Codex clients: show Claude progress-update (narration) blocks as Codex commentary (opt-in)

[lpoto]

Problem
In Codex Desktop with Claude models, long agentic turns look almost silent. Claude does write short progress notes between tool calls (CPA requests thinking.display: "updates"), but they arrive as thinking blocks. CPA maps every thinking block to a Responses reasoning item, and Codex Desktop renders reasoning as a single collapsed line. Users see a truncated first line per step instead of readable updates.

From local session logs: a Claude session had 38 reasoning items, 54 tool calls and 1 assistant message. A comparable GPT session had 43 commentary messages over 231 tool calls.

Why this looks feasible in CPA

• CPA already distinguishes progress notes from reasoning: the Claude signature parser recognizes block kind "narration" vs "thinking" (internal/signature/claude_validation.go, CAQS envelope).
• v8.0.49 added the reverse direction: client.claude.show-codex-commentary (visible Codex commentary with preserved replay). This request is the Claude → Codex counterpart.

Proposal
An opt-in setting, e.g. upstream.claude.codex-progress-commentary: false (default off):

• Response (Claude → Responses): emit each narration block's text as an assistant message with phase: "commentary", keeping what is needed to replay the original block.
• Request (Responses → Claude): restore the original narration thinking block (text and signature, unchanged) in its position, and don't send the commentary text to Claude a second time.
• thinking blocks keep today's reasoning mapping.

The replay mechanism should probably match whatever show-codex-commentary uses; I'd follow your preference there.

I'm happy to open a PR if this direction is acceptable.

Environment
CLIProxyAPI v8.0.51, Codex Desktop (codex-cli 0.166.0), claude-opus-5-5 via Claude OAuth.