Managed PR repair inside disposable E2B sandboxes. Critique reads the review artifact, patches in scope, validates the work, and pushes only when the branch is reachable.

Fixes that land clean.

Start Remedy See credits

remedy sandboxchecks armed

patch artifactready

+ guard stale review state before publish

+ add regression coverage for fork fallback

~ rerun lint, test, build in sandbox

validation

pass

Commit is allowed to touch the PR branch only after project checks clear.

A narrower blast radius

Cloud repair without runaway agent behavior.

The generated reference image leaned into a gapless bento: fewer cards, bigger surfaces, and useful proof instead of decorative widgets. This implementation keeps that structure and lets each panel earn its space.

Isolated execution

Every run gets a disposable cloud workspace. Your laptop never becomes the runtime.

Scope is a gate

If the agent reaches outside the review contract, the run stops instead of improvising.

Validation before branch write

Lint, tests, build, and project checks decide whether a patch is allowed to leave the sandbox.

Two-loop ceiling

The run retries once, then hands back a precise artifact instead of burning time forever.

Model choice stays live

Pick economical lanes for simple repairs and frontier reasoning for structural work.

Remedy patch artifact

findingCQ-2087·Critical · Tenant isolation

remedy/fix-2087·Qwen3 Coder Next·18.4s · 1 / 2 loops

Drag across the same handler to see the patch Remedy generated, ran in the sandbox, and only then pushed back to the branch.

Drag the handle

Head @ pre-patch

invites.ts

app/api/workspaces/[workspaceSlug]/invites.tsTS · 5 issues

36

export async function POST(req: NextRequest, { params }: RouteContext) {

37

  const session = await requireSession(req)

38

  const input = await parseInviteBody(req)

@@ -39,1 @@

-

40

  const targetWorkspaceId = input.workspaceId ?? params.workspaceSlug

// Request body can override the authenticated route slug

-

41

  const membership = await requireWorkspaceMember(session.user.id, input.workspaceId)

// Auth check follows the caller-supplied workspace id

-

42

  await listPendingInvites(targetWorkspaceId)

@@ -43,20 @@

63

  const nextSeatCount = computeSeatTotal(input.members)

@@ -64,2 @@

-

66

  await stripe.subscriptionItems.update(itemId, { quantity: nextSeatCount })

// No idempotency fence · retries double-apply the mutation

-

67

  await db.workspace.update({ data: { lastAppliedSeatCount: nextSeatCount } })

@@ -68,2 @@

70

  return NextResponse.json({ ok: true })

71

}

verification failed·3 failed · 2 issues blocked pushpushback · declined

Head @ remedy/fix

invites.ts

app/api/workspaces/[workspaceSlug]/invites.tsTS · 0 issues

36

export async function POST(req: NextRequest, { params }: RouteContext) {

37

  const session = await requireSession(req)

38

  const input = await parseInviteBody(req)

@@ -39,1 @@

+

40

  const workspace = await requireManagedWorkspace(session.user.id, params.workspaceSlug)

// Authority bound to the authenticated route slug

+

41

  const revision = await nextBillingRevision(workspace.id)

+

42

  const operationKey = `workspace:${workspace.id}:seats:${revision}`

// Stable key ties retries to one billing revision

@@ -43,20 @@

63

  const nextSeatCount = computeSeatTotal(input.members)

@@ -64,2 @@

+

66

  await createSeatOperationFence(tx, operationKey)

// Fence persists before the Stripe call

+

67

  await stripe.subscriptionItems.update(itemId, { quantity: nextSeatCount }, { idempotencyKey: operationKey })

// Retry-safe external mutation

@@ -68,2 @@

70

  return NextResponse.json({ ok: true })

71

}

verification passed·47 tests · lint · tsc · buildpushback · ready

Finding source

id: "CQ-2087"
severity: "critical"
invariant: "requireManagedWorkspace(session, slug)"
lines: [40, 41, 66, 67]

Machine-readable finding from the Critique review — Remedy reads the structured record, not a chat comment.

