Qwen이 로그라이크 게임 DCSS를 플레이하다! (그리고 꽤 잘함)

DCSS가 ~/.crawl/log/에 깔끔한 실시간 턴별 텍스트 로그를 작성한다고 가정하지 마십시오. 그 방식은 로컬 DCSS에서 잘못되었거나 신뢰할 수 없는 것으로 보입니다. DCSS는 curses/tiles 게임이며 stdout/stderr 캡처는 유용한 턴 로그가 아닙니다.

Use the official DCSS-supported mechanisms instead:

1. Use screenshots as the primary visual state source.

- After every player action, capture a screenshot of the DCSS window.

- This gives the bot the actual map, messages, HP/MP, monster positions, inventory popups, etc.

2. Use character dumps as the primary text state source.

- In DCSS, pressing "#" writes a character dump to the morgue directory.

- Configure DCSS init/crawlrc so dumps are useful for bot parsing.

- The options to set/check are:

- dump_on_save = true

- dump_message_count = 100 or higher

- morgue_dir = /home/snoop/.crawl/morgue

- dump_order should include at least:

header, stats, misc, inventory, skills, spells, overview, mutations, messages, screenshot, monlist, notes

- The bot should press "#" after relevant turns, then read the newest .txt file from the morgue directory.

3. Use Ctrl-P only as a fallback for message history.

- Ctrl-P opens previous messages in-game.

- If the dump does not contain enough recent messages, capture a screenshot of the Ctrl-P screen and parse it visually.

4. Recommended hybrid loop:

- Send a key/action to DCSS via xdotool.

- Wait briefly for the game to update.

- Capture screenshot to /tmp/dcss_hermes/screen.png.

- Press "#" to generate/update a character dump.

- Find the newest dump file in /home/snoop/.crawl/morgue/.

- Copy it to /tmp/dcss_hermes/char_dump.txt.

- Extract the last messages and key status from the dump.

- Return both:

a) the screenshot

b) a concise text summary:

- HP/MP

- XL / level / branch

- visible threats

- last messages

- inventory-relevant discoveries

- suggested safe actions

5. Do not rely on OCR as the only source.

- Prefer parsing the character dump for text.

- Use screenshot/vision for map and tactical layout.

6. Build a small test script first.

- It should create /tmp/dcss_hermes/

- It should capture the screenshot.

- It should trigger "#".

- It should locate the newest morgue dump.

- It should copy the dump and create a short tail summary.

Example script:

#!/usr/bin/env bash

# Capture a hybrid DCSS state for bot-controlled remote play.

set -euo pipefail

OUT_DIR="/tmp/dcss_hermes"

MORGUE_DIR="$HOME/.crawl/morgue"

mkdir -p "$OUT_DIR"

# Capture the current DCSS screen.

DISPLAY=:0 flameshot full -p "$OUT_DIR/screen.png" >/dev/null 2>&1 || true

# Ask DCSS to write a character dump.

# In DCSS, "#" is the character dump command.

DISPLAY=:0 xdotool key numbersign

sleep 0.4

# Find newest character dump.

LATEST_DUMP="$(ls -t "$MORGUE_DIR"/*.txt 2>/dev/null | head -1 || true)"

if [ -n "$LATEST_DUMP" ]; then

cp "$LATEST_DUMP" "$OUT_DIR/char_dump.txt"

tail -120 "$LATEST_DUMP" > "$OUT_DIR/summary_tail.txt"

echo "OK"

echo "Screenshot: $OUT_DIR/screen.png"

echo "Dump: $OUT_DIR/char_dump.txt"

echo "Summary tail: $OUT_DIR/summary_tail.txt"

else

echo "WARN: no character dump found in $MORGUE_DIR"

echo "Check DCSS morgue_dir setting and whether '#' worked inside the game window."

fi

7. Before implementing the Telegram/Discord gameplay loop, first verify:

- Which DCSS binary is used: /usr/games/crawl or another path.

- Whether the game window receives xdotool keys.

- Where the actual morgue directory is.

- Whether pressing "#" updates a dump file during a live game.

- Whether dump_message_count is large enough.

Expected final architecture:

- Screenshot = tactical map source.

- Character dump = structured text/status source.

- Ctrl-P screenshot = fallback for extra message history.

- No fake ~/.crawl/log live-log dependency.