К основному содержанию
скилл CX Codex

orca-replay — скилл для Claude Code: анализ и повтор прошлых запусков

Скилл orca-replay для Claude Code. Отвечает на вопросы о прошлых запусках агента по записи, а не по памяти, и повторяет тот запуск. Помогает понять, почему агент что-то сделал, или воспроизвести сбой.

Apache-2.0 ★ 1 109 обновлён 2026-10-01

Установить в Claude Code · Через Skills CLI (Vercel) Все способы и платформы ↓
npx skills add Continuum-AI-Corp/OrcaReplay --skill orca-replay -a claude-code
Открыть на GitHub

Что делает orca-replay

Скилл ищет запись нужного запуска через orca_list_runs, а цепочку событий, которая привела к конкретному результату, строит через orca_graph. Ответ различает recorded и inferred: первое записано при выполнении, второе выведено по правилу. Это важно, потому что след не поручается за вывод.

Перед объяснением скилл воспроизводит запуск через orca_replay. Повтор отдаёт записанные ответы модели и не ходит в сеть, но команды внутри запуска выполняются заново. worktree: true изолирует файлы репозитория, а всё, что запуск трогал за его пределами — /tmp, Docker, локальная база, другой хост, — меняется повторно. Поэтому перед первым повтором стоит прочитать команды запуска через orca_show_run.

Для сравнения моделей есть orca_compare: он форкает один запуск на несколько моделей с одной контрольной точки. Запись при этом уходит чужим моделям, и тратятся деньги. Каждый форк — живой агент: команды выполняются по-настоящему.

Какие задачи решает

  • Узнать, почему агент удалил или перезаписал файл
  • Найти шаг, который сломал сборку
  • Воспроизвести вчерашний сбой
  • Проверить, как другая модель ответила бы на той же точке
  • Отличить записанные события от выведенных

Примеры запросов агенту

  • Почему ты удалил config.json в запуске?
  • Какой шаг сломал сборку? Воспроизведи его.
  • Сравни, как claude-sonnet-4-20250514 прошёл бы этот запуск с той же точки.

Как установить orca-replay

1Через Skills CLI (Vercel)

npx skills add Continuum-AI-Corp/OrcaReplay --skill orca-replay -a claude-code

2Вручную: положить файлы

~/.claude/skills/orca-replay/SKILL.md

Папка скилла целиком; для одного проекта — .claude/skills/orca-replay/

По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов

Вопросы об orca-replay

Как установить orca-replay в Claude Code?

Выполните npx skills add Continuum-AI-Corp/OrcaReplay --skill orca-replay -a claude-code. После этого скилл появится в ~/.claude/skills/orca-replay/SKILL.md.

Что делает orca-replay?

Читает запись прошлого запуска агента, чтобы отвечать на вопросы о нём, и повторяет запуск или делает форк. Запись — это след с результатами инструментов, кодами выхода и изменёнными файлами.

Почему повтор может менять систему?

Повтор заново выполняет команды запуска. Всё, что запуск трогал за пределами рабочего дерева, изменится снова. Проверьте команды через orca_show_run до первого повтора.

Исходный текст

SKILL.md · Apache-2.0
---
name: orca-replay
description: Answers questions about a past agent run from its recording rather than from memory, and replays or forks that run. Use when asked why an earlier run did something, or to reproduce a failure.
license: Apache-2.0
compatibility: Requires the `orcareplay` npm package (Node 20+) with its MCP server registered as `orca`, and at least one recorded run in the project's .orca/runs directory.
metadata:
  author: Continuum-AI-Corp
  version: "0.1"
  homepage: https://github.com/Continuum-AI-Corp/OrcaReplay
---

# Reading a recorded agent run

A recording is evidence. Your memory of a session is not, and neither is a transcript you were
handed — both are missing the tool results, the exit codes, and the files that changed without
anyone mentioning it.

**The rule: when a question is about something that already happened, read the trace before you
answer.** Do not reconstruct it. If a recording exists, guessing is the wrong move even when the
guess would have been right.

## When to Use This Skill

- "Why did you delete/overwrite/move X?"
- "What changed this file?" / "Which step broke the build?"
- "Can you reproduce yesterday's failure?"
- "Does this still reproduce?" (see the limit on that in step 4 — replay cannot tell you
  whether a *fresh* run would fail again)
- "Would a different model have got this right?"

## Workflow

### 1. Find the run

`orca_list_runs` — newest first, and it names the run each fork came from. Skip this only when the
user clearly means the most recent one; every other tool defaults to `run: "last"`.

### 2. Narrow to the chain that produced the thing being asked about

`orca_show_run` gives the whole timeline: model turns with token counts and stop reasons, tool
calls with arguments and results, shell commands with exit codes, and every file the run changed.
Good for orientation, long for a specific question.

`orca_graph` is usually the better tool. It returns causal edges — which event produced which. Pass
`to: <event seq>` to get **only** the chain that produced one event. That is the shape of an answer
to "why did this happen", where the full timeline is the shape of an answer to "what happened".

