v3.0 AutoFix · runs overnight · now live
AutoFix — Developer Intelligence Platform

Fix while
you sleep.

A production signal or Jira bug becomes a triaged, localized, self-verified, tenant-safe draft PR — automatically. You wake up to work that's already done and reviewed.

Last night, 02:1402:17while you slept, 1 P1 fixed
autofix — night shift02:14
Listening for signals… while you sleep
Tracebit · runtime errorswatching
PostHog · user signalswatching
Jira · hourly sync02:00
01 · Signal · evidence
Jira PROD-2841 · P1
Null price on catalog sync
exception TypeError · null is not an object
frame catalog/price.ts:84
impact 412 devices · v2.4.1
TracebitPostHogZydus · Emami
02 · Localized
catalog/price.ts:84 · top_frame · Tracebit
82function unitPrice(item) {
83 return item.price.toFixed(2);
84}
blast radius6 callers · checkout · invoice · zydus override
03 · Patch · null-guard
82function unitPrice(item) {
83- return item.price.toFixed(2);
83+ return (item.price ?? 0).toFixed(2);
84}
self-verifyred → green · analyze ✓ · CI-gated
04 · Self-review
SEC ✓
LOG ✓
ARC ✓
TST ✓
CTX ✓
TEN ✓
⬡ Judge · dedup → rank → cite → READY · 0 blocking
Draft PR #142 opened
✓ analyze✓ tests✓ tenant-safe
+1 −1 · 1 file · Zydus, Emami verified
REC
01 · Listening
The problem today

A “small” bug. A whole afternoon.

Toggle to see why — and what changes when AutoFix runs the lifecycle for you.

🕒

A slow human round-trip

Triage, find the code, understand the blast radius, write the fix, open a PR, wait for review. The work is minutes; the waiting is hours.

// effort spent in every phase, by hand
💥

One fix breaks another tenant

A shared interface or config change ripples to Zydus or Emami silently — and mock-based tests never catch it until prod does.

// no blast-radius awareness
🔁

The same bugs keep coming back

Fixes aren't captured as reusable knowledge. Every incident starts from a blank page — even when we've solved it before.

// nothing compounds
signaltriagelocalizefixreviewverifyship
✗ The manual path
~3h · every phase by hand
✓ With AutoFix
~3m · straight to verify
Same bug, two paths. The manual path crawls phase by phase. AutoFix localizes and fixes autonomously, then drops the work at verify — you review a finished PR, you don't write one.
The AutoFix line

A bug goes in. A reviewed PR comes out.

One bug rides the line — red at intake, green as a draft PR. At the review station the full agent gate opens; then the part loops back to teach the line for next time.

existingbuilt this sessionroadmap
ENGINEListening…webhook · signal → PR
🐞
01 · Intake
Signal & Evidence
Tracebit · PostHog · Jira hourly
02 · Triage
Localize
stack-frame · vector · code graph
03 · Fix
Fix Engine
editor gpt-4.1 · architect
04 · Review
Agent Gate
6 agents + judge · self-verify
05 · Gate
Tenant + CI
READY / WARN / BLOCK
06 · Ship
Draft PR
PR agent · preview app
04 · Review station — agent gatejudging…
SEC
Security
✓ pass
LOG
Logic
✓ pass
ARC
Architecture
✓ pass
TST
Tests
✓ green
CTX
Context
✓ pass
TEN
Tenant
✓ safe
⬡ Judge · dedup → rank → cite → READY
↺ Learning Loop · every merge re-trains the line → faster next time
The unfair advantage

Three things a generic bot can't copy.

Anyone can wrap an LLM around a diff. Our edge comes from data and structure a code-only tool will never have.

📡

It fixes what actually broke in the field

Tracebit and PostHog hand us the real crash — exact file, line, session, app version, and how many devices hit it. The fix starts from evidence, not a guess at the repo.

Repo-only bots never see what your reps' phones actually did.
🏢

It knows which tenant a “safe” fix will break

We model the monorepo's DI overrides and config inheritance, so a one-line change to shared code is scored READY / WARN / BLOCK per tenant — and every PR ships with a tenant-impact report.

