StackMap
Subscribe

OrcaReplay vs shepherd

Time travel for coding agents: a local proxy records any run (Claude Code, Codex, SDKs, even TLS-only harnesses), replays it byte-for-byte offline, and forks from any step onto a different model. — versus — Runtime substrate that records agent runs as reversible, Git-like execution traces — meta-agents can observe, fork, replay and revert any run before outputs are applied or released.

The curated verdict

Both record agent runs as replayable, forkable traces; shepherd is a runtime substrate meta-agents program against, OrcaReplay is a zero-integration proxy for existing coding CLIs with a CLI/UI for humans.

OrcaReplayshepherd
Stars2142.4k
Forks66206
LanguageTypeScriptPython
LicenseApache-2.0MIT
Last activityyesterday2 days ago
Topicsevals, codingagents, orchestration
Curated connections55

OrcaReplay — the curator's take

This is the debugger, not the dashboard: `orca record claude` captures the full loop below the agent (requests, streamed responses, tool calls, shell exit codes, file writes), `orca replay last` reproduces it with zero tokens, and `--from N --model X` makes the model the only variable so A/B comparisons actually mean something. Use it to answer 'why did it delete my migration' and to benchmark models on a real run. Skip it for cost/latency observability (failproofai, numbat, agentsight do that) or for agents you can't put behind a proxy or TLS intercept. Small and young (~200 stars); the trace spec is CC-BY, the default endpoint is the maintainers' OrcaRouter — swap it if that matters.

shepherd — the curator's take

The missing 'version control for agent execution': every run becomes a durable trace you can inspect, fork from any step, replay deterministically, or revert — and workspace outputs are held for review before they're applied. If you're building supervision or meta-agents (agents that manage agents), this is the substrate that makes it tractable. NOT mature: early alpha with a moving API and a fresh paper behind it, and the reversibility guarantee only covers what runs inside its coupled agent+environment model — side effects that escaped to the real world don't revert.