Problem
The streaming feed loop in crates/core/src/controller/live_preview.rs was reworked so the dedicated PreviewTyper thread owns the on-screen state and computes its own common-prefix delta. As a result, the older "direct delta" plumbing that apply_streaming_transcript and PreviewTyper::finish still expose is now dead surface area.
1. The publish closure ignores old and never fails. run_streaming_loop (live_preview.rs:265) builds a closure whose only job is to forward the latest text to the typer (live_preview.rs:295-298):
let mut publish = |_old: &str, new: &str| -> Result<(), String> {
typer.publish(new.to_string());
Ok(())
};It discards _old and always returns Ok(()). Real delta errors are now handled inside the typer thread (PreviewTyper::run → clipboard::apply_preview_delta, live_preview.rs:516-522), so the Result here is a permanently-Ok path.
2. apply_streaming_transcript carries an old/new/Result contract for a dead failure path. Its signature still requires the old text and an error branch (live_preview.rs:390-401):
fn apply_streaming_transcript<F>(last_applied: &mut String, text: &str, apply_delta: &mut F)
where
F: FnMut(&str, &str) -> Result<(), String>,
{
let text = text.trim();
if text != last_applied {
match apply_delta(last_applied, text) {
Ok(()) => *last_applied = text.to_string(),
Err(err) => log::error!("Live preview streaming delta failed: {}", err),
}
}
}The only production caller passes the always-Ok publish closure (live_preview.rs:300-306 and live_preview.rs:330-334), so the Err arm is unreachable in production.
3. PreviewTyper::finish takes a final_text handoff that production always calls as finish(None). Signature at live_preview.rs:472:
fn finish(mut self, final_text: Option<&str>) -> String {Both production call sites pass None: the streaming path at live_preview.rs:347 and the batch path at live_preview.rs:232. The final text is already published via apply_streaming_transcript before finish (live_preview.rs:330-334), so the Some(text) branch inside finish (live_preview.rs:475-477) is never exercised by the app.
Why it matters
- Maintainability: three coupled signatures (
publish closure, apply_streaming_transcript, PreviewTyper::finish) advertise an old-text + Result + final-text-handoff contract that production no longer uses. A reader must trace all three to confirm the failure/handoff paths are dead, which obscures the real design (typer owns state + diff).
- Risk surface, not hot-path cost: this runs once per utterance, not per audio chunk, so there's no perf angle. The cost is the dead
Err/Some branches that invite incorrect assumptions during future edits to the finalization flow.
Proposed change
Collapse to a latest-text-only API, in a single focused PR:
- Change
apply_streaming_transcript to take apply_delta: &mut F where F: FnMut(&str) (drop the old param and the Result); apply unconditionally on change and update last_applied.
- Simplify the
publish closure in run_streaming_loop to |new: &str| typer.publish(new.to_string()).
- Make
PreviewTyper::finish no-arg (fn finish(mut self) -> String) and drop the Some(text) pre-seed branch — unless review finds a real caller that needs the final_text handoff. Update the streaming (live_preview.rs:347) and batch (live_preview.rs:232) call sites.
Tests that must change (they currently encode the old contract and act as the guard):
streaming_transcript_helper_records_deltas (live_preview.rs:864) and streaming_transcript_helper_skips_duplicate_text (live_preview.rs:894) use an apply_delta of |old, new| { ...; Ok(()) } and assert on (old, new) tuples — rewrite for the FnMut(&str) signature while preserving the dedup and delta-recording assertions.batch_append_accumulates_text_with_spaces (live_preview.rs:724) already uses FnMut(&str) for the batch path — leave as-is, it's a model for the new streaming-helper test shape.
Tests that must stay green unchanged (the correctness guard for the is_final-after-typer fix):
pending_cap_drop_forces_unfinalized_streaming_result (live_preview.rs:629)confirm_streaming_finalization_returns_true_when_text_landed (live_preview.rs:919)confirm_streaming_finalization_falls_back_when_text_not_applied (live_preview.rs:928)streaming_preview_reports_unfinalized_when_reset_fails (live_preview.rs:847)
Recommendation / triage
DEFER / do carefully. This area was deliberately reworked for the recent is_final-after-typer correctness fix (commit 62a9d3f, "fix(live-preview): emit is_final only after typer confirms text landed"). Any collapse must keep the confirm_streaming_finalization behavior intact — the final text is published, then the typer must prove it landed on screen before streaming_finalized stays true. Do this as a focused PR with the live_preview streaming tests as the guard, not a drive-by. If review surfaces any real future need for the final_text handoff in finish, keep that param and only collapse the closure + apply_streaming_transcript.
Acceptance criteria
Traceability
- clawpatch finding ID:
fnd_sig-feat-service-39545238a6-ce5f_91fb601faf
- Revalidate:
clawpatch revalidate --finding fnd_sig-feat-service-39545238a6-ce5f_91fb601faf --reasoning-effort high
- Coupled context: builds on the is_final-after-typer correctness fix (commit
62a9d3f); this is the maintainability follow-up to retire the now-dead direct-delta API left behind by that move.
