@local/dsh-token-saver

Token saver

Two halves of one DSH bundle:

• Host half (index.js) — the peak-hour guard. It listens to the agent/pre-step waterfall and returns { kind: 'reject' } for every proposed model request made while the provider bills peak rates.
• Browser half (client.js) — the Settings page that switches the harness's token-saving machinery on and off, plus the savings strip under the composer that prices the session and steers the guard.

Install

dsh plugin --profile web add github:vitkuz573/dsh-token-saver
dsh web

Restart the host process afterwards: host code is not hot-reloaded, so an edit to index.js only takes effect on a restart. Client-only changes appear on the next page load.

Compatibility

Built against DSH 0.2.0-rc.2. The bundle declares its own patch, so DSH's installer refuses it on a runtime whose API does not match — that refusal is the intended behaviour, not a bug to work around with allow-version.

The rate schedule follows DeepSeek's published pricing: peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Chinese public holidays are not modelled, so those days are refused even though DeepSeek bills them off-peak.

What the guard does

A turn that is already running is not torn apart: it finishes the step it is in (its tools complete) and only the next model request is refused, so the turn ends as blocked. Two consequences worth knowing before a peak hour:

• The rejected step has already claimed its user message, and a blocked turn discards that message. The strip therefore warns before a send; if the text matters, grant an override first.
• The guard applies to every agent in this profile that goes through the agent/pre-step waterfall — this GUI session, subagents, workflow children, and any other client. It does not cover requests that bypass that waterfall, such as automatic condensation.

The rate schedule is modelled as two weekday windows in UTC minutes — [[60, 240], [360, 600]], that is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; every other instant is billed at half price. Chinese public holidays are not modelled, so a holiday is refused until an override is granted.

Steering it

Control	Effect
/peak-allow [minutes]	Allow model requests for that many minutes (default 15, max 720)
/peak-block	Drop the override and refuse again
/peak-guard [on|off|status]	Disarm, arm, or describe the guard
Strip button Allow 15m	The same as /peak-allow 15, one click, only shown while refusing
Strip detail panel	Allow 15m / Allow 60m, Block now, Guard on/off, Host status

The commands are commands, not tools, so an agent can never grant itself an override. An override lives in Host memory and disappears on restart.

Developing it

npm test        # schedule equivalence, guard decisions, and a strip render

The schedule is duplicated verbatim in both halves so the browser can price an instant without a round trip; test/pricing.test.mjs asserts the two copies agree minute by minute, and that the guard admits, refuses, and expires correctly. test/guard-ui.test.mjs mounts the strip against a stand-in React and walks its element tree.

Host code is not hot-reloaded. dsh-hmr watches only the profile's patch files, and re-composition reuses the already imported plugin module, so edits to index.js take effect only after restarting dsh web. Edits to client.js appear on the next page load.