✠ The Chronicle

✠ ENTRY LI

A UI-heavy window: two large fresh-eyes swarm sessions worked through a…

A UI-heavy window: two large fresh-eyes swarm sessions worked through a combined 358+ candidate rows down to a fully-dispositioned ledger, a 24-agent content-pitch session filed a 34-row three-tentpole roadmap ladder, the standing three-digest-old flake in combat-round-e2e got hardened, and the content-lifecycle rotation held clean across all five stewards. Baseline re-measured fresh — numbers unchanged, confirming the two mechanics-source commits that triggered staleness were text-only.

UI

Two UI fresh-eyes swarms clear a combined 358+ candidate rows

What changed

A dedicated sweep (plan/2026-09-12-ui-fresh-eyes.prompt.md, PR #300/#301) walked all 27 routes at two viewports, shipped 21 verified fixes with guarding tests, and filed 309 more candidate rows unverified. A follow-up swarm (plan/2026-09-12-ui-fresh-eyes-swarm.prompt.md, PR #303/#304, 96 agents across five workflows) then gave every one of those 309 rows a disposition: collapsed to 187 semantic clusters, 149 reached a three-lens adversarial panel, and 30 shipped as FE-029..FE-058. Two regression readers caught ten self-inflicted regressions (three major, including a hand-fan fix that collapsed the fan it meant to protect) and a repair fleet closed all ten with tests that fail against the damaged version. A completeness critic found eleven bookkeeping gaps, rescuing two owner decisions the panel had wrongly refuted away.

Why it changed

sweep #301's own candidate table was explicit that 309 rows shipped unverified because the container's 2-way concurrency split-voted 87 clusters; the swarm was dispatched specifically to close that gap before it went stale.

Residue for the next sweep: the capture driver dismisses the combat tutorial coach (so no screenshot shows a first-time player's first fight), and the container lacks the Playwright browser revision the repo pins — both logged in plan/AUDIT.md, not fixed here.

commits 8d8a5e41 · ee · 40 · c · 5282cc0a · a · 20 · c

CONTENT

Content roadmap pitch session — 3 tentpoles, 34-row ladder

What changed

A 24-agent session (5 grounding readers, 9 lane designers, 3 KB prior-art scouts, 3 judges, an adversarial auditor, a synthesist, a completeness critic) ran the full pitch pipeline end to end: 45 raw pitches down to 14 survivors, grouped into three tentpoles — The Wages of the Road (make THE PATH's dormant progression axes real choices), The Parish Keeps Accounts (the world reads what the player owes), and Nine Words and No Mouth (keyword legibility + missing card-text register) — sequenced into build-plan Phases 79-110 plus W7-W8.

Why it changed

plan-only deliverable per /oversight's promotion gate; the session also banned citing the stale "4% late-campaign" figure going forward, since the live baseline already read 82%/8.8% at filing time — confirmed again by this digest's own re-measure below.

commit 0f31c61a

CONTENT

Lifecycle rotation — five stewards, one real fix

What changed

critique pass 36 (no findings), adjust-keywords pass 7 (one stale test-title fix, no CREATE/UPDATE), adjust-npcs pass 7, adjust-cards pass 8, adjust-equipment pass 8, and adjust-enemies pass 8 all returned zero-diff re-audits against their prior pass.

Why it changed

standing rotation cadence; six consecutive clean re-audits (five zero-diff plus one single stale-test fix) is itself a signal the content pools aren't drifting under the UI-focused week.

commits 27987751 · 3c39acb9 · 99cac84e · be34ca43 · 030e26ae · 832e3e5e

INFRA

combat-round-e2e's die-power drag flake hardened

What changed

verify-mobile failed the same crash-strict step CI had passed hours earlier with no intervening code change — the power-drag retry fired immediately after staging, with no settle time for the stage-enter animation, so resolveDrop's live measureInWindow could read a stale box mid-animation. Added a 150ms settle wait, bumped the retry budget 3→4, widened the per-attempt wait 120ms→180ms.

