The export stops at 500 rows and I have 912 contacts.
Bug capture
console
ExportRowMismatch: 912 expected, 500 sentnetwork
GET /api/export 200 · 214 msThe right AI model for every job 
Your support teamand your engineers,as one AI workforce.
Tore resolves customer conversations, answers pre-sales questions, catches bugs, writes the fix and preps the pull request for your approval. It does the heavy lifting of support and development, knows when to hand over, and lets your team step in anytime.
app.tore.ai/beacon-data/TCK-482
Example data What Tore reads
ConversationPriya: export stops at 500 rowsto Customer
Sentry errors38 events · export jobto Internal notes
Grafana metricsrows per export · since 09:05to Internal notes
GitHub codebasebeacon/app · mainto Internal notes
Export stops at 500 rowsTCK-482 · Beacon Data · Priya Shah
Claude OpusGemini ProCodex AstraOpenIn progressWaiting for approvalResolvedCustomer 0Internal notes 0 Priya sees thisOnly your team sees this
Writing to Priya…
Claude Opus · mediumThanks Priya, I can see it in your replay. I’m looking into it now and will update you right here.
Fix deployed · 10:02
Writing to Priya…
Claude Opus · mediumFixed. Your export now includes all 912 contacts. Could you try it once more?
All 912 are there. Thank you!
Reading Sentry…
Gemini Pro · mediumvia SentryHer report matches a Sentry error that started this morning.
ExportRowMismatchexport-worker · first seen 09:07
error38events
6users
Reading Grafana…
Gemini Pro · mediumvia GrafanaRows per export have been flat at 500 since the 09:05 deploy.
Rows per exportbeacon-prod · last 3 h
flatReading the code…
Codex Astra · highvia GitHubRoot cause: line 57 still reads cursor, but the API now returns nextCursor, so paging stops after the first 500.
src/export/paginate.tsbeacon/app · main
56 rows.push(...page.items);57 const next = page.cursor;58 if (!next) break;59 page = await fetchPage(next);
last commit a3f9c1e · refactor paging · 09:02Drafting a fix…
Codex Astra · highFix drafted and tested.
paginate.tstore/fix-export-cursor
+1 −157- const next = page.cursor;57+ const next = page.nextCursor;
48 tests passed · 2.1 s
Preparing the pull request…
Codex Astra · highvia GitHubPull request ready. Alex, it needs your approval.
#1294 fix(export): follow nextCursortore/fix-export-cursor → main · 1 file · checks passed
Waiting for Alex to approveApproved by AlexApproveRequest changes
You merge it. Your pipeline deploys.
MergedDeployed 10:02Errors back to normal
From first error to deployed fix.One thread, six steps.
The same export bug from above, all the way through. Scroll, and the product window follows along.
09:21 · diagnosis
Three signals, one issue
The Sentry spike, three customer bug captures and the Grafana latency jump all point at report export. Tore links them into a single issue, so nobody chases the same bug three times.
09:23 · root cause
The cause, down to the line
Tore reads the repository you connected and finds the loop that resets the page cursor. It names the file, the line and the commit that introduced it in the 09:00 deploy.
09:26 · fix
A fix, with the tests already run
It writes the smallest change that solves the problem, adds a test that reproduces the bug, and runs your suite. Everything passes before anyone is asked to look.
09:28 · pull request
A pull request, waiting for your approval
The summary, the evidence and the test results are in one pull request. Your engineer reads it, approves it or sends it back.
10:12 · deployed
Approved, deployed, confirmed
Once approved it ships through your normal pipeline. Tore watches Sentry and Grafana afterwards and confirms the errors stopped and latency is back to normal.
10:14 · customers told
The people who reported it hear back
Priya, Alex and the third reporter each get a reply in their own conversation saying it is fixed. Nobody has to remember who asked.
app.tore.ai/beacon-data/issues/ISS-311
Example dataReport export returns an empty fileISS-311
Sentry error spikeTypeError: cursor is null · 214 eventslinked
3 customer bug capturesPriya R. · Alex C. · 1 more, with replaylinked
Grafana p95 latency380 ms → 2.4 s after the 09:00 deploylinked
Root causebeacon/app
54export async function* paginate(query) {55 while (true) {56 const res = await fetchPage(query, { cursor })57 cursor = null58 if (!cursor) break
Fix written1 file · +1 −1
56 const res = await fetchPage(query, { cursor })57- cursor = null57+ cursor = res.nextCursor58 if (!cursor) break
paginate.test.ts12 passed
export.e2e.ts · new: 900-row export6 passed
rest of the suite30 passed
48 passed · 0 failed
Fix export cursor reset #418
tore/fix-export-cursor → main
- Summary
- The page cursor was reset on every loop, so exports stopped after the first page.
