The Agent-Native Index

agent-native · Observability & Tracing · Verified 2026-10-20

OrcaReplay

Listed on The Agent-Native Index

Your agent broke something at 2am. Replay it at 9am — exactly, offline, as many times as you like

Website ↗ Repo ↗ Full dossier ↗

What it does

OrcaReplay records a coding agent below the harness and gives the run back. Capture happens at the process and socket boundary rather than through an SDK, so the agent binary stays unmodified. Five layers cooperate on one timeline: a loopback proxy for model traffic, a PATH shim for shell commands and exit codes, an MCP JSON-RPC tee, a shadow git index for per-turn file state, and a fetch hook (or opt-in --tls-intercept) for hardcoded origins. That shared timeline is what lets you ask which tool call sent the run down the wrong branch.

orca replay serves recorded HTTP bodies back with the network off (egress=blocked), so the harness runs again against a fixed transcript. orca fork / orca replay --from N --model <other> sits a cursor on a derived checkpoint: earlier turns stay on disk, later turns go live on a different model against the same conversation prefix and restored git tree. orca compare repeats that fork across named models and grades each with a --verify command.

Bounds, stated rather than implied. The byte-for-byte guarantee holds when the prompt was recorded through argv (orca record claude -- -p "…"). A hand-driven terminal session replays the conversation but not the run byte for byte — the harness makes calls for itself, and interactive-only tools such as AskUserQuestion are absent without a person. Replay proves the recorded decisions still reproduce; it cannot prove a fresh run would fail the same way, because the model is not asked again. Replay re-executes recorded tool calls for real and is not a sandbox. orca compare uploads recorded context to third-party providers and spends real money.

Supported capture paths include Claude Code, Codex CLI (API key or ChatGPT-subscription login via --tls-intercept), opencode, goose, OpenClaw, grok-cli, Cursor-in-the-IDE, plus framework traffic that reads a base-URL variable (OpenAI Agents SDK, Vercel AI SDK, LangGraph, CrewAI, Aider, and others listed in the README).

Maturity, stated plainly. The repository was created on 2026-08-29. As of 2026-09-12 it is on npm as orcareplay@0.2.4, listed in the official MCP registry, and ships a published Agent Skill — and it is still early. There is no tagged 1.0, no stability guarantee on the v0 trace schema, and no adoption figure worth claiming. The catalog lists it because the operator-surface is real and documented, not because it is mature.

Agent-facing interfaces

How agents onboard

Interaction pattern: CLI wrapper (process + socket interception; operator-surface track)

npm i -g orcareplay
orca doctor
orca record claude
orca replay last
orca fork last

See the full onboarding section in the dossier.

Protocol surface

Interface Detail
CLI orca doctor, record, replay, fork, compare, show, checkpoints, graph, export, scrub, gc, attach, setup, models, quickstart; --json on the read/replay commands
MCP stdio orca mcp / npx -y orcareplay mcp; six tools; official registry io.github.Continuum-AI-Corp/orcareplay
Agent Skill skills/orca-replay/SKILL.md; ClawHub search orca-replay
TypeScript API import { Orca } from 'orcareplay' — same source of truth the CLI renders
On-disk format v0 trace + JSON Schema; spec at spec/orca-trace-v0.md (CC BY 4.0)
Viewer Optional single-file HTML (orca replay --ui / orca export -o run.html); not required for the operational loop

Verification

Last verified 2026-10-20.

Sources: