Why add this instead of letting compaction summarize at 80%?

Late one-shot summaries can drop paths, errors, and constraints with no chance to reduce incrementally. This keeps or truncates tool results with bounded markers and no generated summary, so the backend compacts less often.

When does a middling retain score still get truncated?

Below 0.2 it truncates even when the global signal disagrees; below 0.5 it also needs should_reduce at or above 0.6. Pinned recent nodes and missing or invalid scores always stay.

How do I recover a pruned tool result mid-session?

Original events stay in the append-only log with a prune audit event. Workspace mode writes text to disk for explicit re-read; unload mode leaves a marker telling the model to rerun the tool.

What do I install and what stays enabled alongside it?

Add it with dsh plugin add, set TYPESAFE_API_KEY, and keep dsh-compaction-basic enabled because this never registers ctx.compaction. Set JEV_COMPACTION_DEBUG=1 for evaluation traces.

When exactly does the advisor evaluate?

Only when all three hold: the context window is known, token pressure is inside the lookahead band, default 70-80% of the compaction threshold, and tokens have grown enough since the last evaluation. The required growth adapts from 25% at the band start down to 10% near the threshold.

Which tool results are candidates for pruning?

Only tool/result surface nodes with a matching earlier tool-call, non-empty text and no error status, and not among the most recent preserveRecentNodes. Candidates are deduplicated by callId keeping the newest, then sorted by token density descending so noisy outputs are evaluated first.

How does workspace mode archive a pruned result safely?

It writes the original text to /jev-compaction/ /tool-result- .txt using exclusive creation with restrictive 0600 and 0700 permissions, and replaces the surface node with an archive pointer. Secrets are detected and rejected, so sensitive content is never archived.

CommunityJev Agents & MCP

DeepSeek Harness Jev Pre Compaction

Pre-compaction advisor for DeepSeek Harness that asks Jev which tool results can be pruned in the 70-80% lookahead band, keeping the log intact and leaving real compaction to the backend.

See highlights

Open github.com

Stars: 5
Language: TypeScript
License: MIT
Published by: Community
Author: wjw66
Last verified: Oct 2026

wjw66/deepseek-harness-jev-pre-compaction

User repository on GitHub

View on GitHub

A pre-compaction advisor for DeepSeek Harness. Runs before the standard `compaction-basic` backend, using TypeSafe JEV to safely prune low-value tool results from model context. Original session events stay in the append-only log; only the model-visible view is replaced with compact markers or archive pointers to reduce context bloat.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A pre-compaction advisor for DeepSeek Harness. It runs before the normal compaction-basic backend and uses TypeSafe JEV to decide which low-value tool results can be safely pruned from the model context. The original session events remain in the append-only log; only the model-visible surface is replaced with a short marker or an archive pointer.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Evaluates only inside the lookahead band with an adaptive growth gate to avoid thrashing.
Asks should_reduce plus per-candidate retain questions in concurrent fail-open batches.
Keeps verbatim or swaps a bounded marker; workspace mode archives to disk with 0600 permissions.
Never registers compaction and falls through to next() on key, network, or archive errors.

Watch out

Error results and the most recent nodes are always preserved, and secrets are redacted from requests and rejected from archives. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

JevLoop

Runnable agent harness that moves pick, score and yes-no loop choices to Jev or Laya in milliseconds while reserving the language model for writing, compiling DECISION.md into executable questions.

Jev Agents & MCP#community#jevbest#catalog-import

Community

zjunlp/JevLoopTypeScript

patdown

Node CLI that lints a file tree against fuzzy natural-language rules in one Markdown file, judging each matched file with a swappable TypeSafe-backed judge.

Jev Agents & MCP#community#jevbest#catalog-import

Community

tyler-dot-earth/patdownTypeScript

jev-cli

Command-line toolkit that wraps Jev judgments into scriptable verify, screen, classify, extract, find, rerank, match, route, ask, compact, and batch commands.

Jev Agents & MCP#community#jevbest#catalog-import

Community

Nasrallah-AL/jev-cliTypeScript

jevwire

Wires Jev into agents as seven MCP tools, an embeddable JevDecisionModel library with run functions, and Claude Code hooks at tool and turn boundaries.

Jev Agents & MCP#community#jevbest#catalog-import

Community

Brainwires/jevwireTypeScript

jev-recipes

TypeScript recipe library with 248 focused functions that call Jev for routing, reranking, gating and labelling, returning structured ready-or-review decisions for Node backends.

Jev Agents & MCP#community#awesomejev#catalog-import

Community