[memtomem]

Fixes #254.

What changes

Without a terminal, syncmill start and syncmill start -a <agent> (no --task) refuse and list the commands to use instead. Both used to raise a plain ClickException whose text embedded \n before _PROJECT_STEPS. The CLI error boundary escapes a ClickException message to one line, so the guidance came out joined by a literal \x0a. Both now raise _MultilineError with the sentence and each step as separate lines. Each line is still escaped on its own, so a value inside a line cannot start another.

Tests

• tests/test_start.py checked only "start needs --task" in output, which passed on the broken output. Both refusals are now compared line by line against the exact expected output: Error: <sentence> followed by every _PROJECT_STEPS line, with no \x0a anywhere.
• New test for the bare-start refusal, which had no text assertion.
• Both tests fail with the cli.py change reverted and pass with it.

Verification

• Full suite: suite-run CLEAN on macOS. The helper was tightened afterwards; tests/test_start.py still passes, 34/34.
• The real binary, run through the [product] Guide the first verified change with an offline trial and task-first entry #171 smoke harness (stdin is a pipe that never closes): both refusals print on separate lines and stdin is never read.
• Codex review: round 1 SHIP. Its one in-scope Minor was adopted: assert the exact ordered lines rather than membership. Round 2 was unprimed (diff only): SHIP, no findings.

Not in this PR

start --trial failures flatten the same way. TrialError messages join subprocess output and the "trial folder was kept" note with \n, and cli.py wraps them in a plain ClickException. That message carries external output, so each line needs its own escaping; it is filed as a separate issue.

🤖 Generated with Claude Code