Summary

The claude-cli live-session runtime and the Codex app-server runtime give very different live UX for the same kind of turn, on the same message-based channel preview:

• Codex app-server: reasoning and tool/command activity stream into a live progress preview in real time; on turn/completed the draft is cleared and the final answer is delivered.
• claude-cli live-session: a single reasoning bubble appears, is immediately removed, then a long silent gap during generation, then the final answer arrives in one piece.

Root cause (source-level)

In the Claude stream-json parser (claude-live-session-*.js, parseClaudeCliStreamingDelta), only:

• content_block_delta → text_delta (assistant text), and
• tool_use start/result events

are forwarded. There is no thinking_delta path — extended-thinking content blocks are not streamed incrementally, so reasoning surfaces as a single block (shown as one transient /reasoning stream preview that is deleted after final delivery, per docs/concepts/streaming.md). And during answer synthesis the channel preview isn't driven incrementally, so it reads as "reasoning blip → silence → final."

By contrast, the Codex harness emits a rich progress event stream (raw reasoning, reasoning progress, agentMessage progress, tool/command lifecycle), which OpenClaw renders as a continuously-updating progress preview.

The data is already available

The bundled claude-cli backend launches with --output-format stream-json --include-partial-messages, so Claude already emits content_block_delta events including thinking_delta and incremental text_delta. The parser just doesn't consume the thinking deltas (and doesn't appear to drive an incremental preview from text deltas) the way the Codex runtime does.

Proposed

1. Parse content_block_delta where delta.type === "thinking_delta" and forward it as a reasoning-preview delta (respecting /reasoning mode: stream → transient preview, etc.).
2. Drive the incremental answer preview from text_delta so edit-capable channels (Telegram/Discord) get a continuously-updating draft during generation.

Net effect: live reasoning/answer progress for claude-cli at parity with the Codex progress streaming experience. This is purely a matter of feeding OpenClaw's existing message-based preview with deltas Claude already sends — no token-delta channel transport required.

Repro

1. Provider anthropic/* with model-scoped agentRuntime.id: claude-cli, live-session enabled, /reasoning stream.
2. Send a prompt that triggers extended thinking.
3. Observe: one reasoning bubble → removed → silence → final. Compare to a Codex-runtime model (live tool/reasoning progress).

Related

Same claude-cli live-session integration area as #97489 (auth-gate doesn't recognize apiKeyHelper) and the picker-visibility contract #83342.

Environment: OpenClaw 2026.6.10.