Orcareplay README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Orcareplay listing page.
OrcaReplay
English · 简体中文 · 日本語 · 한국어 · Deutsch · Français · Español · العربية
Your agent broke something at 2am. Replay it at 9am — exactly, offline, as many times as you like.
Record any coding agent. Reproduce the run byte-for-byte with no model called. Fork it from any step onto a different model and see who gets it right.
Built by the team behind OrcaRouter — one API key and one endpoint
for Claude, GPT, Gemini, Grok, DeepSeek, Qwen and the rest. It is what orca setup points at by
default, and what makes orca compare a single command instead of four provider accounts.
Find us: OrcaRouter All model APIs
Github Repos: OrcaCode Review · OrcaRouter Lite
Connect: X · Discord · Hugging Face · Ollama
Real output from one session — a Claude Code run recorded, replayed against the recording, then
forked at checkpoint 4 onto two models and graded by npx tsc --noEmit. Nothing here is mocked up.
Try it in three commands
One npm package named orcareplay, which puts one command named orca on your PATH. There is no
separate tool called "orca", and nothing is installed into your agent.
The third line is the one people stay for: same files, same conversation prefix, different model from step 4 onward. The model is the only variable, which is what makes the answer mean anything.
The three commands at the top need an agent installed, a key, a network and real tokens. If you have none of those yet, one command brings its own:
It writes a small project with a genuine bug in it and a recording of an agent fixing that bug,
then replays the recording against the project with no model called: two failing tests before,
four passing after, three turns served from the trace and nothing spent. --full prints the whole
timeline and the replay as it happened.
Read your agent's own system prompt
A proxy that sees the whole loop also sees the prompt the harness assembled before it sent anything. One command captures it, scrubs the machine out of it, and files it by model:
Interactive prompts and -p prompts are not the same prompt, and neither is the same across
models. See capture/README.md for the measured differences, the pitfalls,
and the sanitising rules.
Why this exists
Agent debugging today is archaeology. You scroll a terminal, you re-run and get a different failure, you add print statements to someone else's harness. The tools that exist are observability tools: they tell you a run cost $4.12 and used 61k tokens, which is not the question you have. The question you have is why did it delete my migration file.
OrcaReplay answers that by giving you the run back.
| Observability tools | OrcaReplay | |
|---|---|---|
| Tells you what a run cost | ✅ | ✅ |
| Tells you which tool call deleted the file | sometimes | ✅ |
| Runs the agent again and gets the same answer | ❌ | ✅ from the recording, byte-for-byte |
| Lets you change the model and re-run from step 4 | ❌ | ✅ |
| Needs you to modify your agent | usually an SDK wrapper | ❌ two env vars |
| Works after you close the terminal | ❌ | ✅ it is a file |
| Sees past the model API — shell exit codes, file writes | ❌ | ✅ every turn |
| Records an agent with no API endpoint to redirect | ❌ | ✅ opt-in --tls-intercept |
The last two rows are the ones an SDK wrapper structurally cannot reach. Capture happens below the agent — at the process and socket boundary — so it does not matter whether the agent is yours, whether you can edit it, or whether it even holds an API key: a Codex CLI signed in with a ChatGPT subscription talks to its own backend over TLS, with a credential that only that backend accepts, so it cannot be pointed elsewhere and stay the same session — and orca can still record it. See when the harness will not be redirected.
How it works
Model APIs are stateless, so on every turn an agent resends the entire conversation — including the previous turn's tool results. A proxy in front of the model therefore sees the whole loop: each request, each streamed response, every tool call the model emitted, and every tool result the harness produced. That one property is what the tool is built on, and it is why OrcaReplay does not patch your agent — it stands up a local proxy, sets two environment variables, and gets out of the way.
What "egress blocked" means, exactly. On replay the proxy refuses to forward anything it cannot serve from the trace, so no model is called and no tokens are spent — that run prints egress=blocked. --loose lifts it deliberately: an unmatched request is then answered by the provider and recorded as a major divergence, and the same line reads egress=live-on-unmatched. A replay still executes the recorded tool calls for real, and a tool that opens its own socket — a shell command running curl, an MCP server fetching something — is outside the guarantee either way. By default it never reaches the proxy and goes to the network as usual. Under --tls-intercept it does reach the proxy, because that sets HTTPS_PROXY for the whole child: a host off the intercept list is tunnelled through untouched, though its hostname, port and byte counts still land in the trace, and a host on the list is refused there like any other unmatched call. Replay is not a sandbox; if you need one, run it inside one.
Three more layers catch what the protocol cannot see: an exit code, a real duration, which stream a byte came out of, a file written without telling anyone. A fifth exists for the agents that read no base-URL variable at all — see which agents. A sixth reads the agent's own account of its structure, for a harness that has one: which sub-agent ran, which handed off to which, whether a guardrail tripped, which graph node produced what and which of them called no model at all — none of which reaches the wire.
