--offline swaps the live model client for a ReplayClient that answers from canned fixtures — no network, no API keys, no cost, fully deterministic. The demo scaffold uses it for the free first run, and it is the cheapest way to iterate on config plumbing, evaluators, and slices before spending money on live calls.

evalshift run --offline --fixtures fixtures.jsonl

## Fixture format

Fixtures are JSONL — each line pairs a model with a prompt-substring matcher:

{"model": "gemini/gemini-2.5-flash",
 "match": "User account_42 had 5 failed login attempts",
 "kind": "tools",
 "result": {"calls": [{"tool_name": "notify_security_team",
                        "arguments": {"severity": "high", "summary": "..."},
                        "sequence_index": 0}],
            "final_text": null,
            "input_tokens": 240, "output_tokens": 48,
            "cost_usd": 0.0011, "latency_ms": 760}}

Matching rules:

• +The canonical model id must match, and match must be a substring of the rendered prompt — for multi-turn examples, of the last user message only.
• +Exactly one fixture may match; zero or several is an error (an ambiguous match names the offending fixture line numbers).

## Recording fixtures

Record fixtures from a live run:

python scripts/capture_fixtures.py <run-id> --suite golden.jsonl --out fixtures.jsonl

## Showcase scenarios

The CLI repo’s examples/showcase/ ships four offline report-showcase scenarios (pass-clean, fail-dropped-tool, fail-argument-drift, mix); scripts/run_showcase.sh --offline --only pass-clean --open runs one end to end.