@galda/cli 0.10.38 → 0.10.40

package/app/index.html

@@ -853,20 +853,15 @@
   .reqdesc .mdsp{height:6px}
   .reqdesc code,.gdtaskresult code{font-family:var(--mono);font-size:11px;background:rgba(var(--tint),.07);padding:1px 4px;border-radius:4px}
   .reqdesc strong{color:var(--ink);font-weight:700}
+  /* an image the AI pasted into its own report: shown full-width at its own aspect,
+     hairline edge only (§5.5 縁は1px以下のヘアライン), never a fixed-ratio crop box */
+  .mdimg{display:block;max-width:100%;height:auto;margin:8px 0;border-radius:8px;border:1px solid var(--hair)}
   .reqcond{font-family:var(--mono);font-size:11px;color:var(--ink3);margin-top:4px}
   .reqvideo{margin:10px 0}
   .reqnovideo{font-size:12px;color:var(--ink3);border:1px dashed var(--hair2);border-radius:10px;padding:16px;text-align:center;margin:10px 0}
   .beforeafter{display:grid;grid-template-columns:1fr 1fr;gap:8px}
   .ba-col{min-width:0}
   .balabel{display:block;font-family:var(--mono);font-size:9px;letter-spacing:.06em;color:var(--ink3);margin-bottom:2px}
-  /* diff box + zoomed inset (HANDOFF-review-card-diffbox-2026-07-18) — only emitted
-     by beforeAfterBoxedHtml(), so these never touch the plain pair above or the
-     older per-task renderReviewChecklist before/after (out of scope, untouched). */
-  .ba-imgwrap{position:relative;margin-top:8px}
-  .ba-imgwrap .proofimg{width:100%;height:auto;max-height:none;object-fit:initial;margin-top:0;display:block}
-  .ba-box{position:absolute;border:1.5px solid var(--danger);border-radius:2px;pointer-events:none}
-  .ba-zoomhd{margin-top:14px}
-  .lg-ev .ba-zoomimg{display:block;width:100%;height:auto;max-height:none;object-fit:initial;margin-top:8px;border-radius:9px;border:1px solid var(--hair2)}
   .reqbtns{display:flex;gap:8px;margin-top:4px}
   .reqbtn{height:30px;padding:0 14px;border-radius:9px;border:1px solid var(--hair2);background:transparent;
     color:var(--ink2);font:inherit;font-size:12.5px;font-weight:600;cursor:pointer;transition:.15s}
@@ -1620,6 +1615,14 @@
   .lbc.is-quiet .lbc-t{color:var(--ink2)}   /* Done / Later / To Do: visible, not shouting */
   .lbc-w{font:600 8.5px/1 var(--mono);letter-spacing:.11em;text-transform:uppercase;margin-left:auto}
   .lbc-w.bad{color:var(--danger)} .lbc-w.intr{color:#ffca16}
+  /* P0b: the AI's own To-do breakdown, scaled down to the card's density (§1.5-8 fold, not
+     scroll — capped + "+N more" text rather than an inner scrollbar). Reuses the .todo/.box/
+     .tdtext vocabulary from the chat feed (todosHtml) so the ✓/▸/○ language stays one system. */
+  .lbc-todos{display:flex;flex-direction:column;gap:3px}
+  .lbc-todos .todo{font-size:10.5px;line-height:1.4;gap:5px}
+  .lbc-todos .todo .box{width:11px;height:11px;font-size:9px;line-height:11px;margin-top:1px}
+  .lbc-todos.prep{font-size:10.5px;color:var(--ink3);font-style:italic}
+  .lbc-more{font:500 9.5px/1 var(--mono);color:var(--ink3)}
   .lbc-meter{display:flex;align-items:center;gap:7px}
   .lbc-meter .bar{flex:1;height:2px;border-radius:2px;background:color-mix(in srgb,var(--ink) 11%,transparent);overflow:hidden}
   .lbc-meter .bar i{display:block;height:100%;border-radius:2px;background:var(--green);transition:width .3s}
@@ -1712,33 +1715,12 @@
   /* review row number — #NN PREFIX on the title (HANDOFF-v45 §11), same form as the To Do
      rows (PRD §5: numbering matches the sidebar) */
   #fsRoot .rcard .rid{font-size:10.5px;font-weight:600;color:var(--ink3);margin:0 5px 0 0;display:inline-block;font-variant-numeric:tabular-nums}
-  /* PRD §6.5 fold note — on the carrier ("also carries #13") and on the folded card
-     ("merged into #12"). Deliberately NOT the .pr treatment: no tinted ground, no
-     badge (§1.5-9), no red — folding is not a failure. Mono + --ink3 = the same quiet
-     meta tone the rest of the lane uses (§5.5「品は罫線とトーン差・monoで出す」). */
-  /* display:block, not inline-block: it shares .rc with the PR pill, which is itself
-     inline-block — side by side the pill sat on top of the last "#NN" and ate it. */
-  #fsRoot .rcard .fnote{font-family:var(--mono);font-size:9.5px;color:var(--ink3);margin-top:3px;display:block}
-  /* In rv3 the meta sits on the near-black photo scrim whatever the theme, where --ink3
-     is effectively invisible (§5.5「見えるが騒がない」— near-invisible is a fail). Match
-     .rid's white-on-scrim treatment. The folded card opts out of the scrim (rule above),
-     so it keeps --ink3 on the panel where that tone is correct. */
-  #fsRoot[data-rv="3"] .rcard:not(.fold) .fnote{color:rgba(255,255,255,.66);text-shadow:0 1px 3px rgba(0,0,0,.55)}
-  /* A folded card is settled, not failed: title sinks to the settled tone, and it
-     keeps no ✓/✕ because the thing to approve lives on the goal it was folded into. */
-  #fsRoot .rcard.fold .t{color:var(--ink2)}
-  /* A folded card is a pointer, not a review item: no proof to show, nothing to approve.
