billion-context-dsh
Active Context Pruning (ACP) for the DeepSeek Harness — model-driven context management as a CompactionEngine backend.
72 results
Active Context Pruning (ACP) for the DeepSeek Harness — model-driven context management as a CompactionEngine backend.
AgentFrame compaction backend for DeepSeek Harness: semantic + physical KV compression (28.4x) replacing LLM summarization
Adaptive task and risk routing, loop budgets, and progressive tool-result pruning for DeepSeek Harness
High-fidelity, faithful, bilingual, recursive compaction backend for DeepSeek Harness — a drop-in upgrade over dsh-compaction-basic.
Model-scoped DSH plugin: GPT Responses lifecycle, Native V2 compaction and Hosted/Alpha Search; Grok native Web/X Search
Idle-triggered context compaction for DeepSeek Harness: compact a session once it has gone quiet past a token floor
Keeps a local model's context at a size the GPU handles well: measured prefill speed, hard ceiling, early compaction for DeepSeek Harness
Session health for DeepSeek Harness: real token-meter data, configurable continue-vs-new decision model. Reactive header badge (compaction-aware), /compass command, context_compass tool.
Workspace History: journals every compaction summary to <workspace>/.dsh/history/<unix-ts>.<short-session-id>.md, and adds a History subtab to the Workspace Overview tab for reading the journal. Per-session toggle lives in the granular Settings tab — requires dsh-granular-settings (hard dependency).
Chat branch tree plugin for DeepSeek Harness: mind-map of conversation branches, group folders, archiving and compaction preview.
Durable DeepSeek Harness bundle: the local-model performance tuning carried as a re-installable plugin layer (compaction, tool-result pruning, time context, cloud title routing, text-toolcall guard) — survives dsh updates
自动上下文策展:tool 输出瘦身 + 无效日志清理 + 工作流/记忆钉扎保护 + 先归档后压缩
Composer control for DeepSeek Harness auto-compaction: set a percent (30–80) or absolute token threshold that persists in settings.yaml.
DeepSeek Harness plugin: aggressive token optimization without affecting task quality. Trims tool results, manages context window pressure, optional LLM-summary compaction, stable prompt-cache layout.
Persistent task state and context-aware compaction for DeepSeek Harness coding sessions.
DSH plugin: Codex-style context management — token-budget notice, model-facing new_context, and handoff reset instead of summary compaction.
Keep native vision, drop historical images before they hit relay providers. Global DeepSeek Harness bundle for every agent.
Summary-free context window rollover, model-authored notes, and scoped raw history lookup plugin for DeepSeek Harness (DSH)
Codex-style lossless-retrieval context manager for the DeepSeek Harness: token-budgeted sliding window, persistent working notes, and cold-history search over the append-only session log
Official-basic Codex compaction with an optional account-owned native summarization seam and legacy structured reader; no core patches or duplicate login.
Context Compaction plugin for DeepSeek Harness: tool-result pruning, LLM or offline-rule summarization with a user-facing compression report, dead-loop guard, and /restore rollback.
Keep DSH tool results cache-prefix safe: truncate a result before it first enters the session surface, and never rewrite a tool result a provider request has already delivered.
CWL — Context Window Lifecycle for DeepSeek Harness: structured context eviction (arXiv:2606.11213). Graduated, deterministic, zero-LLM eviction of exploration/action episodes when context pressure exceeds budget — no summarization lossiness, no hallucination, user messages never evicted.
DSH compatibility guard: repairs compaction limits, fail-safe sandbox escalation, and missing tool descriptions for third-party models; capability lookups are cached and the deprecated prompt-injection path is off by default.