Summary
The Ultrareview (/code-review ultra / /ultrareview) docs at https://code.claude.com/docs/en/ultrareview describe two output/management surfaces that don't exist in shipped v2.1.96:
/tasks slash command — docs say "Use /tasks to see running and completed reviews, open the detail view for a review, or stop a review that is in progress." In v2.1.96 there is no such slash command; /tasks is not in /help's output.claude ultrareview standalone subcommand with --json flag — docs say "Use the claude ultrareview subcommand to start an ultrareview from CI or a script…" with a --json flag for raw bugs.json. In v2.1.96 there is no ultrareview subcommand; running claude ultrareview 707 parses ultrareview as the prompt and --json errors with error: unknown option '--json'.
Combined, these two gaps make it hard to read a finished Ultrareview's findings.
Reproduction
Tested on Claude Code 2.1.96 (output of claude --version).
$ claude --version
2.1.96 (Claude Code)
$ claude ultrareview 707 --json
error: unknown option '--json'
$ # Inside Claude Code interactive session:
/tasks
# Slash command not found.
The slash command /code-review ultra <PR> and its alias /ultrareview <PR> do work — they launch a cloud sandbox session, display a panel with Setup → Find → Verify → Dedupe phases, and end with "Review complete — N issues found" and a list of red-text finding previews. The previews are truncated and not directly readable (one-line ellipsis e.g. "Bucket arrays from $alloc free-list reuse contai…").
What "Review complete" surfaces (or doesn't)
After "Review complete — 3 issues found" the docs say "the verified findings appear as a notification in your session. Each finding includes the file location and an explanation of the issue so you can ask Claude to fix it directly."
Observed v2.1.96 behaviour: no notification arrived in the local Claude Code session. The full text of the findings was only readable by opening the cloud session URL (https://claude.ai/code/session_<id>?from=cli) in a browser — which the launch banner does print, but it's framed as the "Track" link for live progress, not the canonical "see your findings" surface. The browser page itself has the file:line locations and finding titles visible directly (no truncation), but no input field for further interaction — the session is read-only after completion.
A spurious task-notification event also landed mid-run with status: failed, summary: remote session exceeded 30 minutes — but the panel continued past that point and completed normally. The session ID in that event matched the still-running review. May or may not be related; flagging in case it's useful.
Suggested fixes (in priority order)
- When the cloud review completes, surface the full finding text as a session message — the docs' described "notification" channel. Whatever broke that delivery in v2.1.96 vs the docs' described version.
- Make the red-text finding previews in the panel clickable to expand into the full text. The previews truncate at the panel width with no visible affordance for "click to see more."
- Either ship the
claude ultrareview subcommand the docs describe, or update the docs to remove it. Same for /tasks.
- Treat the browser cloud-session page as a documented output surface, not just a "Track" link. The launch banner could say e.g. "View findings: " in addition to the "Track" framing.
Environment
- Claude Code:
2.1.96
- macOS: 15.7.3 (Apple Silicon)
- Login: Max plan, Claude.ai auth (not API key)
- Triggered via
/ultrareview 707 (the alias form)
- Cloud session: completed successfully, 3 issues found+verified+deduped
The substantive findings themselves were excellent (correctly identified two real GC bugs not caught by the parallel CodeRabbit review — verified and fixed). The feedback here is purely about the find-text-after-completion UX in v2.1.96, not about review quality.
Summary
The Ultrareview (
/code-review ultra//ultrareview) docs at https://code.claude.com/docs/en/ultrareview describe two output/management surfaces that don't exist in shipped v2.1.96:/tasksslash command — docs say "Use/tasksto see running and completed reviews, open the detail view for a review, or stop a review that is in progress." In v2.1.96 there is no such slash command;/tasksis not in/help's output.claude ultrareviewstandalone subcommand with--jsonflag — docs say "Use theclaude ultrareviewsubcommand to start an ultrareview from CI or a script…" with a--jsonflag for rawbugs.json. In v2.1.96 there is noultrareviewsubcommand; runningclaude ultrareview 707parsesultrareviewas the prompt and--jsonerrors witherror: unknown option '--json'.Combined, these two gaps make it hard to read a finished Ultrareview's findings.
Reproduction
Tested on
Claude Code 2.1.96(output ofclaude --version).The slash command
/code-review ultra <PR>and its alias/ultrareview <PR>do work — they launch a cloud sandbox session, display a panel with Setup → Find → Verify → Dedupe phases, and end with "Review complete — N issues found" and a list of red-text finding previews. The previews are truncated and not directly readable (one-line ellipsis e.g."Bucket arrays from $alloc free-list reuse contai…").What "Review complete" surfaces (or doesn't)
After "Review complete — 3 issues found" the docs say "the verified findings appear as a notification in your session. Each finding includes the file location and an explanation of the issue so you can ask Claude to fix it directly."
Observed v2.1.96 behaviour: no notification arrived in the local Claude Code session. The full text of the findings was only readable by opening the cloud session URL (
https://claude.ai/code/session_<id>?from=cli) in a browser — which the launch banner does print, but it's framed as the "Track" link for live progress, not the canonical "see your findings" surface. The browser page itself has the file:line locations and finding titles visible directly (no truncation), but no input field for further interaction — the session is read-only after completion.A spurious
task-notificationevent also landed mid-run withstatus: failed, summary: remote session exceeded 30 minutes— but the panel continued past that point and completed normally. The session ID in that event matched the still-running review. May or may not be related; flagging in case it's useful.Suggested fixes (in priority order)
claude ultrareviewsubcommand the docs describe, or update the docs to remove it. Same for/tasks.Environment
2.1.96/ultrareview 707(the alias form)The substantive findings themselves were excellent (correctly identified two real GC bugs not caught by the parallel CodeRabbit review — verified and fixed). The feedback here is purely about the find-text-after-completion UX in v2.1.96, not about review quality.