v2.14.0: sidebar pagination + named LLM provider profiles

Pagination for the sidebar so large databases stay usable in the TUI, plus named LLM provider profiles so switching between OpenAI / Ollama / a remote inference rig is one keystroke instead of a config edit.

Highlights

Sidebar cursor pagination (#15)

The sidebar previously fetched a fixed 200 rows and stopped, so on large DBs older entries were unreachable from the TUI. Now each tab fetches one page (DEFAULT_PAGE_SIZE = 200) at a time using cursor-based keyset pagination.

• Header shows conversations · N loaded · more (Ctrl+L) when more pages remain
• New binding Ctrl+L (and /load-more analogue via the action) appends the next page in place
• Switching mode/search resets the cursor and reloads from page 1
• Search uses cursor mode too

Also fixed a latent bug in db.search_conversations: when FTS5 found no matches AND cursor mode was requested, it was returning a bare list instead of a PaginatedResult. Wiring the sidebar through cursor mode exposed it.

Named LLM provider profiles (#16)

Define multiple endpoints in ~/.ctk/config.json and switch between them at startup or live in the TUI:

{
  "providers": {
    "default": "muse",
    "openai":  {"base_url": "https://api.openai.com/v1", "default_model": "gpt-5"},
    "muse":    {"base_url": "http://muse.lan:8000/v1",   "default_model": "qwen3-omni"},
    "ollama":  {"base_url": "http://localhost:11434/v1", "default_model": "llama3.1:70b"}
  }
}

• TUI: /provider lists profiles, /provider <name> rebuilds the active provider live; /model now shows profile + base_url too
• CLI: ctk --provider <name> flag at startup
• API keys: reads <PROFILE>_API_KEY first (so MUSE_API_KEY works), falls back to OPENAI_API_KEY, then config file (with warning)
• Backward compat: legacy configs with only providers.openai keep working

Pilot test harness for modals (#14)

Three releases (2.13.1, 2.13.2, 2.13.3) shipped this month because nothing actually opened the modals headlessly before tagging. New harness in tests/unit/test_textual_tui.py uses Textual's Pilot to push every modal, force a render via pause(), and dismiss via Esc / y / n. Would have caught the _render and _bindings shadowing crashes at unit-test time.

Documentation refresh (#17)

• README rewritten around the 2.14.x surface (citation bumped, dropped references to ~17 removed subcommands and the Views system)
• CLAUDE.md gained the Textual _render / _bindings shadowing trap in Gotchas, refreshed test counts, documented the Pilot harness pattern

Stats

• 1700 unit tests pass
• ~1300 net lines added across the four PRs

Install

pip install --upgrade conversation-tk

Full Changelog

v2.13.3...v2.14.0