fix: restore live thinking streaming on 2.1.225/226; add answer streaming + verbose Read content

[asgeirtj]

TL;DR

Live thinking streaming is broken in the current 2.1.225/2.1.226 releases — the thinking-streaming reducer/event sub-fix silently stopped applying, so thinking blocks only render when they land instead of streaming as deltas arrive. This PR fixes that, adds a new answer-streaming patch, hardens the thinking request, and restores full verbose Read output.

The regression (fafd40f)

2.1.225 changed the inner stream handler's options destructure by one character:

// before
function X(e,t,r){let{onSetStreamMode:n,...}=t;
// 2.1.225/226
function X(e,t,r){let{onSetStreamMode:n,...}=t,d=t.authoringProgressSurface===!0;

The comma tail made missingStreamingThinkingHandlerPattern skip the entire reducer branch, while every render-side sub-fix still hit — so thinking-streaming reported 5 patched and looked healthy. The playbook's own warning ("patch count still looks nonzero but live thinking is broken") was the exact failure mode. message_stop also gained a comma-if wrapper (if(dt?.finalize(),...,d)bnt?.resetAuthoringProgress();return) that needed a third needle.

Fix: the destructure matcher accepts either tail, the injection needle drops the trailing separator, and the new message_stop shape has its own needle. On 2.1.226 the module now reports 13 patched and the __cc_prevStreamingThinking markers land inside the stream handler itself. Verified at runtime (tmux + timestamped pane captures): thinking text on screen 11s into a 31s turn, growing continuously.

Debugging rule now in the playbook: verify the __cc_ markers inside the handler — module counts alone prove nothing.

New: answer-streaming (fa24cc7 + fd2d135)

Upstream ships a complete live answer-text pipeline (text_delta → throttled store → bottom-of-transcript preview), but gates all of it on sk = !prefersReducedMotion && !isWindowsTerminal(). With prefersReducedMotion: true, answers only appear after the message lands.

The patch enables just the two consumers answer streaming needs — the delta-apply callback and the preview-visible flag — leaving sk itself (and the deferMessages render-coalescing it controls) untouched. An earlier iteration flipped the whole sk definition; don't do that — it disables the deferred-messages path and degrades live transcript rendering under load. The Windows Terminal guard is preserved.

Hardening: thinking display request (fd2d135)

2.1.226 also split the request builder's env-flag read and put display behind B1()&&Xkn(model) gates:

ii=flag(env.CLAUDE_CODE_DISABLE_THINKING),di=cfg.type!=="disabled"&&!ii,vr=di&&B1()&&Xkn(u)?cfg.display:void 0,Gr=void 0;

The old display-defaulting sub-fix no longer matched. A new matcher rewrites this shape to vr=di?cfg.display??"summarized":void 0. Without the display field the API streams no thinking text at all, so this is request-side insurance against those gates flipping.

New: read-verbose-content (7005bf5)

Completes the tool-call-verbose intent on current builds: upstream's verbose Read result renderer collapsed to "Read N lines" even with verbose:!0 forced. The patch replaces the Read tool's renderToolResultMessage getter with a wrapper that renders the full line-numbered content (via the upstream {content,startLine} formatter, discovered by shape) when verbose, falling back to the original renderer otherwise. All helpers are discovered by shape per the matcher design rules.

Validation

• node patch-claude-display.ts --file content.js --dry-run on pristine 2.1.226: all modules at expected counts (thinking-streaming 13/13, answer-streaming 2/2, read-verbose-content 1/1)
• patched a real darwin-arm64 native binary, re-signed, --version shows (patched)
• runtime-verified in tmux with timestamped pane captures: live thinking mid-phase, live answer text, verbose Read content
• playbook updated for every new shape, plus the marker-verification rule