@allenyolk/pi-compaction 0.2.0 → 0.2.1
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/README.md +27 -3
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -71,9 +71,33 @@ silently lowers your summary quality. Choosing a cheaper or faster summarizer is
|
|
|
71
71
|
The chosen model must be one this host offers with configured credentials; if it is not, the package
|
|
72
72
|
says so and hands compaction back to Pi rather than failing the compaction.
|
|
73
73
|
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
74
|
+
### When `off` does not take effect
|
|
75
|
+
|
|
76
|
+
`thinking: "off"` is a request, not a guarantee. A model's `thinkingLevelMap` decides which levels it
|
|
77
|
+
accepts, and a level mapped to `null` is unsupported — so a model that marks `off` unsupported is
|
|
78
|
+
clamped to its lowest level instead and reasons anyway. This package sends no output ceiling, so that
|
|
79
|
+
reasoning has nothing bounding it.
|
|
80
|
+
|
|
81
|
+
`/compaction` reports the level that will actually run, not the one you asked for:
|
|
82
|
+
|
|
83
|
+
```text
|
|
84
|
+
Thinking: low (clamped from off; this model has no off, so its reasoning phase still runs)
|
|
85
|
+
```
|
|
86
|
+
|
|
87
|
+
**This is a per-model property, not a property of the provider.** Two models behind one gateway can
|
|
88
|
+
differ: on the gateway this package was developed against, one model declares `"off": "none"` and
|
|
89
|
+
genuinely stops while its sibling declares `"off": null` and cannot. Check the model you actually
|
|
90
|
+
configured, not the provider.
|
|
91
|
+
|
|
92
|
+
To see what a model declares, read its `thinkingLevelMap` in `~/.pi/agent/models.json`. A model whose
|
|
93
|
+
`off` is `null` cannot be told not to reason; setting it to `"none"` makes the request send
|
|
94
|
+
`reasoning.effort: "none"` instead.
|
|
95
|
+
|
|
96
|
+
Changing it is a request to the **provider**, not to this package. If the upstream model does not
|
|
97
|
+
accept that effort, the summarization request fails and compaction hands back to stock Pi — annoying,
|
|
98
|
+
but reversible: put `null` back. Verify with one real compaction, then re-read `Thinking:` in
|
|
99
|
+
`/compaction`.
|
|
100
|
+
|
|
77
101
|
|
|
78
102
|
## The one thing that is refused
|
|
79
103
|
|