Recall what worked before
SQLite FTS5 searches executed commands using the current prefix and directory context. The measured 2,000-entry path is below 1 ms on the release test machine.
Engineering case study
A terminal completion system does not ask a model to hold a conversation. It asks for one short, context-bound suffix under a tight deadline, then lets the user accept or ignore it.
~ shellclaw status
shellclaw: running
Ctrl+Space accepts the gray suffix
ShellClaw deliberately narrows generation before model size enters the discussion.
Retrieval and interaction design reduce the amount of generative work needed on every keystroke.
SQLite FTS5 searches executed commands using the current prefix and directory context. The measured 2,000-entry path is below 1 ms on the release test machine.
When recall has no valid match, Qwen2.5-Coder 0.5B generates a short continuation through llama.cpp with Metal acceleration.
Users can ignore a weak suggestion by continuing to type. Stale work is cancelled and late responses are discarded.
The daemon can retrieve memory, run inference, and send concurrent responses without blocking Tokio I/O tasks or the Zsh editing widget.
This case study describes a working architecture, not a universal quality claim.
ShellClaw does not yet publish a full test set with cold start, hot latency, p50, p95, memory, and quality measurements. Those results should not be inferred from the memory fast-path number.
Install on an Apple Silicon Mac, inspect the source, and report command patterns that need work.