CU-86exuferc: AI Review Engine — LLM Integration

[miqdadyyy]

Summary

Implements the core AI review engine that sends code diffs to configured LLM providers and returns structured review feedback. Reuses the existing encrypted provider domain for AI credentials and mirrors the established services/vcs factory pattern for services/llm.

ClickUp Ticket

• Ticket: CU-86exuferc: AI Review Engine — LLM Integration
• Priority: high
• Objective: Build the core AI review engine (diff → prompt → LLM → structured review comments).

Changes

LLM service layer (internal/services/llm/)

• ILLMFactory + ILLMService interfaces mirroring services/vcs.
• OpenAI-compatible client (works with any OpenAI proxy: LiteLLM, vLLM, Ollama, Azure OpenAI), Anthropic client, and custom proxy client.
• Shared base.go HTTP helper (Fiber client) with exponential-backoff retry and upstream status mapping.
• Streaming support via StreamComplete.

AI provider integration (reuses provider domain)

• New constants: ProviderCategoryAI, ProviderTypeOpenAI/Anthropic/Custom, ProviderAuthModeAPIKey.
• ProviderAIMetadata (api key / base URL / model) stored via the existing AES-encrypted EncryptedField — keys never stored in plaintext.

AI connector CRUD (internal/usecase/aiconnector/)

• POST/GET/PUT/DELETE /api/v1/connectors.
• API keys are masked in all responses (****<last4>).

Review engine (internal/usecase/review/)

• POST /api/v1/reviews/diff: orchestrates diff → prompt → LLM → parsed ReviewComment list.
• System + user prompt template system; DB-stored, per-repo editable templates (aiprompttemplate repo domain, AutoMigrate).
• Robust JSON extraction from model output; token counting + cost estimation.

Wiring & config

• New repo/usecase/service constructors added to cmd/cli/wire.go (regenerated wire_gen.go).
• LLMConfig added to config entity and config/local.yaml.

Testing

• go build ./... — success
• go test ./... — 245 passed (incl. new table-driven tests for aiconnector and review usecases covering success, not-found, wrapped-error, and JSON-parse-failure paths)
• make lint — 0 issues
• gofmt clean on all changed files

Acceptance Criteria

• Configure OpenAI, Anthropic, or custom LLM endpoint
• Sending a diff chunk returns structured review comments
• Supports any OpenAI-compatible proxy (LiteLLM, vLLM, Ollama, etc.)
• API keys are never exposed in responses

Note

This repo has no develop branch; the branch was created from feat/ai-provider (not yet on remote) per request. PR is targeted at main — retarget to the intended integration branch before merge if needed.

[github-actions]

✅ Coverage Report

Metric	Value
Total Coverage	70.5%
Threshold	60.0%
Status	Passed

[miqdadyyy]

So basically the AI review workflow is not correct yet
It should be handled under worker / background nats process
The usecase and several code already correct, but the coverage doesn't meet the requirement

[miqdadyyy]

Thanks for the review — all points addressed in the latest commit (4fad062b4ad4c29b81435b74c86cd9438e2d9748):

Architecture: background worker

• POST /api/v1/reviews/diff now enqueues the job via IPubSub and returns 202 Accepted with a job id.
• New worker handler HandleReviewDiff (internal/handler/worker/review.go) consumes review_diff_topic and runs the LLM review under the NATS worker process (subscribed in cmd/cli/runner/worker.go with concurrency 5).

Inline comments

• review.go validation message → now uses fiber.ErrUnprocessableEntity.Error() (applied to all handlers for consistency).
• OpenAI structs → moved to internal/services/llm/models/openai.go.
• Anthropic structs → moved to internal/services/llm/models/anthropic.go.
• firstTextBlock/firstNonEmptyStr (+ emitStream) → moved into base.go.
• models/models.go → renamed to llm.go.
• maskSecret removed → replaced by reflection-based internal/pkg/mask package driven by mask:"start=N" / mask:"end=N" struct tags. ConnectorResponse.APIKey is now masked declaratively (e.g. ******1234).

Coverage
Added tests: mask pkg (97%), llm openai/anthropic via httptest + factory, worker handler (100%), aiconnector List/Update, review Enqueue + prompt-template path. Full suite: 275 passed, make lint 0 issues.