Problem
The streaming feed loop in
crates/core/src/controller/live_preview.rswas reworked so the dedicatedPreviewTyperthread owns the on-screen state and computes its own common-prefix delta. As a result, the older "direct delta" plumbing thatapply_streaming_transcriptandPreviewTyper::finishstill expose is now dead surface area.1. The publish closure ignores
oldand never fails.run_streaming_loop(live_preview.rs:265) builds a closure whose only job is to forward the latest text to the typer (live_preview.rs:295-298):It discards
_oldand always returnsOk(()). Real delta errors are now handled inside the typer thread (PreviewTyper::run→clipboard::apply_preview_delta,live_preview.rs:516-522), so theResulthere is a permanently-Okpath.2.
apply_streaming_transcriptcarries anold/new/Resultcontract for a dead failure path. Its signature still requires the old text and an error branch (live_preview.rs:390-401):The only production caller passes the always-
Okpublishclosure (live_preview.rs:300-306andlive_preview.rs:330-334), so theErrarm is unreachable in production.3.
PreviewTyper::finishtakes afinal_texthandoff that production always calls asfinish(None). Signature atlive_preview.rs:472:Both production call sites pass
None: the streaming path atlive_preview.rs:347and the batch path atlive_preview.rs:232. The final text is already published viaapply_streaming_transcriptbeforefinish(live_preview.rs:330-334), so theSome(text)branch insidefinish(live_preview.rs:475-477) is never exercised by the app.Why it matters
publishclosure,apply_streaming_transcript,PreviewTyper::finish) advertise anold-text +Result+ final-text-handoff contract that production no longer uses. A reader must trace all three to confirm the failure/handoff paths are dead, which obscures the real design (typer owns state + diff).Err/Somebranches that invite incorrect assumptions during future edits to the finalization flow.Proposed change
Collapse to a latest-text-only API, in a single focused PR:
apply_streaming_transcriptto takeapply_delta: &mut F where F: FnMut(&str)(drop theoldparam and theResult); apply unconditionally on change and updatelast_applied.publishclosure inrun_streaming_loopto|new: &str| typer.publish(new.to_string()).PreviewTyper::finishno-arg (fn finish(mut self) -> String) and drop theSome(text)pre-seed branch — unless review finds a real caller that needs thefinal_texthandoff. Update the streaming (live_preview.rs:347) and batch (live_preview.rs:232) call sites.Tests that must change (they currently encode the old contract and act as the guard):
streaming_transcript_helper_records_deltas(live_preview.rs:864) andstreaming_transcript_helper_skips_duplicate_text(live_preview.rs:894) use anapply_deltaof|old, new| { ...; Ok(()) }and assert on(old, new)tuples — rewrite for theFnMut(&str)signature while preserving the dedup and delta-recording assertions.batch_append_accumulates_text_with_spaces(live_preview.rs:724) already usesFnMut(&str)for the batch path — leave as-is, it's a model for the new streaming-helper test shape.Tests that must stay green unchanged (the correctness guard for the is_final-after-typer fix):
pending_cap_drop_forces_unfinalized_streaming_result(live_preview.rs:629)confirm_streaming_finalization_returns_true_when_text_landed(live_preview.rs:919)confirm_streaming_finalization_falls_back_when_text_not_applied(live_preview.rs:928)streaming_preview_reports_unfinalized_when_reset_fails(live_preview.rs:847)Recommendation / triage
DEFER / do carefully. This area was deliberately reworked for the recent is_final-after-typer correctness fix (commit
62a9d3f, "fix(live-preview): emit is_final only after typer confirms text landed"). Any collapse must keep theconfirm_streaming_finalizationbehavior intact — the final text is published, then the typer must prove it landed on screen beforestreaming_finalizedstays true. Do this as a focused PR with the live_preview streaming tests as the guard, not a drive-by. If review surfaces any real future need for thefinal_texthandoff infinish, keep that param and only collapse the closure +apply_streaming_transcript.Acceptance criteria
apply_streaming_transcripttakesFnMut(&str); nooldparam, noResultbranch.run_streaming_looppublishclosure forwards only the latest text (no_old, noOk(())).PreviewTyper::finishis no-arg, OR thefinal_textparam is retained with a documented real caller (decision recorded in the PR).confirm_streaming_finalizationstill gatesstreaming_finalizedon the typer's confirmed on-screen text; pending-cap drop still forces batch fallback.pending_cap_drop_forces_unfinalized_streaming_result, bothconfirm_streaming_finalization_*,streaming_preview_reports_unfinalized_when_reset_fails.streaming_transcript_helper_records_deltas,streaming_transcript_helper_skips_duplicate_text) assert dedup + delta application against the newFnMut(&str)signature.clipboard::apply_preview_deltaorPreviewTyper::rundiff logic.Traceability
fnd_sig-feat-service-39545238a6-ce5f_91fb601fafclawpatch revalidate --finding fnd_sig-feat-service-39545238a6-ce5f_91fb601faf --reasoning-effort high62a9d3f); this is the maintainability follow-up to retire the now-dead direct-delta API left behind by that move.