- Evidence
- Sentry issue, 3 bug captures, Grafana p95 panel
- Tests
- 48 passed, including a new 900-row export test
Waiting for your approval
Approved and deployed10:12
Builddone
Staging checksdone
Production100%
2.4 s → 390 msp95 latency, Grafana
0 new eventssince deploy, Sentry
Customers told3 replies sent
Hi Priya, the empty export is fixed. Exports now include all of your contacts. Thanks for sending the replay, it pointed us straight at the cause.
Tore AI · 10:14 · to Priya R.PPriya R. · sentAAlex C. · sent+11 more · sent
Step 1 of 6 · three signals linked to one issue
Know who did the work.Most of it wasn't a person.
Every resolution is attributed to AI, AI-assisted, a person or self-serve, so you can see how much of the load Tore carries.
app.tore.ai/insights/resolutions
Example dataIllustration with example data.
Meet your AI team.Six roles, one bug, no hand-off lost.
Each agent does one job on the right model, then hands the ticket on with everything it found.
app.tore.ai/beacon-data/team
Example data Example: six AI agents work in parallel. One bug report from Priya at Beacon Data passes from the support agent to triage, investigation, the engineer, review and release, and is deployed with Priya told it is fixed.
The right modelfor every job.
Tore routes each task to the model that does it best. A quick answer gets a fast model, a hard bug gets deep reasoning, the fix gets a coding model, and a second model checks the work before your engineer sees it.
ClaudeCodexGrokGeminiMMuseAnswer a how-to question from your help centerFastGemini Flash reasoning low
Gemini Flash reasoning lowHandle a pre-sales question about plansConversationalClaude Sonnet reasoning medium
Claude Sonnet reasoning mediumGroup 1,284 errors into one issueFastGrok Fast reasoning low
Grok Fast reasoning lowFind the root cause in your codeDeep reasoningClaude Opus reasoning high
Claude Opus reasoning highWrite the fixCodingCodex reasoning high
Codex reasoning highCheck the fix before your engineer sees itSecond opinionMuse reasoning medium
Early access
Bring your own AI keys during early access. Early teams can plug in their own Claude, OpenAI, Grok or Gemini keys while Tore still routes every task.
Four jobs across the divider,and you can watch each one.
Two of them happen where your support team works, two where your engineers do. Pick one and it plays.
left half · support
It answers, and it shows its working
The AI replies from your own help center and cites the article it used. When it is not confident, it hands over instead of guessing.
- Search understands meaning, not just keywords
- Every AI answer cites the article it came from
- Email and widget land in one shared queue
crossing over · capture
The bug arrives with its evidence
A real bug comes in with the screen recording, the console output and the network trail attached. Nobody has to ask which browser they were on.
- Session replay, console and network, attached automatically
- Secrets are stripped before anything leaves the browser
- Repeating reports are detected and promoted to a ticket
right half · engineering
It reads the actual code
The investigation reads the repository you connect: the stack frame, the commit that touched those lines, the files around it.
- Works from the real frame, not a guess from the ticket text
- Checks the recent commits and the call sites
- Connect GitHub or GitLab
right half · your engineer decides
A change, and a person who approves it
The AI proposes a change with the checks it ran. Your engineer approves or rejects it. Anything near authentication, billing or the database escalates instead.
- A readable diff with its reasoning beside it
- By default it waits for your engineer to approve or reject
- Opens the pull request on its own branch for you to approve
app.tore.ai/acme/inbox
Example data Tom W. · email09:40
I can't get into my account, it keeps saying the link expired.
09:40searching help center by meaning
match"Resetting your password" · sign-in help
Tore AI · replycites 1 article
Reset links last a short while for security. Request a fresh one from the sign-in page and open it on the same device.
help center · Resetting your passwordHana S. · widget09:44
Can you move our plan to annual and merge two workspaces?
09:44not confident · handed to Maya on your team
Sarah K. · widget · Safari 17 · iOS 1710:42
Blank white screen when I log in on my iPhone.
session replay · 0:31attached
tap sign inconsole errorrequest failed
0:17TypeError: undefined is not an object (evaluating 'el.inert')
0:24GET /api/session 401 · 112 ms
strippedauthorization header, session cookie
patternlast 24 hours
3 similar reports, all iOS Safari 17. Promoted to ticket TCK-479.
10:46investigation opened · repo acme/web · main
reading the repository4 files
app/login/page.tsx
components/auth/LoginForm.tsx : 88
lib/session.ts
styles/login.css
frameLoginForm.focusTrap() · line 88
blamec72e1f0 · 2 days ago · "add focus trap to login"
calls2 other call sites checked, both guard the element
findingfor your engineer
The new focus trap reads a property older Safari does not support, so the form throws before it renders.
components/auth/LoginForm.tsx+1 −1
87function focusTrap(el) {88- el.inert = false88+ if ('inert' in el) el.inert = false89 trap.activate()
scope: 1 file, 1 line
no authentication, billing or database code touched
report contents treated as data, never as instructions
change proposed · waiting for your approvalApprove changeReject
pull request open on Tore’s branch · you merge it
Most of it startsbefore anyone writes in.