-     So it opts OUT of the rv3 thumbnail treatment (whose .rc is absolutely positioned
-     over an 84px photo ground) and stays the same quiet line in every view — the form
-     Masa drew. Giving it a fake dark thumbnail would claim it has something to look at. */
-  /* .fold.noimg, not .fold: the placeholder-ground rule below is `.rcard.noimg`, which
-     ties this on specificity and would win on order — a folded card would keep the fake
-     dark thumbnail ground. Matching both classes settles it regardless of order. */
-  #fsRoot[data-rv="3"] .rcard.fold.noimg{display:flex;position:static;align-items:center;gap:10px;
-    padding:9px 10px;min-height:0;border:1px solid var(--hair);border-radius:11px;background:none;box-shadow:none}
-  #fsRoot[data-rv="3"] .rcard.fold.noimg .rc{position:static;padding:0;background:none}
-  #fsRoot[data-rv="3"] .rcard.fold.noimg .t{color:var(--ink2);font-size:12.5px;text-shadow:none}
-  #fsRoot[data-rv="3"] .rcard.fold.noimg .rid{color:var(--ink3)}
+  /* PRD §6.5 → Masa 2026-07-21: the fold note ("also carries #13" on the carrier card,
+     "merged into #12" on the folded card) no longer renders in the Review list at all —
+     it read as the same report-in-the-list clutter the feature was meant to avoid. Where
+     folded work went is now logged in the opened detail's "Also carries" section only, so
+     the .fnote / .rcard.fold* rules that used to style these list-level notes are gone
+     with them (dead CSS orphaned the moment nothing emits those classes any more). */
   /* reference numbers everywhere else (HANDOFF-v45 §11): a leading #NN prefix on every
      To Do item's name — same goal = same number (goalReviewNumber), distinct from the
      PR's own #id shown separately. */
@@ -2010,6 +1992,17 @@
   .redesign .rmeta-detail.open{display:block}
   .redesign .rmeta-detail .k{font:600 8px/1 var(--mono);letter-spacing:.1em;text-transform:uppercase;color:var(--ink3);margin-bottom:5px;display:block}
   .redesign .rmeta-detail .v{font:500 11.5px/1.6 var(--ui);color:var(--ink2)}
+  /* Open it section: the address is the control. Mono so it reads as an address, plain
+     text weight so it doesn't shout, and hover moves the GLYPH colour only — never a
+     filled surface (§5.5 アイコンボタンのhoverは面を塗らない). */
+  .redesign .ropen .olink{display:inline-flex;align-items:baseline;gap:6px;padding:0;border:0;background:none;cursor:pointer;
+    font:400 13.5px/1.62 var(--mono);color:var(--ink);letter-spacing:-.004em;text-align:left;overflow-wrap:anywhere;transition:color .14s}
+  .redesign .ropen .olink:hover{color:var(--green)}
+  .redesign .ropen .olink[disabled]{color:var(--ink3);cursor:default}
+  .redesign .ropen .olink .ar{font-size:.85em;color:var(--ink3)}
+  /* the note is body prose like every other section's — same 14px / −.012em / --ink as
+     .rbody (Masa 2026-07-22: it was a shade lighter and read as a different kind of text) */
+  .redesign .ropen .v{margin:4px 0 0 !important}
   /* sections — Linear-docs bold sentence-case headings + high-contrast body.
