dsh-optimize

Long-session resource governor for DeepSeek Harness.

A conversation that keeps growing gets expensive long before anything compacts it: every step re-sends the retained history, the session log grows without bound, and the web UI has more history to draw. dsh-optimize measures that growth and reports it, so a heavy session is a visible fact rather than a surprise.

What it does

• Measures what every outgoing request actually carries (message count, characters, estimated token range) on each agent/pre-step.
• Reads the harness's own contextPressure projection on session events — the authoritative token figure, and O(1) — so the estimate above is never the only number kept.
• Keeps a per-session record — requests, events, last and maximum request weight, projected tokens, compaction count — in $DSH_HOME/storages/dsh-optimize/state.json.
• Logs a warning the first time a session crosses warnTokens, and again at criticalTokens, with the measured range.
• Optionally (off by default) runs the session's own compaction engine once the measured weight passes compactAtTokens, through the sanctioned ctx.agentPresets.serviceFor(agent, 'compaction') handle. The plugin decides when; the harness engine keeps deciding what to summarise.
• Provides the optimize host service (snapshot(id), list(), flush()).

It does not change the system prompt, the tool schemas, or the request prefix, does not write session events, and does not touch the session log.

Install

dsh plugin --profile <profile> add dsh-optimize

Or from a checkout:

dsh plugin --profile <profile> add /path/to/dsh-optimize

Configure

The bundle patch ships the defaults; override them in your profile's patch layer (~/.dsh/profiles/<profile>/cordis.patch.yml):

- id: dsh-optimize
  config:
    warnTokens: 120000
    criticalTokens: 300000
    autoCompact: false
    compactAtTokens: 200000
    compactCooldownMs: 300000
    verbose: false

The estimate is a range: Latin text is near four characters per token, CJK near one and a half. The upper bound is the one compared against the thresholds, so a mixed-language session errs towards reporting early. When the harness's contextPressure projection is available, its number wins for the compaction decision.

Notes on automatic compaction

• It is off by default, because it changes what the model sees. The built-in engine already compacts when the routed model's window is nearly full (ratio 0.8 by default) or when the provider reports a context overflow; this option only makes it fire earlier, at your budget.
• The engine is mounted inside the agent preset's isolate realm, so the only sanctioned handle is ctx.agentPresets.serviceFor(agent, 'compaction').
• Its thresholds are frozen at construction: nothing can retune the engine at runtime, and changing thresholdRatio/retainRatio means editing the preset row and re-mounting.
• Compaction needs an idle agent; a request made mid-turn is refused with busy and retried on a later check.

In the browser

The chat view mounts one DOM row per visible node of the loaded history and lays out every one of them; only the trajectory table virtualises. On a long session that is thousands of live rows.

• Lite rendering (on by default, switchable): a stylesheet puts content-visibility: auto and contain-intrinsic-size: auto 320px on [data-chat-anchor-key], so the browser skips layout and paint for rows outside the scrollport. React still reconciles them exactly as before, and the remembered row height keeps scroll anchors stable.
• Composer chip: shows the current state and how many rows are mounted; clicking it toggles lite rendering.
• Settings row: the same switch, under Settings → General.

The switch is persisted in localStorage under dsh-optimize:lite. Nothing about the session is hidden or unloaded: every row stays mounted and reachable.

Roadmap

• A context-pressure readout in the composer chip, read from the harness's own contextPressure projection.

License

MIT.