Connect your error tracker. Errors group by fingerprint, and a group that crosses your threshold opens an investigation on its own, then prepares a change that waits for your engineer.
app.tore.ai/northwind/errors
Example data error stream · productionSentry
grouped by fingerprint1 group
TypeError: Cannot read properties of undefined (reading 'rows')
0
identical events, counted as one problem
threshold crossed · one investigation
daily investigation budgetset by you
ignore list
kill switch
investigationno human involved yet
Opened from the error groupreports/summary.ts · line 42
Read the repository as it is nowframe, last change, 3 call sites
Prepared one reviewable change1 file · guard the empty result
Waiting for your engineerapprove or reject · the default flow
Sentry GitHub GitLab Datadog
Everything both halves need,on the same record.
The conversation, the evidence, the investigation and the reply all hang off one ticket, so nobody pastes anything between tools.
Priya R.Export stops at 500 rows
bug
Tom W.Link expired when I try to sign in
email
Sarah K.Blank screen on login, iPhone
widget
Hana S.Move our plan to annual
with MayaOne shared queue
Email and the widget land in one inbox, with live updates so two people never answer the same thing twice.
confidence before replying
this questionyour threshold
Not sure enough to answer. Handed to Maya with the conversation so far.
Hands over instead of guessing
Below your threshold, it routes to a person rather than inventing an answer.
similar reports · 24 hours
all iOS Safari 17 · promoted
Example dataPatterns become tickets
Repeating reports are detected and promoted to one ticket for engineering.
report · R-1193
replay · console · network
tokens stripped in the browser
The bug report writes itself
Everything an engineer would ask for in the next four replies is already attached.
when TCK-482 is fixed
reply drafted for 3 reporters
wording approved by Alex C.
sent in the conversations they started
They hear back
The people who reported it are told when it is fixed, once a person approves the wording.
how each conversation was resolved
AIAI-assistedpersonself-serve
IllustrationCredit where it is due
Every resolution is attributed to the AI, AI-assisted, a person, or self-serve.
daily budgetyou set it
ignore list
thresholdper group
kill switch
10,000 identical errors from one bad deploy are one group, and at most one investigation.
It will not go berserk
Errors group by fingerprint before anything acts on them. You set budgets, an ignore list and a kill switch.
And it doesn't stop there.The rest of the product, covered.
The same AI team runs the parts around the queue too, from the first pre-sales question to the customer who almost left.
Knows your product
It learns from your code, help center and past conversations, so it knows the product better than a new hire from day one.
Tore AIAnswers pre-sales questions
Plans, features, fit. Prospects get an answer at 2am instead of waiting for someone in sales.
Who it's forBuilds what customers ask for
The most-requested feature arrives as a pull request, ready for your engineer to approve.
Feature requestsRequests, roadmap, changelog
One board instead of forty scattered threads, and everyone who asked hears when it ships.
RoadmapPortal and status page
Customers check their own requests and your system status instead of emailing to ask.
Customer portalUptime and job checks
Know an endpoint or a scheduled job has stopped before a customer tells you.
Status pageSecurity and data separation
Each organization's data is kept apart by the database itself, with an audit log of who did what.
SecurityKeeps customers who were leaving
Spot accounts drifting away while you can still help, and make one save offer when someone cancels.
Insights3
things attached to a bug reportSession replay, console and network, without asking the customer.
4
ways a resolution is creditedAI, AI-assisted, a person, or self-serve.
1
ticket holds the whole storyThe conversation, the evidence and the investigation, on one record.
14 days
free on the Growth planJoin the waitlist and we'll email you the day you can start.
Your help desk answers.Tore also reaches the code.
| What happens | Help desk alone | Help desk + Tore |
|---|---|---|
| Email and widget in one shared queue | Yes | Yes |
| Bug report arrives with replay, console and network | ask the customer | Yes |
| Repeating reports promoted to one ticket | by hand | Yes |
| An investigation that reads your repository | not there | Yes |
| A proposed change for your engineer to approve | not there | Yes |
| Error tracker events grouped and investigated | separate tool | Yes |
| Reporters told when the fix lands | if someone remembers | Yes |
| Pull request opened on its own branch, for you to approve | not there | Yes |
Teams outside Codas Labsgo here when they say so.
Added as beta customers come on board. Until then the frame stays empty rather than borrowed.
Answer on the left.Fix on the right.
Connect your inbox, your help center and a repository. By default, the change waits for your engineer.