CODE REVIEW CONTROL PLANE

HUBOLT

Context-rich reviews, analyzer-backed evidence, routed model spend, and an audit trail that makes every finding traceable.

Open console GitHub

semantic context budget guardrails inline GitHub review

app.hubolt.local / overview

db ok . 18ms

reviews stored 1,284

queue active 12

month spend $428

high src/server/services/llm-gateway.ts budget reconciliation

medium src/queue/review-processor.ts memory retrieval gap

low web/js/views/reviews.js empty state coverage

Analyzer signal Secret scanner

Hardcoded credential candidates are promoted only with file evidence and verification notes.

Gateway guardrail Budget held

Provider spend is reserved, settled, refunded, and audited as separate states.

How it works

From changed code to a review your team can defend.

Hubolt keeps the model inside a bounded review workflow: gather context, check machine signals, generate findings, and record the outcome.

01 / Intake

Build the review packet.

Hubolt expands the diff with related files, semantic regions, ownership hints, and repository rules.

02 / Evidence

Run machine checks first.

Analyzers produce grounded signals before the LLM writes a finding, keeping guesses out of the review.

03 / Review

Route the model by risk.

Budget gates, provider routing, and rule memories shape the prompt before a finding reaches a reviewer.

04 / Settlement

Write the audit trail.

Usage, final states, posted comments, and reviewer feedback are stored so each review can be traced later.

review-runner / pull/284

stage 01

Context 14 files mapped Analyzers 7 signals checked LLM pass risk route selected Audit usage reconciled

Context map Review packet assembled

Changed code is connected to contracts, tests, and budget services before the model sees it.

14 files mapped 6 contracts linked 3 rules recalled

Evidence layer Machine checks gate the review

Signals are grouped by confidence so the model argues from evidence instead of vibes.

secret scan 3 candidates

TypeScript clean

dependency audit 1 advisory

rule memory 3 cards

promoted finding authorization header requires verification notes

Model route Risk controls choose the pass

Budget is held before the provider call, then the route is selected from repository risk.

fast skip low risk

balanced selected gateway code

thorough hold security critical

budget reservation $0.18 held refund only after terminal failure

Traceable result Review output becomes an audit record

Usage, final states, posted comments, and reviewer feedback stay tied to the same run.

GitHub review 1 high-confidence finding posted with file evidence

00:00 ingest diff from pull request 284 00:02 expand llm-gateway.ts with queue contracts 00:05 secret scanner marks credential-like values as candidates 00:07 TypeScript and dependency audit complete 00:10 reserve budget before provider call 00:14 generate findings with file evidence and confidence 00:18 reconcile usage and write audit event 00:20 post GitHub-ready review summary

Control surfaces

An operations console for the review loop.

The web panel keeps active work one glance away: recent reviews, queue pressure, provider credentials, budget limits, and audit history.

Findings, analyzer signals, confidence, model usage, and file locations in one scan path.

high Auth header used before verificationllm-gateway.ts:134 - high confidence - hybrid

med Query result is unboundedrequest-queue.ts:88 - medium confidence - analyzer

low Prefer const assertionmodel-catalog.ts:12 - low confidence - llm

Integrations

Your stack, in the review loop.

Hubolt reviews where your code lives, routes spend to the models you pick, and pushes findings to the tools your team already watches.

Hubolt 9 live integrations

Source control
Models
Notifications
Issue tracking

Live: GitHub, OpenAI, Anthropic, Google, Slack, Teams, Jira, ClickUp, Asana. On the roadmap: GitLab, Bitbucket.

Gateway costs

Route by risk. Keep spend visible.

Estimate monthly review cost using the gateway catalog rates and a context-heavy planning figure of 60k tokens per review.

Balanced route estimate $54

300 reviews / month fast $1.35 . thorough $90