They all land in the same timeline, ordered by when they actually happened rather than when orca got around to reading them.
Exact, fork and compare are one thing
They are not three subsystems. They are the same proxy with a cursor — the position in the recorded stream where it stops answering from disk and starts answering from the network.
| command | where the cursor sits | what you get |
|---|---|---|
orca replay last | at the end | the whole run again, network blocked — no tokens, no charge, no variance |
orca replay last --from 4 --model X | at checkpoint 4 | turns up to 4 identical, then a different model takes over |
orca compare last --from 4 --models a,b | at checkpoint 4, several times | one table, one variable — the model |
A checkpoint is not recorded; it is derived — any point where the conversation prefix is complete and the workspace was snapshotted. Every fork therefore starts from a state that provably existed.
What a bug hunt actually looks like
Your agent was supposed to fix a failing auth test. It exited 0 and the test still fails. Start with what it actually did:
Three facts the model's own transcript could not have told you, and the run's exit code hid: the
file really changed (seq 6, +1 −3), the check the agent ran failed (seq 12, exit 1), and it
finished anyway. The run exited 0 because the agent exited 0.
That last fact is the one worth a command of its own. orca show gives you the order things
happened in; orca graph gives you what produced what:
Two kinds of edge, and the difference matters. A recorded edge was written when the run
happened, because a tool_use block is physically inside the response that emitted it. An
inferred edge was worked out just now by the rule it names — a filesystem snapshot is taken
once per turn rather than once per tool call, so attributing a file change to a particular call
is a good guess and not a fact. Inferred edges are never written back into the trace, the same way
checkpoints are derived and never recorded, so a field a third-party reader trusts never contains
something orca made up.
--graph-card draws the whole run that way — time left to right, kind of thing top to bottom, with
the chain that produced the failure lit against everything else:
The shape is the point. A run is one motif repeated — request, response, call, effect, result — so anything that breaks it is worth a look, and an event with no edge leaving it is an absence a list cannot show at all.
orca export last --card bug.svg draws just that chain, which is the version that fits in an issue
or a message:
Nothing picked the subject by hand — --to was not passed. The card carries its own legend because
a dashed line travelling without its trace would otherwise launder a guess into a fact, and it
prints the command that reproduces it.
SVG renders in a GitHub issue and almost nowhere else that matters — X will not take it as an
upload, and Slack and Discord give it no preview — so name the file .png and you get one, or
.gif and the chain builds a hop at a time. That path needs a browser, and orca does not depend on
one: docs/media/README.md keeps the render toolchain out of package.json so nobody running
npm ci pays for a Chromium download, and a picture command is not a reason to reverse that. Ask
for a raster without it and orca says the one line that fixes it; orca doctor reports it either
way, and .svg never needs anything.
Now reproduce it as often as you like, for nothing:
No network, no tokens, no variance. Then ask the question you actually have — would a different model have got this right?
Both pass. One costs 15× less. Same files, same conversation prefix, same checkpoint — the model is the only thing that changed, which is the only reason that number means anything.
The timeline
orca replay last --ui (or orca ui) opens the run as one self-contained HTML file — no server
to keep running, no network, nothing to install. Filter it, step it, or press space and watch the
run play back at the pace it actually happened.
Every layer lands in the same timeline, so you can read the run as one story rather than four: the model turns and their token counts, each tool call with its arguments and result, the shell commands with their exit codes and timing, and the filesystem changes with the tree they produced.
orca export last -o bug.html writes exactly that page to a single file you can attach to an
issue. It carries no external reference of any kind — CI asserts that — so it renders from a
download folder, on a plane, in five years.
Same task, different model
orca compare forks one recorded run onto several models from the same checkpoint, with the same
files and the same conversation prefix, and grades each one with a command you choose. The model
is the only variable, which is what makes the answer mean anything.
Pointing it at several models
Comparing models means reaching several providers, and doing that by hand means knowing that
--upstream-anthropic and --upstream-openai exist, that one gateway can serve both wire formats,
and where the key goes. All of that is real and none of it is discoverable, so there is a command
that asks instead:
orca setup asks the gateway what it actually serves rather than just writing the file, so a wrong
URL or a dead key is an answer now instead of a 401 in the middle of a comparison. It also stores the
models you picked, so after that orca compare last --verify "npm test" needs no model list and no
upstream flags at all. orca models prices what it recognises and shows a
dash for what it does not, because inventing a number for an unknown model is how a comparison
table ends up quoting a cost that was never real.
OrcaRouter is the default answer to that first question — press
Enter and you have one origin and one key serving Claude, GPT, Gemini, Grok, DeepSeek, Qwen and the
rest, which is exactly the shape orca compare wants. Its model ids are namespaced by provider
(anthropic/claude-sonnet-4.6, openai/gpt-4o-mini), which orca handles: the namespace picks the
wire format and is stripped before pricing.
