omp packages

@dereknex/pi-jev-compaction0.4.1

Verbatim context pruning for Pi using TypeSafe Jev decisions.

GitHub Actions Published 7 days ago MIT ~206 downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

@dereknex/pi-jev-compaction (Pi Extension & Package)

Continuous, verbatim context pruning extension for the Pi coding agent, powered by TypeSafe Jev and Cloudflare Workers AI Clef.

Features

  • Zero Lost Text: Only unneeded tool calls and results are pruned or truncated. User prompts, assistant responses, thinking blocks, and message ordering remain verbatim.
  • Append-only Ledger: Pruning decisions are recorded as fast-jev-decisions custom entries in Pi's session history; the raw JSONL on disk is never modified.
  • On-the-fly Filtering: Modifies LLM context dynamically via Pi's context hook (applyLedger).
  • Proactive & Threshold Compaction: Runs at turn_end when context usage exceeds compactAtPercent (with cooldown guard), and hooks into session_before_compact.
  • Fail-open Safety: During Pi compaction, an API failure or insufficient reduction delegates to Pi's built-in summary. Proactive and manual failures leave pruning unchanged.
  • Durable Diagnostics: Failed and cancelled runs are recorded as fast-jev-diagnostics custom entries, separate from the decision ledger.

Installation

pi install @dereknex/pi-jev-compaction

Or run directly from a local clone:

pi -e ./packages/pi/extensions/pi-jev-compaction.ts

Configuration

In ~/.pi/agent/settings.json or .pi/settings.json:

{
  "fastJevCompaction": {
    "enabled": true,
    "backend": "typesafe",
    "model": "jev-latest",
    "compactAtPercent": 60,
    "minReductionRatio": 0.25,
    "cooldownTokens": 8000,
    "preserveRecentMessages": 6,
    "truncateHeadChars": 300
  }
}

For Cloudflare Workers AI Clef:

{
  "fastJevCompaction": {
    "backend": "cloudflare",
    "model": "clef-flash",
    "accountId": "your-cloudflare-account-id"
  }
}

backend selects the provider

Field backend: "typesafe" backend: "cloudflare"
model jev-latest (default) clef, clef-flash, @cf/cloudflare/*
accountId not allowed required unless baseUrl is set (else CLOUDFLARE_ACCOUNT_ID)
baseUrl optional, complete endpoint URL, used verbatim optional, same meaning
key TYPESAFE_API_KEY CLOUDFLARE_API_TOKEN → CLOUDFLARE_AUTH_TOKEN → TYPESAFE_API_KEY

An unsupported combination is rejected at session start: the extension logs fast-jev: disabled; <reason> and leaves Pi's own compaction untouched. Omitting backend falls back to inference from model / baseUrl / accountId so older configs keep loading, but setting it explicitly is recommended.

Commands

  • /fast-jev: Show current ledger status and statistics.
  • /fast-jev run: Force a decision run immediately.
  • /fast-jev on / /fast-jev off: Enable or disable for the current session.

Failure diagnostics

Inspect fast-jev-diagnostics custom entries in the session JSONL after a failure. Each entry records outcome (failed or cancelled), backend, model, trigger mode, execution stage, and elapsed milliseconds. Fitted state tokens, state fitting stage, and planned batch count are included once known; missing fields mean they had not been calculated. Batch count describes planned batches, not successful responses.

Failures also record an error type, a redacted message limited to 300 characters, and HTTP status when available. HTTP response bodies, stacks, credentials, and request payloads are excluded. Notifications show the same brief reason and only claim a Pi fallback when the run was triggered by Pi compaction. Cancellation is recorded without a failure warning, does not advance the cooldown, and cancels built-in compaction when encountered in the compaction hook.

Requests retain their existing concurrency and are not automatically retried.

License

MIT