A generic bot has no idea your monorepo is multi-tenant.
🧠

Every merge makes the next fix faster

Merged and reverted outcomes flow back to the Learning Loop. The rule and the solved case enter the knowledge base, so the next similar bug retrieves the proven fix instead of starting from a blank page.

One-shot fixers forget every fix the moment it ships.
faster time-to-merge
20h → 4.8h average
95%
of bugs caught
before human review
0.11
War Room Brier score
verdicts are calibrated
2-click
install on GitHub
or self-host in your VPC
Everything in one place

More than a fix. A platform.

AutoFix is the entry point — every pillar rides the same engine, evidence, and Learning Loop.

FIX
AutoFix
signal → reviewed PR
REV
Multi-Agent Review
6 agents + judge
BRN
Codebase Brain
live repo graph
SPR
Sprint Intelligence
severity-ranked queue
LP
Self-Improving Loop
failures → guardrails
ENTRY POINT · AUTOFIX

From signal to reviewed PR.

A Jira bug or prod signal becomes a localized, self-verified, tenant-safe draft PR — in under an hour. You review, you don't write.

autofix · 02:17

LOCALIZED · catalog/price.ts:84
→ top_frame · 412 device sessions
→ tenants: Zydus, Emami

SELF-VERIFIED · red → green
DRAFT PR · tenant-safe · CI-gated
MULTI-AGENT REVIEW

Catch bugs, not nits.

Six specialized agents run in parallel; a judge dedupes overlapping findings and ranks comments by real impact.

CRITICAL · Security Agent
validateAmount() has 4 callers at risk
→ processPayment():142 · HIGH

Logic Agent · 2 issues
item.price can be null for deleted products
REQUEST CHANGES · 2 critical · 1 warning
CODEBASE BRAIN

The whole graph, not the diff.

Functions, types, callers, configs and tenant overrides — reasoned across the full repo so cross-file regressions surface before merge.

# blast radius · PR #2841

unitPrice() → 6 callers
→ checkout/cart.ts
→ invoice/pdf.ts
→ tenant/zydus/override.ts

2 tenants in blast radius
SPRINT INTELLIGENCE

A queue that reflects reality.

Bugs ranked by real severity — affected tenants, device count, app version — pulled straight from field telemetry.

# triage queue · live

P1 null price · 412 devices · 2 tenants
P2 retry storm · 88 devices
P3 layout shift · 12 devices

ranked by real-world impact
SELF-IMPROVING LOOP

Failures become guardrails.

Outcomes feed back as patterns. Production failures are mined, clustered, and converted into permanent rules.

# learned · this week

+ rule: tenant filter in SQL
+ rule: null-guard catalog reads

14 cases solved · 6 rules mined
the moat widens with every merge
How it runs

The brain is always on. The hands appear only when needed.

An always-on control plane watches for signals, triages and localizes — no toolchain, scales cheap. When a fix is warranted it spins a private, throwaway sandbox per bug that builds, tests, and opens the PR, then self-destructs. Many run in parallel; idle nights cost nothing.

🧠

Control plane · always-on

Ingest, dedup, triage, and graph-localize run 24/7 with no compiler — light, cheap, and continuous. It never blocks; heavy work is queued.

Signal → reviewed PR, with a human only at approval.
📦

Ephemeral sandboxes · per bug

Each fix gets a private, network-restricted box that holds the toolchain, builds only the affected package, runs the tests red→green, opens the PR — then is torn down. One run can never touch another's.

10 bugs at once → 10 isolated runs in parallel.
🔒

Run where you run

Diflux Cloud for the fastest start, hybrid where the control plane is managed but your data stays in your infrastructure, or fully self-hosted in your own VPC. Your code never trains external models.

Your data. Your rules. Your environment.
Capabilities

Less noise. More signal.

AutoFix reasons over far more context than a linter or a single-pass LLM — the real crash, the whole call graph, your tenant rules, and every fix it's shipped before.

Codebase Brain

A live graph of your repo, spinning in your head.

Functions, types, callers, configs and tenant overrides — reasoned across the whole graph, not just the diff. The hot nodes are exactly what your fix actually touches.

