Triage

When something breaks in Sessionwire, Triage finds out why and shows the evidence. On 22 real past problems, it gave the right answer to 20.

I've led development at Sessionwire since 2017. I built Triage there in October 2026.

what it does

Finding why something broke meant checking six places by hand: error reports, logs, the database, call records, payments and the code. I had the order written down. Triage does that checking and comes back with a likely cause and the evidence for it.

It can look at everything and change nothing. Anything that changes the product still goes through me.

The detail: Triage is an MCP server, a set of 23 read-only tools that Claude Code can call.

signalan error, a log line or a user report
reasonClaude Code
readTriage, 23 read-only tools across six sources
reporta likely cause and the evidence
actme, for anything that changes the product

how it was tested

I tested it on 22 real problems from the past where I already knew the answer: real bugs, false alarms, and old problems whose evidence is gone. Each investigation was recorded once, then replayed. The replays never touch live customer data.

Before trusting the score, I tested the test. An answer sheet that is always right scores 100%. Giving the same answer every time scores 33%. A second model grades each answer, so Triage never grades itself.

Right answer: 20 of 22 real past problems. The 22 were 9 real bugs, 8 false alarms and 5 with no evidence left.
Cause found: 8 of 9 real bugs, with the cause in the code named correctly.
Against guessing: 90% with bugs, false alarms and no-evidence cases counted equally (balanced accuracy), against 33% for giving the same answer every time.
Cost: About 24 cents and 17 seconds a problem, at list price.
Detail: Run on 7 October 2026. Claude Opus 5.5 investigated, Claude Sonnet 5.5 graded. A 23rd recording could not be replayed, so it isn't counted.

when the test was wrong

In an earlier run on 6 problems, all 3 misses came from one problem. A page broke in an old version of Safari before the page's error tracking had loaded, so it left no trace.

The honest answer was “I can't tell”. Triage said that 3 times out of 3. My answer key expected the real cause, so it marked each one wrong.

I fixed the answer key, not Triage. “I can't tell” is sometimes the right answer, and a test should reward it.

what it found

In its first live investigation, Triage found a real payment bug that no customer had reported. I checked its evidence and confirmed it.

a part I removed

I also built a version for Sessionwire's admin page, so support staff could ask it directly. It ran on the server, where it could not read the code. I measured it: it scored the same as answering “I can't tell” every time. I deleted it, with the automation around it, about 11,000 lines. The version that reads the code is the one that works.

limits

I'm the only user so far. 22 problems is a small set, and each problem in this run ran once, so the numbers have wide margins. Triage runs when I ask. It does not yet start on its own when something breaks.

Sessionwire: The product, year by year.
Call rebuild: How Call rebuild changed Sessionwire's calls in July 2026.
Search: How Search finds people and music from a description.
Remote control: How Remote control lets you run someone else's computer in a Sessionwire call.