dsh-token-cost

Shows estimated cost for the current DSH Web session and all of its subagents, never an invoice. DSH logs each Assistant settlement's provider route, model ID and disjoint input/output/cache-read/cache-write token counts. The Host projection folds that durable log and prices it using the physically resolved pi-ai catalog and calculateCost belonging to the shipped @deepseek-ai/dsh-llm-pi-ai adapter. It makes no paid inference, opens no HTTP endpoint, and needs no dsh-model-pricing service or account credentials.

Readout

The bottom-row cost button shows the combined total only. Click it to open a popover with Main agent, Subagents (including nested children), and Total. Escape, outside click, or clicking the button again closes it. The existing hover tooltip stays short: “Estimated cost” (and “subscription” on subscription routes). Subscription estimates use the short note “(sub) · Subscription estimate”.

Each session contributes only requests it incurred itself: fork-inherited history is excluded, descendants are deduplicated, and ordinary user forks are not subagents. Child discovery uses the durable subagent catalog, so completed and unopened children are included. The standard non-activating projection read hydrates cold children; subsequent updates are pushed without polling or paid inference. Missing child usage, incomplete discovery, and unpriced requests are flagged in the popover instead of being treated as free. Independently rounded rows may differ slightly from the rounded total.

Installation

Install the latest tagged GitHub release:

dsh plugin --profile web add 'https://github.com/yoggu/dsh-token-cost.git#v0.2.2'

Or clone and link a checkout:

git clone --branch v0.2.2 --depth 1 https://github.com/yoggu/dsh-token-cost.git
cd dsh-token-cost
dsh plugin --profile web add "link:$(pwd)"

Keep linked checkouts in place while installed. Restart DSH after replacing an installed package version, then refresh the browser. Uninstall with dsh plugin --profile web remove dsh-token-cost; this removes the readout, not session logs or credentials.

The shipped llm-pi-ai adapter and its pi-ai dependency must be resolvable from the running Harness process. No process-wide search, cwd fallback, consumer-local pi-ai copy or arbitrary filesystem override is accepted; the old DSH_PI_AI_DATA override is no longer used. If owner resolution or catalog validation fails, settlements show unpriced, not free. The current implementation is tested with pi-ai 0.99.1, without changing the shipped adapter or its dependency.

Price mapping and provenance

• Exact openrouter routes use the installed pi-ai openrouter catalog. Custom gateway aliases are not inferred.
• Exact built-in openai-codex routes use openai-codex; displayed USD is an API-equivalent subscription estimate, not a ChatGPT subscription invoice.
• Historical codex-business and codex-personal log routes retain legacy subscription estimates from today's shipped pi-ai openai-codex catalog after the old account adapter is removed. This is a lookup of old event IDs only, not new-route admission or attestation that the historical adapter used this exact version.
• Exact deepseek-official routes use the deepseek catalog as an independent list-price estimate. Its Messages endpoint is not served by pi-ai and its billing parity is not attested by this plugin. An unknown model ID or gateway route stays unpriced.

For each valid entry, the Host report retains pi-ai package name, version, catalog-file SHA-256 digest, exact route binding label, and the number of contributing settlements; the Web tooltip only says “Estimated cost” (and “subscription” on subscription routes). The displayed cost is calculated using the currently installed catalog; the DSH log stores counts, not an original USD charge or historical price. If prices change after a pi-ai upgrade, intentionally increase the projection stateVersion and cold-replay events before claiming newly repriced history; a catalog revision alone does not invalidate checkpoints. This version supports pi-ai 0.99.1's typed chat:<model-id> catalog keys (as well as legacy bare keys), while looking up the bare model IDs recorded by DSH. Image/classifier entries are never used for chat pricing. It uses stateVersion: 7 to replay checkpoints from versions 1–6, recover estimates previously skipped by the new catalog format, and remove fork-inherited usage before descendant aggregation. There is no catalog filesystem I/O in the fold.

calculateCost selects the applicable request-wide tier using uncached input plus cache reads and writes. Reasoning is already part of output. Incomplete/non-finite/negative prices and ambiguous duplicate model entries fail closed. DSH lacks the cacheWrite1h bucket; models with duration-specific write rates or Anthropic-protocol cache writes are unpriced when their accounting cannot be faithfully reconstructed. The readout dims estimates with partial coverage rather than interpreting unpriced calls as $0.

Limits: Explicit route IDs do not prove custom endpoint ownership or actual billing; profile owners must not repoint reviewed routes silently. The estimate doesn't include non-model charges. DeepSeek parity and historical subscription prices are specifically caveated.

Run npm test for the complete offline suite. MIT; see LICENSE.