dsh-deepseek-budget

How it finds usage

DSH keeps a dedicated adapter, because its log is zstd-compressed and carries stream timing. Everything else is found by content discovery: the plugin walks the home directory, keeps any .json/.jsonl file whose bytes contain deepseek, and parses it with one generic record reader that understands the provider/model/usage shapes every harness writes.

That means a harness nobody anticipated is still counted, and a log that merely mentions DeepSeek in prose is not — the parser only accepts records that carry token usage on a DeepSeek route.

Harness seen on this machine	DeepSeek calls
DSH	✅ counted (dedicated adapter)
Pi (incl. every subagent transcript)	✅ counted (discovery)
Claude Code, Codeg, Copilot, opencode	❌ installed, but no DeepSeek calls

Only DeepSeek routes are counted. A harness that also drives Claude or a local model contributes just its DeepSeek calls; everything else is reported separately and kept out of the budget.

Duplicates

A harness may write the same call into two files — a subagent's transcript and its parent session. Copies are collapsed globally by the provider's own response id, falling back to a strict fingerprint of time, route and token counts. On this machine that removed 593 of 1,419 Pi records; the raw count would have overstated spend by 72%.