CLI evidence snapshots

Golden-file regression tests for the exact text the CLI prints.

File	Role
`evidence_cases.ts`	The registry (`REGISTRY`, ~110 cases): command line, description, category, the mock that makes it offline, golden file name. Also `normalize`, `formatResult`, `runCase`.
`test_evidence_snapshots.test.ts`	Port of `test_evidence_snapshots.py`. One `test_evidence_snapshot [<golden>]` per offline-deterministic case; compares the normalised stdout byte for byte with `golden/<name>.txt`.
`capture_cli_evidence.capture.ts`	The capture driver behind `scripts/capture-cli-evidence.mjs` (Markdown evidence documents).
`golden/`	The committed snapshots, copied unchanged from the Python repository.

Category A cases need nothing but the CLI; category B cases drive a mocked backend (Sandbox.create/connect, template API, MCP IDE config map, API-key lookup, ...). The mocks are installed per case and undone afterwards, so the snapshots never touch the network, the real home directory or real credentials.

Running

npx vitest run tests/test_cli_evidence          # compare every case with its golden file
ABX_UPDATE_EVIDENCE=1 npx vitest run tests/test_cli_evidence   # (re)write golden files instead of asserting

Only refresh a golden file when the output change is intentional and reviewed: the files are the byte-level contract with the Python CLI.

Capture script

node scripts/capture-cli-evidence.mjs                      # offline + mocked; writes .agents/evidence/cli/*.md
node scripts/capture-cli-evidence.mjs --command create     # only cases whose golden name starts with "create"
node scripts/capture-cli-evidence.mjs --out <dir>          # write somewhere else
node scripts/capture-cli-evidence.mjs --live               # also run un-mocked category-B cases (needs E2B_API_KEY)

The cases rely on in-process spies, so the script runs them inside vitest (scripts/capture-cli-evidence.vitest.config.mjs) and passes its options through ABX_CAPTURE_OPTIONS. --live talks to the real platform with the credentials in the environment and is never part of npm test. If the output directory already holds a hand-curated README.md, the generated index is written to README.generated.md instead.