Changelog¶

The format is based on Keep a Changelog, and this project uses semantic versioning.

Unreleased ¶

Subcommands run, config show, models, export, and analyze, with echochain --start kept as echochain run --start.
Settings from a TOML file and ECHOCHAIN_* variables, with CLI flags taking precedence.
Seed files with per-seed claims, selected by id with --seed.
Schedules alternate, swap, factorial, latin, and fixed:<model>, plus same-model, identity, and paraphrase controls.
--repeats for batches, and --resume to continue an interrupted run or batch.
A manifest.json and one JSONL row per step for every run, described by the JSON Schemas in schemas/.
Claim fidelity, an anchored 1 to 5 score, and a comparison with the previous answer on every answer step.
Judge-free length, token-survival, and generic-word metrics, plus cosine and entailment with the baseline extra.
echochain analyze for bootstrap curves, Kaplan-Meier survival, restricted mean time to degradation, and log-rank tests.
--dry-run for call, token, and cost upper bounds, with --yes required above confirm_above_usd.
--budget to stop a run at a USD cap with exit code 3.
Provider pins with order, allow_fallbacks, data_collection, and zdr.
--sampling-seed and top_p are sent on every call and stored on the manifest.
Exit codes 0, 1, 2, 3, and 130, listed in --help.

Question steps are scored for answerability and answer steps for fidelity.
Every chat call requires the requested parameters and disables provider fallbacks unless a pin allows them.
Chat calls share one HTTP client, retry transient failures, and stop at a per-call deadline.

Runs no longer write chain.csv, so use echochain export --format csv instead.
The backoff dependency.

An unparsable judge reply is stored as no score instead of ending the run.
An empty or truncated reply is retried once instead of being recorded as an answer.

The first PyPI release, installed with pip install echochain.
OPENROUTER_API_KEY from the environment or .env, redacted in logs, errors, and run files.

The project is renamed from llm-telephone to echoChain.
Rotating models are current OpenRouter snapshots, and the judge is GPT-4.1 mini.
Each saved cycle records the requested model and the model that served it.

A fresh install no longer fails on a missing httpx import.
--help and --version work without an API key.
Configured model ids are checked against the OpenRouter catalog before a paid call.