Docs update mcp models - #331

…te-settings

Split remote settings into OpenWork and OpenCode sections

…mouse overlay (anomalyco#331)

PR #24 fixed #N rendering by rewriting refs to markdown links fed through
<markdown>, but that killed the hover-card preview: OpenTUI inline link
chunks are TextNodes that can't hold mouse listeners (only block-level
Box/Text are on the mouse hit-grid).

Restore hover (Model A) without touching PR #24's rendering: wrap the
existing TextPart <box> with onMouseMove/onMouseOut, translate absolute
mouse coords -> local row/col via the box's screenX/Y + width, then map to
the nearest ticket ref using a source-text scan. Drives the existing
(intact-but-untriggered) TicketHoverCard context. No change to wrap layout,
so the anomalyco#250 fix is preserved.

New: ticket-ref-scan.ts (scanTicketRefs + resolveRefAt + nearestRefId).

…overlay (anomalyco#331)

Nik reported two bugs on the prior approach: (1) the markdown rewrite
[#N](url) leaked the literal URL into the visible text, (2) hover was 'all
over the place'.

Root cause confirmed by build-test (test harness, no tree-sitter — same as
the real app's <markdown>): a [#N](url) link renders its URL literally; NO
markup form (link/bold/code) conceals it. Only PLAIN #N is leak-free. The
hover drift came from estimating wrap geometry against the rewritten content,
whose [#N](url) strings were far longer than the visible text.

Fix (Option 1 — keep PR #24's wrap-correct <markdown> flow, drop the rewrite):
- Render source text directly as plain #N. No URL leak, wrap unchanged.
- Line-based hover resolver: cursor row -> source line -> ref on that line
  (single = exact; multi = nearest by column). Now accurate because the
  scanned/estimated text IS the visible text.
- Click-to-open via overlay onMouseUp (replaces the lost OSC-8 click).
- HIVEMIND_UI_BASE default -> https://hivemind.grunt.si.

Tests: 13 resolver unit tests + full TUI suite (146 pass) green; typecheck clean.

…cursor leaves ref (anomalyco#331)

Reported: hover card never closed (snapped to nearest ref) and moving far
just switched to a neighbor ticket, even ~30cm from the actual #N.

Cause: resolveRefAt unconditionally returned the nearest ref (incl. a global
fallback for ref-free lines), so onMouseMove never got null → never closed.

Fix (Option b — strict tolerance everywhere):
- Vertical: ref-free line → null (no cross-line snapping). Wrapped ref lines
  stay hoverable across their visual rows.
- Horizontal: gate the cursor's VISIBLE column against each on-line ref's
  visible span ±HOVER_COL_TOLERANCE (2). Beyond tolerance → null → card closes.
- Accurate columns: subtract the concealed leading-marker length (- , ## , > )
  so a ref's visible column on list items/headings matches the rendered text.

17 resolver unit tests (incl. far-on-line→null, ref-free→null, between-refs→
null, list-prefix shift, wrapped 2nd row) + TUI suite (150 pass) + typecheck green.

…ate native hover+click (anomalyco#331)

The estimate-based mouse overlay (cursor→char column math) was fundamentally
fragile: hover drifted and the proximity gate made it close/misfire. OpenTUI
exposes no public coord→char API, so any JS-side estimate drifts.

Switch to the accurate path: render each #N as an inline <a href> LinkRenderable.
Build-verified: renders a clean 'anomalyco#331' label with the URL carried as OSC-8
hyperlink metadata (NO url leak, unlike markdown [#N](url)). The terminal owns
the link geometry, so hover-highlight + ⌘-click are pixel-accurate with zero
estimation.

To honor OpenTUI's inline-flow-vs-markdown tradeoff with minimal loss, split the
message into paragraph blocks: ref-free paragraphs (and all headings/lists/code)
keep full <markdown>; only a paragraph that mentions a #N renders as inline
<text> of <span> prose + <TicketRef> (<a>) refs (correct char-wrap per anomalyco#250).
Fenced code is never split or linkified.

- ticket-ref.tsx: clean inline <a href> (theme link color, OSC-8 url).
- ticket-ref-scan.ts: splitTextIntoBlocks (paragraph + fence-aware) + splitOnTicketRefs.
- TextPart: per-block markdown-or-inline rendering; removed the overlay + scan resolver.
- Dead-but-inert: TicketHoverCard/hivemind hover context no longer triggered.

12 scan unit tests + TUI suite (145 pass) + typecheck green.

…gle <markdown> (no flicker) (anomalyco#331)

Definitive fix after exhausting the alternatives (all build-tested):
- single <markdown> is the ONLY flicker-free renderer; per-block split flickered
  on stream + lost formatting.
- markdown [#N](url) links always leak the URL in this app (no tree-sitter);
  inline <a> can't live inside <markdown>.
- JS-side column estimation drifted (hover 'all over the place').

So: keep ONE <markdown> rendering plain  (clean, full formatting, wrap-correct,
flicker-free), and resolve hover/click by reading the renderer's actual FRAMEBUFFER —
renderer.currentRenderBuffer.buffers.char is a public Uint32Array indexed y*width+x.
ticketIdAtCell() inspects the characters painted under the cursor and returns the
#N id (with the same boundary rules as the text regex: rejects foo#12, ##12, #fff).
Pixel-accurate (uses rendered cells, not a reconstruction), null off-ref so the
hover card closes. Drives the existing TicketHoverCard; onMouseUp opens hivemind-ui.

10 framebuffer hit-test unit tests + TUI suite (143 pass) + typecheck green.

…fixes (anomalyco#331)

…-process (anomalyco#331)

Refs render as plain #N inside <markdown> (a markdown link leaks its URL), so
they were base text color. Recolor the painted #N cells in a renderer
post-process (renderer.addPostProcessFn): single buffer scan re-drawing matched
#N runs in theme.markdownLink + underline. Same boundary rules as the hit-test
(skips foo#12 / ##12 / #fff). Theme proxy is read each frame so it tracks theme
switches.

4 recolor unit tests (+ existing 10 hit-test) + TUI suite (147 pass) + typecheck green.

….md (anomalyco#331)

Add a 'TUI enhancements' section covering the clickable/hoverable/colored #N
ticket refs, the framebuffer-hit-test + post-process-recolor implementation, and
why refs aren't markdown links (URL leak / inline-<a>-in-markdown limitation).
Refresh the 'two buckets' framing + rebase-chain example to include the TUI patch
and the loop-primitive patch.

…co#331)

…#331)