Upgrading to 0.4

0.4 removes what 0.3 had marked for removal. If your code ran on 0.3 without warnings, it runs on 0.4 unchanged.

Old module paths are gone

0.3 moved every module into a layered package and kept the 0.2 paths working with a warning. In 0.4 the old paths no longer import; use the new ones:

Old path	New path
`counterfix.adapters`	`counterfix.integrations`
`counterfix.agents`	`counterfix.detect.agents`
`counterfix.background`	`counterfix.runtime.background`
`counterfix.benchmarks`	`counterfix.evals.benchmarks`
`counterfix.capabilities`	`counterfix.integrations.capabilities`
`counterfix.code`	`counterfix.detect.code`
`counterfix.confidence`	`counterfix.detect.confidence`
`counterfix.context`	`counterfix.runtime.context`
`counterfix.detectors`	`counterfix.detect`
`counterfix.engine`	`counterfix.fixes.engine`
`counterfix.mcp_server`	`counterfix.integrations.mcp_server`
`counterfix.metrics`	`counterfix.evals.metrics`
`counterfix.otel`	`counterfix.integrations.otel`
`counterfix.perf`	`counterfix.runtime.perf`
`counterfix.presets`	`counterfix.core.presets`
`counterfix.profiles`	`counterfix.fixes.profiles`
`counterfix.rag`	`counterfix.detect.retrieval`
`counterfix.report`	`counterfix.runtime.report`
`counterfix.rules`	`counterfix.fixes`
`counterfix.rules.store`	`counterfix.core.rules`
`counterfix.safe_eval`	`counterfix._internal.safe_eval`
`counterfix.text`	`counterfix._internal.text`
`counterfix.verify`	`counterfix.fixes.verify`

counterfix.context(), counterfix.evaluate, counterfix.otel, counterfix.perf, counterfix.background and counterfix.metrics remain as short names.

Evaluation features other tools already provide are gone

counterfix explains why a call failed and verifies fixes; it works alongside evaluation tools rather than replacing them. These are removed in 0.4:

Removed	Use instead
metrics `bleu`, `chrf`, `rouge_l`	sacrebleu, rouge-score, or your evaluation tool
metrics `bertscore`, `embedding_similarity`	bert-score, sentence-transformers, or your evaluation tool
metrics `faithfulness`, `answer_relevancy`, `answer_correctness`, `context_precision`, `context_recall`, `citations`	Ragas or DeepEval
metric `pii`, `counterfix.evals.safety`	Microsoft Presidio
metric `toxicity`	Detoxify
benchmarks `mmlu`, `gsm8k`, `arc`, `truthfulqa`, `ifeval`, metric `ifeval`	lm-evaluation-harness
benchmarks `humaneval`, `humaneval+`, `mbpp`, `mbpp+`, metric `evalplus`	EvalPlus
`counterfix.evals.perplexity`	lm-evaluation-harness
extras `metrics`, `semantic`, `safety`, `benchmarks`, `code`	(not needed)
the ignored `need=` option of `fix()` and `speed_fix()`	(drop it)

Staying: the simple checks (exact_match, contains, f1, number, choice, json_valid, json_match, latency, ttft), the label-based retrieval metrics, tests (F12), tool_call (F14), injection (F13), confidence, and the injection and tools probe sets.

To keep a removed tool's verdict in the loop, bring its results file (DeepEval, Ragas and Promptfoo) or wrap it as your own check (Fix with your own test).