Files
pipelines/docs/notes/chat-injected-content.md
alan 07d4740d0c Show injected Claude input as agent reports and notices, not operator bubbles
Claude Code records subagent/peer reports, task notifications, reminders and
skill bodies as user-role entries. Classify them server-side (history and live
stream) by origin/isMeta/queued_command mode: reports become kind agent_report,
task notifications kind notice, harness meta is hidden, and the mid-turn
wrapper is unwrapped to the operator's text. The app draws a collapsed Agent
report card and a one-line system notice on the agent's side.
2026-11-02 16:44:07 +00:00

19 lines
1.7 KiB
Markdown

# Claude chat: harness-injected user entries
Claude Code records more than the operator's prompts as user-role input. The
server (`internal/claude/injected.go`) classifies each one for history
(`transcript.go`) and the live stream (`translate.go`):
| Entry | Identified by | Shown as |
|---|---|---|
| Subagent / peer report | `origin.kind == "peer"` (user entry or `queued_command` attachment, often `isMeta`), or text opening `<agent-message` | `ChatMessage{role: assistant, kind: "agent_report"}`; text is `origin.handbackReportText`, else the framed body with the 2-space indent removed; part `name` is the sender |
| Background task notification | `origin.kind == "task-notification"`, `queued_command` with `commandMode: "task-notification"`, a `<task-notification>` block, or the stream's `system/task_notification` | `kind: "notice"`, one line `Background task finished: <summary>`; id `task-notice-<tool-use-id>-<status>` dedupes repeats |
| Mid-turn wrapper | `<system-reminder>` + `The user sent a new message while you were working:` | the operator's text only, as a user turn |
| Reminders, skill bodies, image placeholders, interrupt markers, local commands | `isMeta`, or a harness prefix (`<system-reminder>`, `<local-command-`, `[Request interrupted by user`, `Base directory for this skill:` …) | hidden |
Report and notice ids are derived from content, so the live frame
(`chat_event` text with `message_kind`) and the history entry are one message.
Both carry role `assistant`, so a client that ignores `kind` still draws them
on the agent's side. The app renders them as a collapsed "Agent report" card
and a one-line system notice (`ui/transcript/chat_injected_rows.dart`).