Prompts: prevent GPT-5.1/5.2 from ending turns after tool preambles

[4erdenko]

Summary

With gpt-5.2 (especially at higher reasoning effort), Codex CLI can emit a short progress/preamble message (e.g. “Next, I’ll …”) and then yield back to the user without emitting any tool calls. Because the CLI only continues an agent turn when tool calls are emitted, this stalls multi-step work and forces manual “continue” nudges.

This PR updates the GPT-5.2 (and GPT-5.1 for parity) system prompts to explicitly forbid “dangling” preambles and require issuing tool calls in the same assistant turn.

Problem

Observed behavior:

• The agent prints a preamble like “Ok, now I’ll run … / Next, I’ll …”
• No tool call follows
• The turn ends and control returns to the user mid-task

Expected:

• If the agent says it will run a command/apply a patch, it should immediately emit the tool call(s) in the same response (or ask the user directly if input/approval is required).

Root cause (why this happens)

Codex CLI’s agent loop continues a turn only when the model emits tool calls. Pure text updates are treated as “end of turn”.

The current GPT-5.x prompt strongly encourages short “next I’ll …” style progress updates, but doesn’t clearly state that such preambles must not be a standalone turn. gpt-5.2 sometimes treats the preamble as a separate phase and stops before emitting tool calls.

This appears consistent with a broader class of similar “agent stops while it still has a plan” issues, and matches recent reports specifically for v0.71.0 + GPT‑5.2.

Changes

• codex-rs/core/gpt_5_2_prompt.md: clarify “heads-down” guidance and add an explicit rule to avoid standalone preambles and to emit tool calls immediately in the same response.
• codex-rs/core/gpt_5_1_prompt.md: mirror the same clarification to keep behavior consistent across GPT-5.1 and GPT-5.2.

Related issues

Fixes #7900
Related: #5264, #6277

Test plan

N/A (prompt-only change)

Notes

• If you don’t want auto-closing, use “Addresses GPT-5.2 not clearly prompted they shouldn't end conversation turn for tool preambles. #7900” instead of “Fixes GPT-5.2 not clearly prompted they shouldn't end conversation turn for tool preambles. #7900”.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[4erdenko]

I have read the CLA Document and I hereby sign the CLA