dsh-cost-ledger

Cross-session persistent cost ledger for DeepSeek Harness.
Auto-logs every LLM token-usage event to SQLite and exposes record_cost / query_cost / set_budget agent tools. Built-in DeepSeek pricing, overridable via plugin config. Install-and-go, no extra wiring.

Status

Phase 1 (host side) — complete & live-verified in DSH 0.1.0-rc.6 (last verified 2026-08-13). 28 data-layer checks pass (pnpm selftest); the plugin loads and runs in a real dsh web profile.

The token-usage source and tool registration API are confirmed from DSH 0.1.0-rc.6 source and wired for real:

Integration	API	Status
Token usage capture	ctx.on('llm/stream', (options, next) => AsyncIterable<StreamChunk>) — Cordis waterfall around every model call	✅ wired
Tool registration	ctx.tools.register(ToolDefinition) — real { name, description, parameters, output:{schema,render}, execute(args, exec) }	✅ wired
Persistent store	SQLite via better-sqlite3 (prebuilt win32 binary, no compile)	✅
Pricing	DeepSeek official (CNY/1M tok), cache-aware; config-overridable	✅
Live load in dsh web	installed via dsh plugin --profile web add ., host loads the bundle, apply() runs, SQLite opens	✅ verified
End-to-end token capture	dsh --profile headless "..." → real LLM call → llm/stream listener fires → usage chunk harvested → SQLite row written	✅ verified

Live-verified: a one-shot dsh --profile headless "reply with exactly: hi" produced real rows in ledger.db ({model:"glm-5-2", inputTokens, outputTokens, cacheReadTokens}) — the full llm/stream → usage-chunk → SQLite pipeline works against the real host.

Module loading uses .ts extension specifiers (./store.ts) — the Cordis loader rewrites these and loads TypeScript directly, so no build step is needed; the plugin runs from source after install.

Two issues found & fixed during live integration:

• { global: true } on ctx.on('llm/stream') — without it the listener only sees calls from its own fiber; the agent loop dispatches from a different scope. (Mirrors the host's own invariant.js usage.)
• No uninjected service access — ctx.session/ctx.workspace throw "cannot get property without inject" unless declared in inject. Session id now comes from the llm/stream event's options.sessionId instead.

Phase 2 (WebUI dashboard panel) — researched, not implemented. See Phase 2.

---

What it does

• Auto-logs token usage: subscribes to the llm/stream waterfall and writes {timestamp, session, project, model, inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, cost} per real provider call (including retries).
• Computes cost: built-in DeepSeek pricing (deepseek-chat / deepseek-reasoner, CNY per 1M tokens). inputTokens is uncached input; billed input = input + cacheRead + cacheWrite, each priced at its own rate. Override or add models via config.
• Three agent tools:
  • record_cost — manually record a usage entry (backfill/test).
  • query_cost — total spend + per-model breakdown within a range (today/week/month/all/ISO); reports current daily-budget status.
  • set_budget — create/update/remove a spending budget (daily/weekly/monthly/never window, label-scoped: daily, model:<name>, project:<name>).

Install (community / standard)

dsh plugin --profile web add <path-or-package>

The host discovers the plugin from package.json (dsh.bundle.patch) — no absolute path, no manual wiring. The prepare script builds lib/ from source, so a git checkout installs self-contained. Restart dsh web after install.

Git install note: pnpm ≥10 blocks a git dependency's prepare script until allowed. If the first add fails, copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml (allowBuilds: dsh-cost-ledger: true) and re-run — exactly as the official docs describe.

Local development

pnpm install
pnpm selftest              # verify the data layer (28 checks)
pnpm typecheck
pnpm build                 # build host (lib/*.js) + client (lib/client.js)

# load into a running dsh web via the patch (absolute path, this machine):
dsh web --patch ./cordis.patch.yml

Pricing — built-in official rates

Built-in model pricing (元/百万 tokens, from each platform's official pricing page):

• deepseek-v4-flash — input 0.10 / output 3.0 / cacheRead 9.0 (peak tier)
• deepseek-v4-pro — input 0.30 / output 9.0 / cacheRead 27.0 (peak tier)
• deepseek-chat (V3) / deepseek-reasoner (R1) — legacy aliases
• glm-5-2 — default 0 (override in the dashboard ⚙️设置 tab or config; fill from Zhipu's pricing)

Unknown models record tokens only (cost = 0) until you set a price. Override anytime via the ⚙️设置 tab in the dashboard or the WebUI settings card — changes apply to new calls immediately.

