🪷 शरण

sharana

Ramayana — Vibhishana's surrender

Divine Refuge — Sharanagati to the Lord

Fallback strategies for LLM agents. Model fallback, response fallback, default values.

Formerly agent-fallback — Part of the Vedic Arsenal: 100 production-grade Python libraries for LLM agents, each named from the Vedas, Puranas, and Mahakavyas.

When Vibhishana, brother of Ravana, came to Ram's side — renouncing Lanka, renouncing his own blood — Ram accepted him with these words: "Whoever comes to me for refuge, I shall protect."

This is Sharana — divine refuge, unconditional fallback.

sharana implements the fallback chain for LLM systems. When the primary model fails, sharana catches you. When GPT-4 is rate-limited, Claude catches you. When Claude is unavailable, the graceful degradation catches you. Like Ram's promise to Vibhishana, the fallback chain never abandons the task — it finds another way forward, every time.

flowchart LR
    A[Task] --> B[Primary LLM]
    B -- Fail --> C[sharana fallback]
    C --> D[Secondary LLM]
    D -- Fail --> E[Tertiary LLM]
    E -- Fail --> F[Graceful Degradation]
    style C fill:#ec4899,color:#fff

pip install sharana

Or from source:

git clone https://github.com/darshjme/sharana.git
cd sharana && pip install -e .

from sharana import *

# See examples/ for full usage

sharana is one of 100 libraries in darshjme/arsenal — each named from sacred Indian literature:

Each library solves one problem. Zero external dependencies. Pure Python 3.8+.

1. Fork the repo
2. Create feature branch (git checkout -b fix/your-fix)
3. Add tests — zero dependencies only
4. Open a PR