      margins are !important: the app's pre-existing .lg / .sa-item rules load first and
      otherwise tie/beat these on specificity, flattening the Linear-docs rhythm. */
@@ -2019,6 +2012,36 @@
      13px read thinner/lighter for Japanese than the surrounding app). §2. */
   .redesign .rbody{font-size:14px;line-height:1.62;letter-spacing:-.012em;color:var(--ink);max-width:70ch;margin:0 !important}
   .redesign .rdesc,.redesign .rrisk{color:var(--ink2)}
+  /* Result (P1, Masa 2026-07-21): the AI's own report, rendered as written. Same 14px
+     body rhythm as .rbody — it is prose the reviewer reads, not a log. Headings/lists/
+     code inside it are the report's own structure, so they get the section's type scale
+     rather than the smaller .reqdesc one used in the older checklist dialog. */
+  .redesign .rreport{font-size:14px;line-height:1.62;letter-spacing:-.012em;color:var(--ink);max-width:70ch;overflow-wrap:anywhere}
+  .redesign .rreport .mdh{font-weight:680;color:var(--ink);margin:14px 0 4px}
+  .redesign .rreport .mdh:first-child{margin-top:0}
+  .redesign .rreport .mdul{margin:2px 0;padding-left:20px}
+  .redesign .rreport .mdul li{margin:2px 0}
+  .redesign .rreport .mdsp{height:8px}
+  .redesign .rreport code{font-family:var(--mono);font-size:12px;background:rgba(var(--tint),.07);padding:1px 4px;border-radius:4px}
+  .redesign .rreport strong{color:var(--ink);font-weight:700}
+  .redesign .rreport a{color:var(--ink);text-decoration:underline;text-underline-offset:2px}
+  /* "Open it" is waiting on a real process to boot (up to ~15s), so the chip says so
+     rather than looking dead. Dim + no pointer, no spinner — §5.5 薄く・主張しない. */
+  .redesign .mx.mwait{opacity:.55;pointer-events:none}
+  /* the AI's own screenshots: full width, own aspect, hairline edge (§5.5 縁は1px以下).
+     No grid, no crop — one shot is often a comparison sheet and dies when shrunk. */
+  .redesign .rshot{margin:0 0 16px}
+  .redesign .rshot:last-child{margin-bottom:0}
+  .redesign .rshot img{display:block;width:100%;height:auto;border-radius:10px;border:1px solid var(--hair)}
+  .redesign .rshot figcaption{margin-top:6px;font-family:var(--mono);font-size:11px;color:var(--ink3)}
+  /* before/after pair + the red box over what actually changed (pixel-diffed here from
+     the worker's own two shots — no box appears unless the pixels really differ) */
+  .redesign .beforeafter .proofimg{width:100%;height:auto;max-height:none;object-fit:initial;display:block;border-radius:8px;border:1px solid var(--hair)}
+  .redesign .ba-imgwrap{position:relative}
+  .redesign .ba-imgwrap .proofimg{margin-top:0}
+  .redesign .ba-box{position:absolute;border:1.5px solid var(--danger);border-radius:2px;pointer-events:none}
+  .redesign .ba-zoomhd{margin:14px 0 8px;font-family:var(--mono);font-size:11px;letter-spacing:.04em;text-transform:uppercase;color:var(--ink3)}
+  .redesign .ba-zoomimg{display:block;width:100%;height:auto;border-radius:8px;border:1px solid var(--hair)}
   .redesign .rfiles{margin:12px 0 0;font-family:var(--mono);font-size:12px;color:var(--ink2);line-height:1.75}
   .redesign .rfiles .a{color:#3fd07a;font-style:normal}.redesign .rfiles .d{color:#ff7b72;font-style:normal}
   /* Request accordion (spec §3) — chevron › rotate 90, body indented, no left rule */
@@ -2091,22 +2114,25 @@
      in at the bottom of the opened body. */
   .redesign .ract{margin-top:24px !important}
   .redesign .ract>.rsec-h{margin-bottom:10px !important}
-  .redesign .lacc>summary{list-style:none;cursor:pointer;display:flex;gap:9px;align-items:baseline}
+  .redesign .lacc>summary{list-style:none;cursor:pointer;display:flex;gap:14px;align-items:baseline}
   .redesign .lacc>summary::-webkit-details-marker{display:none}
-  .redesign .lacc .ltog{flex:0 0 auto;color:var(--ink3);font-style:normal;line-height:1;transition:transform .14s}
+  .redesign .lacc .ltog{flex:0 0 auto;color:var(--ink2);font-style:normal;font-size:1.15em;line-height:1.25;transition:transform .14s}
   .redesign .lacc[open]>summary .ltog{transform:rotate(90deg)}
   .redesign .lacc .lfirst{flex:1;min-width:0}
   .redesign .lacc .lbody{margin-top:12px;padding-left:18px}
   /* transcript rows — Linear-clean, borderless (You / tool mono / res green / Claude) */
   .redesign .rtx{border:0;background:none}
-  .redesign .txr{display:flex;gap:12px;padding:5px 0;font-size:13px;line-height:1.55}
-  .redesign .txr .w{flex:0 0 46px;font-size:11px;color:var(--ink3);padding-top:1px}
+  .redesign .txr{display:flex;gap:12px;padding:5px 0;font-size:13px;line-height:1.55;align-items:baseline}
+  .redesign .txr .w{flex:0 0 46px;font-size:11px;color:var(--ink3)}
   .redesign .txr.you .w,.redesign .txr.cl .w{color:var(--ink);font-weight:600}
   .redesign .txr .m{flex:1;min-width:0;color:var(--ink2)}
   .redesign .txr.cl .m{color:var(--ink)}
   .redesign .txr.tool .m{color:var(--ink2);font-family:var(--mono);font-size:12px}
   .redesign .txr.tool .m b{color:var(--ink);font-weight:600}
   .redesign .txr.res .m{font-family:var(--mono);font-size:12px;color:var(--green)}
+  /* Manager's own progress notes inside the implementation log: commentary, not results.