### 3. Report `recorded` and `inferred` differently

Every edge from `orca_graph` is labelled:

- **`recorded`** — the recorder watched it happen and wrote it into the trace.
- **`inferred`** — derived just now from a rule the edge names. The trace does not vouch for it.

Carry that distinction into your answer. "The trace shows the `rm` at step 14 removed it" and "this
looks like the `rm` at step 14, going by timing" are different claims, and flattening them into one
confident sentence is the specific failure this tool exists to prevent. Name the rule when you lean
on an inferred edge.

### 4. Reproduce it before explaining it

`orca_replay` re-runs the recording and reports what could not be reproduced — divergences, and
requests the recording could not serve.

**What "offline" covers, and what it does not.** Every model response comes from the trace and the
proxy forwards nothing upstream, so no provider is contacted and no tokens are spent. An unmatched
request halts the replay rather than falling through to the network, unless `--loose` was asked for.

That covers the model traffic. It does not cover the agent's own subprocesses: unless the recording
used `--tls-intercept` — in which case replay re-establishes interception for the hosts it recorded
— a `curl`, `npm install`, `git push` or database call inside a recorded shell command goes
straight out. Replay is not a sandbox; only a network-isolated container makes it one.

**What a matching replay proves, and what it does not.** It shows the recorded decisions reproduce
against today's environment. It cannot show the failure is deterministic, because the model is not
being asked again — the same recorded responses are served back. If the user wants to know whether a
fresh run would fail the same way, say that replay cannot answer it; that needs real runs.

**Replay re-executes the agent, not just its model traffic.** The recorded model responses are
served from the trace, but the agent process runs again for real — so every shell command it issued
runs again too. `worktree: true` isolates repository files and nothing else. Anything the run
touched outside the tree — `/tmp`, Docker, a local database, a package manager, another host — is
mutated a second time.

**So check before the first replay of a run, not after.** Read its shell commands with
`orca_show_run` and tell the user what will re-execute. If any of it reached outside the working
tree, get approval for that specifically or replay inside a container; do not treat the earlier
`worktree` answer as covering it. A run that only read files and edited the repository is free and
repeatable, and worth replaying before committing to any explanation.

**Pass `worktree: true`.** It replays into a scratch copy and leaves the working tree alone.

Without it, replay is destructive for as long as it runs: it restores the recorded filesystem over
the working tree and puts the tree back when the replay ends. Uncommitted work is absent in the
meantime, and stays absent if the replay is interrupted before it can restore. Run an in-place
replay only when the user has been told that and has agreed to it. "They do not appear to be
typing" is not consent.

A replay reporting `reused=3/5` on an interactive recording is not a partial failure. Harnesses make
calls for themselves — a quota probe, a session-naming request — and a replay does not repeat them.

### 5. Only then consider comparing models

`orca_compare` forks one run onto several models from the same checkpoint: same files, same
conversation prefix, so the model is the only variable. Pick the fork point with `orca_checkpoints`
and pass it as `from`.

Grade with `verify` — a shell command whose exit code is the verdict. Use something the repository
already declares (`"npm test"`, `"npm run typecheck"`) or an explicitly local binary
(`"./node_modules/.bin/tsc --noEmit"`), not `npx <tool>`: with no local install, npx runs whatever
the registry has under that name, and `npx tsc` resolves a package deprecated in 2016 that is not
TypeScript.

**`orca_compare` uploads the recording to other people's models, and spends real money doing it.**
Each model named receives the same files and conversation prefix the original run had — so whatever
that run touched (source, prompts, configuration, anything a credential was pasted into) is sent to
every provider behind those model ids.

**And each fork is a live agent, not a replay.** From the fork point onward the model is really
being asked, and whatever it decides to do, it does — its shell commands execute for real, and so
does the `verify` command you pass. Each fork gets its own worktree, so repository files are
isolated per model; nothing outside the tree is. A fork can also take actions the original run never
took, because it is a different model making fresh decisions.

So the approval has three parts, and they are not the same question:

1. **Disclosure** — what context is uploaded, and to which providers. Approving a bill is not
   approving a disclosure, and the two need separate answers when the recording is from a private
   codebase. `orca scrub` is for when the comparison is worth running but the trace is not safe to
   send as-is.
2. **Side effects** — what the recorded run did outside its worktree, since each fork may repeat it
   and may go further. Same check as step 4, `orca_show_run`, and the same answer if it reached
   Docker, a database, a deployment or another host: get approval for that specifically, or run the
   comparison in an isolated environment.
3. **Cost** — how many models times how many forks.

Never run it to satisfy curiosity the user did not express.

## If there is no recording yet

Say so plainly rather than falling back to guessing, and offer to start one.

