REPOSITORY ANALYSIS

fast-jev-compaction-pi

This Pi extension asks Jev whether each completed tool call and its full result should be kept, then locally retains, truncates, or drops verbatim tool evidence in the compaction summary.

E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
Stars
0
Forks
0
License
MIT
Last commit
2026-09-20

What Jev does here

Jev supplies separate keep-call and keep-result probabilities; local code fits selected verbatim evidence alongside prose summary text, with downgrade when the final budget is insufficient.

Preserves selected exact output inside a conventional Pi summary while making its final-budget limitations explicit.

Fixed-version documentation and implementation material were inspected on 2026-10-24. No Pi session or live call was run. A core renderer benchmark does not establish live model retention quality.

What task does this project fit?

Keep selected tool evidence inside a Pi compaction summary while allowing normal prose summarization.

Material analysed:

From input to action

  1. The extension listens before Pi compaction and pairs completed tool activity.
  2. Two Nouls evaluate whether the call and its verbatim result remain useful.
  3. Local code combines generated summary text with evidence under a final output budget.

Implementation details that matter

A keep verdict is not a guarantee of full inclusion

The README states that even a keepResult verdict can be downgraded to truncation or omission when the final summary budget is insufficient. Check the emitted evidence, not only the model’s answer.

Inspect the supporting source ↗

Transport validation covers both decisions

src/jev.ts reads call_<id> and result_<id> for every item, rejects malformed JSON and distinguishes timeout from other failures.

Inspect the supporting source ↗

When to choose it

Choose it when combining a conventional summary with selected exact tool evidence is acceptable. Prefer a different design if the original session surface itself must remain directly addressable.

How to inspect it yourself

Inspect the README’s decision table and final-budget caveat, then the readAnswer calls in src/jev.ts. Follow a nominal keep verdict through the documented budget downgrade before relying on original error text being present.

Limits of this review

  • No Pi session or live call was run.
  • A core renderer benchmark does not establish live model retention quality.
Primary material used in this analysis