It is a default, not a destination: type over it, or pass --gateway <url>, and anything that
speaks the OpenAI-compatible /v1/models and chat endpoints works just as well — another hosted
gateway, or something you run yourself.
It is also only ever a default for traffic you asked to send somewhere. With no gateway
configured, orca record proxies your agent's own calls straight to whatever provider it was
already talking to, on the agent's own key. Orca does not reroute a recording you never configured:
that would post your source code to a third party as a side effect of pressing record.
Non-interactive: orca setup --key <k> takes the default, orca setup --gateway <url> --key <k>
names another, and --key-env <VAR> reads the key from the environment rather than keeping a
credential on disk.
The key never reaches a trace. It is attached to the outbound request only, while what gets recorded is built from the incoming request with auth stripped — so it is invisible to the recording by construction, not by a rule someone has to remember. It is withheld entirely if a flag sends that traffic somewhere other than the gateway that issued it.
The other half of the run: orca push and orca pull
A local recording and a gateway recording see different things, and neither is the whole picture.
orca record sits on your machine, so it sees the shell, the filesystem and MCP — the side effects.
An OrcaRouter gateway sits on the wire, so it sees what no laptop can: every key, every colleague,
every CI job, and the routing decision behind each call. Both write the same orca-trace v0, so the
interesting direction was never one or the other. It was moving a run between them.
Until now that round trip was a browser download and a form upload. Three commands:
Real output. The card is generated by scripts/render-sync-card.mjs from
docs/media/sync-transcript.txt, a captured round trip — orca record, then push, then pull into a
different workspace, then orca show reading the pulled copy. No line is written by hand. The
gateway there is a local stub speaking the same two endpoints, which is why its address is a
loopback one.
push defaults to the last run, like every other run-taking command. pull will not: "last" means
nothing on a machine that has not seen the run yet, so it takes an id.
Which is what orca list --remote is for. An id you do not have is not a default you can supply,
and the only place to read one used to be the console in a browser — so the command that needs an
id lived on a tab the terminal could not open. --remote asks the gateway the same question
orca list asks this directory, over the GET on the path push already POSTs to:
LAYERS is the column to read. It is what the two halves of this page are about: model and
route come from the gateway, which sees the routing decision no laptop can; fs, shell, mcp
and net come from a run recorded here. It says what is in a run before you spend a download on
it. SOURCE sits beside it because where a run came from is a different question from what it holds.
STARTED is the run's first event, not the moment the gateway stored it — for a pushed run those
differ by however long you waited before pushing. OUTCOME is the gateway's word for a run it
recorded (end_turn, error) and the exit for one pushed from here. Anything the gateway does not
report reads — rather than blank; an older deployment sends no timestamp at all, and STARTED is
left unset rather than dated 1970. --source gateway narrows to what the gateway recorded;
--source upload to what was pushed to it — those two, and a typo is refused rather than answered
with an empty list; without the flag you get both. --limit sets how many (20 by default). --gateway
stays what it is for push and pull — an optional override of which host — rather than doubling
as the switch, so a destination you have already named needs no url typed at it. All three belong
to --remote: given without it they are refused, not ignored, because orca list --gateway <url>
answering with this directory's runs would read as that host's answer.
--json asks the same question: orca list --remote --json emits the gateway's runs as one
document on stdout, which is where the id in orca pull <run> wants to come from when nobody
is reading a table.
What it needs
A key with the replay scope, from Console → API Keys → replay. That scope can read and push
Sessions runs and cannot relay, so it is not a model-traffic credential wearing another hat — a
replay key is refused on the relay routes outright. Uploads are also off per workspace by default
and turned on under Settings → request logs; a gateway that is only recording, not accepting,
is the normal state.
Both exports persist for the shell session, which is the point of the rule above: a later
orca push --gateway <somewhere-else> in that same shell is refused rather than quietly handed the
key you exported for your own gateway.
What it will not do
- Never a default destination.
orca setupmay default the model gateway to OrcaRouter, because proxying a call your agent was already making is not a disclosure. A run is: it holds source, shell output and workspace snapshots, which travel with it as content-addressed blobs. So push has no default host. A gateway counts as a destination only when you NAMED it —--gateway,ORCA_GATEWAY_URL, or a URL you typed atorca setup— and the oneorca setupfills in for you when you press Enter does not count, however well it serves your model traffic. Running plainorca setupand thenorca push lastis refused, and says why.push.packedalso reports the file count and byte size before the request goes out rather than after. - Never the workspace snapshots, unless you ask.