+     Green would read as "passed" (§5.5 UIクロームは薄く・優しく・主張しない). */
+  .redesign .txr.note .m{font-family:var(--mono);font-size:12px;color:var(--ink3)}
   /* Discussion (review back-and-forth) folded at the bottom of Activity's body */
   .redesign .lbody .rthread{display:flex;flex-direction:column;gap:16px;margin:0}
   .redesign .lbody .rmsg .who{font-size:12px;font-weight:600;color:var(--ink);margin-bottom:4px}
@@ -3361,16 +3387,6 @@ function reviewCardConfigFor(projectId){ return { ...DEFAULT_REVIEW_DEFINITION.r
 // goal's own request language, same as the planner/worker; ja/en pin the
 // project to one language regardless of what a given goal was typed in.
 function reviewLangFor(projectId){ const l = reviewDefinitionFor(projectId).language; return (l === 'en' || l === 'ja') ? l : 'auto'; }
-// PRD §6.5 fold note language (Masa 2026-07-17: "ユーザーの言語に合わせて" — not
-// hardcoded English). Mirrors the server's own 'auto' resolution exactly
-// (reviewDefinition.language === 'auto' ? detectRequestLanguage(goal.text) : ...,
-// engine/server.mjs) so this note reads in the same language as the review
-// headline/PR body for the same goal: the project's fixed override wins if set,
-// otherwise match the language the goal's own request was written in.
-function foldLangFor(goal){
-  const l = reviewLangFor(state.active);
-  return (l === 'ja' || l === 'en') ? l : (/[぀-ヿ㐀-鿿]/.test(goal?.text || '') ? 'ja' : 'en');
-}
 
 function esc(s){ const d = document.createElement('div'); d.textContent = s ?? ''; return d.innerHTML; }
 
@@ -3522,30 +3538,28 @@ function saReviewTitle(g){
   // reviewSummary.title, so they fall back to the request text unchanged (no regression).
   return String(g?.reviewSummary?.title || g?.text || g?.plan?.[0] || g?.reviewSummary?.changed || '').replace(/\s+/g, ' ').trim().slice(0, 90);
 }
-// H23 (2026-07-18) evidence-kind detection. Mirrors engine/lib.mjs isUiChange() /
-// hasTestRelevantChanges() — kept in sync manually since the browser script isn't a
-// module. flagship's demo data carries an explicit `kind` field
-// ('image'|'logic'|'docs'|'missing'); real goals/tasks have no such field, so the
-// review card derives one from real signals it already has (a captured proof image,
-// whether changed files look like UI source, whether they're all docs-only). This
-// is a HEURISTIC, not a stored fact on the goal — flagged as a judgment call in the
-// H23 implementation PR for a human to correct if it misclassifies real data.
-const UI_TOUCH_RE_LOCAL = /^app\/|\.(?:html|css|tsx|jsx|vue|svelte)$/i;
-function isUiChangeLocal(files){ return (files ?? []).some((f) => UI_TOUCH_RE_LOCAL.test(String(f ?? ''))); }
+// H23 (2026-07-18) evidence-kind detection. flagship's demo data carries an explicit
+// `kind` field; real goals/tasks have no such field, so the review card derives one
+// from real signals it already has (a captured proof image, whether the changed files
+// are all docs-only).
+//
+// P1 (Masa 2026-07-21): the 'missing' kind is gone, and with it the app-local copy of
+// isUiChange(). It said "this touched the UI, so a screenshot SHOULD exist" — a rule
+// from the era when we photographed the work ourselves. That capture never ran outside
+// this one repo, so the only thing the rule ever produced was an apology on every UI
+// goal. We don't take the photo any more, so there is nothing to apologise for.
 const DOC_ONLY_RE_LOCAL = /(^|\/)(README|START-HERE|CHANGELOG|LICENSE)(\.[\w.-]+)?$|(^|\/)docs\/.*\.(md|mdx|txt)$|\.mdx?$|\.txt$/i;
 function hasTestRelevantChangesLocal(files){
   const fs = (files ?? []).map((f) => String(f ?? '').trim()).filter(Boolean);
   if (!fs.length) return true;
   return fs.some((f) => !DOC_ONLY_RE_LOCAL.test(f));
 }
-// Priority: image (a real capture exists) > missing (UI touched, should have a
-// capture per PRD §4(C) but doesn't) > docs (every changed file is docs-only) >
+// Priority: image (a real capture exists) > docs (every changed file is docs-only) >
 // logic (verified only by tests, no UI). Falls back to 'image' (today's generic
 // fact-row + no image) when no signal is conclusive — never claims a stronger kind
 // than the data actually supports.
 function saEvidenceKind(filePaths, hasProofImg, testResult){
   if (hasProofImg) return 'image';
-  if (isUiChangeLocal(filePaths)) return 'missing';
   if (filePaths.length && !hasTestRelevantChangesLocal(filePaths)) return 'docs';
   if (testResult?.ran) return 'logic';
   return 'image';
@@ -3596,11 +3610,28 @@ function requirementSummaryLine(task){
 // Worker reports are written in Markdown (## headings, **bold**, - lists, `code`).
