@eslaykey/dsh-breeze-tts 1.0.4 → 1.0.7

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -5,6 +5,61 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), version
5
5
 
6
6
  ## [Unreleased]
7
7
 
8
+ ## [1.0.7] - 2026-11-19
9
+
10
+ ### Changed
11
+ - Model-tagged replies use emoji Markdown markers in DSH. Both the new markup and legacy HTML comments are recognized for speech.
12
+ - The settings page shows connection management on demand and keeps voice, tone, and preview controls together.
13
+
14
+ ### Fixed
15
+ - The chat quota indicator refreshes after speech synthesis and picks up tool usage without restarting DSH or querying the Breeze account balance.
16
+ - Refreshing the balance and today's usage preserves unsaved settings. The save control stays visible and reports pending changes.
17
+
18
+ ## [1.0.6] - 2026-11-15
19
+
20
+ ### Changed
21
+ - One normalised segment plan now drives requests, daily budget and the cache key, so all three can no longer disagree
22
+ - Auto-speak waits for the harness turn-end signal (`legacy.turnEnds`); when that signal is missing it stays silent and leaves reading to the manual button
23
+
24
+ ### Fixed
25
+ - A single leading emotion marker stayed in the spoken text
26
+ - Sentences longer than the API limit were sent whole; they are now split without breaking Unicode
27
+ - Mixed Chinese/English text was synthesised as one request with one language
28
+ - Partial failures left already-synthesised segments unaccounted
29
+ - Concurrent requests for the same text called the API twice; synthesis now runs at most two calls in parallel
30
+
31
+ ### Fixed (quota and language)
32
+ - Synthesis now runs one call at a time, so two concurrent requests can no longer both pass the daily check; the settings file also has a single write lock
33
+ - `languageMode: 'en'|'zh'` works again (a non-auto mode is the language itself); the language actually sent is what the cache key and the budget use
34
+
35
+ ### Fixed (auto-speak lifecycle)
36
+ - The play call is checked against its generation and pending instance after it resolves, so a session switch can no longer write the old reply into the new session's state
37
+ - The synthesis/play request is never replayed after it was sent (a retry could charge and speak twice); such a failure is recorded as skipped and left to the manual button, while the read-only `/status` keeps a bounded retry
38
+ - A `/status` already in flight is abandoned when the same session starts a new stream, the signal disappears or the candidate is replaced; only identical re-renders keep waiting
39
+ - A message is marked spoken only after playback succeeds, so a failed `/status` or play is retried with backoff (three attempts) instead of being recorded
40
+ - Replies skipped by settings (auto-speak off, over the limit, background) are recorded as skipped and never replayed later
41
+
42
+ ### Fixed (auto-speak completion)
43
+ - Speech now requires the recorded turn end reason to be `completed`; error, aborted, interrupted and unknown endings never speak
44
+ - A pending 400 ms timer is cancelled on every render and on session switch, and a late `/status` reply is discarded, so the previous session can no longer be spoken
45
+ - The played-message table follows the active session
46
+
47
+ ### Fixed (tool surface)
48
+ - `tts_speak` ignored its public `mood` argument; it now maps to the matching tone and its instructions
49
+ - `play:true` with a `.mp3` save path synthesised WAV and handed it to the player; it is now refused before any request
50
+
51
+ ### Added
52
+ - `plan.js` with offline tests: `plan`, `concurrency`, `tool`, `autospeak`, `autospeak-timer`, `autospeak-integration`, `tts_speak` (50 checks)
53
+
54
+
55
+ ## [1.0.5] - 2026-10-31
56
+
57
+ ### Fixed
58
+ - After a client remount (or a fresh session) the "already spoken" marker could prime on an empty message list, so the previous reply was spoken again
59
+ - Auto-speak now reads only the message's own text blocks instead of the broad fallback walk, which could pick up unrelated text
60
+ - Added a one-line diagnostic to ~/.dsh/breeze-tts/autospeak.log recording what was spoken
61
+
62
+
8
63
  ## [1.0.4] - 2026-10-31
9
64
 
10
65
  ### Fixed
package/README.md CHANGED
@@ -56,6 +56,8 @@ API key 在这里申请:<https://breezeblue.ai/app/developer/api-keys>
56
56
  - `dailyChars`:**硬闸门**,当天合成字数超了就拒绝,防止额度意外消耗
57
57
  - `autoSpeakMaxChars`:只有不超过这个字数的回复才会被自动朗读
58
58
 
59
+ 「模型标注」会让模型在需要变换语气的段落前插入 Markdown 表情标记,例如 `[😄](breeze-mood:happy)`。在 DSH 聊天里它显示为 😄,朗读时插件按标记选择语气,只合成正文。历史回复中的 `<!--mood:happy-->` 仍可朗读。
60
+
59
61
  ## 费用
60
62
 
61
63
  按字符计费,10 个计费单位 = 1 credit,中日韩字符算 2 个单位。实测约 **0.19 credits / 汉字**。