feat(nextedit): add /api/nextedit/completions endpoint

[kilo-code-bot]

Summary

Adds a new proxy endpoint at /api/nextedit/completions to support next-edit (NES) requests, complementing the existing /api/fim/completions route. For now only Inception (Mercury Edit) is supported — the same provider used by Kilo-Org/kilocode#10536.

The route mirrors the structure of /api/fim/completions (auth → balance/org checks → BYOK lookup → upstream proxy → usage logging), but speaks the chat-style messages shape that Inception's /v1/edit/completions endpoint expects. Only models prefixed inception/ are accepted; anything else returns unsupported_nextedit_model.

Changes

• apps/web/src/app/api/nextedit/completions/route.ts — new POST handler. Validates the request, resolves provider from the inception/ model prefix, enforces org policy (only / ignore), proxies to https://api.inceptionlabs.ai/v1/edit/completions, and reports usage.
• apps/web/src/lib/ai-gateway/llm-proxy-helpers.ts — adds extractNextEditPromptInfo and countAndStoreNextEditUsage (reuses the Inception FIM per-token cost formula since it's the same model family).
• apps/web/src/lib/proxy-error-types.ts — adds unsupported_nextedit_model.
• packages/db/src/schema-types.ts — adds nextedit_completions to GatewayApiKindSchema so usage rows can be tagged distinctly from fim_completions.

Notes

• /v1/edit/completions only accepts a single role: "user" message (Mercury bakes its system prompt server-side and returns 400 on role: "system"), so the schema enforces that shape implicitly via messages: Array(min(1)) and the route forwards messages verbatim.
• Streaming is not supported by the upstream endpoint today; the schema rejects stream: true.
• Pricing reuses computeInceptionFimMicrodollarCost (0.25/0.75 microdollars per input/output token) because Mercury Edit 2 sits in the same Inception family as the existing FIM model.

Verification

• Manual review of the route against the existing /api/fim/completions route to ensure the auth/balance/org-policy/BYOK flow is identical, with only the request/response parsing and upstream URL swapped.

Visual Changes

N/A — backend-only.

Reviewer Notes

• Full repo pnpm typecheck was skipped (dependencies not installed in this sandbox; full typecheck takes long enough to disrupt the loop). CI will run it.
• The new nextedit_completions enum value is additive; existing consumers of GatewayApiKind accept it as a string without further changes.
• Future work: route through the Kilo gateway instead of calling Inception directly, and add martian/other providers once they expose an equivalent endpoint.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

Both previously flagged warnings have been resolved: max_tokens is now validated as a positive integer at the Zod schema boundary, and countAndStoreNextEditUsage is now guarded with a .catch() that reports to Sentry and returns null cleanly.

Resolved Issues (previously flagged)

File	Issue	Status
apps/web/src/lib/ai-gateway/llm-proxy-helpers.ts	parseNextEditUsageFromString unguarded rejection inside after()	✅ Fixed — .catch() with captureException added
apps/web/src/app/api/nextedit/completions/route.ts	Negative max_tokens bypassing security guard	✅ Fixed — schema changed to z.number().int().positive().optional()

Files Reviewed (4 files)

• apps/web/src/app/api/nextedit/completions/route.ts — no issues
• apps/web/src/lib/ai-gateway/llm-proxy-helpers.ts — no issues
• apps/web/src/lib/proxy-error-types.ts — no issues
• packages/db/src/schema-types.ts — no issues

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-4.6-sonnet-20260217 · 227,578 tokens}

_{Review guidance: REVIEW.md from base branch main}