orca scrubrewritesevents.jsonland the blobs, but it cannot rewrite the shadow git store underfs/— its objects are deflated and addressed by the hash of their contents, so it searches them, tells you what is still in there, and offers--drop-fs. Push therefore leavesfs/at home by default: a scrub that reports "objects still hold the material" must not be followed by a push that ships them anyway.--fssends them for the case that needs them — a teammate reproducing against the tree — and says so on the way out, because that is the one part of the payload the scan on either end cannot read (the objects are compressed, so a plaintext scan sees nothing). - Never a key to a host it was not set up for. Every credential has a home — the stored key's
is
orca setup's gateway, an exported key's isORCA_GATEWAY_URL, and a key exported with no URL beside it has no home at all — and it is attached only when the destination matches that home.--gatewaychanges where the run goes, not what the key was issued for, so pointing it elsewhere refuses the push rather than authenticating it with someone else's credential. The one case left through is a key whose only destination is the one this invocation names: nothing earlier associated it with a host, which is the CI shape. - Never anonymously. None of the three is attempted without a key. An unauthenticated request
is exactly what a misconfigured public endpoint answers, and a
200is a poor way to learn either that your run went somewhere with no owner, or that what a listing showed you was never scoped to you. - Never a partial replace.
pull --forcewrites the new copy beside the old one and swaps by rename, so a crash leaves a complete run either way, and the next pull finishes or rolls back whatever the last one left behind.--forceasks to replace a recording, which is not a licence to destroy it and then fail.
What the gateway does on arrival
It re-computes the integrity root over events.jsonl and refuses a push whose root does not match
its manifest — which is why push sends the bytes already on disk and converts nothing. It also
scans the archive for plaintext secrets and refuses it on a hit. That refusal is the point of
--force, and why --force is opt-in: the default is what stops a key reaching a shared server.
Redaction itself is not re-run here. TraceWriter puts every event through spill, redact,
validate, append on its way to disk precisely so no caller can bypass it, so the bytes push reads
are already redacted; re-scrubbing them would rewrite the bytes the integrity root covers, which is
the one thing that makes an archive unpushable. The gateway's scan is a second, independent line,
not a repeat of the first.
Which agents
Two things decide whether a harness can be recorded: whether it can be pointed at the proxy, and whether orca understands the wire format it speaks once it arrives.
| Agent | How it is captured | State |
|---|---|---|
| Claude Code | ANTHROPIC_BASE_URL | works — validated against a real bug fix, in detail |
| Codex CLI (API key) | OPENAI_BASE_URL → Responses API | works |
| Codex CLI (ChatGPT login) | --tls-intercept → Responses API | works, with a decision to make |
| OpenAI Agents SDK | OPENAI_BASE_URL → Responses API | works — the SDK itself, on the Responses API it defaults to, records and replays at exact=1, in CI; its own tracing is a second egress |
| Vercel AI SDK | OPENAI_BASE_URL via @ai-sdk/openai (since 2.0.41) — orca record generic-openai --; fetch hook (orca record node --) when the origin is compiled in | works — env-aware SDK clients via generic-openai at exact=1, in CI; an agent posting to an origin compiled into its source records and replays at exact=1 |
| grok-cli (and its Telegram bot) | orca record grok — GROK_BASE_URL, plus the hook for its sub-agents | works |
| OpenClaw | orca record openclaw — the hook for the gateway, inherited variables for the agents it spawns | works |
| opencode | orca record opencode | adapter shipped, both origins redirected |
| goose (Block) | orca record goose — OPENAI_HOST and OPENAI_BASE_URL, ANTHROPIC_HOST → Responses API | works — driven end to end against goose 1.49.0, what is different about it |
| LangGraph / LangChain | OPENAI_BASE_URL, ANTHROPIC_BASE_URL | works — a two-node graph, streaming and with a tool, records and replays at exact=2 and forks live, in CI; pip install orcareplay-langgraph also records the graph — which node ran, in which superstep, including the ones that call no model |
| Haystack | orca record generic-openai -- python your_pipeline.py — its OpenAIChatGenerator leaves base_url to the OpenAI SDK, which reads OPENAI_BASE_URL | works — a Pipeline with ChatPromptBuilder and OpenAIChatGenerator records and replays at exact=1, in CI |
| OpenHands | orca record generic-openai -- python your_agent.py — its SDK wraps LiteLLM and reads OPENAI_API_BASE | works — the SDK's own LLM layer records and replays at exact=1, in CI |
| CrewAI | orca record generic-openai -- python your_crew.py — since 1.x its own provider, reading OPENAI_API_BASE and OPENAI_BASE_URL | works — a real Agent, Task and Crew records and replays at exact=1, in CI; what 1.x changed |
| Aider | orca record generic-openai -- python your_agent.py — routes through LiteLLM, which reads OPENAI_API_BASE | works — the LiteLLM layer records and replays at exact=1, in CI |
| browser-use | orca record generic-openai -- python your_task.py — its ChatOpenAI passes an unset base_url straight through | works — records and replays at exact=1, in CI; LLM layer only, the browser is not driven |
| Hermes (Nous Research) | ORCA_BASE_URL_VARS=… orca record generic-openai -- hermes … | should work — it overrides per provider; name the variable |
