exec resume: input_tokens grow by ~18K per turn regardless of message size (system prompt re-injected every resume)

Bug Report

Summary

When resuming a session with codex exec resume --last, the input_tokens count grows by approximately 18,000 tokens on every turn — regardless of how short the actual user message is. After 3 turns with single-sentence messages, the token count has tripled.

Steps to Reproduce

# Turn 1 — start a new session
echo "the cat said meow" | codex exec --json -

# Turn 2 — resume and ask a short follow-up
echo "what did the cat say?" | codex exec resume --last --json -

# Turn 3 — another short follow-up  
echo "and what was the cat's name?" | codex exec resume --last --json -

Observed Output (from turn.completed usage field)

| Turn | input_tokens | Message length |
|------|-------------|----------------|
| 1 | 18,080 | ~5 words |
| 2 | 36,197 | ~5 words |
| 3 | 54,341 | ~6 words |

Each turn adds ~18,000 tokens. The user messages are 5–6 words each (~10 tokens). The delta cannot be explained by conversation history alone.

Root Cause (from session JSONL)

Looking at ~/.codex/sessions/.../rollout-*.jsonl, the session file stores the system/developer instructions as response_item entries inline in the conversation history. When exec resume loads this file, it replays all stored items (including the system instructions) and then prepends a fresh copy of the system instructions for the new turn.

The result is that each resumed turn contains one additional copy of the full system prompt (~18K tokens) compared to the previous turn:

Turn 1 input: [system_instructions, user1]
Turn 2 input: [system_instructions_fresh, system_instructions_from_history, user1, assistant1, user2]  
Turn 3 input: [system_instructions_fresh, system_instructions_from_history×2, user1, assistant1, user2, assistant2, user3]

Expected Behavior

The system prompt should appear once per API call, regardless of how many turns have been resumed. A 3-turn conversation with short messages should cost roughly:

Turn 1: ~18K tokens
Turn 2: ~18K + message_delta (~100 tokens)
Turn 3: ~18K + message_delta (~200 tokens)

Not 3× the base cost.

Impact

• Token usage grows linearly with number of resumed turns, not with conversation content
• The repeated system prompt also breaks prompt caching — each turn has a different prefix so cached tokens from prior turns cannot be reused
• Long sessions become disproportionately expensive

Environment

codex-cli 0.117.0
Windows 11 / x86_64
model: gpt-5.4 (reproduced, also gpt-4o-mini)