Warnings are emitted as agent_message_chunk and indistinguishable from assistant replies

[roshangautam]

Summary

codex-acp emits adapter-generated diagnostics as ACP agent_message_chunk updates, the same update kind used for genuine assistant-authored replies. Downstream consumers therefore cannot distinguish a warning from model prose without unsafe text matching.

Current behavior

In 1.1.0 and the current 1.1.7 release:

• createTextEvent() emits genuine model deltas through createAgentTextMessageChunk(...).
• createConfigWarningEvent() emits Config warning: ... through the same helper.
• createWarningEvent() emits Warning: ... through the same helper.
• Other adapter-generated notices, including review-mode and lifecycle messages, also use agent_message_chunk.

A downstream ACP runtime such as ACPX consequently produces the same structural event for both:

{"type":"text_delta","stream":"output","tag":"agent_message_chunk","text":"..."}

The unknown-model warning visible in paperclipai/paperclip#1465 is one concrete trigger.

Impact

Consumers that relay assistant replies to another surface can misattribute adapter diagnostics as if the assistant authored them. Failing closed by dropping every agent_message_chunk also drops genuine replies, so neither available choice is correct.

Expected behavior

Non-assistant diagnostics should not be emitted as agent_message_chunk. Please either:

1. route warnings/status/lifecycle notices through a non-assistant ACP update kind, or
2. add a stable machine-readable origin/provenance discriminator and coordinate with runtimes so it survives normalization.

The discriminator must be structural; matching warning text is not safe because an assistant may legitimately quote or explain the same warning.

Versions verified

• @agentclientprotocol/codex-acp 1.1.0
• @agentclientprotocol/codex-acp 1.1.7 (latest on npm at filing)