Replay a decision
The selected-step lab compares original and edited context across repeated model responses. Repeat rates describe observed behavior; they do not recover private reasoning.
OBSERVABILITY FOR CODING AGENTS
See what your coding agent did.
Understand what changed.
See what your agent saw, what it proposed, and what actually ran. Then change the context and explore what happens next.
Local-firstOpen sourceMIT licensed
raytrace / session inspectorEXAMPLE
SESSION / 0013 events
“Fix the failing authentication test.”
01
Context received INPUT
src/auth.ts · tests/auth.test.ts
expect(session.expiresAt)
.toBeGreaterThan(Date.now());
.toBeGreaterThan(Date.now());
02
Tool call proposed INTENT
Bash
$ npm test -- auth
03
Process observed EVIDENCE
gVisor exec checkpoint · process exit
npm test -- authexit 0
[ ] A proposal and an observed process are different facts.
THE QUESTIONS BEHIND EVERY TRACE
What did it see? / What did it try? / What actually ran? / What would change?
01 / INSPECT
A trace you can
take apart.
A final answer only tells part of the story. RayTrace connects prompts, context, tool calls, and runtime evidence so you can inspect the steps that led there.
MODEL INPUTauth-fix / step 01
Start with what the model saw.
Inspect the input snapshot for a model request: the prompt, tool definitions, prior results, and file excerpts included in its context.
context / tests/auth.test.ts
// Excerpt included in the model request
test('session has not expired', () => {
expect(session.expiresAt)
.toBeGreaterThan(Date.now());
});A file in the repository is not necessarily a file the model saw.
02 / EXPERIMENT
Change one thing.
Follow a new branch.
Was a piece of context useful? Select a step, edit or remove what the agent saw, and compare the continuation with the original run.
CAPTURED STEPOriginal contextPrompt + tool results + files
BASELINEKeep context unchangedRepeat the same step
FORKEdit context or project filesContinue in a separate sandbox
COMPAREWhat changed?Tool calls · answers · file diffs · checks
Continue from a snapshot
For eligible sandboxed Claude Code sessions, Playground restores the project before a step and runs a fork. Compare file changes, steps, and an optional check command.
Compare against a baseline
Models can vary even when nothing changes. Playground can interleave edited runs with unchanged runs, so you can examine that variation alongside your experiment.
KNOW THE BOUNDARY Forks need a recorded sandbox session with snapshots. Decision replays and Codex continuations have different setup and provider requirements.
03 / UNDER THE HOOD
Local capture.
Connected evidence.
A local proxy records model exchanges. SQLite connects agent intent with runtime evidence. The dashboard gives you a way to inspect both.
Agent / clientOpenAI-compatible traffic
request / response
RayTrace proxy127.0.0.1:8797
LOCALRoutes to OpenAIAnthropicOpenRouter*
captured exchanges
SQLite evidence store.raytace/evidence.db
inspect / experiment
Dashboard + experiment enginelocalhost:3000
* OpenRouter is optional experiment routing. Claude Code sandbox sessions also use managed hooks to capture their transcripts.
Keep the history close.
Trace data stays in a local SQLite database. Content-addressed payloads reuse repeated context instead of storing the same history for every request.
Observe beyond the transcript.
A gVisor sandbox inside a Lima Linux VM independently collects process-start and exit events. The forwarder links them to tool calls using markers or command and timing evidence.
Observed execution ≠ correctness.
A process starting or exiting successfully does not prove the task was done correctly. Missing evidence means unknown. RayTrace does not reconstruct private chain of thought.
04 / GET STARTED
Your agent.
Your machine.
Your trace.
Start with local request capture. Add sandbox execution when you need independent process evidence.
REQUIRESNode.js 22.13+ · npm · compatible client
Built for experimentation. RayTrace is an early prototype for local debugging and evaluation. Captured prompts and file contents can be sensitive; the local services do not yet provide authentication or encryption at rest.
LOCAL SETUP
- Install in your RayTrace checkout
npm install - Start the capture proxy
npm run proxy - Start the dashboard in another terminal
npm run dev
Add the optional gVisor sandbox
Requires Lima and Docker. From your RayTrace checkout:
npm run sandbox:setup
npm run dev:allFor a Claude Code sandbox session:
npm run dev:all -- --claudeOpenRouter-backed decision experiments also require your own API key and provider configuration in the project’s local environment file.