Manually trim, compact, persist, and repair Claude Code conversation sessions stored as JSONL files. Use when the user wants to: (1) fix a session hitting 'Prompt is too long' errors, (2) manually compact or strip a session to reduce context, (3) analyze a session's token usage and message chain, (4) strip tool call inputs and results from a session, (5) strip thinking blocks from a session, (6) compact part of a session before a specific message, (7) verify or repair a broken session chain, (8) persist and summarize tool results to reduce context intelligently, (9) perform any JSONL-level surgery on Claude Code sessions. Triggers on mentions of session surgery, session trimming, session stripping, context compaction, session repair, prompt too long, session compact, manual session manipulation, summarize tool results, or persist tool output.
Local peer-to-peer chat and remote control between Claude Code sessions running concurrently on the same machine, with mIRC-style rooms: register under a name, broadcast to your project room or other rooms, DM other agents, get unread messages auto-delivered by hooks, wake idle peers (nudge), watch their terminals (screen), drive them with keys (escape/enter/...), and spawn brand-new agents into fresh iTerm windows or tmux sessions. Use when: (1) the user asks to message, notify, coordinate with, watch, stop, or control another running agent/session, (2) joining the agent chat ('register as summarizer'), (3) announcing schema/format changes or claiming work areas between parallel agents, (4) checking/reading agent chat messages or browsing rooms, (5) spinning up a new worker agent in a new window or headless tmux session, (6) rebinding an agent identity after a session fork/strip/restart, (7) installing agent-chat on a new machine. Triggers on agent chat, message the other agent, tell the other session, coord
Restart the current Claude Code CLI in its own terminal (iTerm2 window or tmux pane) and resume a session — the external relaunch step an agent cannot perform on itself. Use when: (1) the session's context is nearly full and was just stripped in place with session-stripper, so the CLI must restart to load the smaller transcript, (2) resuming a forked stripped session under a new session id, (3) the user or the agent says respawn, restart yourself, restart the CLI, reload the session, or resume after strip. Backends: iTerm2 (macOS) and tmux (works headless, e.g. Linux servers).
Convert AI conversation transcripts (Claude Code, Codex CLI, ChatGPT JSONL; Cursor IDE SQLite) to readable Markdown. Use when the user wants to: (1) convert .jsonl conversation transcripts to .md, (2) strip binary/base64 data from AI conversation logs, (3) make AI conversation history human-readable, (4) process session exports from ~/.claude/ or ~/.codex/, (5) batch-convert a folder of conversation JSONL files, or (6) list, search, or export Cursor IDE chat/composer conversations. Triggers on mentions of JSONL conversations, transcript conversion, conversation export, readable conversation logs, or Cursor chat history.
Export a claude.ai (Claude Desktop / web) conversation and convert it to a Claude Code CLI JSONL session so it becomes resumable via `claude -r <sessionId>`. Use when the user wants to: (1) continue a claude.ai/Claude Desktop chat in Claude Code CLI, (2) import a web conversation into CC, (3) archive a claude.ai conversation locally with images and attachments, (4) unfreeze a claude.ai chat that hit its context limit, or (5) convert conversation.json from the claude.ai API to CC's JSONL envelope format. Triggers on mentions of exporting/importing Claude Desktop chats, resuming a web chat in CLI, conversation.json → JSONL, or claude.ai chat archival.
Build a complete mental map of a project through phased exploration so you can work effectively without rediscovering things. Use when the user wants to: (1) Onboard onto a new or unfamiliar codebase, (2) Get a full project overview or briefing, (3) Understand project architecture, structure, and open work, (4) Initialize context at the start of a deep work session, (5) Explore and summarize a repository end-to-end. Triggers on requests like "explore this repo", "get familiar with this project", "give me a project overview", "onboard me", "init context", or any request to deeply understand a codebase before working on it.
Create hierarchical, bottom-up "layered summaries" by generating `AGENTS.md` files per directory in a subtree (e.g., `reth/crates/engine`). Deep directories summarize files and key APIs; parent directories roll up children. Also use to incrementally update an existing `AGENTS.md` subtree after content changes. No snippets; keep summaries navigational and checkable.
Pack repositories with Repomix for whole-codebase, cross-file analysis. Default to a whole-repo measurement pass, then a filtered whole-repo pack; if it fits under ~1M tokens, stop and hand off to Gemini. Only if it doesn’t fit after the initial noise filter, reduce scope at folder granularity (avoid file-by-file selection), then iterate based on Gemini feedback.