 // The review cards showed them raw, so "## " and "**" leaked as literal text.
 // Render the common subset to HTML (escaping first) so they read cleanly.
-function mdLite(src){
+// P1 (Masa 2026-07-21): the AI's report IS the deliverable, so an image it pasted
+// (`![alt](url)`) has to render as that image — this is the whole "paste it and it
+// shows up" wiring. Only src values we can vouch for are allowed through: http(s),
+// a same-origin path, or an inline data:image. Anything else (javascript:, data:text
+// /html, …) renders as the literal text it was, never as an <img src>.
+const MD_IMG_SRC_OK = /^(?:https?:\/\/|\/(?!\/)|data:image\/(?:png|jpe?g|gif|webp|svg\+xml);base64,)/i;
+// `images: false` renders a pasted image as nothing at all. Masa 2026-07-22, after
+// seeing one in a real card: the review card is not a gallery — the way to see the work
+// is Open it, and a lone blurry still of "one of the three" is worse than no still. The
+// renderer keeps the capability (chat and the older checklist still use images); the
+// review card is simply not the place.
+function mdLite(src, { images = true } = {}){
   const esc = (s) => s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;');
   const inline = (s) => esc(s)
     .replace(/`([^`]+)`/g, '<code>$1</code>')
     .replace(/\*\*([^*]+?)\*\*/g, '<strong>$1</strong>')
+    // images before links — ![alt](url) would otherwise match the link rule below
+    // and render as a bare "!" plus an <a>.
+    .replace(/!\[([^\]]*)\]\(([^)\s]+)\)/g, (m, alt, url) => (!images ? ''
+      : MD_IMG_SRC_OK.test(url)
+      ? `<img class="mdimg" src="${url.replace(/"/g, '&quot;')}" alt="${alt}" loading="lazy">`
+      : m))
     // Feedback: make links tappable — markdown [text](url) and bare URLs (e.g. a
     // "open the ReadMe/PROOF.md" link in the worker report).
     .replace(/\[([^\]]+)\]\((https?:\/\/[^)]+)\)/g, '<a href="$2" target="_blank" rel="noopener">$1</a>')
@@ -3740,30 +3771,22 @@ function proofMediaHtml(taskId, proof, altLabel){
   }
   return imgTag;
 }
-// reviewCard.beforeAfter: side-by-side comparison when a "before" capture
-// exists. Proof today never carries one (only a single post-change
-// snapshot/run is captured) — so this falls back to the normal single-shot
-// view rather than showing an empty/broken "Before" pane.
-function beforeAfterMediaHtml(taskId, proof, altLabel){
-  if (!proof) return '';
-  const beforeImg = proof.beforePng ?? proof.beforeGif;
-  if (!beforeImg) return proofMediaHtml(taskId, proof, altLabel);
-  const beforeTag = `<img class="proofimg" src="${withKey(`/proof/${taskId}/${beforeImg}`)}" alt="before: ${altLabel}" loading="lazy">`;
-  return `<div class="beforeafter">
-    <div class="ba-col"><span class="balabel">Before</span>${beforeTag}</div>
-    <div class="ba-col"><span class="balabel">After</span>${proofMediaHtml(taskId, proof, altLabel)}</div>
-  </div>`;
-}
-// Diff box + zoomed inset (HANDOFF-review-card-diffbox-2026-07-18): on top of the
-// plain before/after pair above, isolate the region that actually changed and draw
-// a red box over it on both images, plus a cropped/enlarged row below. Computed
-// client-side from the two PNGs via <canvas> pixel comparison at render time — not
-// baked in at capture time — so it works retroactively on every proof pair already
-// on disk with no new server-side image dependency (see the handoff for the full
-// rationale). Reference algorithm, verified against a real before/after pair
-// (composer icon swap, 640x200): correctly isolated a 57x52px box centered on the
-// changed icon — 0.13% image coverage, no false positives elsewhere in the frame.
-function computeDiffBox(beforeImg, afterImg, { threshold = 24, minCoverage = 0.0005, maxCoverage = 0.7, pad = 0.03 } = {}) {
+// (P1, Masa 2026-07-21) beforeAfterMediaHtml() was deleted here. The pair was staged by
+// us, not by the work: the "before" shot came from the same photo-staging pass that P1
+// removed on the engine side (#228), and in practice it never existed at all, so every
+// card fell through to the single shot anyway. Callers now ask for that shot directly.
+// Red box (restored 2026-07-22 on Masa's ask). This machinery was deleted in #230 with
+// the staged before/after pair, because the 'before' was a photo WE took of a screen
+// that had nothing to do with the change. It comes back on a different footing: the two
+// images are both the worker's own shots, taken by the agent that did the work, and the
+// box is computed here from those two PNGs by pixel comparison — nothing is invented, a
+// box only appears when the pixels really differ in one region.
