Fix log

Fixes vendors shipped after our reports

We run real AI agents through a product’s key jobs and report what breaks. When the vendor ships a fix, we run the same test again on their live release. Every entry links to the vendor’s own record.

4 fixes from 3 vendors2 to 6 days from our report to their releaseLatest fix shipped October 12, 2026
Microsoft CopilotShipped October 12, 2026

Lucid

What broke. In Copilot Studio the org-chart tool never ran. Its rows had no declared type, so Copilot stopped before the call and showed “Error code: SystemError”. Typing the rows exposed a second problem: Copilot sent every row with only the first row’s fields. The person at the top has no manager, so every row arrived without one and no chart was made.

The fix. Lucid typed the row fields and asked for managerId on every row, with an empty value for the person at the top. That is the change we had tested through a proxy.

ReportedOctober 6, 2026 · GitHub issue #13
ShippedOctober 12, 2026 · 6 days later
Before → after0 → 6 of 6runs in Copilot

Through Lucid’s connector and the MCP URL, three runs each. Claude and GPT: 6 of 6 before and after.

“…your detailed reproduction and proxy tests helped us identify the two Copilot behaviors.”

Lucid, on GitHub issue #13 →

Microsoft CopilotShipped October 7, 2026

monday.com

What broke. Through the MCP URL, Copilot keeps only the first 70 tools. monday’s search tool was 85th, so Copilot cut it. Search also required 8 of its 13 inputs, and the GraphQL tools required variables even for a query that has none. GPT-5.5 in Copilot could not look up the board.

The fix. monday moved search to the top of the tool list, made its filters optional and gave GraphQL variables a default of {}.

ReportedOctober 5, 2026 · GitHub issue #508
ShippedOctober 7, 2026 · 2 days later · PR #509 and the hosted server
Before → after9 → 12 of 12runs in Copilot

Through monday’s connector and the MCP URL, with Claude and GPT, three runs each.

“Thanks for the detailed report and repro.”

monday.com, on GitHub issue #508 →

Microsoft CopilotShipped October 4, 2026

monday.com

What broke. In Copilot Studio the tool that creates items never ran. The model built a correct call. Copilot turned the item rows into its own table type, treated their required fields as missing and asked the user to fill them in. The call never reached monday.

The fix. Item rows no longer declare required fields, and monday checks the item name in its own code. It was the smallest of the three fixes we had tested.

ReportedSeptember 29, 2026 · GitHub issue #499
ShippedOctober 4, 2026 · 5 days later · PR #504
Before → after0 → 9 of 12runs in Copilot

The other three failed on the next problem, fixed on October 7. Claude and GPT through the API: 6 of 6 before and after.

“The reporter tested several fixes through a proxy. Removing required from the row object took Copilot Studio from 0 of 6 to 6 of 6 runs…”

monday.com, in PR #504 →

Agent sign-upShipped September 10, 2026

Headless Domains

What broke. Every agent got a working API key, but there was no endpoint to check it. One agent in 11 looked for a “who am I” endpoint and got a 404. Three “checked” the key on a public endpoint that works without one, which proves nothing. The revoke address they published returned a 404 page.

The fix. Headless Domains shipped GET /api/v1/agents/me and POST /oauth2/revoke, and added one sentence to their auth.md telling agents to check a new key with /api/v1/agents/me.

ReportedSeptember 8, 2026 · a LinkedIn thread
ShippedSeptember 10, 2026 · 2 days later
Before → after1 → 11 of 11agents checked their new key the right way

Before the fix, 3 agents “checked” the key on the endpoint that needs no key. After it, none did.

“Docka showed us that agents could get through signup and pointed out where they ran into trouble afterward.”

Headless Domains, on their blog →

Want your MCP server checked?

We run your main jobs through Copilot, ChatGPT and Claude, check the result in your product, and test the fix. The first report covers one job and is free.

Get a free reporteugene@docka.ai