flowctx-dsh

When an AI coding agent runs a long session, the context window fills up fast with tool output and conversation history. DeepSeek Harness ships a three-layer compaction mechanism, but it has real gaps: middle content is discarded permanently, generic summaries blur away failure paths, function signatures, and file paths into vague prose, and synchronous summarization injects a full LLM request into the critical path on every trigger. flowctx-dsh is an additive extension over dsh-compaction-basic that closes those gaps with a three-tier memory model.

The core idea is to tier context by recency: the current task is kept verbatim so freshly read code and error output never vanish; recent history is reversibly projected, compressing bulky logs and file snippets into hash-referenced markers while the originals live in a local CompressionStore and can be retrieved byte-exact on demand; older history is folded in the background into a six-section engineer handoff note that preserves failed approaches, verbatim identifiers, and file artifacts in dedicated sections. Layered summaries run fire-and-forget so they never block the current step. Every extension is individually toggleable; with all toggles off the plugin behaves exactly like dsh-compaction-basic with a different summary style. The KV-cache prefix stays byte-stable, keeping cache hit rates intact.

It is designed for engineers and teams building AI coding agents on DSH, running long multi-file multi-stage coding tasks, dealing with heavy tool output that keeps inflating prompt tokens, and wanting a local-first, auditable, and fully recoverable context management layer.