Tyan66666/billion-context-dsh
Model-driven context compression (Active Context Pruning) for DeepSeek Harness: the model decides when and what to compress.
Install
via npm
Resolves a published tarball rather than cloning the repository, and installs without any extra setup. Swap `web` for your profile name if you run another one.
via GitHub · npm package
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/web/pnpm-workspace.yaml.
What happened when we ran it
Installed cleanly when we ran it
Every command here is run in a throwaway container against a clean profile, and the result is whatever the harness recorded — not a guess from the source. Last run 6d ago.
Due diligence
What it is — Model-driven context compression plugin for DeepSeek Harness where the model decides when to compress.
Who it is for — Developers working on long conversations that exceed the model context limit in DeepSeek Harness should install this plugin. Users whose conversations remain within the model context limit may not need it.
Watch out — Sandbox test passed: registers successfully in a new profile. The plugin executes shell commands. It is a test version and not for production.
The verdict — I would install it because it supports persistent and recoverable.
Generated by grok-4.6, and a starting point rather than a verdict. Where it says a plugin installs or does not, that is from a real run in a clean profile — everything else is read off the repository. Trust the source over this.
billion-context-dsh is a DeepSeek Harness plugin that adds model-driven context compression through Active Context Pruning. Rather than applying a fixed truncation threshold or automatically generating a replacement summary, DeepSeek Harness gives the model a compress tool and lets it choose when to compress and which sequence range to cover. The model writes the summary, while an automatic policy only injects a nudge through the agent/pre-step hook.
billion-context-dsh maps compression onto DeepSeek Harness's append-only session surface. A compression becomes a persistent surfaceOp: { op: 'replace' }: the summary remains visible as a checkpoint, while the original messages stay in the session log. decompress restores hidden content, search_context searches summaries and source text, and acp_status reports context composition, block accounting and checkpoint mappings. The plugin also supports repeated compression for tiered distillation and rebuilds its block ledger from the log after restart.
This is intended for long-running agent tasks where selective, recoverable context matters. It is not a replacement for every setup: DeepSeek Harness's built-in compaction-basic must be disabled wherever this backend is enabled, because two providers of ctx.compaction conflict. The README marks the project and DeepSeek Harness as test versions. The plugin reaches the terminal/durable surface to read and update the append-only session record, so it should be reviewed before installation in an agent with sensitive session or terminal permissions.
DeepSeek Harness loads the bundle into a profile such as web; the bundle applies the compaction-acp composition entry at startup. Custom deployments can add the entry to ~/.dsh/profiles/web/cordis.patch.yml:
| Key | Behaviour |
|---|---|
modelContextLimit |
Optional model context limit. If omitted, the plugin detects the model window and falls back to 128000. |
prompts |
Optional prompt overrides for nudge text, context decomposition, growth and batch notices, tier distillation, range tables, the ACP system-prompt section and tool descriptions. |
Prompt templates accept documented named placeholders such as {pct}, {philosophy} and {surface}. Placeholder spelling is checked during construction and causes startup failure if invalid. An empty string or null has the documented slot-specific meaning.
DeepSeek Harness exposes four model tools:
compress replaces a startSeq–endSeq range with the model's summary. Boundaries are balanced to tool-call/result pairs. Compressing an existing summary creates tiered distillation.decompress restores a compressed block as read-only content. It accepts an acp_status block reference such as bN or a compaction id.search_context searches block summaries and hidden source text, then links matches back to their blocks.acp_status reports visible tool, text and summary proportions, the compression ledger, nudge decision and active checkpoint bN → seq mappings. It supports scope, view, tool, sort and limit filters.The bundle also adds the /acp command.
DeepSeek Harness derives each model request from the append-only session surface. The agent/pre-step hook injects a nudge based on context pressure; compactIfNeeded returns null, so the policy does not force an automatic summary. Block state is held in memory and rebuilt from the session log, with no sidecar file. Shadowed-token accounting uses DeepSeek Harness's ctx.tokenMeter.measure when available, with the mirrored fallback in src/host-tokens.ts.
Enable only one provider of ctx.compaction in each realm. For global use, disable compaction-basic before inserting compaction-acp; for a single agent preset, disable or remove dsh-compaction-basic in that realm. The README identifies version v0.2.11 as a test release and advises against production use. It does not document an operating-system restriction or environment variables.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub
Same category
volcengine
OpenViking memory and context bundle for DeepSeek Harness: pre-step auto-recall and profile injection, session capture, `viking://` URI guarding, and recall/write memory tools backed by an OpenViking server.
No one-line install — this plugin lives inside a larger repository and publishes no npm package.
vectorize-io
Hindsight, agent memory that learns: long-term project memory with auto recall and retain, knowledge pages, deep reflection, and per-repo memory banks.
No one-line install — this plugin lives inside a larger repository and publishes no npm package.
bowenliang123
Context insight panel: see what the model's context window is made of and how it evolves — composition vs. window size, per-request history, compression/injection events, and per-message token stats.
Installed cleanly when we ran it
vshulcz
Memory for coding agents — Claude Code, Codex, Cursor and 17 others. Indexes the sessions they already wrote to disk, including months from before you installed it, and recalls them in any of them. No LLM, no embeddings, one local Go binary.
Installed cleanly when we ran it
csyangwen
为 DeepSeek Harness 带来「跨会话长期记忆 + 后台自我进化」能力的纯插件实现：五轨记忆 · git 分支感知 · 回合内自我审查 · 技能自我进化与技能管理器 · 四轨待办 · COI 调度 · 会话广播 · 会话搜索 · 提示词管理器 · 临时信息便签——零核心修改、零运行时依赖，随装随用、卸载即净。
Installed cleanly when we ran it
tinqiao-oss
The Engramory curated-memory discipline as an installable plugin ([npm: dsh-engramory](https://www.npmjs.com/package/dsh-engramory)): a deterministic 200-line / 25 KB cap on the `MEMORY.md` index via `ctx.tools.guard()` — growth denied, a shrinking rewrite always passes — plus the protocol registered as a runtime skill. The store is plain markdown, one file per fact, shared with Claude Code, Codex, Kiro, and OpenClaw.
Installed cleanly when we ran it