Consult Fable (bounded Claude Fable 5 advisory)

Ask Claude Fable 5 a question and get an answer back through bounded backend
attempts. This replaces ad-hoc timeout 120 claude -p "..." calls, which hang
or expire with no output. The default backend order is claude-fable-5 then
claude-opus-5 through the local claude CLI. CLI attempts pass the prompt
via stdin and use an empty MCP configuration. Explicit
ARAGORA_MODEL_TRANSPORT=vibeproxy-prefer tries those exact models through
VibeProxy before the CLI; vibeproxy-required fails closed instead of using
CLI or paid API fallbacks.

--timeout is the maximum for one backend attempt. By default, the total
consult budget is derived from every enabled attempt, so a full-timeout
VibeProxy attempt cannot consume the CLI fallback budget. Pass
--overall-timeout to set an explicit cap shared across all attempts.
Prompts from inline args, --prompt-file, and stdin are capped at 512 KiB and
are rejected before any backend attempt, so oversized context cannot
silently burn API tokens when --api-fallback is enabled. The cap also applies
to programmatic consult() calls after any --system/system preamble is
combined with the user prompt.

API fallback is off by default because it can consume Anthropic API credits or
billing when ANTHROPIC_API_KEY or aragora secrets are configured. Use
--api-fallback only when paid API fallback is acceptable; unsupported
CLI-only model ids such as claude-fable-5 are skipped for direct API calls.
If no API key is available, the API backend records a normal failed attempt in
the JSON envelope instead of being omitted.

How to invoke

From the repo root (any worktree):

# Inline question, defaults: model claude-fable-5, timeout 600s per attempt
# Default total budget is timeout multiplied by enabled backend attempts.
python3 scripts/consult_claude.py "One-paragraph question with the live state inlined."

# Long prompt from a file, machine-readable result
python3 scripts/consult_claude.py --prompt-file /tmp/question.md --json

# Require the local proxy and prohibit transport fallback
ARAGORA_MODEL_TRANSPORT=vibeproxy-required \
  python3 scripts/consult_claude.py --json "Reply with exactly READY"

# Prefer the local proxy, then fall back to the default CLI path
ARAGORA_MODEL_TRANSPORT=vibeproxy-prefer \
  python3 scripts/consult_claude.py --json "Reply with exactly READY"

# Bigger total budget for a hard question
python3 scripts/consult_claude.py --timeout 300 --overall-timeout 1200 --prompt-file /tmp/question.md

# Explicit paid API fallback after CLI attempts fail
python3 scripts/consult_claude.py --api-fallback --prompt-file /tmp/question.md

--json returns {ok, model, backend, elapsed_s, text, attempts}. Exit codes:
0 ok, 2 timed out, 3 no prompt, 4 all backends failed or the explicit
overall budget was exhausted before all fallback attempts could run, 64
usage/config error. In JSON, budget exhaustion is reported as
budget_exhausted: true, distinct from backend timed_out: true.

If scripts/consult_claude.py does not exist in your checkout yet, use an
installed user skill copy such as ~/.codex/skills/consult-fable/consult_claude.py.
The tracked repo skill itself expects the tracked repo script.

Writing a good consult prompt

Fable has no access to your session state. Inline everything it needs:

• The decision you are facing and the 2-4 options you see.
• Live facts, verified this cycle: PR numbers with head SHAs, gate/quorum
  state, dissent findings, owner state. Never paste stale transcript claims.
• Constraints that bind you (operating contract tier limits, no --admin,
  shared-root read-only, anti-treadmill rules).
• Ask for a single recommendation with reasoning, not a survey.

For "what should I work on next?" consults, a good shape is: list the
candidate progress units with one line of live state each, state the
anti-treadmill rules that apply, and ask Fable to pick exactly one and say why.

Hard rules

1. Advisory only. The answer is one input to your decision. It is never
   authority to merge, settle, post evidence, or bypass a gate. Verify any
   factual claim in the answer against live repo state before acting on it.
   VibeProxy does not create a reviewer family, and this skill's output is not
   countable merge-quorum evidence.
2. One consult per decision. Do not loop consults on the same question;
   if the answer doesn't unblock you, the blocker is information, not advice.
3. Bounded always. Keep the default 600s per-attempt timeout or raise it
   explicitly; never wrap this tool in an outer retry loop.
4. No secrets in prompts. Inline repo state, not credentials or tokens.

Failure handling

• Exit 2 (timeout) or 4 (all backends failed): report the consult as
  unavailable and proceed with your own judgment. Do not retry more than once.
• Exit 64 is a usage or configuration error. Correct the invalid argument or
  ARAGORA_MODEL_TRANSPORT value before retrying; do not treat it as backend
  unavailability.
• In vibeproxy-prefer, proxy unavailability is recorded in attempts before
  the CLI fallback runs. In vibeproxy-required, it is the final failure.
• If only the API backend fails with a missing-key error, the local claude
  CLI is the expected path; check which claude before concluding anything.