eve-openai-compaction

eve-openai-compaction provides a remote Codex-style compaction strategy for Eve. It sends the current transcript to OpenAI's stateless /responses/compact endpoint and carries the returned opaque encrypted checkpoint into the next Eve model call.

There is deliberately no local prose-summary fallback. A remote failure fails the compaction attempt.

Compatibility

The package targets [email protected], ai@7, and @ai-sdk/openai@4. It requires the included Eve patch to expose custom compaction strategies.

The package also ships patches/[email protected], the reviewable TypeScript source patch with focused Eve tests. The installable pnpm patch targets Eve's published dist files.

The continuation model must be an OpenAI Responses model configured with providerOptions.openai.store: false. AI SDK then replays the encrypted checkpoint instead of trying to reference a server-stored item by ID.

Install

pnpm add eve-openai-compaction @ai-sdk/openai
mkdir -p patches
cp node_modules/eve-openai-compaction/patches/[email protected] patches/[email protected]

Register the patch in pnpm-workspace.yaml:

patchedDependencies:
  [email protected]: patches/[email protected]

Then run pnpm install. pnpm permits only one patch entry per package version; combine unified diffs if the application already patches [email protected].

Use

import { createOpenAI } from "@ai-sdk/openai";
import { defineAgent } from "eve";
import { codexRemoteCompaction } from "eve-openai-compaction";

const openai = createOpenAI();

export default defineAgent({
  model: openai.responses("gpt-5.3-codex"),
  modelOptions: {
    providerOptions: {
      openai: { store: false },
    },
  },
  compaction: {
    strategy: codexRemoteCompaction(),
    thresholdPercent: 0.9,
  },
});

By default, the compaction request reads OPENAI_API_KEY only when compaction runs. To use workload identity or a secret manager, resolve credentials at the last responsible moment:

strategy: codexRemoteCompaction({
  apiKey: async () => await resolveShortLivedOpenAIToken(),
  headers: async () => ({ "OpenAI-Project": await resolveProjectId() }),
})

Do not put credentials in source, logs, argv, persisted Eve context, or static provider options.

Strategy

The package follows Codex CLI 0.144.6's remote-v2 replacement shape:

1. Use AI SDK's OpenAI Responses serializer to turn Eve's full ModelMessage history and current system instructions into Responses input items.
2. POST that window to the official stateless /responses/compact endpoint.
3. Require exactly one valid opaque compaction item. Never decrypt or convert it to prose.
4. Retain the newest genuine user messages under Codex's 64,000-token budget, preserving images and midpoint-truncating the text at the budget boundary.
5. Replace Eve history with those user messages and an AI SDK openai.compaction custom part containing the encrypted checkpoint.

The 64k client-retention behavior is pinned to Codex compact_remote_v2.rs. The remote checkpoint semantics and canonical output are documented in the official OpenAI compaction guide.

Eve re-applies its current framework-owned post-compaction state after the strategy returns, and its current system instructions remain outside durable message history.