- Drop redundant !lastToolError check from the messaging-tool clean-stop
early return; the earlier lastToolError early return already handles
that case, so the extra condition was dead and misleading.
- Update the CHANGELOG entry to reference only stopReason=stop; the pi-ai
StopReason type does not include end_turn, so the earlier mention was
a documentation-only discrepancy.
The agent runner was surfacing a '⚠️ Agent couldn't generate a response'
warning even when the assistant had already sent user-visible content
through a messaging tool and the turn ended cleanly. Treat that path as
a successful delivery and skip the warning while keeping real failure
modes (tool errors, stopReason=error, interrupted tool use) intact.
Fixes#70396.
The tts tool previously returned a fixed "Generated audio reply."
string in its content, so session transcripts lost what was actually
spoken. Across every channel, a voice-only reply left no text record
for future turns, forcing users to recover transcripts from the
provider's API. Echo the synthesized text back in the tool result
content (audio still delivered via details.media).
Sanitize the transcript before embedding so crafted utterances cannot
inject reply directives when tool output is rendered in verbose mode:
MEDIA: at line start and [[…]] markers are interrupted with a
zero-width word joiner (U+2060) that defuses parseReplyDirectives
without altering the visible text.
Replace full-table scan via vec_distance_cosine() + ORDER BY LIMIT with
sqlite-vec's native MATCH + k = ? KNN operator. Keep vec_distance_cosine()
in the SELECT so score = 1 - dist preserves the existing cosine [0,1]
semantics the downstream merge pipeline depends on.
Fixes#69666.
Benchmark on 10,827 chunks, 4096-dim embeddings:
- Before (full scan): ~8490 ms/query
- After (KNN + join): ~50 ms/query
No behavioral changes: returned ids and ordering are identical to the
previous query on all tested queries. The LIMIT ? binding is replaced by
k = ? which caps sqlite-vec's candidate set to the same count.
