/ultrareview: task-notification surfaces only a subset of verified findings (2 of 8); dashboard text not selectable

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Hi — wanted to flag two issues with the /ultrareview integration
discovered during a code-review session today.

Setup

• Claude Code (CLI), repository mlehmhk-create/legal-document-anonymizer
• Free tier (Free ultrareview 2 of 3 in the launch banner)
• PR #8 in that repo (a draft PR, base=main, head=feature branch, ~13 commits diff)
• Invoked: /ultrareview 8
• Cloud session: claude.ai/code/session_01FB4ePvSqiyTp3gMuhvFHGj

Issue 1 — task-notification surfaces only a subset of findings

The cloud dashboard reported:

Setup ✓
Find: 23 candidates
Verify: 8 confirmed · 15 refuted
Dedupe: 8 issues

So 8 verified, corroborated, deduped findings. However, the
task-notification Claude Code received on completion contained
only 2 of those 8 (bug_007 and bug_008 in the internal
numbering). The structured JSON delivered to the in-IDE channel
had corroborated_by: null on both, suggesting a corroboration-
or-confidence filter is being applied between "verified by the
dashboard" and "delivered to Claude Code." Whatever the filter
is, it's not visible to the user and not documented in the
launch banner.

Impact: a user who triages findings via Claude Code (the
recommended workflow per Claude Code docs) silently misses 75%
of the review's output unless they remember to manually open
the cloud dashboard and cross-check the count. For non-developer
users this is especially easy to miss — the in-IDE response
looks complete.

Suggested fix options:
(a) Send all verified findings to the task-notification, not
just the highest-confidence subset.
(b) If the filter is intentional, include a count
("2 of 8 surfaced; see dashboard for the rest") in the
notification so users know to follow up.
(c) Post the full set to the GitHub PR as inline review
comments (in my case the PR had 0 comments, 0 reviews
after /ultrareview completed — confirmed via gh api).

Issue 2 — Cloud dashboard text isn't selectable / copyable

After noticing the gap above, I tried to copy the missing 6
findings out of the cloud dashboard to paste back into Claude
Code for triage. Cmd+C / select-text didn't work — the
findings list appears to be in a component that blocks text
selection. Opening the dashboard URL in a fresh browser tab
went to the Claude.ai login wall rather than the session
view.

Workaround that did work: screenshot the dashboard, drag the
PNG into Claude Code, let it read the text via vision. Fine
for one-off; not great for systematic triage at scale.

Suggested fix:
(a) Make the findings list copyable as text, or
(b) Provide a "Copy all" / "Export as JSON" button, or
(c) Make the dashboard URL shareable so a session can be
reopened in a separate browser tab without re-auth
friction.

Why this matters

The integration's value proposition is "AI reviewer findings
flow back to Claude Code so the user can triage and fix
in-place." Issue 1 breaks the flow-back; Issue 2 blocks the
manual workaround. Together they meaningfully reduce the
free-tier /ultrareview value.

Both issues are reproducible — happy to re-run or share more
details (session IDs, raw task-notification JSON) if helpful.

Thanks for the great tool overall.

What Should Happen?

Hi — wanted to flag two issues with the /ultrareview integration
discovered during a code-review session today.

Setup

• Claude Code (CLI), repository mlehmhk-create/legal-document-anonymizer
• Free tier (Free ultrareview 2 of 3 in the launch banner)
• PR #8 in that repo (a draft PR, base=main, head=feature branch, ~13 commits diff)
• Invoked: /ultrareview 8
• Cloud session: claude.ai/code/session_01FB4ePvSqiyTp3gMuhvFHGj

Issue 1 — task-notification surfaces only a subset of findings

The cloud dashboard reported:

Setup ✓
Find: 23 candidates
Verify: 8 confirmed · 15 refuted
Dedupe: 8 issues

So 8 verified, corroborated, deduped findings. However, the
task-notification Claude Code received on completion contained
only 2 of those 8 (bug_007 and bug_008 in the internal
numbering). The structured JSON delivered to the in-IDE channel
had corroborated_by: null on both, suggesting a corroboration-
or-confidence filter is being applied between "verified by the
dashboard" and "delivered to Claude Code." Whatever the filter
is, it's not visible to the user and not documented in the
launch banner.

Impact: a user who triages findings via Claude Code (the
recommended workflow per Claude Code docs) silently misses 75%
of the review's output unless they remember to manually open
the cloud dashboard and cross-check the count. For non-developer
users this is especially easy to miss — the in-IDE response
looks complete.

