Fast Jev Compaction — leonaaardob

A write-up, by the author's own account: the README opens with a warning against using it, calls it an experimental proof of concept rather than a tool to stand behind, and points you at Codex or Claude Code default compaction for real work. It also links a detailed critique arguing that pruning on a probability threshold misunderstands compaction and cache rewrites.

No text is ever removed or shortened: only tool calls and tool results are candidates, user and assistant text stays verbatim and in order, and calls in the first or newest preserveRecentMessages messages are pinned. The critique's counterpoint is that reasoning traces frontier models never expose over the API cannot be protected by any of this.

Both, in that order. Codex hooks cannot substitute the compacted history the way Claude Code function hooks can, so the plugin wraps compaction: PreCompact writes the pruned transcript to disk, Codex runs its built-in summarizer, then SessionStart hooks matching source compact re-inject the verbatim keep as additionalContext, capped at FAST_JEV_CONTEXT_CHARS with a pointer to the full file.

It steps aside. A Jev failure, a missing TYPESAFE_API_KEY, a transcript that cannot be fitted, or an estimated reduction below FAST_JEV_MIN_REDUCTION all make the hooks exit cleanly and let Codex's built-in summary run unchanged. Installation still needs Node 18 or newer on PATH, and plugin hooks are not auto-trusted, so you review the three definitions first.

PreCompact reads transcript_path, replays the rollout JSONL including compacted records so the live history is what Codex sees rather than the whole log, and converts function_call, custom_tool_call, local_shell_call, tool_search_call and web_search_call items into tool uses paired to their outputs by call_id. The pruned transcript, decisions and stats are written under $PLUGIN_DATA. SessionStart hooks matching source compact then emit the verbatim keep as additionalContext, and PostCompact reports kept, truncated and dropped counts as a systemMessage.

Environment variables replace the Claude version's userConfig: TYPESAFE_API_KEY, FAST_JEV_KEY_FILE defaulting to ~/.typesafe_key, FAST_JEV_MODEL, FAST_JEV_BASE_URL, FAST_JEV_KEEP_THRESHOLD at 0.5, FAST_JEV_PRESERVE_RECENT at 6, FAST_JEV_MAX_STATE_TOKENS at 25000, FAST_JEV_MAX_REQUEST_TOKENS at 30000, FAST_JEV_TRUNCATE_HEAD_CHARS at 300, FAST_JEV_MIN_REDUCTION at 0.25 and FAST_JEV_CONTEXT_CHARS at 60000. The desktop app does not source ~/.zshrc, so hooks read the first readable key file, with the environment variable winning.

It is a Codex port, not a rewrite: the compaction engine in src/ and the Jev scoring design are by tamaratran and unchanged here, and this repository swaps the Claude Code function-hook adapter for Codex lifecycle hooks, portable plugin manifests and skills docs. The practical consequence is that Codex hooks cannot substitute the compacted history, so the kept history is re-injected as additionalContext after the built-in summary instead, and the compactAtPercent and turn.complete triggers have no Codex equivalent.