Config

Edited in the dashboard ⚙️设置 tab (runtime, not persisted across restart) or under the WebUI plugin config card (durable) / cordis.patch.yml:

config:
  dbPath: 'dsh-cost-ledger/ledger.db'      # relative to DSH cwd
  defaultDailyBudget: 0                     # CNY; 0 = no limit
  pricing:
    deepseek-chat:                          # override or add any model
      input: 2
      output: 8
      cacheRead: 0.5
      cacheWrite: 8

Architecture

src/
  pricing.ts   built-in DeepSeek price table + computeCost()
  store.ts     SQLite ledger: insert + summaryByModel + spentSince + budgets
  tools.ts     three ToolDefs (pure handlers over store + config)
  config.ts    Schema (schemastery) config
  index.ts     apply(ctx): llm/stream listener + registerTools + cleanup
  client/      Phase 2 WebUI entry (stub)
cordis.patch.yml   local-dev profile patch (absolute path)
scripts/selftest.ts   standalone data-layer verification

Token capture (confirmed API)

llm/stream is a Cordis waterfall wrapping every streaming model call (retries included). The listener transparently forwards the underlying stream while harvesting the usage chunk:

ctx.on('llm/stream', (options, next) => (async function* () {
  let usage: TokenUsage | undefined
  for await (const chunk of next()) {
    if (chunk.type === 'usage') usage = chunk.usage
    yield chunk
  }
  if (usage) record(options, usage)   // persist one row
})())

Key billing facts (from official TokenUsage):

• inputTokens = uncached input only. Billed input = input + cacheRead + cacheWrite.
• reasoningTokens is already included in outputTokens — never double-count.
• usage is not guaranteed (early abort/error) — absence is logged at debug.

The TokenUsage / StreamChunk / GenerateOptions types live in @deepseek-ai/dsh-llm (an unpublished host-injected package). This plugin declares minimal local types to typecheck standalone; swap for a real import type once the host package is resolvable.

Phase 2 (WebUI panel)

Researched against the dsh-web-ui reference. Key finding: DSH exposes no general-purpose sidebar-panel slot for external plugins. sidebar.workspaces / sidebar.settings are single-occupant and taken by the shell. Options:

• DOM-level bypass (the dsh-task-board / dsh-ssh approach): MutationObserver + createRoot into [data-pane="sidebar"] / [data-pane="conversation"]. Most flexible.
• Settings card via web-ui.plugin.item (lightest — a summary card under Settings → Plugins).
• No chart-library precedent in the ecosystem. Recommended: inline SVG, or vendor recharts (strict CSP, no CDN).

Compatibility

• DSH version: verified against 0.1.0-rc.6 (the API surface — ctx.on('llm/stream'), ctx.tools.register, ctx.inject, agentDefaultModel — is confirmed from DSH 0.1.0-rc.6 source).
• Last verified: 2026-08-13.
• Profiles: loads under both dsh web (full UI + HTTP API + dashboard) and dsh --profile headless (token capture + tools only; HTTP API and dashboard are web-profile-only and silently skipped).
• OS: the SQLite backend ships a prebuilt win32 binary via better-sqlite3; other platforms build from source at install time (requires a C++ toolchain). Verified on Windows 11 / Node 22.
• Pre-release caveat: DSH is in developer preview with expected compatibility-breaking changes. The token-usage source and tool registration API are confirmed against rc.6; a future mainline may rename them, which would require a plugin update.

Install

dsh plugin --profile web add <path-or-package>

The host discovers the plugin from package.json (dsh.bundle.patch) — no absolute path, no manual wiring. The prepare script builds lib/ from source, so a git checkout installs self-contained. Restart dsh web after install.

Git install note: pnpm ≥10 blocks a git dependency's prepare script until allowed. If the first add fails, copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml (allowBuilds: dsh-cost-ledger: true) and re-run — exactly as the official docs describe.