| IndexRAG | orca record indexrag -- --data-dir dataset/<name>/documents — the whole pipeline as one run: OPENAI_BASE_URL for chat, INDEXRAG_EMBEDDING_BASE_URL for a second origin | works — driven end to end against the real pipeline: indexing, embeddings and answering replay offline at exact=11/11 retrieval=3/3, and a fork keeps the recorded index, what is different about a pipeline |
| a RAG pipeline (LlamaIndex, GraphRAG, LightRAG) | orca record generic-openai -- <cmd> — embeddings and rerank replay without a dialect | works — a concurrent index build plus an embedding batch records and replays at exact=5 retrieval=2, in CI |
| Codex-in-the-IDE | orca record exec --tls-intercept -- code . | works — the extension spawns the agent, and it inherits the capture |
| a bot with a hardcoded origin | orca record exec --tls-intercept -- <cmd> | works — a Grok bot posting to a URL in its own source, in detail |
| an agent in a sandbox or on another machine | orca attach | works — orca is reachable and prints what to export, in detail |
| anything else | orca record generic-openai -- <cmd> | works if it reads a base-URL variable; orca record node -- <cmd> if it does not |
Claude Code, Hermes and goose have been driven end to end against the real harness, and each of them broke something. Claude Code broke four things; Hermes found a streaming exchange the proxy was dropping entirely; goose broke two more, and neither was in the adapter — one was the replay matcher, one was a run reporting success over a trace of nothing but errors. The rest are held to the adapter contract and to fixtures that record the exact variables each one sets, so a harness that renames the variable it reads turns a check red instead of producing an empty trace.
The harness that reads different variables
Every other OpenAI-shaped client in this table reads OPENAI_BASE_URL. goose reads it too, but it
reads OPENAI_HOST first — and for Anthropic it reads ANTHROPIC_HOST and nothing else.
That combination is worse than it sounds, because both halves fail silently:
- Setting only
OPENAI_BASE_URLworks right up until the user already hasOPENAI_HOSTexported for something else. Then their value wins, the run goes to their origin, and orca prints a clean recording over an empty trace. ANTHROPIC_BASE_URL— the variablegeneric-openaisets, and the one the rest of the ecosystem reads — does nothing at all here. A sink addressed only by it receives no request.
So the adapter sets OPENAI_HOST, OPENAI_BASE_URL and ANTHROPIC_HOST, all at the same proxy,
and the fixture pins all three. The traffic that arrives is the Responses API — POST /v1/responses
plus one GET /v1/models on start-up — which the proxy's openai-responses dialect already claims.
GOOSE_PROVIDER and GOOSE_MODEL are passed through, never invented: goose has no default for a
custom endpoint, and choosing one for you would launch a different agent than goose does.
Two limits worth knowing before you rely on it. goose runs its shell without going through orca's
PATH shim, so shell tool calls are in the trace with their output but without the real exit code,
duration or stdout/stderr split — record with --no-shell to claim nothing rather than that. And
goose asks a second model for a session title while the conversation is already under way, so
that call and the first real turn race; the replay matcher handles them arriving in either order,
which it did not before goose was the first harness to do this.
A gateway that launches the coding agent
OpenClaw does not do the coding: it runs Claude Code, Codex or opencode as child processes and
drives them from a chat app, so one run carries two kinds of model traffic. The gateway's own calls
are caught by the fetch hook. The coding agent's calls are caught by the ordinary variables — not
because OpenClaw reads them, but because a child process inherits its parent's environment, so
the Claude Code it spawns sees ANTHROPIC_BASE_URL exactly as it would if you had run it yourself.
That inheritance is a property of the operating system rather than of orca, which is the kind of thing that stays obviously true right up until some layer in between sanitises the environment. So it has a test: a gateway fixture that makes no model call of its own, spawns an agent that does, and is recorded and replayed offline through the grandchild's traffic.
An agent that reads nothing at all
A bot with https://api.x.ai/v1 typed into its source reads no variable, so nothing can be
redirected. It is also, by some distance, the most common shape of agent people write. Capture it
below the agent instead, at the socket:
exec (or any) launches your command and sets nothing — no origin and no credential — so the bot
still talks to whoever it was already talking to, and orca reads the conversation on the way past.
api.x.ai is on the default decrypt list, so a Grok bot needs no host named.
The same command reaches an agent orca did not launch, as long as something orca did launch is its ancestor. That is the shape of Codex-in-the-IDE: the extension spawns the agent as a child, so recording the editor captures the agent underneath it.
Replay needs no flags repeated. A run recorded through interception writes down which hosts it
decrypted, and orca replay reads that back and re-establishes the same policy — --no-tls-intercept
if you would rather it did not.
To decrypt an endpoint the default list does not cover, add it rather than replacing the list:
A plain --tls-hosts a,b still means "decrypt exactly these", as it always has. Mixing the two
forms is refused rather than guessed at.
