One Idea, Four Agents: Jev Compaction for Codex, Pi and omp

When a technique fits a real pain point, it gets ported fast. Context compaction using Jev, TypeSafe AI's decision model that returns typed scores instead of text, is a clean example. The first tool targeted Claude Code. Within days, separate authors shipped equivalents for OpenAI's Codex, the Pi agent and omp.

Comparing them is useful because the core idea is identical while the engineering choices are not. What happens when the API is down? Can dropped output be recovered? Does the tool touch your session file? Those answers tell you more than a star count does. All four projects are listed on the Awesome Jev radar, and all were created within days of this writing.

Key Takeaways

• Same core loop everywhere: score old tool calls and results with Jev, then keep, truncate or drop them with local code.
• Failure behavior is the differentiator. Codex returns the full input on failure. Pi falls back to its built-in summary compaction.
• Evidence is thin. Three of these repos had 1 to 2 stars as of September 2026, and the site's own review labels them auto-extracted and not independently tested.
• omp's README publishes an unflattering number. Reduced context answered 63 to 75% of test questions versus 100% with full context.
• Privacy applies to all of them. Session content is sent to a decision endpoint.

Codex: a traceable handoff packet

codex-jev-compaction (aiskill.market page) installs through the Codex plugin marketplace and needs Node.js 22+, Codex CLI 0.153.4+ and a TYPESAFE_API_KEY.

Instead of shrinking a live transcript, it produces a handoff package: retained source blocks, per-ID decisions and content metrics, as JSON or Markdown. Only complete, read-only tool pairs with verified results are eligible. Pairs scoring below 0.2 are dropped, and uncertain or protected content stays.

Failure is handled conservatively. If the key is missing, the state exceeds 24,000 bytes, or the network errors, the plugin returns the complete input with a reason code. The README reports an 88.02% source reduction on its fixture tests (7,081 bytes down to 848). That is a fixture, not a real session, so do not read it as a typical result. It also warns that it cannot reliably infer unfinished-work markers in every language, so pin critical context explicitly.

Pi: a ledger, not a rewrite

pi-fast-jev-compaction is a Pi 0.85.1 extension. It sends a compact representation of tool calls to TypeSafe's /v1/systemone endpoint and applies three monotonic decisions: keep, drop the result (keeping the call with truncated output), or drop the call entirely.

Two design choices stand out. It "fails open": errors, timeouts, malformed responses or insufficient reduction hand control back to Pi's built-in summary compaction. And it never rewrites the JSONL session file. It filters only what is sent to the model, storing decisions as append-only entries that are rebuilt after resume, fork or reload. The site's review notes that no savings percentage has been verified.

There is a second Pi package, pi-jev-compaction (listing), by the author of the Codex tool. It describes the same aim (keep critical instructions and tool history, prune noise, fall back gracefully) but was auto-summarized from its README, so treat details as unconfirmed.

omp: sticky decisions and an honest recall number

omp-jev-compaction (listing) works with either the TypeSafe API or OpenRouter. Tool output Jev marks as no longer needed is truncated to a short head plus a recoverable note, and dropped output is stored in spill files. User and assistant text is preserved verbatim.

Details worth noting from the README:

• Sticky mode applies decisions consistently so the provider's prompt cache is not invalidated between requests.
• Reduction only activates above 150,000 characters by default.
• Each scoring pass costs approximately $0.0005, according to the author.
• Measured recall: reduced context answered 63 to 75% of questions, against 100% with full context.

That last figure matters. It says the trade-off is real and measured by the author, which is more than most tools in this space offer. The privacy caveat is stated too: session state travels to the decision endpoint.

The north-star variant

hermes-jev-north-star (listing) is not compaction, but it belongs to the same family of "let Jev judge, let code enforce." It is a Hermes agent skill that turns an intention into a checkable finish line, then has Jev judge whether requirements are met (against a 0.60 probability threshold) and which requirement is weakest. Its README concedes that numbers fluctuate between calls on identical evidence and that deterministic checks are the load-bearing part. It had 1 star as of September 2026.

What to do next

• Match the tool to your agent, then read its failure behavior before its feature list.
• Test recall on your own tasks. omp's 63 to 75% shows why: run a session you know well and check what the reduced context can still answer.
• Keep sensitive repos out until you are comfortable with session data going to an external endpoint.
• Compare with the Claude Code originals in Context GC, and use the checklist in How to Evaluate Jev Projects Honestly.