Verification bundle

eslint .1.4s

tsc --noEmit3.2s

vitest run api/workspaces6.1s

next build7.7s

Every step ran inside the e2b sandbox. Nothing touched your branch until the full bundle returned green.

Patch artifact

--- run_8a2f91c3b4e7.patch
diff app/api/.../invites.ts
- 5 lines removed
+ 6 lines added
signed: "remedy-sandbox-7f3b"

Pushed to PR5 cr · 1 loop

run_8a2f91c3b4e7/Pre-patch verification blocked 3 regressions before the patch was authored.sandbox closed · artifact archived

From finding to branch

The fix loop becomes inspectable.

The reference direction uses a pinned narrative: one stable promise on the left, and a stack of artifacts moving on the right. In code, each frame scales in and fades with ScrollTrigger.

Blueprintartifact

Blueprint

The machine-readable review output becomes an execution contract: files, findings, scope, and validation steps.

finding.security[2]scope: app/api/**validation: lint + test

Sandboxartifact

Sandbox

A clean E2B runtime clones the PR head and runs the fix away from your machine and shared host state.

vm: ephemeralrepo: clonedagent: opencode

Patchartifact

Patch

The selected model writes the bounded fix, updates tests where needed, and leaves a reviewable artifact.

diff: preparedtests: expandedfiles: in scope

Pushartifact

Push

Passing runs commit back to the PR branch when the GitHub App can access the head repository.

checks: greencommit: signedbranch: updated

Remedyisnotasuggestionbox.Itisaboundedrepairsystemthatturnsreviewevidenceintocode,validatestheresult,andleavesaclearaudittrailwhenautomationshouldstop.

Execution choice

Use Remedy, or hand off the blueprint.

Teams with existing agent subscriptions can reserve Critique credits for review and queue Cursor, Claude, or Codex from the same blueprint — or paste a Fix prompt into Copilot and Claude Code.

BYOA

Cursor Cloud Agents

Queue from a review run with your Cursor API key — same blueprint, your plan.

BYOA

Claude Managed Agent

Server-side Anthropic key launches a managed session on the PR.

BYOA

OpenAI Codex

Queue OpenAI Responses runs or export JSON for Codex cloud / CLI.

BYOA

Fix prompt & Copilot

Paste the blueprint into Claude Code, Copilot, or any agent workflow.

Model routing

Spend reasoning where the patch deserves it.

Simple lint and nullability fixes do not need the same execution lane as a behavioral security change. Remedy exposes model floors clearly, then prices the full managed run around execution and verification.

“The useful part is not that it writes code. It knows when the fix is outside the contract.”

Model selector

17 lanes

Gemma-4-31B

Small, cheap patches with an open Google lane.

0.5

model floor

Safety cap

2 loops max

Loop onewaiting

Loop twoarmed

Credits

Transparent floors, visible execution overhead.

The model floor is only the starting point. A managed Remedy run also includes sandbox setup, repository clone, validation, git operations, and optional re-review when loop two is needed.

ActionCreditsNotes

Standard /ai reviewReview costLead plus specialist costs apply

/ai security onlySub-agent onlySingle specialist pass

/ai suggest fix (Remedy)2.75+Model floor plus managed execution

Remedy run - low-floor model3+1 model floor plus 2 managed execution

Remedy run - frontier model22+Model floor plus managed execution

Verification re-review+ Review costApplies only if loop two is needed

ModelModel floor

Gemma-4-31B

0.5 cr

MiMo v2.5

0.5 cr

Ling-2.6-Flash

0.5 cr

Selected model

Run outcome

Selected model floor0.5 cr

Managed execution2 cr

Re-review loop0 cr

Estimated total

2.5 cr

Close the loop where review starts.

Connect GitHub, run Critique on a pull request, and let Remedy turn validated findings into a tested branch update.

Connect repo Open Workspace