dsh-project-context

Long dsh development sessions inevitably hit context limits; once compression or disconnect strikes, the design decisions, pitfalls, and critical paths accumulated over dozens of log files become hard to reuse. dsh-project-context archives every session event stream into the project directory, distills them into structured memory and context on a schedule, and automatically hands off a summary to a fresh session when context approaches its ceiling, so project knowledge survives beyond any single conversation.

The pipeline runs in four automated stages: archiving writes each event to JSONL and renders human-readable Markdown with an index; consolidation folds scattered dialogue into MEMORY.md and CONTEXT.md during idle periods; skill distillation produces reusable SKILL.md files once sufficient evidence has accumulated; and handoff creates a new session carrying a summary plus recent raw messages when context usage crosses the threshold and no background subagent is running. All four stages share a single config and lifecycle manager, with crash recovery, cross-process locking, and historical session backfill all handled.

Built for engineers who run multi-day, many-round development sessions in dsh and want project-level knowledge to live alongside the code, versioned and continuously accumulated across sessions. No external services required—install, tune a few parameters, and it runs.