[adamXbot]

Scored polish hill climb of the mantis CLI's human-readable output, run with the ui-polish method: eight real screens captured in light and dark at 100 columns against a fixture API with a frozen clock, judged per category by three independent reviewer models against GitHub CLI and Wrangler, fixed where the judges agreed, re-captured and re-judged. The target, 9.0 in every category from every judge in the same round, was never shown to the judges. It cleared in round 8. Round 9 then applied changes the maintainer chose after a comparison against twelve more CLIs and a resize test; Opus and Sonnet re-cleared it, and Fable could not be run that round (usage limit).

--json, --id-only, --url-only, flags and exit codes are unchanged. Every change is to text, layout, glyphs and ANSI attributes. 37 test files, 275 tests pass; four tests were added in round 9 for the middle cut and the error-hint arrow, and two assertions were updated to the new wording (a since cell without "ago", the escaped host rather than the source in a table row).

Score history (overall per judge)

round	Fable	Opus	Sonnet	lowest category
0	6.0	6.0	6.3	5.0
1	8.2	7.5	7.8	7.5
2	8.6	8.1	8.4	8.0
3	8.9	8.3	8.5	8.0
4	9.0	8.6	8.7	8.0
5	9.2	8.8	8.8	8.5
6	9.2	8.9	8.8	8.5
7	9.3	9.0	9.0	8.5
8	9.5	9.1	9.2	9.0
9	not run	9.2	9.2	9.0

Round 8 per category: Fable 9.5 in all eight; Opus 9.0 in all eight; Sonnet 9.0 to 9.5. Round 9 per category: Opus and Sonnet 9.0 to 9.5 in all eight, no regressions. References on the same scale, same judges: GitHub CLI 8.0 to 8.5, Wrangler 5.5 to 6.5.

Round 9: the maintainer's picks and the resize fixes

After round 8 the same judges scored mantis beside Homebrew, Ansible, nala, AWS CLI, Azure CLI, fzf, witr, Docker, kubectl, cargo, npm and uv (mantis ranked first at 8.85, ahead of cargo 8.2, uv 8.2 and fzf 8.1). The maintainer picked from what the judges agreed on, and a resize test added three fixes:

• Error hints print as an indented → line, the same as empty states (formatError).
• URLs and file paths are underlined at the text colour (c.link); cyan is kept for commands.
• A failed delivery keeps a red ✗ in its own column at the left edge (the maintainer's call, given to the judges as a constraint), with times right-aligned after it; the ssh arrow sits in a gutter with the ip header over the addresses; since right-aligns like age.
• Failure reasons hang under their row behind a dim └ (childLines).
• The new-key wizard reads as a form: questions, a dim default column, every answer bold and starting at column 44; choices and rejections sit in the answer column; inline fallback below 70 columns.
• Bug fix: the masked webhook redraw moved up one row only, so a real-length Discord URL (about 121 characters) or Slack URL (about 77) that wrapped left the prompt twice and the start of the token on screen. It now clears every row the answer used.
• Narrow windows: memos and identities cut in the middle on word edges (cutMiddle), headlines fit the window, and times, addresses and delivery are never cut.
• status --watch redraws 150 ms after a resize instead of waiting for its next tick.

Changes by category

Hierarchy (cli/src/lib/out.ts: heading, kv, tripped; status.ts, hits.ts, doctor.ts)

• Verdict lines open status (● 2 of 6 monitored keys tripped), doctor (⚠ 5 passed, 1 skipped, 3 warnings), hits and list.
• Tripped rows bold and first; failed-delivery rows carry ✗ in the when cell; bot, duplicate and disabled rows dim their identity cells.
• One heading (bold memo, dim short id) and one label block (kv, 12-column label gutter) on show, status detail, hit detail and the wizard summary; hit detail leads with who, where and when, then delivery.
• The home screen is a 12-row launcher instead of a 119-row command dump; --help is grouped.

