orca-replay

Files (1)

• SKILL.md

SKILL.md (first lines)

# 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.

From Continuum-AI-Corp/OrcaReplay, Apache-2.0.

More from Continuum-AI-Corp/OrcaReplay