+// minCoverage lowered 2026-07-22: the old 0.05% floor was tuned on a 640x200 proof pair.
+// A real worker shot is a whole 900-1440px page, where a genuine change (a dot going
+// green, a label swapping) is well under that floor and was being thrown away as noise.
+// 30 pixels differing by more than `threshold` is a change someone made on purpose;
+// compression noise does not clear a 24-point RGB delta.
+function computeDiffBox(beforeImg, afterImg, { threshold = 24, minPixels = 30, maxCoverage = 0.7, pad = 0.03 } = {}) {
   const w = beforeImg.naturalWidth, h = beforeImg.naturalHeight;
   if (w !== afterImg.naturalWidth || h !== afterImg.naturalHeight) return null; // different-shaped images: don't guess
   const c1 = document.createElement('canvas'); c1.width = w; c1.height = h;
@@ -3785,7 +3808,7 @@ function computeDiffBox(beforeImg, afterImg, { threshold = 24, minCoverage = 0.0
     }
   }
   const total = w * h;
-  if (maxX < 0 || diffPx / total < minCoverage) return null; // no real difference — don't box a no-op
+  if (maxX < 0 || diffPx < minPixels) return null; // no real difference — don't box a no-op
   if (diffPx / total > maxCoverage) return null; // diff too diffuse to be "a spot" — a full redesign shouldn't get a red box either
   const padPx = Math.round(Math.max(w, h) * pad);
   minX = Math.max(0, minX - padPx); minY = Math.max(0, minY - padPx);
@@ -3811,34 +3834,45 @@ function baZoomDataUrl(img, box, maxW = 420, maxH = 200){
   c.getContext('2d').drawImage(img, box.x, box.y, box.w, box.h, 0, 0, outW, outH);
   return c.toDataURL('image/png');
 }
-// Async: loads the before/after PNGs, runs computeDiffBox(), and — only when a real
-// box is found — renders the before/after pair with a red box overlay on each image
-// (percent-positioned so it tracks the image's displayed size at any viewport width)
-// plus a "Zoomed on the boxed area" row below. Returns null on every fallback
-// condition (no PNG pair, load failure, no clean box) so the caller keeps showing
-// the plain beforeAfterMediaHtml() pair unchanged — never a guessed or empty box
-// (HANDOFF-review-card-diffbox-2026-07-18). Only kind==='image' Ledger evidence
-// (saEvidenceHtml) calls this; the older per-task renderReviewChecklist before/after
-// is explicitly out of scope for this handoff and stays on beforeAfterMediaHtml.
-async function beforeAfterBoxedHtml(taskId, proof, altLabel){
-  const beforeSrc = proof?.beforePng, afterSrc = proof?.png;
-  if (!beforeSrc || !afterSrc) return null; // the box needs two real stills, not a GIF frame
-  let beforeImg, afterImg;
-  try {
-    [beforeImg, afterImg] = await Promise.all([
-      baLoadImage(withKey(`/proof/${taskId}/${beforeSrc}`)),
-      baLoadImage(withKey(`/proof/${taskId}/${afterSrc}`)),
-    ]);
-  } catch { return null; } // a missing/broken file is a fallback, not a crash
-  const box = computeDiffBox(beforeImg, afterImg);
+// A before/after pair exists only when the AI SAYS two of its shots are the same view
+// before and after (task.run.compare = [beforeName, afterName]). Masa 2026-07-22: an
+// earlier version of this guessed the pair from file names containing "before"/"after",
+// which is this layer inventing a relationship it cannot know — the same mistake as the
+// intake AI writing the pass condition. A request for three options has no "after"; only
+// the agent that did the work knows which images relate, so only it can say so.
+//
+// Nothing declares `compare` yet (the engine keeps cmd/url/note/check), so no card draws
+// a box today. The machinery below stays wired: the day the worker can declare a pair,
+// it lights up with no further app change.
+function saShotPair(shots, declared){
+  const list = shots ?? [];
+  const pair = Array.isArray(declared) ? declared : null;
+  if (!pair || pair.length !== 2) return null;
+  const find = (n) => list.find((s) => s.name === n || String(s.url || '').endsWith(String(n)));
+  const before = find(pair[0]), after = find(pair[1]);
+  return before && after && before !== after ? { before, after } : null;
+}
+function saPlainPairHtml(pair){
+  const col = (sh, label) => `<div class="ba-col"><span class="balabel">${label}</span><img class="proofimg" src="${esc(sh.url)}" alt="${esc(label)}" loading="lazy"></div>`;
+  return `<div class="beforeafter">${col(pair.before, 'Before')}${col(pair.after, 'After')}</div>`;
+}
+// Renders the pair with the changed region boxed on both, plus a cropped enlargement.
+// Returns null on every fallback (not two images, load failure, no clean box, different
+// sizes) so the caller keeps showing the plain shots — never a guessed box.
