Subagent model: gpt-5.4-mini
Subagent reasoning: medium
Orchestrator model: gpt-5.4
Orchestrator reasoning: medium
Search: codex --search
Session root: .agent-docs/qa-session/

codex --help
codex exec --help

codex --search \
  -m gpt-5.4-mini \
  -c model_reasoning_effort=medium \
  -a never \
  exec \
  --skip-git-repo-check \
  --sandbox read-only \
  -o /tmp/codex-web-check.txt \
  "Use native live web search. What is the headline and URL of the newest item currently listed on the OpenAI News page? Answer in one sentence with the date if shown."

mkdir -p .agent-docs/qa-session

.agent-docs/qa-session/001-question.md
.agent-docs/qa-session/002-question.md

.agent-docs/qa-session/001-answer-correct.md
.agent-docs/qa-session/002-answer-not-clear.md

undefined

Answer template:

```md

Keep answers short. Orchestrator context is finite; verbose answers pollute the
merge loop.

Search up to 50 URLs if needed.
Search up to 50 keyword/search-term variants if needed.
Use fewer when the evidence is already strong.

library X changelog feature Y
library X GitHub issues feature Y
library X documentation Y API
library X Y workaround OR alternative
library X Y release notes

cat > .agent-docs/qa-session/001-prompt.txt <<'PROMPT'
You are a Codex research subagent using native Codex web search only.
Do not use MCPs, browser plugins, custom scrapers, or external research tools.

Read .agent-docs/qa-session/001-question.md.
Choose exactly one status suffix from: correct, findings, incorrect,
not-clear, timeout.
Write exactly one answer file:
.agent-docs/qa-session/001-answer-<chosen-status>.md
Do not include the angle brackets literally.

Before searching, decompose the question into 3-5 narrower sub-questions. Do
not search the top-level question directly as your only search. Search up to
50 URLs and up to 50 keyword variants if needed; you may need fewer. Prioritize
the highest-signal search angles based on your own experience.

Use the required answer template. Keep it concise. No preamble, no extra
commentary, no markdown headers beyond the template.

Your answer file must contain these fields:
Question: <copy from question file>
Status: <chosen status>
Answer: <single line>
Confidence: <high|medium|low>
Sources:
- <URL or exact search term> - <what it proved>
Notes:
<optional, only if material>

If the answer likely does not exist, stop spinning and write not-clear or
incorrect with the best evidence.
PROMPT

codex --search \
  -m gpt-5.4-mini \
  -c model_reasoning_effort=medium \
  -a never \
  exec \
  --skip-git-repo-check \
  --sandbox workspace-write \
  -C "$PWD" \
  -o .agent-docs/qa-session/001-last-message.txt \
  - < .agent-docs/qa-session/001-prompt.txt

cd /path/to/project
mkdir -p .agent-docs/qa-session

for q in .agent-docs/qa-session/*-question.md; do
  base="${q%-question.md}"
  n="$(basename "$base")"
  prompt=".agent-docs/qa-session/${n}-prompt.txt"
  last=".agent-docs/qa-session/${n}-last-message.txt"

  cat > "$prompt" <<PROMPT
You are a Codex research subagent using native Codex web search only.
Do not use MCPs, browser plugins, custom scrapers, or external research tools.
Read ${q}.
Choose exactly one status suffix from: correct, findings, incorrect, not-clear, timeout.
Write exactly one answer file: ${base}-answer-<chosen-status>.md
Do not include the angle brackets literally.
Before searching, decompose the question into 3-5 narrower sub-questions.
Search up to 50 URLs and up to 50 keyword variants if needed.
Your answer file must contain:
Question: <copy from question file>
Status: <correct|findings|incorrect|not-clear|timeout>
Answer: <single line>
Confidence: <high|medium|low>
Sources:
- <URL or exact search term> - <what it proved>
Notes:
<optional, only if material>
PROMPT

  (
    codex --search \
      -m gpt-5.4-mini \
      -c model_reasoning_effort=medium \
      -a never \
      exec \
      --skip-git-repo-check \
      --sandbox workspace-write \
      -C "$PWD" \
      -o "$last" \
      - < "$prompt"
  ) >> .agent-docs/qa-session/_dispatch.log 2>&1 &

  while [ "$(jobs -pr | wc -l | tr -d ' ')" -ge 8 ]; do
    sleep 5
  done
done

wait

delay=60
while jobs -pr | grep -q .; do
  sleep "$delay"
  echo "[progress] Checking answer files..."
  ls .agent-docs/qa-session/*-answer-*.md 2>/dev/null | sort || true
  echo "[progress] Pending questions:"
  for q in .agent-docs/qa-session/*-question.md; do
    base="${q%-question.md}"
    ls "${base}"-answer-*.md >/dev/null 2>&1 || echo "PENDING: $q"
  done
  case "$delay" in
    60) delay=120 ;;
    120) delay=240 ;;
    240) delay=480 ;;
    *) delay=600 ;;
  esac
done

for q in .agent-docs/qa-session/*-question.md; do
  base="${q%-question.md}"
  ls ${base}-answer-*.md 2>/dev/null || echo "MISSING: $q"
done

out=.agent-docs/qa-session/_summary.tsv
printf 'id\tstatus\tanswer\tconfidence\tfile\n' > "$out"

for q in .agent-docs/qa-session/*-question.md; do
  base="${q%-question.md}"
  id="$(basename "$base")"
  answer_file="$(ls -t "${base}"-answer-*.md 2>/dev/null | head -1)"
  if [ -z "$answer_file" ]; then
    printf '%s\tmissing\t\t\t\n' "$id" >> "$out"
    continue
  fi
  answer_status="$(basename "$answer_file" | sed -E 's/^[0-9]+-answer-(.*)\.md$/\1/')"
  answer="$(grep -m1 '^Answer:' "$answer_file" | sed 's/^Answer:[[:space:]]*//')"
  confidence="$(grep -m1 '^Confidence:' "$answer_file" | sed 's/^Confidence:[[:space:]]*//')"
  printf '%s\t%s\t%s\t%s\t%s\n' "$id" "$answer_status" "$answer" "$confidence" "$answer_file" >> "$out"
done

echo "questions: $(ls .agent-docs/qa-session/*-question.md | wc -l | tr -d ' ')"
echo "answers:   $(ls .agent-docs/qa-session/*-answer-*.md 2>/dev/null | wc -l | tr -d ' ')"
grep -E $'\t(timeout|not-clear|missing)\t' .agent-docs/qa-session/_summary.tsv || true
grep -L '^Answer:' .agent-docs/qa-session/*-answer-*.md 2>/dev/null || true