[JacobStephens2]

Summary

Setup now has its own module. Its rules sit in front of a private Outside seam. The User config module keeps parsing, loading, the defaults and completing.

 src/
 ├── config.rs           # parse, load, DEFAULTS, complete, home and path
-│                       #   + Setup flow, writing answers, suggested address
-├── questions.rs        # Setup's questions, on the terminal directly
+├── setup.rs            # setup() and offer(): the flow, writing answers, suggested address
+│                       #   trait Outside: OnMachine (real), Scripted (tests)
+└── setup/
+    └── questions.rs    # the questions, asked through Outside

setup() / offer()                 # same two entry points, no arguments
  OnMachine::new(home)            # terminal, echo-off, claude test call, codex catalog,
                                  # Credentials, gh, git, User config file, Resend
  run_setup / run_offer(outside)  # rules: re-ask loops, Harness default, Codex settling,
                                  # noreply filter, write order, the offer's warnings

Evidence

• Before: about 70 pseudo-terminal end-to-end tests across tests/setup.rs (56) and tests/setup_offer.rs (14) were the only way to check a Setup rule.
  After: 59 in-process tests in src/setup.rs drive Setup through Scripted. That adapter checks each prompt in order and records every line said and every write, save and send. The end-to-end suites keep only the 16 tests for wiring and the real terminal that the issue lists. cargo test passes in full, and cargo clippy --all-targets is clean.

Scripted::answering([(OFFER,"y"), …, (KEY_PROMPT, KEY), (TEST_EMAIL,"y")])
offer() effects == [WriteNew, Step("wrote the User config …"), SaveKey(KEY),
                    Step("wrote the Credentials …"), SendTest{to: me@example.com}, Step(SENT)]

Two rules from deleted end-to-end tests live outside Setup, so they moved to unit tests in the module that owns them:

• every default key carries a comment giving its default (config.rs);
• a key is saved into existing Credentials in place (resend_key.rs).

Merge Danger

Door: two-way

This is a refactor with no change to prompts, lines, files or exit codes, so a revert is clean.

Blast Radius: Setup

The risk is a behaviour drift on a real terminal that the deleted end-to-end tests would have caught. The kept tests cover Ctrl-C, hidden key input, Enter throughout, and the offer's wiring into a Run.

Unaddressed findings

Standards

• Outside is wide (16 methods); read/read_hidden, say/step and the two User config reads could merge. Skipped: the issue lists these exact methods, and it keeps hidden input as its own method on purpose.
• Feature Envy: set/set_email_to use config's TOML helpers, which became pub. Skipped: the issue says writing the answers, setting a key in place and setting email.to move to Setup, while completing stays in config.
• E2E with_no_terminal_and_no_user_config_setup_writes_the_defaults and on_a_terminal_pressing_enter_throughout… duplicate in-process names. Skipped: the issue's keep list names both. Each checks the real file and terminal, not only the rule.
• questions calls the parent's suggested_address. Skipped: the lookup must stay lazy, made only when notifications are on and there is no email.to. A child module using its parent's private helper is ordinary in Rust.
• Scripted uses &'static str for scripted failures. Skipped: this is test-only and matches the other scripted adapters.

Spec

• Claude's test call prints "checking the Model … with a test call" through progress::step, not Outside::step. Skipped: the real adapter makes the call "as the code does today" through Choice::test_call, which a Run shares. The line is unchanged.
• No in-process check that the test email uses the key just entered. Skipped: the scripted adapter covers the order (save, then send). The key reaching Resend is the real adapter reading the Credentials, which is not a Setup rule.

Closes #411

🤖 Generated with Claude Code