Punchbacks captures what went wrong in the browser, reproduces it in a sandbox, lets your chosen model repair the code, deploys an isolated preview — and asks the original reporter to confirm the fix.
Your cart
Coffee Mug
₹799
Watch a real report travel the loop — hover to hold a step, click the rail to jump.
The closed loop
Every report travels the same evidence-backed gauntlet. Nothing ships on vibes — each gate must pass before the next one opens.
A user hits “Report a bug”, re-runs the failing interaction, and submits. Punchbacks records the action trace, console, DOM and safe network metadata — masked before it ever leaves the browser.
no account needed
The trace becomes a deterministic Playwright test. Only when it fails against the base revision for the reported reason does the report move forward.
no blind repairs
Your configured model investigates an isolated checkout inside OpenSandbox: deny-by-default egress, resource limits, bounded attempts and cost budgets. It patches the root cause, never the test.
policy-enforced tools
A punchback/<report> branch becomes a draft pull request carrying the root cause, the diff, and the full verification checklist. It is never auto-merged.
human merge authority
Vercel builds an isolated preview of the branch. Punchbacks replays the original interaction against it — the exact clicks that failed must now succeed.
never production
The person who hit the bug gets a tracking link: “Your fix is ready to test.” Their “It’s fixed” is the final gate — the acceptance test is human.
loop closed
The new development loop
The old loop turns a production bug into a ticket. Punchbacks turns the same moment into evidence, a tested change, and a decision-ready pull request.
One continuous system
From user evidence to human-verified code
Report
production truth
Reproduce
failing proof
Repair
bounded agent
Preview
isolated replay
Verify
human sign-off
Observed
Real interaction, not a reconstructed ticket
Contained
Repair runs behind enforceable boundaries
Accountable
A person still decides what ships
Built like a rig, not a demo
Capture, isolation, evidence, and human sign-off — engineered as one pipeline instead of four disconnected tools.
Passwords are never captured. Elements marked data-punchback-private are masked, request bodies stay off by default, and reporters preview exactly what leaves their browser.
masked client-side · v1 redaction
Every repair runs in an ephemeral OpenSandbox container: deny-by-default egress, CPU/memory/TTL limits, credentials vaulted, workspace destroyed after the run.
The generated regression test must fail on the base revision for the reported reason — and pass on the patch. The existing suite stays green too.
Root cause, minimal diff, artifacts and the verification checklist — delivered as a draft pull request your team reviews like any other.
Original failure reproduced
Regression test passes
Existing suite passes
Replay verified on preview
Reporter confirmation → posted to PR
The dashboard streams each run over SSE: actions, console, agent steps and test results on one timeline, with the original rrweb replay beside the fixed one.
Docker Compose brings up the whole control plane. A read-only MCP server exposes reports and repair evidence to your favourite agent client.
$ docker compose up --build
Bring your own intelligence
Connect OpenAI, Anthropic, or OpenCode Go, choose a model per project, and let Punchbacks route every repair through the provider you trust.
Encrypted at rest
Versioned AES-256-GCM encryption keeps credentials write-only.
Decrypted just in time
The worker opens the key only when the selected repair starts.
Never inside the sandbox
Provider credentials stay in the control plane, away from repository code.
OpenAI
credential + model
Anthropic
credential + model
OpenCode Go
credential + model
Encrypted credential vault
selected at project level
repair agentPunchbacks verified
A fix only earns the stamp when technical, behavioural, and human verification all pass. Green CI alone doesn't count.
The original failure is recreated against the base revision.
The new regression test and the existing suite pass on the patch.
The original interaction succeeds against the isolated preview.
The person who hit the bug confirms it is actually gone.
Hard guardrails
Draft PRs only. Merge authority stays with your engineers, always.
Repairs and verification happen on isolated previews and sandboxes.
Attempt caps, wall-clock limits, token budgets, denied paths — enforced by the host, not the model.
If the bug can't be reproduced, it's marked and handed back with evidence — no speculative patches.
Add one script tag to your site. The next bug your users hit comes back as a reproduced, tested, preview-deployed, reporter-verified pull request.
