Skip to content

Point the answer at the page, because the drawing cannot be relied on - #214

Merged
Parshkov merged 1 commit into
mainfrom
claude/link-to-the-working
Sep 7, 2026
Merged

Point the answer at the page, because the drawing cannot be relied on#214
Parshkov merged 1 commit into
mainfrom
claude/link-to-the-working

Conversation

@Parshkov

@Parshkov Parshkov commented Sep 7, 2026

Copy link
Copy Markdown
Owner

Tested in both clients — the thing that had never actually been done.

client model receives the image person sees it
claude.ai ⚠️ only after expanding the tool block — two clicks
ChatGPT ✅ ("да, я вижу её как радарную диаграмму") no <img> in the page at all

So "images work in chat" was true of the protocol and false of the person. A picture can be a bonus; it can never be the carrier.

What changes

The words already carried the answer. What they lacked was somewhere to look.

resonance_discover now returns see_on_the_page — the person's own radar on the page with the match selected, /people?thought=<mine>&select=<theirs>, a deep link the page has supported all along. Every match row carries its own, and the sentence ends with it:

…4 contradictions ("Shared landlord history before signing"). The same thing drawn, and every match on it: https://resonance.parshkov.com/people?thought=ses-…

Live data, every client, one click — and from there the person can ask for an introduction.

What it does not do

  • Nobody is sent to an empty page. No matches means nothing is drawn, so no link.
  • A deployment with no canonical origin emits none. A half-built URL sends a person nowhere.
  • The bridge learns its origin from the runtime, the way the notifier already does. allowed_origins is a set and cannot say which of several hosts is canonical — and this deployment serves two, so a link to the wrong one lands a person where their session is not.
696 tests OK (1 skipped)   both gates pass, gold unedited

🤖 Generated with Claude Code

Tested in both clients, which is the thing that had never been done: the image
reaches the MODEL everywhere -- ChatGPT said plainly "yes, I see it as a radar
diagram" -- and reaches the PERSON almost nowhere. ChatGPT renders no <img> at
all; claude.ai renders it, but inside a collapsed tool block, two clicks away.
So "images work in chat" was true of the protocol and false of the person, and
a picture can be a bonus but never the carrier.

The words already carry the answer. What they lacked was somewhere to look.

`resonance_discover` now returns `see_on_the_page`: the person's own radar on
the page, with the match selected -- /people?thought=<mine>&select=<theirs>,
a deep link the page has supported all along. Every match row carries its own.
The sentence ends with it. Live data, every client, one click, and from there
the person can ask for an introduction.

Nobody is sent to an empty page: with no matches there is nothing drawn, so
there is no link. A deployment with no canonical origin emits none either --
a half-built URL sends a person nowhere.

The bridge learns its origin the way the notifier already does, from the
runtime, which now resolves it once at build time. `allowed_origins` is a set
and cannot say which of several hosts is canonical, and this deployment serves
two -- a link to the wrong one lands a person where their session is not.

696 tests OK, both gates pass with gold unedited.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@Parshkov
Parshkov merged commit c09bfb1 into main Sep 7, 2026
3 checks passed
@Parshkov
Parshkov deleted the claude/link-to-the-working branch September 7, 2026 04:09
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