If `orca` is already installed:

```console
orca record claude           # or codex, opencode, openclaw, grok
```

If it is not, ask before installing it — a global install changes the user's machine, and that is
their call, not a detail of your task:

```console
npm i -g orcareplay     # ask first
```

`orca record <agent>` runs the agent unmodified behind a local proxy. Nothing about the agent
changes; two environment variables get set. Recording a session now is what makes the next "why did
it do that" answerable.

For a run started with a prompt in argv — `orca record claude -- -p "…"` — the replay is exact. A
session someone typed into replays approximately, because the prompts were never on the wire and
are recovered from the harness's own transcript; `orca replay` says which is which rather than
papering over it.

## Sharing a run with someone else

`orca export last -o run.html` writes one self-contained file. A trace holds whatever the run held,
so run `orca scrub` before sending one anywhere.

Scrubbing is best-effort, not a guarantee. It matches known key shapes and high-entropy strings; it
cannot know that a particular internal hostname, customer name, or unreleased feature is
confidential to this user. So scrub, then have the user look at what is actually going out, and get
their agreement — do not describe a scrubbed trace as safe on the strength of the scrubber alone.

## Limitations

- **It only sees what was recorded.** Runs started without `orca record` leave no trace, and
  nothing here recovers them. The answer to "why did it do that" in an unrecorded session is
  honestly "there is no recording", not a reconstruction.
- **A typed session replays approximately, not exactly.** Prompts entered at a terminal were never
  on the wire; orca recovers them from the harness's own transcript. Only a run started with the
  prompt in argv (`orca record claude -- -p "…"`) replays byte-for-byte.
- **Some turns are not repeated.** A harness makes calls for itself — a quota probe, a
  session-naming request — and a replay steps over them. Tools that need a person
  (`AskUserQuestion`, plan mode) are absent when the same agent runs without one, which can make a
  replayed request differ from the recorded one by enough to halt.
- **`inferred` edges are not evidence.** They are derived from a named rule at query time. Treat
  them as a reading of the trace, never as something the recorder witnessed.
- **Not every harness is recordable.** Agents that read no base-URL variable and pin their own
  origin need `--tls-intercept`, and some cannot be reached at all. A recording that came back
  empty means the harness was not captured, not that nothing happened.
- **Replay is not a time machine, and not a sandbox.** It reproduces the agent's side of the run
  against today's world. External state the run depended on — a database row, a remote branch, the
  clock — is whatever it is now, and the run's own shell commands reach it for real.
- **A matching replay is not a determinism result.** The model is not re-asked; its recorded
  responses are served back. Whether a fresh run would fail the same way is a different question
  that replay cannot answer.

## Tools

| tool | arguments | notes |
|---|---|---|
| `orca_list_runs` | — | newest first, names the parent of each fork |
| `orca_show_run` | `run` | the full timeline |
| `orca_checkpoints` | `run` | where a fork can start |
| `orca_graph` | `run`, `to` | causal edges; `to` narrows to one chain |
| `orca_replay` | `run`, `worktree` | offline, free, repeatable |
| `orca_compare` | `run`, `models`*, `from`, `verify` | **spends real tokens** |

`run` accepts a run id or `"last"`, and defaults to `"last"`. Replay traces are skipped when
resolving `"last"`, so it means the newest run you actually recorded.

© Continuum-AI-Corp · Apache-2.0 · Continuum-AI-Corp/OrcaReplay

Похожие скиллы

Все скиллы по теме «Настройка агента: память, контекст, субагенты» →
browser-replay ruvnet Скилл browser-replay для Claude Code повторяет записанную сессию браузера на том же или изменённом адресе. Используется для регрессионных тестов после деплоя. CX Codex +1 721установок competition-custom-protocol-replay zhaoxuya520 Скилл competition-custom-protocol-replay для Claude Code восстанавливает бинарные и текстовые протоколы и воспроизводит состояние сессии в CTF. Запускается при разборе кастомных кадров, контрольных сумм и порядка пакетов. CX Codex +1 109установок vibe-replay tuo-lei Плагин vibe-replay для Claude Code превращает сессию ИИ-кодинга в интерактивный HTML-реплей и артефакты для GitHub PR. Показывает ревьюеру ход работы, а не только дифф. ★ 39звёзд репо investigating-replay PostHog Скилл investigating-replay собирает метаданные записи сессии PostHog, профиль пользователя, события и ошибки в один связный отчёт. Пригодится при расследовании сбоев по идентификатору сессии. CX Codex +1 424установок finding-replay-for-issue PostHog Скилл finding-replay-for-issue находит среди сессий, привязанных к ошибке в PostHog, ту запись, по которой видно, что пользователь делал до сбоя. CX Codex +1 407установок agent-strace Siddhant-K-code MCP-сервер agent-strace для Claude Code сохраняет действия ИИ-агентов, вызовы инструментов и ошибки на локальный диск. Помогает пошагово разбирать сессии и проверять готовность веток к ревью. CX Codex +1 ★ 108звёзд репо

Из того же набора2

Спонсоры проекта

Компании, которые помогают держать аналитику открытой