CR-2026-146 — Retain Audio for Live-Dictation Voice Captures — v0_7

Version: 0.7 (adds to v0.4–v0.6 — see below) Status: Scoped, ready for a follow-on build brief. v0.1–v0.6 already built and merged. Repos: loomworks-engine, loomworks

New gap, distinct from v0.5

v0.5 fixed recordings lost to failed transcription. This is a different, structural gap: recordings are silently lost even on successful transcription, whenever the classified intent does not create a new assertion. Traced by code, confirmed:

• A good transcript takes the onTranscriptReady path (not v0.5's onCaptureIncomplete path).
• dictationRecorder.stop() resolves a real Blob. The transcript is sent to /converse and classified.
• Only add_knowledge and change_assertion's supersede path (CR-2026-145) create an assertion and return a resource_kind: "note" reference. Every other intent — general_conversation, recall_personal, ask_about_past_input, orient, tune_setting, and any future intent that doesn't create an assertion — returns no such reference.
• With no reference, the frontend's assertion-id lookup resolves to null. The code path that would attach the Blob (if (blob && assertionId)) is skipped. The Blob is a local variable inside a closure; once the closure returns, it is garbage-collected. No held item, no error, no signal to the Operator.

This is not specific to one intent — it is structural: any conversational turn whose classification does not produce a new assertion loses its recording, regardless of whether the speech was clearly and correctly transcribed. This directly undermines the reason CR-2026-146 exists: for content where verifying the original speech matters (e.g. a recorded account from an Indigenous Elder), a turn the classifier reads as "just talking" would lose its audio entirely and invisibly.

Fix

Do not gate audio retention on whether the classifier happened to create an assertion. When a live-dictation recording produces a real Blob, give the Operator a way to keep it regardless of classification outcome — options to consider at build time:

• Surface a lightweight "keep this recording?" prompt after any turn where a Blob exists and no assertion was created, independent of intent.
• Or: attach the Blob to the conversational turn itself (not only to an assertion), so it can be replayed from the conversation history even when no Memory item resulted.

Build brief to determine the smaller, more consistent mechanism — this needs a real design decision, not an assumption, given it touches every non-assertion-creating intent, not just one.

Seed check

No conflict. Operator-authority: whether a recording is kept must not depend silently on an internal classification decision the Operator never sees. "Corrections preserved, not smoothed" extends here too — an Operator's actual spoken words should not vanish because the system judged them conversational.

Out of scope

Changing which intents create assertions (a classifier/routing question, not this CR's concern). v0.1–v0.6 mechanics, unaffected.