dsh-plugin-cost-lens

Cost Lens — DeepSeek API rates and per-session costs for dsh web.

• Sidebar rate pill (above Settings): current rate state (Peak hours / Off-peak hours; until the peak/off-peak schedule starts it shows Peak hours with the flat prices), an Ends in: hh:mm countdown to the next change, and a click-to-expand detail with Input: $x.xxx/M · Output: $x.xxx/M. The peak/off-peak windows are shown in the browser's local timezone. A small spinner shows while rates refresh; a warning dot + tooltip appear when the last update failed ("working with data from last update date").
• Chat stats strip (under the prompt box): ≈ $0.35 · 1.2M in · 80k out for the active session.
• Session rows + hover cards (sidebar): a small cost next to the time field and a cost line in the hover card.
• Schedule to off-peak: long-press the send button to open a small menu and queue the draft for the next off-peak window start; scheduled prompts show above the composer (a distinct tint) with a cancel button. The host half keeps the queue and fires each prompt via the session's agent (agent.followup) — it works while the dsh web server is running, with no browser open.
• Workspace cost totals next to each workspace in the sidebar (sum of its sessions' costs).
• Account balance on the pill: shown in the expanded view and tooltip, with the last-updated time on hover; refresh interval configurable (5 minutes default).
• Config card (Settings → Plugins → Configurable): rate refresh interval, balance refresh interval, backfill window, per-surface display toggles (session rows, session cards, chat stats, workspace totals), Refresh now, and Recalculate costs buttons — the plugin configuration page.

Install / rebuild / revert

From GitHub

One block — installs the plugin, registers the loader row, and applies the served-bundle shims (session rows / hover cards and the long-press schedule menu; they have no plugin slot, see Shims below):

dsh plugin --profile web add github:davidgereb/dsh-plugin-cost-lens --config.block-exotic-subdeps=false
PATCH="${DSH_HOME:-$HOME/.dsh}/profiles/web/cordis.patch.yml"
grep -q "name: dsh-plugin-cost-lens" "$PATCH" 2>/dev/null || cat >> "$PATCH" <<'EOF'

- insert:
    - id: ui-cost-lens
      name: dsh-plugin-cost-lens
EOF
node "${DSH_HOME:-$HOME/.dsh}/profiles/web/node_modules/dsh-plugin-cost-lens/scripts/patch-live-ui-workspace.mjs"
node "${DSH_HOME:-$HOME/.dsh}/profiles/web/node_modules/dsh-plugin-cost-lens/scripts/patch-live-ui-conversation.mjs"

Then restart dsh web (or just refresh the page if the host half was already running). The shim scripts are idempotent — re-running is safe, and --revert undoes them (the session-row pills and the long-press menu stop showing if you skip this step).

pnpm ≥ 11 note. The --config.block-exotic-subdeps=false flag lifts pnpm 11's default ban on git-hosted transitive dependencies — this plugin's dependency dsh-lib-context-injection is resolved that way. No config files to edit.

From a local checkout

# 1. build
node scripts/build.js
npm install   # installs the git dependency (dsh-lib-context-injection)
node scripts/smoke-test.mjs
# 2. make the package resolvable by the profile (repeat for the lib dep)
dsh plugin --profile web link /path/to/dsh-plugin-cost-lens
dsh plugin --profile web link /path/to/dsh-lib-context-injection
# 3. register the loader row in cordis.patch.yml (see above)
# 4. apply the served-bundle shims (idempotent; --revert to undo)
node scripts/patch-live-ui-workspace.mjs
node scripts/patch-live-ui-conversation.mjs

Host-half (lib/index.js) changes require a dsh web restart; browser-bundle changes (client.js and the shims) reach the GUI on a page refresh.

Shims

The session-row and hover-card cost surfaces have no plugin slot, so Cost Lens patches the served dsh-client-ui-workspace bundle; the long-press schedule menu patches the served dsh-client-ui-conversation bundle. Apply both after any dsh upgrade that re-serves those bundles:

node scripts/patch-live-ui-workspace.mjs            # session rows / hover cards / workspace totals
node scripts/patch-live-ui-conversation.mjs         # long-press send -> schedule menu
node scripts/patch-live-ui-workspace.mjs --revert   # remove (byte-exact)
node scripts/patch-live-ui-conversation.mjs --revert

Revert

1. Remove the ui-cost-lens row from cordis.patch.yml.
2. dsh plugin --profile web rm dsh-plugin-cost-lens
3. Revert the shims (node scripts/patch-live-ui-workspace.mjs --revert and node scripts/patch-live-ui-conversation.mjs --revert).
4. Restart dsh web.

