Repository navigation
[Bug]: Claude subagent rows show a raw <tools-used-before-summary> tag before the summary #30069
Description
Before submitting
- I searched existing issues and did not find a duplicate.
- I included enough detail to reproduce or investigate the problem.
Area: apps/server
Steps to reproduce
- Use the Claude provider with Claude Code CLI 2.1.316.
- Ask Claude to hand a small job to a subagent, for example: "Use a subagent to list the functions in
math.tswith their line numbers." - Wait for the subagent to finish and look at its row in the thread.
Expected behavior
The settled subagent row shows the subagent's summary:
add (line 1) mean (line 6)
Actual behavior
The row shows Claude Code's internal markup in front of the summary:
<tools-used-before-summary>Read</tools-used-before-summary> add (line 1) mean (line 6)
The subagent's work and the parent agent's answer are correct; only the shown summary is wrong.
Cause: Claude Code 2.1.316 opens a foreground Agent tool result with a <tools-used-before-summary>…</tools-used-before-summary> line. It is a note for the parent model naming the tools the subagent called. claudeSubagentResultText in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts copies the tool result text into the subagent's result as is. The row's plainDetail in apps/web/src/components/chat/V2LifecycleRow.tsx joins the lines but leaves the tag, which gives the text above. Nothing in the repo handles the tag today (git grep tools-used-before-summary finds nothing).
The CLI does not add the line to every result: its own schema describes it as the tools the subagent called that the conversation's requests did not yet name. So some subagent rows look fine and others don't.
Impact
Cosmetic issue
Version or commit
main @ cefb544
Environment
macOS (Darwin 25.6.0), Claude Code CLI 2.1.316 through the Claude Agent SDK.
Logs or stack traces
An adapter test that sends a foreground subagent result opening with the line fails on main:
AssertionError: expected '<tools-used-before-summary>Read, Grep…' to equal 'add (line 1)\nmean (line 6)'
- Expected
+ Received
+ <tools-used-before-summary>Read, Grep</tools-used-before-summary>
add (line 1)
mean (line 6)Workaround
None found.
Note for whoever files this: the row text above was seen in a downstream build that shares this adapter and row code, not in a T3 Code build. On T3 main the defect is shown by the adapter test only.
Activity
Jacksondr5 commentedon Nov 7, 2026
The leading tag isn't the only wrapper. Recent Claude Code versions (seen on 2.1.316 to 2.1.319) can also wrap the Agent tool result in a hand-back frame and a trailer, and T3 stores all of it as the subagent's
result, which also becomes the child thread's last message:That example is from T3's own fixture,
testkit/fixtures/claude_nested_subagent_model/claude_transcript.ndjson. The frame indents every report line by two spaces, and the trailer adds theagentId: …line and a<usage>block. With the frame, the collapsed subagent row shows "[Subagent hand-back] The text below is the final report…" instead of the report.In one install's real threads from 2026-10-24 to 2026-11-01, 34 subagent results were wrapped: 9 had only the leading tag, 19 only the frame, and 6 both.
Since #30070 already unwraps the leading tag in
ClaudeAdapterV2, it may be simplest to unwrap all three there: drop the tag, take the text after "The report follows:" and remove its two-space indent, and drop the trailingagentId:/<usage>lines. @yaacovcorcos, happy to send that as a follow-up PR on top of yours instead, if you'd rather keep #30070 to the tag.Claude Opus 5.5, an AI agent, on Jackson's behalf.
yaacovcorcos commentedon Nov 9, 2026
Thanks, @Jacksondr5, that's a good find, and the counts make the case. Yes, please send the frame and trailer unwrapping as a follow-up on top of #30070. Keeping #30070 to the leading tag keeps both reviews small, and the shared fixture is a good test for yours.
Written with Claude Opus 5.5 in Claude Code, on Yaacov's behalf.
josephv123 commentedon Nov 11, 2026
Also seen in the subagent's child thread, not just its row, on T3 Code (Nightly) 0.1.9-nightly.20261110.4491 for macOS with Claude Code CLI 2.1.322.
When a Claude subagent opens as its own thread, the last assistant message there is the same
result. It is stored as:The markdown renderer hides the unknown tag but keeps its contents, so the message starts with a bare line
Read, ToolSearch, WebFetch, WebSearch, Writethat looks like something the agent wrote.It's the same source:
claudeSubagentResultText→updateClaudeSubagentNode, which also emits the child thread's:resultmessage. So #30070 should fix this view too.