Uninstall

# Remove the bundle from the profile, then restart dsh web
dsh plugin --profile web remove dsh-cost-ledger

This removes the bundle from the profile's dsh.profile.bundles list and unlinks the package. The ledger database (dsh-cost-ledger/ledger.db) and any config overrides in cordis.patch.yml are not removed automatically — delete them manually if you want a clean sweep:

Remove-Item -Recurse -Force dsh-cost-ledger   # the data dir created under the DSH cwd

To disable temporarily without removing: comment out dsh-cost-ledger in the profile's cordis.yml bundles list and restart.

Quick start

# 1. Install into your web profile (see Install above)
dsh plugin --profile web add .

# 2. Restart dsh web, then just use the agent normally
dsh web

Every model call is now auto-logged. After a few prompts, ask the agent to query spend:

> 查一下今天花了多少钱（query_cost today）

Or open the dashboard panel in the Web UI (the cost-ledger tab) to see a live summary, per-model breakdown, and the budget/budget settings — no extra wiring.

Permissions & data

• Files written: one SQLite database at dsh-cost-ledger/ledger.db (path configurable via dbPath, relative to the DSH working directory) plus its WAL/SHM sidecar files. Nothing else touches the filesystem.
• Network: none. The plugin makes no outbound network calls. (The parse-prices endpoint calls the host's own ctx.llm.stream() — it reuses the model call path you already configured, not a new connection.)
• Credentials: none read, stored, or transmitted. The plugin never touches your API keys; it only reads token counts and model names that the host already emits on the llm/stream event.
• Data logged per model call: {timestamp, sessionId, project, model, provider, inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, cost}. No prompt content or completions are ever stored — only aggregate token counts and metadata.
• HTTP API: when running under dsh web, the plugin registers read-only-ish JSON endpoints under /api/cost-ledger/*. They bind to the host's web server and are reachable from the same origin as the Web UI. The cleanup and parse-prices endpoints are write/POST endpoints intended for the dashboard UI.

Troubleshooting

• Cannot find module './store.ts' / load error after a git pull: the Cordis loader resolves .ts specifiers directly, but a stale lib/ can interfere. Run pnpm build (or pnpm install to trigger prepare) and restart dsh web.
• better-sqlite3 install fails on macOS/Linux: the prebuilt binary is win32-only; other platforms compile from source. Install a C++ toolchain (build-essential / Xcode CLT) and re-run pnpm install.
• No cost rows appear: confirm the model call actually emitted a usage chunk (early aborts/errors may not). Token capture is best-effort — absent usage is logged at debug level. Also confirm dbPath is writable.
• parse-prices returns 503 / "default model not configured": the agentDefaultModel service hadn't resolved yet, or no default model is set in Agent settings. Set a default model and retry.
• pnpm blocks the prepare script on git install: add allowBuilds: dsh-cost-ledger: true to the profile's pnpm-workspace.yaml and re-run dsh plugin add.
• Logs: plugin diagnostics go to the host log (dsh web console / dsh-run.log); the SQLite ledger itself is the system of record for cost data.
• Rollback: dsh plugin --profile web remove dsh-cost-ledger + restart; optionally delete the dsh-cost-ledger/ data dir (see Uninstall).

Development

pnpm install
pnpm selftest              # verify the data layer (28 checks)
pnpm typecheck
pnpm build                 # build host (lib/*.js) + client (lib/client.js)

# load into a running dsh web via the patch (absolute path, this machine):
dsh web --patch ./cordis.patch.yml

react / react-dom and the @deepseek-ai/dsh-client-* runtime packages are external to the client bundle (resolved by the host's module loader at runtime), so they are correctly declared as devDependencies / peerDependencies, not dependencies. The only runtime dependency is better-sqlite3.

License & security

MIT — see LICENSE.

Security reporting: this plugin has no network surface and stores no credentials, but if you find a vulnerability (e.g. unsafe SQL handling, path traversal via dbPath), please do not open a public issue. Report it privately via GitHub Security Advisories (Security → Report a vulnerability on the repo). All SQL uses parameterized statements (@named bind params) and all filesystem access is confined to the configured dbPath.