+async function shotsBoxedHtml(before, after){
+  if (!before?.url || !after?.url) return null;
+  let bImg, aImg;
+  try { [bImg, aImg] = await Promise.all([baLoadImage(before.url), baLoadImage(after.url)]); }
+  catch { return null; }
+  const box = computeDiffBox(bImg, aImg);
   if (!box) return null;
-  const w = beforeImg.naturalWidth, h = beforeImg.naturalHeight;
-  const boxStyle = `left:${(box.x / w * 100).toFixed(3)}%;top:${(box.y / h * 100).toFixed(3)}%;width:${(box.w / w * 100).toFixed(3)}%;height:${(box.h / h * 100).toFixed(3)}%`;
-  const frame = (img, label) => `<div class="ba-col"><span class="balabel">${label}</span><div class="ba-imgwrap"><img class="proofimg" src="${img.src}" alt="${label.toLowerCase()}: ${altLabel}" loading="lazy"><div class="ba-box" style="${boxStyle}"></div></div></div>`;
-  const zoomCol = (dataUrl, label) => `<div class="ba-col"><img class="ba-zoomimg" src="${dataUrl}" alt="zoomed ${label.toLowerCase()}: ${altLabel}"></div>`;
-  return `<div class="beforeafter">${frame(beforeImg, 'Before')}${frame(afterImg, 'After')}</div>
-    <div class="lg-evhd ba-zoomhd">Zoomed on the boxed area</div>
-    <div class="beforeafter ba-zoomrow">${zoomCol(baZoomDataUrl(beforeImg, box), 'before')}${zoomCol(baZoomDataUrl(afterImg, box), 'after')}</div>`;
+  const w = bImg.naturalWidth, h = bImg.naturalHeight;
+  const st = `left:${(box.x / w * 100).toFixed(3)}%;top:${(box.y / h * 100).toFixed(3)}%;width:${(box.w / w * 100).toFixed(3)}%;height:${(box.h / h * 100).toFixed(3)}%`;
+  const frame = (img, label) => `<div class="ba-col"><span class="balabel">${label}</span><div class="ba-imgwrap"><img class="proofimg" src="${img.src}" alt="${label.toLowerCase()}" loading="lazy"><div class="ba-box" style="${st}"></div></div></div>`;
+  const zoom = (dataUrl, label) => `<div class="ba-col"><img class="ba-zoomimg" src="${dataUrl}" alt="zoomed ${label}"></div>`;
+  return `<div class="beforeafter">${frame(bImg, 'Before')}${frame(aImg, 'After')}</div>`
+    + `<div class="ba-zoomhd">Zoomed on the boxed area</div>`
+    + `<div class="beforeafter ba-zoomrow">${zoom(baZoomDataUrl(bImg, box), 'before')}${zoom(baZoomDataUrl(aImg, box), 'after')}</div>`;
 }
 // swap the still for the real video only on demand, one at a time
 document.addEventListener('click', (e) => {
@@ -3904,15 +3938,28 @@ function progressMeta(t){
 // checked + struck through, the in-progress one is highlighted, pending ones
 // are dimmed. Empty list → nothing (workers without todos look unchanged).
 const TODO_MARK = { completed: '✓', in_progress: '▸', pending: '○' };
-function todosHtml(t){
-  const list = state.todos[t.id] ?? [];
-  if (!list.length) return '';
-  const rows = list.map((d) => {
+// P0b (Masa 2026-07-21, planner demoted): the AI's own TodoWrite list is now the
+// only breakdown a goal has (1 goal = 1 task, no invented sub-tasks). `state.todos`
+// mirrors it live via the 'todos' SSE event; once the task finishes, the server
+// freezes the same shape into `task.todosFinal` (engine/server.mjs) since the live
+// field gets stripped off finished tasks. This is the one place that resolves
+// "which list to show" — live wins while it has content, else the frozen one.
