dsh-token-optimizer
面向 DeepSeek Harness 的确定性、可恢复工具结果压缩与缓存感知压缩整理
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Snow-ea/dsh-token-optimizer#e4c4c825b7152b16610807107c272005a1ccebbbToken Optimizer 以 dsh-token-optimizer 发布，当前版本为 0.2.5。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Deterministic, recoverable tool-result compression and cache-aware compaction for DeepSeek Harness 0.2.1-rc.5.
run_code programs.retrieve_spill for exact, session-authorized recovery.BasicCompactionEngine adapter with a 62.5% default threshold while preserving DSH's cache-aware replay, branded Session API, and routed image accounting.Install the 0.2.5 bundle into a DSH 0.2.1-rc.5 Web profile:
dsh plugin --profile web add dsh-token-optimizer@0.2.5
Restart the existing dsh web process after installation. The bundle enables recoverable tool-result compression, spill retrieval, projection, the Web dashboard, and the 62.5% automatic compaction engine across all four shipped agent modes. No user preset copy is required.
The root engine observes each agent's pressure and overflow lifecycle. In Standard, PTC, and Creator modes it runs before the isolated stock engine, so the surface is reduced at 62.5% and the stock 80% fallback has no second reduction to perform. Minimal mode has no preset compaction group and uses the root engine directly. The optional preset-cordis.yml fragment remains for deployments that deliberately want to replace a preset's isolated provider itself; do not edit shipped presets. DSH 0.2.1-rc.5 ships the four presets as extra patch layers inside its Web bundle, so a user-copied preset from an older DSH should be revalidated against the current composition before reuse.
DSH 0.2.1-rc.5 ships its own spill subsystem (dsh-spill-local plus dsh-spill-policy at maxInlineTokens: 12500). The optimizer keeps its durable archive as the retrieval authority and mirrors each replacement into that backend when it is present, so the two coexist without compressing the same result twice.
Since DSH 0.1.7-rc.1 the host checks a plugin's @deepseek-ai/dsh* peer ranges at startup. 0.2.0-rc.2 turned that check from fail-soft into fail-closed: a bundle whose ranges do not match the runtime is refused outright rather than disabled in place. This plugin declares >=0.1.5-rc.1 <0.3.0, so one release covers 0.1.5-rc.3 through 0.2.x while explicitly refusing the unaudited 0.3.0. The previous ^0.1.5-rc.1 range could not reach 0.2.0 — a caret on a 0.x version only authorises patch movement, which is why 0.2.2 is skipped on every 0.2.x runtime. 0.2.1-rc.2 adds an optional engines.dsh range that would take precedence over the peers when present; this plugin deliberately keeps the peer form, so exactly one requirement surface stays authoritative. See the 0.2.0-rc.2 peer-gate section and the 0.2.1-rc.2 gate section for the measured matrices.
This plugin's early compaction is controlled by one boolean, compaction, which defaults to on:
# Turn the plugin's 62.5% early compaction off
- id: dsh-token-optimizer
config:
compaction: false
With it off, no root ctx.compaction provider and no pressure/overflow listener is registered, while tool-result compression, retrieve_spill, the session projection, and the dashboard all stay active. Three things to know:
config instead of deep-merging it. Every field left unwritten falls back to the value the shipped bundle sets, so a switch-only override is safe and a regression test locks that equivalence; write the other fields out when you also want to change them.service "compaction" has been registered at <BasicCompactionEngine> rather than degrading silently.The threshold itself defaults to 0.625 and can be changed:
- id: dsh-token-optimizer
config:
thresholdRatio: 0.5
retainRatio: 0.16
retainRatio must stay below thresholdRatio, or the upstream engine refuses to load.
pnpm install
pnpm run check
npm pack --dry-run
The current test suite covers deterministic compression, persistent recovery, session lineage authorization, compaction defaults, the compaction switch, DSH projection behavior, and real tool-runtime integration.
| dsh-token-optimizer | DeepSeek Harness | Status |
|---|---|---|
0.1.9 | 0.1.1-rc.2 | Superseded |
0.2.0 | 0.1.2-rc.1 | Superseded |
0.2.1 | 0.1.5-rc.1 | Superseded |
0.2.2 | 0.1.5-rc.1 … 0.1.7-rc.2 | Superseded |
0.2.3 | 0.1.5-rc.1 … 0.2.x | Superseded |
0.2.4 | 0.1.5-rc.1 … 0.2.x | Superseded |
0.2.5 | 0.1.5-rc.1 … 0.2.x | Current release |
0.2.0 removed the retired dsh-client-runtime dependency, uses the split Conversation/Renderer/Session client graph, folds projections through the on-demand Session log API, and handles PTC durable dispatch logs. See the 0.1.2-rc.1 compatibility research.
0.2.1 adapts to the SpillSource discriminated union that dsh-spill gained — tool producers must now declare kind: 'tool' — and raises every declared DSH range to 0.1.5-rc.1. No other plugin surface changed between the two releases. See the 0.1.5-rc.1 compatibility audit.
0.2.2 adapts to 0.1.7-rc.2: it follows the removal of the tool-result wrapper block, adopts the awaited agent/created contract, mirrors the new upstream headroomTokens field behind two compile-time drift assertions, and declares plugin metadata for Plugin Manager. The compaction switch can also now turn the plugin's early compaction off. It keeps supporting 0.1.5-rc.3, because the host's peer check accepts ^0.1.5-rc.1 for the whole 0.1.x line. See the 0.1.7-rc.2 compatibility audit.
0.2.4 adapts to 0.2.1-rc.2. No plugin-facing declaration changed in a way this plugin can observe, and no source line changed: thirteen of the sixteen audited packages differ at the .d.ts level, but every difference is additive, internal, or a doc-comment rewrite, the three core contract packages (dsh-compaction-basic, dsh-session-projection, dsh-spill) are byte-identical, and cordis 4.0.5 / schemastery 3.18.5 emit identical declarations. The release re-pins the tested closure and certifies compress → archive → retrieve on a real 0.2.1-rc.2 headless session. See the 0.2.1-rc.2 compatibility audit.
0.2.5 adapts to 0.2.1-rc.5. Fifteen of the sixteen audited packages are byte-identical at the .d.ts level; the only change is dsh-client-ui-conversation, and it is purely additive (View tabs and pins, the header More menu, draft handling, MessageSubmissionState.newSession/agentPreset), so no source line changed and all 24 tests pass on the real 0.2.1-rc.5 closure. The launcher's gate implementation is byte-identical to 0.2.1-rc.2, the existing >=0.1.5-rc.1 <0.3.0 peer window still loads through 0.2.1-rc.5 and still refuses 0.3.0, and a real headless 0.2.1-rc.5 session was certified end to end: a 34,532-character tool result compressed to a mode=large spill with a matching SHA-256 artifact, retrieve_spill recovered it in a later turn, and the tokenOptimizer projection folded the saving. See the 0.2.1-rc.5 compatibility audit.
0.2.3 adapts to 0.2.0-rc.2. No plugin-facing declaration changed in a way this plugin can observe — fourteen of the sixteen audited packages are byte-identical at the .d.ts level, and the two that moved (dsh-session's repair internals, dsh-client-ui-conversation's submission analytics) are outside its surface, so the 0.2.2 sources compile and all 24 tests pass unchanged. The release exists because the launcher's peer gate became fail-closed: the declared range moves to >=0.1.5-rc.1 <0.3.0, which is what keeps the bundle loading on 0.2.x while still covering 0.1.x. See the 0.2.0-rc.2 compatibility audit.
The DSH packages remain optional peer dependencies so the profile runtime supplies the matching services.