An agent that is not on this machine
An agent in a dev container, on a VPS, or in someone's CI cannot be launched by orca, so there is
no environment for it to build. orca attach turns it around: orca holds the proxy open, prints
the block to paste on the far side, and records whatever arrives until you stop it.
The variables come from the same adapters orca record uses, so a sandbox recording cannot drift
from a local one. --advertise is required when you bind a wildcard: 0.0.0.0 is a statement about
listening, not an address anything can dial, and orca refuses to print a URL that cannot work.
Replaying such a run has the same problem in reverse — the agent may not exist on this machine — so replay attaches too:
It serves the recording with egress blocked and takes the harness from the recording itself. Point the same remote agent at it and the run happens again, offline, with no model called.
A base-URL variable orca has never heard of
Enumerating them is hopeless. Hermes overrides per provider, and its own .env.example carries
NOVITA_BASE_URL, GLM_BASE_URL, KIMI_BASE_URL, MINIMAX_BASE_URL, HF_BASE_URL,
NEBIUS_BASE_URL and a dozen more. A list baked into orca would be stale the week after it was
written, so name the variable instead:
Each name is pointed at the proxy with /v1 appended, which is what an OpenAI-compatible override
wants; =<path> overrides that, and =/ gives the bare origin.
If you record an agent this way and it works, an adapter is about twenty lines —
docs/plugins.md. If it does not, the trace is the most useful thing you can send:
orca export last -o run.html.
When the harness will not be redirected
Base-URL injection captures every harness that reads a base-URL variable, and the fetch hook covers the Node ones that do not. A Codex CLI signed in with a ChatGPT subscription is neither — but not for the reason this section used to give.
It does read a provider base URL, and an explicit one overrides the ChatGPT backend; both paths
even speak the same wire protocol. What is bound to that backend is the credential: a signed-in
subscription carries a plan token the ChatGPT backend accepts and api.openai.com does not. Point
the URL somewhere else and Codex asks for an API key instead — you are no longer recording the
subscription session, you are recording a different one. Codex is Rust, so there is no fetch of
ours to reach either.
--tls-intercept is the answer to that, and it is deliberately a separate decision you have to
make, because it mints a certificate authority.
The CA is unique to the run, trusted only by the agent orca launches — through that child's own environment, never a system or browser trust store — and deleted when the run ends. Orca will not offer to install it anywhere. Hosts outside the allowlist are tunnelled unread and recorded as an address and a byte count, with no path and no body, because orca never held the plaintext. Asking to intercept everything is refused rather than honoured.
What comes back through it is not a log line. An intercepted request is parsed by the same wire dialects as any other, so it lands in the trace as an ordinary exchange — replayable offline and forkable to a different model, on a run that never had an API key of yours in it.
It works on orca replay --model, orca fork and orca compare too, which launch a live agent for
the same reason.
For an agent, a script, or CI
A trace is a file, which is the one thing an observability dashboard cannot be — so the most useful question about a failed run is one an agent can ask: replay my last run and tell me what diverged. Every command answers as data, and orca serves itself over MCP.
One JSON document on stdout, diagnostics on stderr — including the recorded agent's own output, so
the document stays parseable while a run is talking. Failures answer in JSON too, with a non-zero
exit. --json covers list, show, events, checkpoints, graph, record, replay,
compare and doctor.
As tools. orca mcp serves the trace store to an agent over stdio:
orca_list_runs, orca_show_run, orca_checkpoints, orca_graph, orca_replay and
orca_compare. Replay is free and offline; orca_compare says in its own description that it
spends real tokens, because a model choosing a tool reads that string and nothing else — and
orca_graph spends its description saying what recorded and inferred mean, for the same
reason.
From code, if you would rather not shell out:
It never writes to your stdout and never calls process.exit — both asserted, because a library
that does either cannot be built on.
Status
Early. v0 is the walking skeleton of the three commands above. Everything below is exercised by
1,393 tests, the trace-format conformance check and a plugin-API neutrality check, on Node 20 and 22.
