dsh-infinite-context

Long sessions are the daily reality for many DSH users, but a finite context window means early discussion details, project conventions, and debug context vanish the moment tokens run out. Local models often have only a few K of window, and a single extended-thinking expansion can blow past the entire budget. DSH itself only compacts reactively after the API returns CONTEXT_WINDOW_EXCEEDED, which is too late. dsh-infinite-context elevates memory management to a first-class plugin-layer capability, letting sessions proactively and progressively top up before the ceiling is hit, turning long conversations into a genuinely usable workflow.

The core mechanism is a three-tier memory pyramid: recent turns are kept verbatim, middle-round dialogue is LLM-summarized, and far history is consolidated into refined checkpoints, all persisted in SQLite so memories survive restarts. The compression trigger line is derived dynamically from the routed model REAL window. Local models are probed via Ollama, llama-server, or OpenAI-compatible endpoints to read the server actual runtime context, and the effective window is the minimum of declared and probed values, guaranteeing compaction fires before the true ceiling, not after. The mid-thinking guard is especially critical: the plugin wraps the agent LLM stream, measures input plus output block by block, and when the dynamic line is approached it injects an overflow signal, triggers durable compaction, then retries with headroom. The model gets stopped, compacted, and re-thinks, instead of wasting an entire doomed generation. Semantic retrieval splices the most relevant top-K memories into context each turn, three-layer dedup prevents duplicate ingestion, and a four-class structured index makes memories auditable and forgettable.

Two user profiles benefit most. First, developers who regularly work through large codebases or lengthy documents where conversations easily run dozens of rounds and losing early context means re-explaining everything from scratch. Second, heavy local-model users whose windows are small and whose thinking phases are long, where real-window awareness plus early compression is the critical combo. If you want DSH long sessions to have memory like a human rather than a goldfish, this plugin is worth a look.