Problem. crates/tui/src/tools/registry.rs registers 50+ tools alphabetically (sorted for prefix-cache stability at :196-200). The tool catalog alone is ~8–12K tokens on every request that passes tools. defer_loading already exists for MCP tools (:936) but not for native ones — so the same model sees automation_list, revert_turn, and image_ocr in its context on every turn even though they're called once a session at most.

Shape. Pick the core 14:

read_file, write_file, edit_file, exec_shell, grep_files, file_search,
git_status, git_diff, run_tests, apply_patch, web_search, fetch_url,
list_dir, agent_open

Everything else (diagnostics, pandoc_convert, git_log, git_show, git_blame, recall_archive, revert_turn, automation_*, task_*, checklist_*, finance, image_ocr, notify, validate_data, etc.) gets defer_loading: true and is materialized on the first ToolSearch hit. Config knob (tools.always_load = ["foo", "bar"]) lets users opt back in.

Acceptance.

• Tool catalog token count measured before/after with a one-shot script under scripts/; target ≥ 40% reduction.
• Default behavior unchanged for any tool the model already calls in a typical session.
• Per-tool opt-in via config.

Existing issues this resolves on PR merge

Partially addresses:

• Partially addresses Feature: DeepSeek cache-aware prompt diagnostics and wire payload optimization #1253
• Partially addresses 输入缓存命中率太低了 #1177
• Partially addresses Cache hit problem #1747
• Partially addresses There still seems to be some problems with cache hits缓存命中方面似乎还是有些问题 #1120

When opening a PR for this issue, the PR body must include the Closes #N lines for every issue above that this PR fully resolves, so they auto-close on merge.

Slate item 3/18. Tracking doc: .private/issue-slate.md (maintainer-only).