Agents ✦ 86outlineAuthor-reported — measure on your traces

Verbatim context compaction

Who: Coding-agent / long-session product teams. Steps: (1) Candidate tool-call/result blocks as state. (2) Noul/Choice keep vs delete/truncate. (3) Code splices kept text verbatim (no rewrite summary). Expected effect: shrink context without lossy paraphrase of kept spans (author claim). Caveat: keeping text verbatim ≠ proving omitted history was unnecessary.

1

Treat this as an outline — adapt state and questions to your data.

2

Implement in code — call System One / Jev; compose answers yourself.

3

Gate on confidence — act, confirm, or escalate before side effects.

Outline sketch

Prompt
Audience: coding-agent / long-session products.
Steps:
1) Snapshot tool-call/result blocks.
2) Noul/Choice: keep | delete | truncate.
3) Code splices kept text verbatim — no summary rewrite of kept spans.
Expected effect: smaller context with verbatim retention (author claim).
Caveat: omit ≠ proven unnecessary.

Needs access to: typesafe-sdkcoding-agent

Who it's for

Coding-agent / long-session product teams

Steps / how it's set up

Sources: fast-jev-compaction · cobanov “Context and compaction” in awesome-jev · aliaihub agent-supervision showcase.

Keep verbatim spans the agent must not invent. Sibling pattern: tool-result sieve. Compaction quality claims from community repos are author-reported.

Sources (wave-2 deepen)

Expected effect

Decision model chooses what stays; code owns transcript bytes.

Unofficial outline for learning. Paraphrased from public docs and tutorials — not a production recipe. Review sources before you automate anything.