Fix: strategic_questions block not removed when ExplorationPlanner suggestions unavailable

[francisjervis]

Problem

get_prompt("normal") performs a two-stage template expansion:

1. Stage 1 (inside get_prompt): format_prompt(INTERVIEW_PROMPT, {...}) replaces the uppercase {STRATEGIC_QUESTIONS} token with the STRATEGIC_QUESTIONS string, which itself contains the lowercase placeholder {strategic_questions}.

2. Stage 2 (inside _get_prompt): format_prompt(main_prompt, format_params) fills in runtime values.

The staleness-removal code ran between these two stages:

main_prompt = main_prompt.replace("\n{STRATEGIC_QUESTIONS}\n", "\n")

But by then the uppercase token is already gone — this replace is a no-op. The lowercase {strategic_questions} placeholder remains in the prompt. When the strategic_questions key is absent from format_params (because _should_include_strategic_questions() returned False), format_prompt's missing-key fallback re-inserts it verbatim, and the LLM receives the literal string {strategic_questions} inside the <strategic_questions> block.

This affects every turn before the ExplorationPlanner completes its first run (turns 2–3 by default) and any future stale periods.

Fix

Replace the no-op string replace with re.sub that strips the already-expanded <strategic_questions>…</strategic_questions> XML block:

main_prompt = re.sub(
    r"\n<strategic_questions>.*?</strategic_questions>",
    "",
    main_prompt,
    flags=re.DOTALL,
)

re is already imported in interviewer.py. No other files are changed.

Verification

Confirmed with a unit test that:

• After the fix, the block is completely absent from the prompt when suggestions are unavailable
• When suggestions are available (key present in format_params), the block is still filled correctly

---

🤖 Generated with Claude Code