sybil

Sybil — The Oracle of Escalation

Sybil was the prophetess of Delphi — the voice you consulted when you were truly stuck. In Homer, she is the only layer allowed to make external API calls mid-run. Expensive, precise, rare.

Role

Sybil is Homer's escalation path. Zeus invokes Sybil when:

1. A MUSES worker returns ROI=0 (empty output, timeout, or unusable) on a research / synthesis subtask
2. Zeus Phase 3 synthesize finds a plan-coverage gap no muse addressed
3. Two muses disagree on a factual claim and Zeus cannot adjudicate
4. A correction loop is detected (2+ user corrections on the same problem)

Sybil invokes Anthropic's advisor_20260301 API via the brain advise CLI. The advisor pattern runs Sonnet as executor with Opus as advisor on escalation — all in one /v1/messages request. See research/brain_v2_sota_synthesis_2026-04-10.md Part 1 for the full architecture.

Cost Cap (hard)

Max 2 Sybil escalations per Homer session_id. Enforced by sybil.py via per-session state file at Toke/automations/homer/sybil/.state/session_<session_id>.json. Approximate cost ceiling per session: ~$0.60 worst-case (each call ~$0.15-0.30 depending on advisor iterations).

If Sybil would be the 3rd escalation in a session, it refuses and Zeus escalates to L4 (ask the user) instead.

Preconditions (all must pass)

Check	Why
ANTHROPIC_API_KEY set in env	Required for Anthropic API calls
brain_cli.py reachable at expected path	Sybil delegates through Brain's CLI
brain help mentions advise subcommand	Verifies the CLI path works end-to-end
Task is NOT creative content	Sacred Rule #6 — no advisor for lore / dialogue / GDD narrative / song / verse
Session cap not hit	Hard $ ceiling — 2 escalations per session

Any failure → Sybil refuses silently, Zeus falls through to L4 (ask the user).

Invocation Contract

Zeus invokes Sybil programmatically — Sybil is a Python module, not a Claude-side tool.

# From Zeus or integration code (with sybil/ on sys.path):
from sybil import escalate

result = escalate(
    stuck_task="<what Zeus was trying to accomplish>",
    approaches_tried=["<muse 1 output summary>", "<muse 2 output summary>"],
    blocker="<specific failure symptom>",
    session_id="<Homer session id>",
    max_uses=3,
    max_tokens=4096,
)

if result["ok"]:
    advisor_guidance = result["advisor_stdout"]
    # Feed guidance back into Zeus plan revision (Phase 1 re-plan or Phase 3 re-synthesize)
else:
    # Fall through to L4 (ask the user)
    reason = result["reason"]