up $0" — smaller than the real balance and internally absurd.
  The total is now topped-up plus granted in the primary wallet's currency
  ($48.79 in that example), with the two parts still shown separately.
• Wallets in another currency are surfaced instead of discarded. DeepSeek keeps
  yuan and dollar wallets side by side, and the DSH account panel shows them as
  "$29.53 + ¥1.45"; the plugin used to drop everything but the first currency.
  The status payload now carries an others list (also filled from the extra
  balance_infos of /user/balance), and the chip, the panel row and the
  tooltip append it as "+ ¥1.45". Yuan and dollars are still never added to one
  another.
• The smoke suite pins the account arithmetic against the real wallet shape
  (topped-up 19.26 + granted 29.53 + CNY 1.45 -> 48.79 USD + ¥1.45) and the
  multi-currency /user/balance response; all three READMEs document the suffix
  and the assertion count.

v0.3.1

• DSH 0.2.1 support declared. The 0.2.1 line needs no code change: every host
  contract the plugin touches is unchanged there — session/event with
  request/context and request/header (same payload fields, same
  contextWindow and config.maxTokens), the tokenUsage / contextPressure
  projections (onChanged, stateOf), the session query surface (readSurface,
  filterEvents), credentials.resolve, the account balance shapes
  (credential-stored, value / bonusWallets), webServer.register, the
  model catalog (llm.resolveModelInfo) and both client slots
  (conversation.session.header.utilities, conversation.input.right) — and
  uiWorkspace.startSession only gained an optional argument. The compaction
  formula in dsh-compaction-basic is byte-for-byte identical between
  0.2.0-rc.2 and 0.2.1-rc.1, so the threshold and the bands stay right.
• The compatibility range now lives in engines.dsh. DSH 0.2.1 checks that field
  and ignores peerDependencies once it is present, so this is the declaration
  that actually decides; the peers stay so package managers and older hosts keep
  reading a range. The same range still matches 0.2.1-rc.1 because the host
  compares prereleases included, which is what the plugin quietly relied on
  before.
• On Web, DSH 0.2.1 resolves a stored credential before the environment. The
  balance follows with no code change — it resolves the key through
  credentials.resolve('DEEPSEEK_API_KEY') and only then falls back to
  DEEPSEEK_API_KEY from the environment — so the documented order now matches
  the host's.
• The DSH compatibility tables in all three READMEs name 0.2.1-rc.1, and the
  development section's assertion count now matches the smoke test (101, not the
  stale 58).

v0.3.0

The peak/off-peak tariff now follows Chinese public holidays, rates and the seasonal tariff are per provider, and a provider without a prompt cache can now declare cached: false.

• The peak/off-peak tariff now honours Chinese public holidays. DeepSeek's rule
  is peak on weekdays excluding Chinese public holidays, and off-peak all day
  on those holidays — but the plugin only knew weekdays and hours, so on the
  National Day holidays (1–7 October 2026) it reported peak (×2) and doubled
  the rates and the step cost while DeepSeek was charging the off-peak half.
  The year's calendar is fetched in the background from a maintained source
  (holidayUrl, default chinese-days data), cached on disk under
  <DSH_HOME>/cache/context-governor, and refreshed on its own — no yearly code
  edit. holidays config takes 'YYYY-MM-DD' dates or ranges and wins over the
  source; holidayFetch: false keeps the plugin fully offline. Until the
  calendar is known the weekday rule applies and the panel says so
  (season.holidayKnown); the countdown now skips whole holidays to the next
  working peak, and the tariff row reads "off-peak (Chinese holiday)".
• README.zh.md: a full Chinese translation of the README, since the plugin
  already ships a Chinese interface. The language switcher in both existing
  READMEs links to it, and the stale assertion count in the dev section was
  corrected to match the smoke test.
• The balance is printed to the cent. fmtMoney used three decimals below 100
  (a leftover from the step-cost formatter fmtUsd, where a step can cost a
  fraction of a cent), so a ¥8.95 balance read ¥8.950 while the tooltip
  printed the raw 8.95. Money is now two decimals everywhere — chip, panel and
  tooltip, in every currency.
• The READMEs state plainly that the step cost and the rates are in US dollars
  while the balance is shown in the source currency (USD or CNY) without
  conversion, so the two must not be compared directly.
• At peak the balance and the time indicators are red. The host's season.color
  is now red during peak (it was amber), and the client paints the chip balance,
  the chip tariff marker, and the panel's tariff / countdown / Beijing-time /
  balance rows with it, so the expensive hours read at a glance.
• Fixed a literal "undefined" in the English and Chinese interfaces. The client's
  balanceUnknown string existed only in the Russian dictionary, so on a
  non-DeepSeek provider the panel's balance row and the chip tooltip rendered
  undefined instead of text. A smoke test now compares each language's keys
  against English recursively, so a missing key fails the suite instead of
  leaking into the UI.
• No price UI at all on providers the plugin has no rates for. Previously the
  chip and the panel wrote "price unavailable" for cline and friends, which was
  noise around numbers that do not exist. The status payload now carries
  config.priced, and the client drops the whole price block when it is false:
  the chip loses its cost segment, and the panel loses the step-cost,
  relative-to-base and rates rows. The peak marker stays (it is a clock reading),
  but its ×N multiplier appears only where rates are known.
• Rates and the seasonal tariff are now per provider. providerRates.<id>
  overrides the default freshRate / cacheReadRate / cacheWriteRate /
  outputRate (and peakMultiplier), so a second provider can carry its own
  numbers, and a provider with explicit rates is priced even if it is not in
  pricedProviders. The DeepSeek season — peak/off-peak, the ×N multiplier and
  the Chinese holidays — now applies only to a seasonal provider, by default an
  id starting with deepseek; a foreign provider declares seasonal: true in its
  providerRates entry to opt in. Behaviour change: a non-DeepSeek provider no
  longer has its rates doubled in DeepSeek's peak hours, and it gets neither the
  tariff marker nor the tariff rows.
• providerRates.<id>.cached: false for a provider with no prompt cache. The
  arithmetic already worked (zero cache buckets contribute nothing), but the UI
  and the signals did not know: the rates row still listed cache-read/cache-write
  prices, the panel showed cache rows and a permanent 0% cache-hit, and the
  cacheHitLow warning fired on every step because 0% is below the 90% floor —
  plus a cold-prefill alarm on each request, since every token is fresh by
  definition there. With the flag the cache rows, the cache rates, the chip's
  cache-hit marker and both warnings go away; a caching provider is unaffected.

v0.2.0

First public release.

A session context chip for DeepSeek Harness: prompt size, a cache-aware per-step
cost estimate in USD, cache-hit, bands and the compaction threshold, DeepSeek
balance and the peak/off-peak tariff, plus a one-click handoff.

It never calls a model and spends no tokens.

Highlights

• Token usage comes from the harness tokenUsage projection, with an
  llm/stream fallback for builds without projections.
• The model window and pressure come from the contextPressure projection.
• The compaction threshold is computed per session, exactly as
  dsh-compaction-basic does; a session never inherits another model threshold.
• Cache input is split into read and write, each with its own rate; cache-hit
  counts reads only.
• The balance prefers the harness deepseekAccount service and falls back to
  the DeepSeek API key; it is shown only for DeepSeek providers.
• UI in English, Chinese and Russian.

Install

dsh plugin --profile web add github:mpinaev/dsh-context-governor

Requires DeepSeek Harness 0.1.7-rc.2 or newer on the 0.1.x/0.2.x lines.

Checks

• npm test — 34 assertions, no network, no model calls
• ./scripts/compat-check.sh — boots a spare instance and probes the live API