Why it changed

conclusive CI-only flake, reproduced locally across 5 consecutive runs before the fix and clean after; no mechanics/cards/effects path touched, so no cross-package legs required.

commit 5a75ca8d

BALANCE

Baseline re-measured fresh — numbers unchanged

What changed

session start flagged the standing baseline (b1624da0, reduced-nightly) stale by two mechanics-source commits — both from the fresh-eyes sweep (a threat-preview label rename, a speech-mark fix). Re-measured via the reduced-nightly pass (30 runs/cell) at commit a2ea15ff. The regenerated JSON differs from the prior baseline in only its meta block (timestamp + commit) — the report body is byte-identical, confirming both source commits were text-only and touched nothing the deck matrix reads. Stage terminal win rates: early 97.3%, mid 97.0%, late 82.0%, impossible 8.7%. Overall combat-quality index 0.763.

Why it changed

this digest's own freshness gate exists so a balance question never gets answered against a stale engine; plan/AUDIT.md's own loop-call row on this staleness predicted exactly this outcome ("a regen would spend a full sim run to reproduce the same numbers") — confirmed.

Note: the late-stage cliff (82.0% late / 8.7% impossible) is unchanged and still owned by the standing [score 6.5] plan/PHASE_CANDIDATES.md row ("survived THE PATH tuning pass and needs a named design-level fix, not another tuning pass") — not re-filed here.

commits a2ea15ff · baee · ee · this · c

✠ The marginalia

The night's watch — what ran while nobody looked
when (UTC)verboutcome
09-12 08:32–08:44march → critiquesuccess — pass 36, no findings
09-12 10:26–10:48march → adjust-keywordssuccess — pass 7, one stale test-title fix
09-12 12:35–12:43marchsuccess, no-op
09-12 14:27–14:42march → adjust-npcssuccess — pass 7, zero-diff
09-12 16:28–16:33march → iteratesuccess — audit, top score 1.6, handed to expand
09-12 18:29–18:52march → adjust-cardssuccess — pass 8, zero-diff
09-12 20:29–20:42march → adjust-equipmentsuccess — pass 8, zero-diff
09-12 22:26–22:39marchsuccess, no-op
09-13 01:07–01:11marchsuccess, no-op
09-13 02:36–03:08marchsuccess, no-op
09-13 04:32–04:37marchsuccess, no-op
09-13 06:44–06:48marchsuccess, no-op
09-13 08:33–08:55march → adjust-enemiessuccess — pass 8, zero-diff

13 of 13 march runs succeeded this window. Alongside the cadence, three larger dispatched sessions ran outside march entirely: the two UI fresh-eyes swarms (PRs #300/#301, #303/#304) and the content-pitch session (0f31c61a) — all covered by the cards above. The combat-round-e2e flake fix (5a75ca8d, 08:01 UTC) also landed between ticks, outside the march cadence — matches the profile of a direct /fix-ci response to a red verify-mobile run, not a scheduled tick.

The measure — did the fighting change?

None from this digest's own meta-loop rail this window — 13 of 13 march ticks succeeded, the content-lifecycle rotation stayed clean across all five categories, and the nightly breadth check ran green. No gate/cadence/ceiling mistuning observed.

What is still unfinished
  • Build plan (plan/steps/01_build_plan.md): 0 open phases ([ ]),

2 blocked, 125 shipped.

  • plan/AUDIT.md: ~44 rows pending (93 headers under "Pending" minus 49

resolved-but-unmoved [x] bookkeeping rows).

  • plan/CRITIQUE.md: ~22 rows pending (50 headers minus 28 [x]

bookkeeping rows); 2 standing [HIGH] — see Needs You above.

  • plan/PHASE_CANDIDATES.md: 13 rows pending.
  • No open triage:needs-user or loop:do issues.
  • Deploy: clean — HEAD is a docs/content tick, no gated workflow

triggered yet (grace window).

  • Nightly breadth (e2e:minigames): GREEN — 10/10 sandbox+live

runs, no crashes.

  • Measurement: was STALE at session start (baseline b1624da0, 2

mechanics-source commits behind); re-measured this tick via npm run baseline:regen -- --runs=30 --confidence=reduced-nightly → now FRESH at a2ea15ff. See the [balance] card above.

The nightly entry also carries a panel addressed to the maintainer — open questions, permission grants, things only he can rule on. It is held back from this page: it is correspondence, not a report. The unfinished work it tracks is published above, in plain words.