name	oracle
description	Ask the Oracle — a one-shot Fable answer from a stripped headless session (~1.3k tokens of overhead vs ~6k for the oracle subagent). Use when the user says "ask the oracle", "oracle,", or "/oracle <question>".

Send the question to a bare headless Fable session, then present its answer to the user in your own words, along with the token/cost stats for that call.

Never put a heredoc inside $(...) (i.e. q=$(cat <<'EOF' ... EOF)). On macOS's stock /bin/bash (3.2 — an old pre-GPLv3 build Apple never upgraded), a heredoc nested in a command substitution breaks the parser the instant the body contains a single quote (don't, it's, any contraction) — unexpected EOF while looking for matching. It's a real bash 3.2 bug, not a wording problem, and it reproduces even for a one-word heredoc; it doesn't happen in zsh or bash 4+, which is why it's intermittent depending on what's invoking the script.

Instead, write the question to a plain text file with the Write tool (no shell parsing of the content at all, so quotes/backticks/$ are inert), then read it back with a bare $(cat file) — safe because there's no heredoc involved.

Write the question, plus only the minimal context Oracle needs, to <scratchpad>/oracle_question.txt (use your actual scratchpad path) with the Write tool — plain text, no heredoc, no escaping.
Run with the Bash tool:

Q=<scratchpad>/oracle_question.txt
mode=--safe-mode
if [ -n "$ANTHROPIC_API_KEY" ]; then mode=--bare; fi
out=$(claude -p $mode --model fable --output-format json --tools "WebSearch" --allowedTools "WebSearch" < /dev/null \
  --system-prompt "You are Oracle. Answer like a telegraph operator paying by the word: cut every word that carries no decision-relevant information — filler, throat-clearing, restated context, hedges that would not change what the asker does. Use compact notation where it stays unambiguous: & or + for and, w/ for with, w/o for without, b/c for because, vs for versus, numerals instead of spelled-out numbers, standard abbreviations (e.g., i.e., approx., hrs, min, etc.). Join clauses with a comma instead of and or but where the meaning stays clear. Keep every fact, number, caveat, and recommendation that changes what the asker does. Full sentences only where needed for clarity; fragments fine otherwise. Lead with the verdict. Flag uncertainty only when it would change what the user does next. When a question is ambiguous, answer the most likely reading. Search the web only when the answer needs facts newer or more specific than your knowledge; otherwise answer directly." \
  "$(cat "$Q")")
echo "$out" | jq -r '.result'
echo "$out" | jq -r '(.usage.input_tokens + .usage.cache_creation_input_tokens + .usage.cache_read_input_tokens) as $in | (.total_cost_usd * 1000000 | round / 1000000) as $cost | "[\($in) in / \(.usage.output_tokens) out tokens · $\($cost)]"'

Notes: --tools "WebSearch" stays unconditional — the schema costs a measured ~600 input tokens (evals Phase 0), which is noise; omitting the flag loads the CLI's ~17k-token default toolset instead. < /dev/null is required — claude -p otherwise stalls 3s waiting for stdin and prints a warning into the output. --bare (API key, off-quota) is used when ANTHROPIC_API_KEY is set, else --safe-mode (OAuth, on-quota). --output-format json is required to get cost/usage data back instead of plain text; the two jq calls split the JSON into the answer line and the stats line.

Rules:

Every token in oracle_question.txt is Fable-priced. Pass the question and the few lines of context it genuinely needs — never file dumps or conversation history.
Oracle can search the web but has no file access. If the question needs a file's contents, read the file first and paste the relevant lines into oracle_question.txt.
Rewrite Oracle's answer in your own words before presenting it — keep every fact, number, caveat, and recommendation intact, and do not add anything it did not say or pad it out longer than the original. Follow it, on its own line, with the stats line the script prints ([N in / M out tokens · $X]) unchanged.
If the command fails (auth, sandbox), fall back to the oracle subagent via the Agent tool and note the fallback costs ~6k tokens.

Q=<scratchpad>/oracle_question.txt mode=--safe-mode if [ -n "$ANTHROPIC_API_KEY" ]; then mode=--bare; fi out=$(claude -p $mode --model fable --output-format json --tools "WebSearch" --allowedTools "WebSearch" < /dev/null \ --system-prompt "You are Oracle. Answer like a telegraph operator paying by the word: cut every word that carries no decision-relevant information — filler, throat-clearing, restated context, hedges that would not change what the asker does. Use compact notation where it stays unambiguous: & or + for and, w/ for with, w/o for without, b/c for because, vs for versus, numerals instead of spelled-out numbers, standard abbreviations (e.g., i.e., approx., hrs, min, etc.). Join clauses with a comma instead of and or but where the meaning stays clear. Keep every fact, number, caveat, and recommendation that changes what the asker does. Full sentences only where needed for clarity; fragments fine otherwise. Lead with the verdict. Flag uncertainty only when it would change what the user does next. When a question is ambiguous, answer the most likely reading. Search the web only when the answer needs facts newer or more specific than your knowledge; otherwise answer directly." \ "$(cat "$Q")") echo "$out" | jq -r '.result' echo "$out" | jq -r '(.usage.input_tokens + .usage.cache_creation_input_tokens + .usage.cache_read_input_tokens) as $in | (.total_cost_usd * 1000000 | round / 1000000) as $cost | "[\($in) in / \(.usage.output_tokens) out tokens · $\($cost)]"'