Typography (out.ts: formatStamp, formatAge, formatAgo, wrapProse, proseWidth, cap, memoWidth)

• One timestamp format, 2026-11-06 02:26 UTC, time only for today, seconds only in hit detail; one relative form (4m, 3m40s, 2h, in prose with "ago").
• Prose and failure reasons wrap on word boundaries with a hanging indent; nothing is cut mid-token; memo columns size from the terminal (38 to 60).
• Agents parsed to one form (Chrome 141 · Windows, aws-cli 2.31.4 · linux, Googlebot 2.1).

Spacing and layout (out.ts: table, fitMiddle, maskSecretUrl; hit-format.ts, destination-format.ts)

• Tables right-align ages and counts; headers dim and underlined; the ssh client address sits in the ip column so the who column is the identity alone.
• Notify and delivery rows share one shape (mark, channel, target, result) with results aligned and the failure reason on its own line without the repeated address.
• Webhook tokens masked to https://host/…tail on a terminal (piped and --json output keep the whole target).

Colour and contrast (out.ts: c, mark, hintLines, formatError)

• Strict roles: red for tripped and failed, green for ok, yellow for warning and pending (bold, so it survives light mode), cyan for anything to run or open, dim for labels, ages, ids and legends.
• Instruction lines, hints and state words at full weight; the mark carries the colour, the word carries the meaning.

Iconography and controls (out.ts: mark, arrow, dot, glyph; wizard.ts, prompt.ts)

• One mark set, ✓ ✗ ⚠ … ● –, with ASCII fallbacks + x ! ~ ! -; tripped has its own dot so ⚠ means warning only; legends explain every mark on screen.
• Wizard: defaults dim in brackets, typed answers bold, accepted defaults echoed dim, choices on their own line before every prompt, a rejected URL redrawn dim, an accepted webhook redrawn masked, phase breaks, summary drawn by the same renderer as show.

Consistency (hit-format.ts, destination-format.ts, hostOnly, withoutRepeat)

• One renderer each for hits, deliveries and destinations, shared by list, show, status, hits, watch and the wizard; the destinations label became notify, matching --notify.
• Hosts named for identification drop their scheme (doctor, home); the next command in every footer uses the resolved short id.

Clarity (runner.ts: networkReason; resolve.ts; doctor.ts)

• Errors are error: plus the message plus one hint line; network errors say which server and why (connection refused, host not found); the echoed prefix is bold.
• Doctor marks unconfigured checks as skipped and counts them apart, names causes in words, names the older side and the upgrade path for the install in use, and scopes the shared hint.
• Delivery states in words: delivered on attempt 1, failed after 5 attempts, suppressed, no destinations; Latest N hits when --limit cut the page.

Character

• The green wordmark on home and in the wizard title; the keychain notice is one line that is erased once the read returns, so every screen opens with its own content.

Beyond polish, for the maintainer to decide

• Tripped state in mantis list. All three judges asked for it every round. The list endpoint carries no monitor state and reading it costs one request per monitored key (what mantis status does), so showing it in list is a behaviour change. The footer points at mantis status.
• Latest 2 of 3 hits under --limit. The hits endpoint reports no total.
• Collapsing duplicate hits under --verbose to one line naming the original (all three judges in round 8, Opus and Sonnet again in round 9).
• One footer grammar (count line, then an indented → pointer) and relative-first timestamp pairs on show: both round 9 judges' top shared findings, left for a later round.
• Doctor warning names at default weight and a ⚠ on the warning count itself (single-judge, low).

Judges, references, rounds

• Judges: Claude Opus, Claude Fable and Claude Sonnet, each a fresh subagent that saw only the brief and the screenshots, run in parallel, never each other's reviews, never the target. Fable shares the fixing model but none of its context.
• References: GitHub CLI 2.x (gh pr --help, gh run list, gh pr list, gh pr view, gh pr checks) and Cloudflare Wrangler (--help, kv namespace --help), captured through the same harness and palette.
• Nine rounds after the baseline (round 9 judged by Opus and Sonnet only: the account's Fable limit was reached; the session model was Opus 5.5 for round 9); one commit per round named with that round's scores. Judge disagreements and single-judge findings not taken are recorded per round in the run notes (kept out of git with the captures).
• Output was also checked with colour off and an ASCII-only locale: every mark has a text fallback.

🤖 Generated with Claude Code