SKILL-mode output sometimes leaks wrapper-like text and stray Unicode fragments

Summary

During chat workflows that rely on repo SKILL.md instructions, assistant output sometimes includes leaked transport/meta text and stray multilingual Unicode fragments that the user did not enter.

Observed Symptoms

• Injected lines resembling internal wrappers, e.g. user to=all with code block ... plain text:
• Unexpected Telugu fragment (example seen: ుతోంది / ోంది)
• Appears inline with normal turn output, which makes it look like the model answered for the user

Why this matters

This can corrupt structured workflows (e.g., guided form capture) and undermines trust because the assistant appears to fabricate user input.

Context

• Environment: Codex CLI agent workflow
• Trigger pattern: Seen repeatedly during local SKILL-driven interaction (chat-first guided logging)
• Not observed as frequently in plain, non-skill chat

Repro (high level)

1. Create/use repo-local skill(s) with structured SKILL.md instructions.
2. Run a long multi-turn guided interaction with one-question-at-a-time prompts.
3. After many turns, observe occasional injected wrapper-like text and/or stray Unicode token fragments in assistant replies.

Expected

Assistant output should contain only intended natural-language response text for the current turn.

Actual

Occasional leakage of internal-looking wrapper text and unrelated Unicode fragments.

Notes / Hypothesis

Potential boundary/stitching issue where role/transport scaffolding or partial streamed chunks bleed into user-visible output, more likely under skill-mode prompt scaffolding.