Problem

Time-to-first-prompt (TTFP) — from deepseek invocation to interactive cursor — is a first-impression metric. Slow startup hurts adoption. Today TTFP isn't measured.

Proposed solution

1. Measure. Add a --print-startup-trace debug flag that emits timestamps for every initialization step:

   • process exec
   • clap parse
   • config load
   • keyring probe
   • MCP server connect (if enabled)
   • skill discovery
   • terminal init / alt-screen enter
   • first render

2. Profile a cold launch on each platform. Linux x64, macOS arm64, Windows x64. Different bottlenecks.

3. Optimize the obvious:

   • Lazy keyring probe — don't talk to the OS keyring unless the active provider has no env-var key. Today we probe all 4 providers regardless.
   • Lazy MCP connect — defer MCP server connections to first tool-use, not startup. Saves the connect timeout × N when MCP is configured but unused this session.
   • Lazy skill discovery — don't fs::read_dir ~/.deepseek/skills/ unless /skill or /skills is invoked.
   • Strip release binary — verify [profile.release] strip = true in Cargo.toml. (Possibly already done.)

Acceptance criteria

• --print-startup-trace flag.
• Baseline measurement on a clean install (Linux + macOS at minimum) committed to docs/perf.md.
• At least one optimization landed with before/after numbers.
• TTFP target on macOS arm64: < 200ms cold, < 50ms warm.

Related

• crates/cli/src/lib.rs and crates/tui/src/main.rs — startup paths.
• crates/tui/src/mcp.rs — MCP connect timing.
• crates/tui/src/skills.rs — skill discovery.