| Capability | State |
|---|---|
| Trace format v0 + JSON Schema | working |
| Anthropic / OpenAI-compatible model capture | working |
| OpenAI Responses API capture | working — the format the OpenAI Agents SDK and the Codex CLI default to. Records, replays offline and forks; a fork stays on the wire format the agent speaks |
| Agents that read no base-URL variable | working — orca record node -- <cmd> writes a preload into the run directory and redirects globalThis.fetch for an allowlist of provider hosts. Node and Bun both, since Bun ignores --require in NODE_OPTIONS. Use this when the origin is compiled into source; ordinary @ai-sdk/openai clients that honour OPENAI_BASE_URL go through generic-openai instead |
| A call orca cannot read | working — forwarded rather than refused, and recorded as net.request / net.response: evidence, not a replayable turn. A recording that captured nothing warns instead of exiting clean |
Machine-readable output (--json) | working — one JSON document on stdout, diagnostics on stderr, failures as JSON |
Causal graph (orca graph) | working — what caused what, as a table or as JSON. Every edge says whether the trace recorded it or orca derived it just now, and names the rule either way. --to N narrows to the chain that produced one event |
| Shareable cards | working — orca export --card draws one causal chain, --graph-card draws the whole run with that chain lit, and compare --share draws the verdict table. .svg always; .png and .gif when the optional render toolchain is installed, which orca doctor reports and npm ci never pulls in |
First run without an agent (orca quickstart) | working — the package carries a real recording and the project it was made against, and replays one over the other offline, so the first look costs no key, no network and no tokens |
MCP server (orca mcp) | working — six tools over stdio, so an agent can read, explain and replay its own runs |
Programmatic API (Orca) | working — the commands render what it returns, so the terminal is a view of one source of truth |
| Replaying a session you typed into | working, and approximate — what that means |
| Exact replay with divergence reporting | working — restores the recorded filesystem over your working tree, then puts it back; --worktree for a scratch copy, --in-place to restore nothing. Writes a run of its own recording what the replay discovered — divergences, unmatched requests — and points at the parent for what it merely repeated; --no-trace to skip |
| Fork replay from a checkpoint | working — a fork records its own filesystem snapshots, so it is a run you can fork again |
| Compare across models | working — orca setup stores a gateway (OrcaRouter by default, any URL you name otherwise), key and model list, so orca compare needs no flags |
| Filesystem snapshots and diffs | working |
| Single-file HTML export | working |
| MCP call recording | working — opt in with --mcp-config <path>. Replay and fork re-instrument from the config the recording used, so the layer does not stop at the fork point |
Post-hoc scrubbing (orca scrub) | working |
Shell capture (PATH shim) | working — exit codes, duration and the stdout/stderr split. --no-shell to skip |
| Non-model network capture | working — opt in with --tls-intercept; mints a per-run CA the launched agent alone trusts, decrypts an allowlist of hosts, tunnels the rest unread, and deletes the key when the run ends |
| Codex subscription model capture/replay | working — recognizes the /backend-api/codex/responses HTTPS fallback, decodes zstd request bodies for matching, and serves the recorded SSE response without opening the origin during replay |
| Validated against a real agent | Claude Code, recording a real fix to a real bug: recorded, replayed offline end to end, forked from a checkpoint and exported. It broke four things no fixture could have produced, all since fixed — what a real agent found |
| Subscription-auth harnesses | Claude Code works. A Codex CLI signed in with a ChatGPT subscription carries a plan credential only its own backend accepts, so redirecting the URL would record a different session rather than that one: it needs --tls-intercept. With an API key it needs nothing special |
Replaying a session you typed into
A run started as orca record claude and driven by hand has its prompts nowhere on the wire. Orca
recovers them from the harness's own transcript and replays the session by handing them back
without a terminal — which reproduces the conversation, and does not reproduce the run byte for
byte. Two differences are inherent rather than defects, and orca names both rather than papering
over them:
- The harness makes calls for itself. A quota probe before the first turn, a request to name
the session, a delegation prompt written fresh for a sub-agent. A replay does not repeat them.
The matcher steps over them onto the next real match and reports how many, so
reused=3/5on an interactive recording is a complete replay of what you asked, not a partial one. - A terminal session is offered tools a replay cannot have.
AskUserQuestion,EnterPlanMode,ExitPlanMode,EndConversationall need a person in front of them, so they are absent when the same agent is driven without one. Their schemas are large, so a request carrying them can differ from the recorded one by enough to halt; the halt says which tools are missing and why.
Recording with a prompt in argv — orca record claude -- -p "…" — has neither problem, because the
replay is started exactly the way the recording was. Use that for a run you intend to replay
exactly. Reading a run is unaffected either way: orca show and the viewer are complete for both.
Install
Node 20 or newer. No native dependencies, so there is nothing to compile and nothing fetched at install time beyond the tarballs themselves.
Every release from 0.1.1 onward is published by the tagged workflow in
RELEASING.md, with a provenance attestation naming the commit and the run that
built it — npm shows it on the package page, and npm audit signatures checks it.
From source, to work on it or to run an unreleased commit:
npm install -g . from the repository root installs nothing: the root is a workspace with no
binary of its own, and orca lives in packages/cli.
Node 20+ to run it (the CLI's own engines says >=20.0.0). Contributing needs `^20.19.0 ||
=22.12.0
, because the test toolchain does; the rootpackage.jsondeclares that separately sonpm ci` tells you up front. No account, no signup, no API key changes.
