bug(web): Claude tool-call markup (<function_calls>/<invoke>) renders as raw text instead of a tool card

[0xshawn]

Summary

In the web chat view, a Claude session's assistant turn rendered raw tool-call markup as literal text instead of being parsed into a tool card. The <function_calls> / <invoke name="..."> / <parameter ...> block leaked directly into the message bubble.

What I saw

The chat bubble showed literal text like:

<invoke name="TaskUpdate">
<parameter name="taskId">8</parameter>
<parameter name="status">completed</parameter>
</invoke>
<invoke name="Bash">
<parameter name="command">git add CLAUDE.md && git commit -q -m "docs: add project CLAUDE.md" && git log --oneline -5</parameter>
<parameter name="description">Commit CLAUDE.md</parameter>
</invoke>

Expected: this is a TaskUpdate + Bash tool call and should render as tool cards (like #480), not as raw XML-ish text in the message body.

Environment

• hapi version: 0.20.2 (latest at time of report)
• Agent flavor: Claude (Claude Code CLI)
• Surface: web PWA chat view
• Deployment: self-hosted hub (container) + host runner

Notes / suspected cause

The leaked markup is Claude Code's internal tool-call serialization. Two candidate causes:

1. The CLI streamed the tool-call block into the assistant text channel (partial/mis-tagged event), and the web renderer passed it through verbatim.
2. The web renderer received a well-formed tool-use event but failed to recognize/strip this markup shape, falling back to plain-text render.

Either way, the web side should detect and either render a tool card or suppress the raw markup rather than displaying it as message text.

Repro

Not yet reliably reproduced — appeared during a normal Claude session that emitted TaskUpdate + Bash tool calls. Will add exact steps / session export if it recurs.

Severity

Cosmetic — session state and execution were unaffected; only the rendered transcript was wrong.

[github-actions]

Thanks, this matches an unhandled shape in the current web pipeline.

Verified behavior in the codebase: HAPI renders Claude tool cards when the stored Claude assistant message has structured message.content[] blocks with type: "tool_use". I did not find any parser/normalizer for raw <function_calls> / <invoke> markup, so if Claude Code stores or streams that markup as a text block, the web renderer currently treats it as ordinary assistant text.

So the likely boundary to confirm is: did the affected stored message contain a real tool_use block, or did it contain the XML-ish markup inside text content? If it is the latter, this is not currently normalized into a tool card.

Relevant code:

• cli/src/api/apiSession.ts:458 - non-user Claude log records are sent to the hub as agent output data.
• cli/src/claude/utils/sdkToLogConverter.ts:198 - Claude SDK assistant messages are preserved as assistantMsg.message in the log record.
• web/src/chat/normalizeAgent.ts:236 - string assistant content is normalized as text.
• web/src/chat/normalizeAgent.ts:249 - only structured tool_use array blocks are normalized into tool-call content.
• web/src/chat/reducerTimeline.ts:825 - normalized tool-call content becomes a chat tool block.
• web/src/components/AssistantChat/messages/AssistantMessage.tsx:20 - assistant-ui renders tool-call parts through the tool message component.
• cli/src/claude/utils/__fixtures__/1-continue-run-ls-tool.jsonl:5 - existing fixture showing the structured Claude tool_use shape that should render as a tool card.

Need more info:

• A session export or raw stored message record for the affected assistant turn, including the full message.content around the leaked markup.
• Claude Code CLI version used for that session.

---

HAPI Bot