Snow-ea/dsh-token-optimizer

dsh-token-optimizer is a cordis bundle purpose-built for DeepSeek Harness 0.2.0-rc.2 Web profiles, shipped as an installable plugin rather than a stand-alone agent or generic MCP server. Installation is the single command dsh plugin --profile web add dsh-token-optimizer@0.2.3 followed by a restart of the running dsh web process; once mounted the bundle registers a root-level BasicCompactionEngine adapter that engages DSH's cache-aware replay, branded Session API, and routed image accounting while preserving the complete structured values that run_code programs depend on. The plugin declares a localized title, description, and icon so the Plugin Manager displays it as a first-class bundle alongside the four shipped agent presets.

Inside the DSH agent lifecycle the engine watches each mode's pressure and overflow events: in Standard, PTC, and Creator modes it runs before the isolated stock engine so the surface is already reduced at the 62.5% ratio by the time the stock 80% fallback reaches it, leaving the fallback nothing further to trim, while Minimal mode has no preset compaction group and uses the root engine directly. Successful large tool results are compressed before they repeatedly re-enter the model context, PTC sub-dispatch log copies are compacted without altering the structured payloads, and every lossy replacement is archived and exposed through retrieve_spill for exact session-authorized recovery. An optional preset-cordis.yml fragment remains for deployments that deliberately want to replace a preset's isolated provider itself, but shipped presets should never be edited by hand.

The optimizer coexists with DSH's built-in spill subsystem (dsh-spill-local plus dsh-spill-policy at maxInlineTokens: 12500) by keeping its own durable archive as the retrieval authority and only mirroring replacements into that backend when it is present, so the same tool result is never compressed twice. Configuration is exposed through a single boolean compaction switch plus thresholdRatio (default 0.625) and retainRatio, with the constraint that retainRatio must stay below thresholdRatio or the upstream engine refuses to load. Since DSH 0.1.7-rc.1 the host checks @deepseek-ai/dsh* peer ranges at startup, and 0.2.0-rc.2 turned that check from fail-soft into fail-closed, which is why the declared range is >=0.1.5-rc.1 <0.3.0, covering 0.1.5-rc.3 through 0.2.x while explicitly refusing the unaudited 0.3.0. The test suite is run with pnpm install && pnpm run check && npm pack --dry-run and covers deterministic compression, persistent recovery, session lineage authorization, compaction defaults, the switch, DSH projection behavior, and real tool-runtime integration; the repository is released under the MIT license.