[szymdzum]

S28 retro for the ship skill, plus a fix to the eval harness script. No product code.

Tool bugs

• Briefs: the forbidden block bans URLs another app opens (mailto:, tel:, sms:, facetime: and others). In S28 a mailto: redirect cell made macOS Chrome, even headless, open the owner's mail app again and again. ship brief --check now refuses a brief that contains such a URL. A scheme named as a word, like the ban itself, doesn't count. Checked against every S28 brief and every open issue's text (those briefs quote the issue): no false refusal.
• Briefs: the forbidden block left <worktree> open as a placeholder in the design and verifier briefs. I added that line in chore(ship): S27 retro: cloud red-team kit in the skill, verifier foreign in gate 8, snapshot in ~/.cache/ship, FINDINGS, load warning #898.
• Briefs: every brief runs ls -A ~/Downloads before the first and after the last test or bdg run, not only the tester. In S28 a reviewer's red-commit smoke run left files there.
• Removed the "snapshot is 12 h old" warning from chore(ship): S27 retro: cloud red-team kit in the skill, verifier foreign in gate 8, snapshot in ~/.cache/ship, FINDINGS, load warning #898. It fired during a session that ran overnight, and a really stale snapshot already shows up as a gate 8 failure.
• run-wait now defaults to 90 minutes instead of 60. A ×10 repeat of fix: dom fill reads the value again 2 s after filling; a late revert exits 90 (#892) #900 (2 s per dom fill) took longer than 60.
• ci-wait: a PR waiter that has seen a CI run no longer gives up when one gh poll comes back empty. On this PR it saw "CI queued 4 min", then advised closing and reopening the PR (exit 4).
• scripts/agent-eval.mjs exits 1 and says why when the budget let no run start. It used to exit 0 with an empty summary.
• Not changed: "a TIER=1 reviewer brief keeps the red-commit section" doesn't reproduce. The rendered brief has the intended Red commit. None: tier 1… line.

Process

• Promoted (S27 candidate, S28 found dom fill reports ✓ when the page replaces the value with an unrelated one later (rejection read as formatting) #906 after merge): a PR that changes an action verdict, a hint or a message agents act on runs its area's eval tasks on the branch before merge, next to main (about $0.10 a task).
• Candidates:
  • first-draft repeat removal: 2 of 3 sessions with no catch;
  • /simplify removal: back to 0, since S28's /simplify found real issues;
  • three new rows: the macOS suite time for a change that slows a frequent command, one yes/no question per AskUserQuestion call, and the "unrelated replacement" row for revert heuristics.
• log.md: the S28 rows.

Verified

• Test results:
  • node --test .claude/skills/ship/scripts/ship.test.mjs: 120 pass, with a new test for osHandlerUrls;
  • node --test scripts/agent-eval.test.mjs: 78 pass;
  • eslint clean.
• A real eval with --budget=0.1 --run-budget=0.5 exited 1 with "no run started (budget: …)". It spent nothing.
• Rendered implementer, reviewer, tester and verifier briefs: no <worktree> left, and the ban text passes --check.