Implement RunPod cloud adapter

[caydyan]

Summary

Replaces the RunPod cloud stub with a GraphQL-backed GPU pod adapter:

• verifies credentials with myself
• quotes GPU pods from explicit hourlyPrice config or RunPod gpuTypes pricing
• provisions on-demand pods through podFindAndDeployOnDemand
• lists account pods via myself { pods }
• checks a single pod via pod(input: ...)
• terminates pods with podTerminate
• maps RunPod pod state and public runtime port IPs into sh1pt Instance values
• documents required config, including imageName for real provisioning

RunPod reference used:

• https://docs.runpod.io/sdks/graphql/manage-pods
• https://graphql-spec.runpod.io/

Safety

Real provisioning requires config.imageName. GPU pods still honor spec.maxHourlyPrice before any mutation is sent. Dry-run provisioning returns a local instance and does not call RunPod.

Validation

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts
• git diff --check

[greptile-apps]

Greptile Summary

This PR replaces the RunPod cloud stub with a complete GraphQL-backed GPU pod adapter. It correctly addresses every concern raised in prior review rounds: null guards on myself and podFindAndDeployOnDemand, dry-run path that avoids network calls, conservative on-demand pricing for maxHourlyPrice guardrails, volumeInGb opt-in via optionalPositiveNumber, and gpuTypes ?? [] null coalescing.

• Core operations: connect verifies credentials via myself, quote fetches gpuTypes pricing or uses an explicit hourlyPrice, provision calls podFindAndDeployOnDemand with a full null guard, list/status/destroy map cleanly to RunPod's GraphQL API.
• Price guardrail: priceForGpu uses Math.max(communityPrice, securePrice) for cloudType: ALL, so the guardrail is always checked against the highest possible on-demand rate, not the cheaper community estimate.
• Test coverage: 15 focused unit tests covering every code path including null API responses, dry-run, maxHourlyPrice enforcement, and spot pricing, plus the contractTestCloud suite.

Confidence Score: 5/5

Safe to merge; every previously identified issue has been addressed with tests to confirm.

Every concern from the previous review rounds has been resolved: null guards on myself and podFindAndDeployOnDemand, dry-run path that skips all network calls when no hourlyPrice is set, conservative pricing with Math.max for the ALL cloud type, volumeInGb now opt-in via optionalPositiveNumber, and data.gpuTypes ?? [] preventing a null dereference. No new defects were found in this pass.

No files require special attention; all changed files look correct.

Important Files Changed

Filename	Overview
packages/cloud/runpod/src/index.ts	Full GraphQL adapter implementation replacing the stub; all previously identified null-safety, dry-run, pricing, and provisioning issues are addressed. No new defects found.
packages/cloud/runpod/src/index.test.ts	Comprehensive unit test suite covering all adapter methods, null RunPod responses, dry-run, maxHourlyPrice guardrail, and spot pricing behavior; all previously identified edge cases have dedicated tests.
packages/cloud/runpod/README.md	Updated from stub placeholder to accurate documentation of all config fields, pricing behavior, and provisioning requirements.

Sequence Diagram

sequenceDiagram
    participant C as sh1pt CLI
    participant A as RunPod Adapter
    participant R as RunPod GraphQL API

    C->>A: connect(ctx, config)
    A->>R: "query Myself { myself { id email } }"
    R-->>A: "{ myself: { id, email } }"
    A-->>C: "{ accountId }"

    C->>A: quote(ctx, spec, config)
    alt hourlyPrice in config
        A-->>C: Quote (no network call)
    else
        A->>R: "query GpuTypes(input: { id: gpuTypeId })"
        R-->>A: "{ gpuTypes: [...] }"
        A-->>C: Quote (on-demand price x count)
    end

    C->>A: provision(ctx, spec, config)
    alt dryRun and no hourlyPrice
        A-->>C: synthetic dry-run Instance (no network call)
    else
        A->>A: quote() check maxHourlyPrice
        A->>R: "mutation DeployPod(input: {...})"
        R-->>A: "{ podFindAndDeployOnDemand: pod | null }"
        alt pod is null
            A-->>C: throw pod was not provisioned
        else
            A-->>C: Instance (mapped from pod)
        end
    end

    C->>A: list(ctx, config)
    A->>R: "query Pods { myself { pods { ... } } }"
    R-->>A: "{ myself: { pods: [...] } }"
    A-->>C: Instance[]

    C->>A: status(ctx, instanceId, config)
    A->>R: "query Pod(input: { podId })"
    R-->>A: "{ pod: pod | null }"
    A-->>C: Instance

    C->>A: destroy(ctx, instanceId, config)
    alt dryRun
        A-->>C: log only
    else
        A->>R: "mutation TerminatePod(input: { podId })"
        R-->>A: "{ podTerminate: null }"
        A-->>C: void
    end

Loading

_{Reviews (5): Last reviewed commit: "Guard empty RunPod provision response" | Re-trigger Greptile}

[caydyan]

Addressed the Greptile feedback in 061285b0:

• Guard myself in connect()/list() and return a clear RunPod account-permission error instead of a TypeError.
• Keep dry-run provisioning offline when hourlyPrice is omitted; it now returns a local zero-cost preview without calling RunPod.
• Quote cloudType: ALL with the highest available community/secure price so maxHourlyPrice is not checked against a lower community-only estimate.

Validation after the fix:

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts (19 tests)
• git diff --check

[caydyan]

Follow-up pushed in 01d61381 to cover the remaining storage-cost concern from the Greptile inline review:

• volumeInGb is now omitted unless explicitly requested through config or spec.storage, so RunPod will not attach default network storage on ordinary pod launches.
• Added coverage that real provisioning without storage does not send volumeInGb.
• README now calls out that volumeInGb is opt-in network volume storage.

Validation after this follow-up:

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts (20 tests)
• git diff --check

[caydyan]

Addressed the latest Greptile follow-up in 2013f9a7:

• quoteFromApi() now treats missing/null gpuTypes as an empty result and surfaces the existing clear RunPod GPU type not found error.
• selectGpuType() no longer silently falls back to the first returned GPU when the requested model does not match.
• Added tests for gpuTypes: null and non-matching GPU result sets.
• README notes that unknown GPU types fail clearly instead of selecting another GPU.

Validation after this fix:

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts (22 tests)
• git diff --check

[caydyan]

Addressed the spot-pricing guardrail issue in 0477848e:

• RunPod quotes now use on-demand GPU prices for guardrails because this adapter does not submit spot bidPerGpu values.
• Quote.spot is now false for RunPod quotes unless/until the adapter actually provisions spot bids.
• Added coverage that maxHourlyPrice blocks against on-demand price even when spec.spotOk is true.
• README documents that spot bids are not sent by this adapter.

Validation after this fix:

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts (23 tests)
• git diff --check

[caydyan]

Addressed the latest Greptile finding in 5465626d:

• provision() now guards a null/empty podFindAndDeployOnDemand mutation result and returns a clear capacity/provisioning error instead of letting podInstance() throw a TypeError.
• Added regression coverage for a null RunPod provision response.

Validation after this fix:

• corepack pnpm --filter @profullstack/sh1pt-core build
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod typecheck
• corepack pnpm --filter @profullstack/sh1pt-cloud-runpod build
• corepack pnpm vitest run packages/cloud/runpod/src/index.test.ts (24 tests)
• git diff --check