pi-self-evolve 0.1.0 → 0.1.1
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 +4 -2
- package/package.json +2 -2
package/README.md
CHANGED
|
@@ -1,6 +1,8 @@
|
|
|
1
1
|
# Self-Evolve Pi Extension
|
|
2
2
|
|
|
3
|
-
|
|
3
|
+
[](https://www.npmjs.com/package/pi-self-evolve)
|
|
4
|
+
|
|
5
|
+
**pi-self-evolve** is a context-engineering harness for the [pi coding agent](https://pi.dev). It shrinks the agent's context window on long agentic loops. In the author's A/B measurements (n=3 per cell, twice, one agentic benchmark suite), median context dropped ~50% with no solved-rate loss; a 10-task Terminal-Bench 2.1 run (n=1 per task) showed no task regressions. Your results may differ — see the limits section below. Open source, MIT, one-command install: `pi install npm:pi-self-evolve`.
|
|
4
6
|
|
|
5
7
|
This is a pi extension. It keeps agent context small on long agent loops. It uses five mechanisms:
|
|
6
8
|
|
|
@@ -150,7 +152,7 @@ Limits, stated plainly:
|
|
|
150
152
|
|
|
151
153
|
## Reproduce
|
|
152
154
|
|
|
153
|
-
Adversarial + agentic comparison: `node run-compare.mjs` (env `N_RUNS` to change cell size; results in `results/compare.json`).
|
|
155
|
+
Adversarial + agentic comparison: `node eval/run-compare.mjs` from the repo root (env `N_RUNS` to change cell size; results in `results/compare.json`).
|
|
154
156
|
|
|
155
157
|
Harbor A/B:
|
|
156
158
|
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "pi-self-evolve",
|
|
3
|
-
"version": "0.1.
|
|
4
|
-
"description": "Self-evolving context harness for the pi coding agent
|
|
3
|
+
"version": "0.1.1",
|
|
4
|
+
"description": "Self-evolving context harness for the pi coding agent: evidence reducer with verbatim validation, ObservationPack disk archive with recall, online compaction, action fusion, telemetry. In the author's A/B measurements (n=3, twice, one agentic benchmark suite), median context dropped ~50% with no solved-rate loss.",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"author": "ralph-abejuela",
|
|
7
7
|
"keywords": [
|