On Windows, shell capture writes .cmd shims in front of sh.exe and bash.exe (Git for Windows
provides both), and Windows runs a .cmd only through a shell. So it captures a shell started
through cmd.exe or PowerShell, and not one an agent starts itself: a Node harness's
child_process.spawn('bash') goes straight to the real bash.exe, and Claude Code finds its shell
without consulting PATH at all. orca doctor says which of the two holds on the machine, and
orca record warns shell.ineffective on any run whose tool calls ran commands the shim never
saw — record those with --no-shell. If no POSIX shell is on PATH, doctor warns and
--no-shell is the way to record.
Where your runs are kept
Everything lands in .orca/runs/ inside the project you recorded in — per-project, never a
global store, so a run travels with the checkout it belongs to. One run directory is one
self-describing thing:
Finding an old session:
orca list reads the run directories directly, so it works on a trace someone sent you: drop it in
.orca/runs/ and every command sees it. Nothing indexes, and there is no database to corrupt.
--remote is the one exception, and it is not an index either: it asks the gateway and prints the
answer, holding nothing.
Privacy
Traces are local, mode 0600, and the recorder makes no network connection of its own. Secrets are
redacted in the write path: environment capture is deny-by-default, auth headers are never written,
and known key shapes plus high-entropy strings are replaced with stable placeholders.
Redaction is best-effort mitigation, not a guarantee. Treat a trace as sensitive — roughly as sensitive as a shell history plus a heap dump.
orca scrub rewrites events.jsonl, the manifest and every text blob, re-runs the standard
detectors, refreshes the integrity digest, and leaves binary blobs byte-identical.
It cannot rewrite the filesystem snapshots. Git objects are addressed by the hash of their own
contents, so editing one changes its id, which forces every tree naming it to be rewritten and
every event naming those trees after that — a history rewrite whose failure mode is a run that no
longer restores. So scrub searches the snapshot store and tells you when your string is still in
there, rather than reporting a clean trace it could not clean. --drop-fs deletes the store
outright, at the cost of being able to fork the run.
What is open, and what is not
Always open, under Apache-2.0: the trace format, the core, the CLI, the viewer, the adapters, and the provider interface.
OrcaReplay is built by the people who build OrcaRouter, and that shows
up in two places, both of them things you asked for. orca setup suggests it when you do not name a
gateway — a default you can see and overtype, on a question you chose to answer, not a route
anything takes on its own. And an artefact you explicitly generate — an export, a --share card —
signs itself "built by the OrcaRouter.ai team", the way a chart carries its source.
Every model path stays a plain URL you can point anywhere, there is no code path that treats that origin differently from any other, and a credit line routes nothing anywhere.
What the vendor does not get is privilege. A plugin — OrcaRouter's included — may use only the
public Provider interface in @orcareplay/plugin-api, with no private API behind it. No vendor
plugin exists yet, so the CI job that enforces this (scripts/check-neutrality.mjs) says so and
passes as a no-op; it starts building against the published package rather than workspace source the
moment one lands. If a plugin ever needs a capability, that capability goes into the public
interface first, with a second implementation showing it is not shaped around one vendor.
Documentation
Start here if you have a problem right now:
- My agent broke something. How do I find out why?
- Why did my agent delete that file?
- Would a different model have got this right?
Reference:
spec/orca-trace-v0.md— the normative trace formatdocs/integrations.md— recording a framework, one command each, every number asserted in CIdocs/architecture.md— how capture, replay and fork actually workdocs/validation.md— what broke the first time this met a real agentdocs/launch-path.md— what is built, what is not, and what is nextdocs/plugins.md— writing an adapter or a providerCONTRIBUTING.md— five-minute dev loop- Good first issues — twelve of them, with the file to start in
Help wanted
The format is v0 and the walking skeleton works, which is the interesting point in a project's life: the decisions are still cheap to change and there is a lot of obvious work with the file to start in already written down.
- Twelve good first issues, each naming the file and the test.
- Write an adapter. One file, one fixture. If your harness reads a base-URL variable it is
about twenty lines — docs/plugins.md. If it does not,
nodemay already cover it; a recording that comes back empty from a harness not listed above is worth an issue either way. - Add a framework to the integration checks.
test/integrations/records a real framework against a stub origin, kills the origin and replays it. Eight are covered; AutoGen, LlamaIndex, Pydantic AI and smolagents are not. A new one is a script and a row. - Reimplement the reader. The spec is CC BY 4.0 on purpose. There is already a Python reader; Go and Rust are open.
- Break the replay. The matching ladder is the heart of this and the fastest way to improve it
is a real recording it gets wrong. Open an issue with
orca export last -o bug.htmlattached — it is one self-contained file, andorca scrubis there for anything you need out of it first.
If it saved you an afternoon, a ⭐ helps other people find it.
License
Apache-2.0 for the code. The trace specification is CC BY 4.0, so anyone may reimplement it.
Built by the OrcaRouter team ·
orcarouter.ai ·
all models ·
OrcaCode Review ·
X ·
Hugging Face