[FEATURE] Expose manual /compact command for long-running sessions with local LLMs

[thuliumsystems]

Strix already includes a context compression mechanism (memory_compressor), but there is currently no user-facing way to trigger it on demand. For users running local models (llama.cpp, Ollama, vLLM, OpenAI-compatible servers), context is often the primary bottleneck rather than inference speed.

A manual /compact command would allow users to compress the current conversation and agent state before the context window becomes saturated, reducing token usage and enabling much longer assessments on consumer hardware.

Ideally, this command would reuse the existing compression pipeline instead of introducing a separate implementation.

[scythe-327]

Feature Request: Manual /compact command for on-demand context compression

Context & Rationale

As an interactive, multi-agent AI pentesting tool, Strix often maintains long, complex conversation states with its agents during heavy reconnaissance or exploitation phases. When running local models (via Ollama, llama.cpp, vLLM, or custom LLM_API_BASE endpoints), context window limits are typically a much tighter bottleneck than raw inference speed.

While Strix already implements an internal memory_compressor mechanism to handle context bloat in the background, users currently lack a way to trigger it proactively. Adding a manual /compact (or /compress) slash command would allow users running local hardware to manually trigger a compression cycle before the context window becomes completely saturated—enabling much deeper, extended security assessments without OOM crashes or context truncation.

Proposed Solution

Expose a user-facing command in the interactive CLI loop:

• Command: /compact (or /compress)
• Under the Hood: Reuse the existing memory_compressor pipeline rather than introducing a separate implementation, ensuring that the current conversation loop and agent memory states are safely condensed into an optimized summary state.

Benefits

• Local-First Optimization: Drastically improves the UX for users utilizing local open-source LLMs where VRAM/RAM boundaries are strict.
• Proactive Context Management: Allows users to cleanly pivot or manually trim long terminal/proxy request histories before hitting model token limits.
• Maintainable: Reuses the existing background compression pipeline with zero architecture duplication.