Data model

The host half owns everything data-related:

• Rates: fetched from the official pricing page (api-docs.deepseek.com/quick_start/pricing), parsed into a schedule of pricing generations (the current flat table and the upcoming peak/off-peak table with its UTC windows), cached (TTL configurable, 1 day default), persisted, and served to the browser over /cost-lens/api (JSON RPC) and /cost-lens/events (SSE). A failed update keeps the last good data and marks it stale.
• Per-session costs: a costLens session-projection unit — a pure incremental fold that prices every assistant/message usage record (uncached input, cache-read, cache-write, output) at the rate in effect at that turn's timestamp, using the message's reported model. The fold state is persisted by the existing session-projection-cache, so costs are accumulated, not recomputed per page load; new tokens just extend the fold. Recalculate costs re-registers the unit at a bumped stateVersion, forcing a full refold with the current schedule (and discards stale cache rows).
• Backfill: on activation, sessions active within the backfill window (7 days default) are folded asynchronously in chunks. The projection stateVersion is persisted across restarts, and cache rows written at a stale version (after a recalc or a plugin version change) are re-folded automatically on activation — so per-session costs and their pills survive server restarts.

⚠️ Prices are an approximation. The per-session numbers are computed from the usage fields the session log records (input / cache-read / cache-write / output tokens) priced at the published per-token rates. They are a good estimate for tracking spend, not an exact bill: caching in particular is not directly observable from the logs — a turn's reported cacheReadTokens is the model's own accounting, the published cache-hit/cache-miss split is a simplification, and DeepSeek's actual billing can apply discounts or rounding the published table does not capture. Expect the totals to be close but not identical to the API invoice; the rate pill's per-model prices are the published list prices, not what you were actually charged.

Layout

dsh-plugin-cost-lens/
├── package.json              # dsh.client web plugin declaration
├── lib/                      # generated artifacts
│   ├── index.js              # host half
│   ├── client.js             # browser bundle
│   └── ui-workspace-shim.js  # shim for the served ui-workspace bundle
├── src/
│   ├── node-source.js        # host half source (single file)
│   ├── client-source.js      # browser half source (single file)
│   └── ui-workspace-shim.js  # shim source
└── scripts/
    ├── build.js              # generate lib/ from src/
    ├── patch-live-ui-workspace.mjs  # apply/revert the ui-workspace shim (byte-exact revert)
    ├── patch-live-ui-conversation.mjs  # apply/revert the ui-conversation shim
    └── smoke-test.mjs        # node logic + client registration + shim tests

How it's wired

Policy source

https://api-docs.deepseek.com/quick_start/pricing — peak hours 01:00–04:00 and 06:00–10:00 UTC, off-peak at half the peak rates, taking effect 2026-08-16T16:00:00Z. Before that date the flat table applies; the indicator shows Flat rate and the fold prices by each turn's timestamp, so history stays correctly billed across the transition. If DeepSeek changes the windows/prices, the fetch picks the new schedule up (or use Refresh rates now); the bundled default schedule in src/node-source.js is the fallback.

The parser handles both page layouts: the pre-transition page (flat table, "will be updated to peak" + effective-date sentences, one model per table row) and the post-transition page (peak/off-peak table only — models in a header row, one price column per model). Once the page stops listing the flat era, the flat generation is carried over from the bundled default (bounded by the known transition instant), so historical turns before the switch keep their correct flat pricing after a recalc or on fresh installs.

Compatibility. Tested against dsh 0.1.0-rc.6 on Node.js v24.19.0 (dsh web profile). Older or newer dsh releases may change the internals this plugin hooks into — check the changelog before upgrading.

⚠️ AI-generated, provided as-is. This project was written with the assistance of an AI. It is provided AS IS without warranty of any kind, express or implied. The author cannot be held responsible for any damage, data loss, or misbehaviour that results from using it. Use at your own risk.