Kunj157

Wyrdline — AI skill for Claude Code

AI community

Causal flight recorder and rewind engine for AI coding agents.

How to install Wyrdline

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Kunj157/wyrdline and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Wyrdline does

Causal flight recorder and rewind engine for AI coding agents. Records what an agent intended and what it actually did (eBPF), checkpoints every tool call, and lets you rewind, fork and replay sessions.

Alternatives in AI

Claude Replay — Convert AI coding agent sessions (Claude Code, Cursor, Codex, Gemini, OpenCode) into self-contained, embeddabl 644 ★
Mirofish CLI — Multi-agent AI prediction engine - digital sandbox for scenario simulation (fork of 666ghj/MiroFish) 307 ★
AI Team OS Installation Guide (for Claude Code) — This file is intended to be read by Claude Code (CC) to assist with installation 158 ★

README

Wyrdline

**A causal flight recorder and rewind engine for AI coding agents.**

Every agent leaves a thread. Wyrdline lets you follow it, cut it, and re-weave it.

**[Website and live demo](https://kunj157.github.io/wyrdline/)**: the real Studio replaying a recorded session in your browser.

Coding agents (Claude Code, Codex CLI, Gemini CLI, any MCP client) run real commands on real machines. What they *say* they did and what they *actually* did can differ, and today there is no independent record and no undo for effects that happen outside the agent's own edit tool.

Wyrdline runs an agent inside a **Cell** (a cgroup-scoped copy-on-write sandbox) and records three planes into one tamper-evident **Ledger**:

Plane	Source	Example
Intent	agent hooks, MCP gateway, OpenTelemetry GenAI spans	`Bash("pnpm test")`
Effect	eBPF in the kernel, scoped to the agent's cgroup	`exec pnpm`, `write src/cart.ts`, `connect 1.2.3.4:443`
State	content-addressed checkpoint after every tool call	checkpoint 9 = manifest `b3:…`

From that it can:

Correlate every effect to the tool call that caused it, with confidence.
Rewind the workspace to any checkpoint, fork a new branch, diff branches, and apply the one you want.
Guard sessions with policies written as WebAssembly components and enforced in the kernel (Landlock + cgroup BPF), and backtest a policy against a recorded session.
Flag intent–effect divergence: "all tests pass" with no test run, writes outside the task's scope, a secret read followed by a network connection, effects that never appear in the transcript.
Export a redacted, signed .wyrd bundle that replays offline in the browser (the correlator runs as Wasm), and annotate it with teammates through local-first CRDT sync.
Capture agent runs in CI through a GitHub Action and roll them up in a team Hub.