Suggested fix options:
(a) Send all verified findings to the task-notification, not
just the highest-confidence subset.
(b) If the filter is intentional, include a count
("2 of 8 surfaced; see dashboard for the rest") in the
notification so users know to follow up.
(c) Post the full set to the GitHub PR as inline review
comments (in my case the PR had 0 comments, 0 reviews
after /ultrareview completed — confirmed via gh api).

Issue 2 — Cloud dashboard text isn't selectable / copyable

After noticing the gap above, I tried to copy the missing 6
findings out of the cloud dashboard to paste back into Claude
Code for triage. Cmd+C / select-text didn't work — the
findings list appears to be in a component that blocks text
selection. Opening the dashboard URL in a fresh browser tab
went to the Claude.ai login wall rather than the session
view.

Workaround that did work: screenshot the dashboard, drag the
PNG into Claude Code, let it read the text via vision. Fine
for one-off; not great for systematic triage at scale.

Suggested fix:
(a) Make the findings list copyable as text, or
(b) Provide a "Copy all" / "Export as JSON" button, or
(c) Make the dashboard URL shareable so a session can be
reopened in a separate browser tab without re-auth
friction.

Why this matters

The integration's value proposition is "AI reviewer findings
flow back to Claude Code so the user can triage and fix
in-place." Issue 1 breaks the flow-back; Issue 2 blocks the
manual workaround. Together they meaningfully reduce the
free-tier /ultrareview value.

Both issues are reproducible — happy to re-run or share more
details (session IDs, raw task-notification JSON) if helpful.

Thanks for the great tool overall.

Error Messages/Logs

Steps to Reproduce

Hi — wanted to flag two issues with the /ultrareview integration
discovered during a code-review session today.

Setup

• Claude Code (CLI), repository mlehmhk-create/legal-document-anonymizer
• Free tier (Free ultrareview 2 of 3 in the launch banner)
• PR #8 in that repo (a draft PR, base=main, head=feature branch, ~13 commits diff)
• Invoked: /ultrareview 8
• Cloud session: claude.ai/code/session_01FB4ePvSqiyTp3gMuhvFHGj

Issue 1 — task-notification surfaces only a subset of findings

The cloud dashboard reported:

Setup ✓
Find: 23 candidates
Verify: 8 confirmed · 15 refuted
Dedupe: 8 issues

So 8 verified, corroborated, deduped findings. However, the
task-notification Claude Code received on completion contained
only 2 of those 8 (bug_007 and bug_008 in the internal
numbering). The structured JSON delivered to the in-IDE channel
had corroborated_by: null on both, suggesting a corroboration-
or-confidence filter is being applied between "verified by the
dashboard" and "delivered to Claude Code." Whatever the filter
is, it's not visible to the user and not documented in the
launch banner.

Impact: a user who triages findings via Claude Code (the
recommended workflow per Claude Code docs) silently misses 75%
of the review's output unless they remember to manually open
the cloud dashboard and cross-check the count. For non-developer
users this is especially easy to miss — the in-IDE response
looks complete.

Suggested fix options:
(a) Send all verified findings to the task-notification, not
just the highest-confidence subset.
(b) If the filter is intentional, include a count
("2 of 8 surfaced; see dashboard for the rest") in the
notification so users know to follow up.
(c) Post the full set to the GitHub PR as inline review
comments (in my case the PR had 0 comments, 0 reviews
after /ultrareview completed — confirmed via gh api).

Issue 2 — Cloud dashboard text isn't selectable / copyable

After noticing the gap above, I tried to copy the missing 6
findings out of the cloud dashboard to paste back into Claude
Code for triage. Cmd+C / select-text didn't work — the
findings list appears to be in a component that blocks text
selection. Opening the dashboard URL in a fresh browser tab
went to the Claude.ai login wall rather than the session
view.

Workaround that did work: screenshot the dashboard, drag the
PNG into Claude Code, let it read the text via vision. Fine
for one-off; not great for systematic triage at scale.

Suggested fix:
(a) Make the findings list copyable as text, or
(b) Provide a "Copy all" / "Export as JSON" button, or
(c) Make the dashboard URL shareable so a session can be
reopened in a separate browser tab without re-auth
friction.

Why this matters

The integration's value proposition is "AI reviewer findings
flow back to Claude Code so the user can triage and fix
in-place." Issue 1 breaks the flow-back; Issue 2 blocks the
manual workaround. Together they meaningfully reduce the
free-tier /ultrareview value.

Both issues are reproducible — happy to re-run or share more
details (session IDs, raw task-notification JSON) if helpful.

Thanks for the great tool overall.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Environment - Claude Code 2.1.129 - macOS (Darwin 25.4.0 per uname) - /ultrareview free tier (Free ultrareview 2 of 3)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_