LLMSwap 5.7.0 — One interface. One Best Answer.

LLMSwap 5.7.0 — One interface. One Best Answer.

LLMSwap 5.7.0 adds Best Answer: ask two to five independent LLM drafts, then
synthesize one decision-ready response that keeps agreement, disagreement, and
cautions visible. It works in the Python SDK, CLI, and local Arena.

This release also adds llmswap doctor, a secret-safe readiness check for the
installation, credentials, provider SDKs, configured models, and MCP
prerequisites.

Highlights

• Generate a Best Answer from two to five independent candidate responses.
• Use private-by-default, same-provider synthesis or explicitly consent to
  cross-provider output sharing.
• See normalized token usage, total latency, agreement level, disagreements,
  cautions, judge model, and optional candidate answers.
• Create a Best Answer from an existing Arena comparison with only one
  additional judge request.
• Diagnose Python/CLI mismatches, .env discovery, credentials, dependencies,
  models, MCP requirements, and optional live provider connectivity.
• Reject empty Sarvam and provider-verification responses rather than reporting
  false success.

Install or upgrade

python -m pip install --upgrade llmswap==5.7.0
llmswap --version
llmswap doctor

For the browser Arena:

python -m pip install --upgrade "llmswap[web]==5.7.0"
llmswap web

Try Best Answer

# Three independent drafts and one judge, all using the primary provider/model
llmswap --provider openai best \
  "Which rollout plan has the lowest operational risk?"

# Explicit cross-provider comparison
llmswap --provider openai best "Review this decision" \
  --models openai:gpt-5.6 sarvam:sarvam-105b \
  --allow-cross-provider-sharing --show-candidates

# Automation-friendly output
llmswap --provider openai best "Assess this proposal" --format json

from llmswap import LLMClient

client = LLMClient(provider="openai", model="gpt-5.6")
result = client.best_answer("Which option is most robust and why?")

print(result.best_answer)
print(result.agreement_level)
print(result.disagreements)
print(result.cautions)
print(result.total_usage)

Privacy and cost behavior

Same-provider synthesis is the default. LLMSwap blocks a flow that would send
candidate outputs to a judge on another provider unless the caller explicitly
enables cross-provider sharing. Candidate labels are anonymized before judging,
but prompt and response content are not automatically redacted.

The default Best Answer flow makes four billable provider requests: three
candidate calls in parallel, followed by one judge call. In the Arena, Best
Answer reuses completed comparison responses, so only the judge request is
added. Provider billing and retention policies still apply. Model agreement is
a useful cross-check signal, not proof that an answer is true.

Compatibility

This release is backward compatible with the public 5.6 API. Existing ask,
chat, review, debug, generate, providers, workspace, web, and MCP
workflows continue to work. Best Answer and doctor are additive.

See CHANGELOG.md, MODEL_SUPPORT.md,
and SECURITY.md for the complete details.