@lingjingai/scriptctl 0.47.0 → 0.47.1
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/changes/0.46.0.md +19 -0
- package/changes/0.47.0.md +31 -0
- package/changes/0.47.1.md +13 -0
- package/package.json +1 -1
|
@@ -0,0 +1,19 @@
|
|
|
1
|
+
# 0.46.0
|
|
2
|
+
|
|
3
|
+
## Added
|
|
4
|
+
- `ingest.name-fix` task: unregistered body-sourced speakers are merged into the existing
|
|
5
|
+
registry instead of minting ghost assets. Never folds two already-registered entities, and
|
|
6
|
+
binds unlabeled 龙套 to 常态.
|
|
7
|
+
|
|
8
|
+
## Changed
|
|
9
|
+
- Location normalization: different spellings of the same place (`矿镇广场` / `Cinder Hollow 广场`
|
|
10
|
+
/ platform-prefixed variants) merge into one canonical location rather than scattering as
|
|
11
|
+
separate entries.
|
|
12
|
+
- Serial identity split into `newEntities` (every 叫法 captured as aliases) and `merges`;
|
|
13
|
+
presence is taken from the body and 造型 from sparse `lookChanges` with carry-forward.
|
|
14
|
+
|
|
15
|
+
## Fixed
|
|
16
|
+
- None.
|
|
17
|
+
|
|
18
|
+
## Breaking
|
|
19
|
+
- None.
|
|
@@ -0,0 +1,31 @@
|
|
|
1
|
+
# 0.47.0
|
|
2
|
+
|
|
3
|
+
## Added
|
|
4
|
+
- Accurate LLM token accounting across both ingest pipelines (text and video). Adapters
|
|
5
|
+
extract per-call usage from the provider response (Anthropic `usage`, OpenAI-compatible
|
|
6
|
+
`usage`, Gemini `usageMetadata`); the router accumulates input/output tokens per task
|
|
7
|
+
alongside the existing call/reuse counts.
|
|
8
|
+
- Token spend is persisted to `workspace/usage.json` (`totals` + `byTask`). It is written
|
|
9
|
+
before the deterministic finalize/validate step, so a validation failure never discards a
|
|
10
|
+
run's accounting, and it accumulates across resumed process runs (cached work is `reused`,
|
|
11
|
+
costing no tokens) for a true cumulative total.
|
|
12
|
+
- `view --watch` surfaces token usage: a header badge (total, input/output, cache hit/create,
|
|
13
|
+
calls) and a per-task table sorted by spend.
|
|
14
|
+
- Cache tokens are broken out. `cachedInputTokens` (cache-read, billed cheap) and
|
|
15
|
+
`cacheCreationTokens` (cache-write, billed premium) are reported as subsets of total input,
|
|
16
|
+
so cost can be estimated per tier.
|
|
17
|
+
|
|
18
|
+
## Changed
|
|
19
|
+
- None.
|
|
20
|
+
|
|
21
|
+
## Fixed
|
|
22
|
+
- Video serial parse no longer spins through all retry attempts on a content-filtered
|
|
23
|
+
episode: an empty Gemini body with a `blockReason` or a terminal `finishReason`
|
|
24
|
+
(`SAFETY`, `PROHIBITED_CONTENT`, `RECITATION`, `BLOCKLIST`, `SPII`, `IMAGE_SAFETY`) is
|
|
25
|
+
surfaced immediately as `content_filtered`, since retry and resume never clear it. A
|
|
26
|
+
blank finishReason stays a transient, retryable hiccup.
|
|
27
|
+
- Backfilled the missing 0.46.0 changelog entry (the 0.46.0 bump shipped without one, which
|
|
28
|
+
broke `scriptctl changelog`'s current-version lookup).
|
|
29
|
+
|
|
30
|
+
## Breaking
|
|
31
|
+
- None.
|