Skip to content

docs: the data channel is live - phase 3 verified with a real agent - #28

Merged
kiyeonjeon21 merged 1 commit into
mainfrom
docs/data-channel-live
Jul 25, 2026
Merged

docs: the data channel is live - phase 3 verified with a real agent#28
kiyeonjeon21 merged 1 commit into
mainfrom
docs/data-channel-live

Conversation

@kiyeonjeon21

Copy link
Copy Markdown
Owner

What

Records ROADMAP step 8 phase 3 as built and verified live.

The app spawns querypad mcp as a stdio child, hosts the conversation on a Unix socket it owns, and proxies every frame - so every tool call and result set flows through the app as structure. Claude Code joins via querypad mcp-attach (#27). First panels: a live result table with the SELECT's column order preserved, an activity feed, and the executed SQL as headline.

Verified with a real agent on the hard dataset

  • The fan-out question answered correctly (Acme 2,039/4, all seven rows matching ground truth), with the panel showing the CTE that aggregates each grain separately - the exact pattern the planning pass teaches
  • In open-ended analysis Claude spontaneously caught both planted data-quality traps: the 0.00 open invoice, and invoice 17's net_amt 609 vs amt_txt "600.00 USD" - the stale-legacy-column trap working in the wild

What the first session's failures taught

Three defects found by a human using it, none by the automated wire test: the panel squeezed out by the terminal (no holding priorities), an empty striped table reading as mystery pills, and a crash on the second differently-shaped result (addTableColumn synchronously queries the delegate against stale row indices). All fixed and re-verified - and the lesson recorded: both kinds of check stay.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKZ9UVgw2Mhiu6T5YAG7k6

The app hosts the MCP conversation on a socket it owns, proxies every frame,
and grew its first panels: a live result table (SELECT column order preserved),
an activity feed, and the executed SQL as headline.

Verified on the hard dataset with real Claude Code: the fan-out question
answered correctly with the separate-grain CTE pattern visible in the panel,
and in open-ended analysis the agent spontaneously caught both planted
data-quality traps (the 0.00 open invoice, and invoice 17's net_amt 609 vs
amt_txt '600.00 USD' - the stale legacy column).

Also recorded: the first real session found three defects (panel squeeze-out,
mystery-pill empty state, a crash on the second differently-shaped result)
that the automated wire test could not see. Both kinds of check stay.
@kiyeonjeon21
kiyeonjeon21 merged commit 1379816 into main Jul 25, 2026
1 check passed
@kiyeonjeon21
kiyeonjeon21 deleted the docs/data-channel-live branch July 25, 2026 20:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant