synartesis 0.8.5 → 0.8.7
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 +51 -0
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,57 @@
|
|
|
2
2
|
|
|
3
3
|
What changed, and why it mattered. Dates are release dates.
|
|
4
4
|
|
|
5
|
+
## 0.8.7 — 2026-09-22
|
|
6
|
+
|
|
7
|
+
### Added
|
|
8
|
+
|
|
9
|
+
- **Two presets for people with no GPU.** The window has run free models since
|
|
10
|
+
it had a model picker -- Ollama, LM Studio and vLLM all need no key and
|
|
11
|
+
nothing leaves the machine -- but every one of those is free only if you own
|
|
12
|
+
the hardware to run it. **Groq** and **GitHub Models** are hosted, both have
|
|
13
|
+
a free tier, and neither wants a card: a GitHub token you already have is
|
|
14
|
+
enough for the second. Both speak `/v1/chat/completions`, so they are
|
|
15
|
+
presets rather than code, which is what the note at the head of that list
|
|
16
|
+
has always said any endpoint can be.
|
|
17
|
+
|
|
18
|
+
The note beside each says what it actually costs -- rate limited, and meant
|
|
19
|
+
for trying things rather than for a day's work -- because the failure worth
|
|
20
|
+
preventing here is somebody picking a free option and reading its quota
|
|
21
|
+
refusal as this window being broken. The same reason the local entries say
|
|
22
|
+
to pick a model that can call tools: the window is an MCP client, and a
|
|
23
|
+
model that cannot emit a tool call cannot drive it at all.
|
|
24
|
+
|
|
25
|
+
## 0.8.6 — 2026-09-22
|
|
26
|
+
|
|
27
|
+
A test I added in 0.8.5 failed the release build. The published package was
|
|
28
|
+
never affected -- tests do not ship -- but the tag produced no GitHub release
|
|
29
|
+
and no installers, so this is 0.8.5 with that test written properly.
|
|
30
|
+
|
|
31
|
+
### Fixed
|
|
32
|
+
|
|
33
|
+
- **The new screen budget asserted milliseconds, and milliseconds are a
|
|
34
|
+
property of the machine.** `tests/screen-budget.test.ts` held a console
|
|
35
|
+
frame and a watch frame under a stated number of milliseconds. That passed
|
|
36
|
+
on the machine it was written on and failed on a CI runner twenty minutes
|
|
37
|
+
later, which is several times slower and shared -- the gate in `release.yml`
|
|
38
|
+
went red and 0.8.5 never got its installers.
|
|
39
|
+
|
|
40
|
+
It asserts two machine-independent things now. A ratio, measured back to
|
|
41
|
+
back on whatever is running: answering from the indexes against answering
|
|
42
|
+
by reading every action in every run, which is about 36x here and has to
|
|
43
|
+
clear 4, because going back to reading the rows scores 1. And a query plan,
|
|
44
|
+
which is a fact rather than a duration -- the watch screen's newest-first
|
|
45
|
+
read has to name the index, and must not say `USE TEMP B-TREE FOR ORDER BY`,
|
|
46
|
+
which is exactly what it said before 0.8.5 added `actions_recent`.
|
|
47
|
+
|
|
48
|
+
It also seeds its four thousand actions once for the file rather than once
|
|
49
|
+
per test, which was most of the runtime and bought nothing: all three
|
|
50
|
+
questions are read-only.
|
|
51
|
+
|
|
52
|
+
The correctness half is new and is the part worth keeping. The two ways of
|
|
53
|
+
counting -- from the indexes and from the rows -- must agree, because a
|
|
54
|
+
frame that is fast and wrong is worse than the slow one it replaced.
|
|
55
|
+
|
|
5
56
|
## 0.8.5 — 2026-09-22
|
|
6
57
|
|
|
7
58
|
Three states you could get into and not get out of, two screens that read the
|