PR
api
db
auth
tenant
log
cfg
utils
jobs
ui
Self-verifying

Proves red → green.

Every fix ships with a regression test the agent wrote and ran — failing before, passing after. A diff you can trust, not one that merely looks right.

# verify
before: ✗ fail
after: ✓ pass
✓ red → green · CI-gated
Tenant gate

READY / WARN / BLOCK.

Before a shared-code fix merges, AutoFix scores its blast radius per tenant and attaches a tenant-impact report to the PR.

READYWARN · 2 tenants
Evidence

Localizes from prod.

Tracebit + PostHog give the exact file, line, session and device count — so the fix starts near-pinpointed instead of searching the whole repo.

top_frame412 devicesPostHog
Learning loop

Sharper every merge.

Merged and reverted outcomes train repo-specific rules. Production failures are mined, clustered, and converted into permanent guardrails.

clusterevalimprove
Self-host

Your code never leaves.

Run AutoFix in your own VPC with your own LLM provider. Code never trains external models. SOC 2 Type II and air-gapped deploys available.

SOC 2VPC self-hostBYO LLM
Multi-model

The right model for every step.

Triage runs cheap, the fix runs on a frontier coding model, review on another — each step routes to its best model across frontier, open-source, and local, with automatic fallback. No vendor lock-in, and every step gets the model it deserves.

ClaudeGPTGeminilocalbest per step
Vs. the field

Bots read your repo. AutoFix sees prod.

Capability
AutoFix
Sentry Seer
Datadog Bits AI
Sweep AI
Localizes from real prod telemetry
✓ Tracebit / PostHog
✓ errors + traces
✓ APM + logs
✗ repo / ticket
Bug ticket / prod signal → PR, unattended
✓ Jira + signals
✓ error → PR
~ investigates, few code PRs
✓ ticket → PR
Models multi-tenant blast radius
✓ tenant gate
✗ no
✗ no
✗ no
Multi-agent self-review + red→green verify
✓ judge + test
~ generates a test
✗ alerts only
~ CI loop
Learns from merged outcomes (KB loop)
✓ compounds
✗ no
~ alert patterns
✗ no
Best model per step (no vendor lock-in)
✓ multi-model routing
✗ one model
✗ one model
✗ one model
Self-hostable — code stays in your VPC
✓ docker-compose
~ self-host Sentry
✗ SaaS only
✗ SaaS only

Sentry Seer and Datadog Bits AI also localize from telemetry — the closest on that axis. AutoFix adds the multi-tenant impact gate, the multi-agent self-review + red→green verify, and Jira/Figma context, and stays self-hostable. Sweep AI turns tickets into PRs but reads only the repo — no prod signal. Pricing target $19–$39/dev/mo.

Shipping

Built this quarter. And what's next.

💬

Interactive PR Agent

Comment on the draft PR — the agent re-opens the branch, makes the change, re-runs CI, and pushes. A conversation, not a one-shot patch.

🖥️

Preview patched app

One click builds the PR branch as a live web app and boots it in the dashboard — confirm the bug is gone before merge.

What's next · planned
⚙ Configurable skills
toggle a skill in 2–3 clicks — no code
🔗 Parent-Jira linkage
an order bug → the order epic
🧩 Multi-stack (React)
same engine on the dashboard app
Get started

Install in two clicks. Or zero, if your CTO insists.

Start free on a single repo with the GitHub App. For air-gapped environments, deploy AutoFix into your own VPC.

FREE FOREVER
Recommended

GitHub App

Click “Install”, pick a repo, see your first reviewed fix in under 60 seconds.

$ visit github.com/apps/diflux
$ click “Install”
✓ ready in 6 seconds
Install GitHub App →
ENTERPRISE
Self-hosted

Air-gapped VPC

Run AutoFix in your own AWS, GCP, or on-prem. Bring your own LLM provider.

$ helm repo add diflux https://…
$ helm install diflux/core -n diflux
✓ deployed to k8s cluster
Talk to founders →

Fix while you sleep.

Put your bugs on the line tonight. Wake up to reviewed, tenant-safe draft PRs — and a system that's a little smarter than it was yesterday.