+function todoSourceFor(t){
+  const live = state.todos[t?.id];
+  if (Array.isArray(live) && live.length) return live;
+  return Array.isArray(t?.todosFinal) ? t.todosFinal : [];
+}
+function todoRowsHtml(list){
+  return list.map((d) => {
     const st = ['completed', 'in_progress', 'pending'].includes(d.status) ? d.status : 'pending';
     const label = st === 'in_progress' && d.activeForm ? d.activeForm : d.content;
     return `<div class="todo ${st}"><span class="box">${TODO_MARK[st]}</span><span class="tdtext">${esc(label ?? '')}</span></div>`;
   }).join('');
-  return `<div class="todos"><div class="tdhead">To-do</div>${rows}</div>`;
+}
+function todosHtml(t){
+  const list = todoSourceFor(t);
+  if (!list.length) return '';
+  return `<div class="todos"><div class="tdhead">To-do</div>${todoRowsHtml(list)}</div>`;
 }
 
 function taskBlock(t){
@@ -4386,7 +4433,6 @@ function renderTasks(){
   const digestRows = buildReviewDigest({ goals: pgoals, tasks: list });
   const digestCount = digestRows.length;
   const dGreen = digestRows.filter((r) => !r.testResult || r.testResult.ok !== false).length;
-  const dProof = digestRows.filter((r) => r.proof).length;
   const dPct = digestCount ? Math.round((dGreen / digestCount) * 100) : 0;
   const dBlocked = pgoals.filter((g) => g.status === 'blocked').length; // §10: reflect real blocked count, don't hardcode
   // CDO review 2026-07-08 §6: dGreen/digestCount already exclude blocked goals (they
@@ -4404,7 +4450,7 @@ function renderTasks(){
     ? `<div class="s5sum" id="summaryEntry" title="Open the review summary (ペライチ)">
         <span class="s5exp"><svg width="13" height="13" viewBox="0 0 24 24"><path d="M15 4h5v5M20 4l-7 7M9 20H4v-5M4 20l7-7"/></svg></span>
         <span class="s5ring" style="background:conic-gradient(var(--green) 0 ${dPct}%, var(--hair2) ${dPct}% 100%)"><span class="s5in">${dGreen}/${digestCount}</span></span>
-        <span class="s5tx"><span class="s5hl">${dAllClear ? 'Ready to review' : `${dGreen} of ${digestCount} ready`}</span><span class="s5sub">${dAllClear ? 'all green' : dGreen === digestCount ? 'blocked needs attention' : 'some failing'} · ${dProof} with proof · ${dBlocked ? `${dBlocked} blocked` : 'none blocked'}</span></span>
+        <span class="s5tx"><span class="s5hl">${dAllClear ? 'Ready to review' : `${dGreen} of ${digestCount} ready`}</span><span class="s5sub">${dAllClear ? 'all green' : dGreen === digestCount ? 'blocked needs attention' : 'some failing'} · ${dBlocked ? `${dBlocked} blocked` : 'none blocked'}</span></span>
       </div>`
     : '';
   const REVIEW_LIMIT = 5;
@@ -4418,16 +4464,12 @@ function renderTasks(){
     // Feedback 3: the row carries only the check-line + PR link. Tests/proof are
     // internal gate checks (they still block), NOT list clutter; the Review number
     // is redundant with the title; tapping the row opens the detail (no button).
-    // PRD §6.5: goals that landed on the same files were folded onto this one, so
-    // they ride this PR instead of opening a second one editing the same lines.
-    // Whoever opens this review must know it answers more than one request.
-    // Only rendered when it actually happened (§1.5 art.3 — no empty container).
-    const carried = pgoals.filter((x) => x.foldedInto === g.id);
-    const carryNote = carried.length ? (foldLangFor(g) === 'ja'
-      ? `+ ${carried.map((x) => `#${x.id}`).join(', ')} の仕事もここに入っています`
-      : `also carries ${carried.map((x) => `#${x.id}`).join(', ')}`) : '';
+    // PRD §6.5 → Masa 2026-07-21: goals folded onto this one are NOT announced here
+    // any more ("+ #901 の仕事もここに入っています" read as the same report-in-the-list
+    // clutter as the old "→ #12 にまとめました" row — Masa: review欄にいらない). Where
+    // the work went is logged inside the opened detail's "Also carries" section
+    // (see saItemHtml's carriesBody) instead — the list row stays just the ask.
     return `<div class="task revrow" data-goal="${g.id}"><div class="ic pending"></div><div class="tx"><div class="name">${esc(checkLine)} ${goalSourceBadge(g.source)}</div>
-     ${carryNote ? `<div class="cond carries">${carryNote}</div>` : ''}
      <div class="cond">${g.pr ? `<a class="prbig" href="${esc(g.pr)}" target="_blank">PR #${esc((g.pr.match(/pull\/(\d+)/) ?? [,''])[1])} ↗</a>` : ''}</div>
      <div class="qreplyrow" data-qreplyrow="${g.id}"${state.qreplyOpen.has(g.id) ? '' : ' hidden'}><input type="text" data-qreplyinput="${g.id}" placeholder="直したい点を書いて送信（Todoに戻して再作業）"><button type="button" class="qreplysend" data-qreplysend="${g.id}" aria-label="Send">${ICONS.up}</button></div></div>
      <div class="revquick">
@@ -4437,16 +4479,10 @@ function renderTasks(){
      </div></div>`;
   }).join('')
     + (reviewGoalsAll.length > REVIEW_LIMIT ? `<div class="more" data-reviewmore="1">${state.showAllReview ? 'show less' : `more · +${reviewGoalsAll.length - REVIEW_LIMIT}`}</div>` : '');
-  // PRD §3.5: a folded goal is never deleted — deleting it just moves the question
-  // from "why is my finished work stuck" to "where did #13 go". It keeps its row and
-  // says where its work went. Not a failure: no red, no painted surface (§1.5 art.8);
-  // no ✓/✕ either, because the thing to approve lives on the goal it was folded into.
-  const foldedRows = pgoals.filter((g) => g.status === 'folded').map((g) => {
-    const checkLine = (g.reviewSummary?.check || g.plan?.[0] || g.text || '').replace(/\s+/g, ' ').slice(0, 66);
-    const mergedNote = foldLangFor(g) === 'ja' ? `→ #${esc(String(g.foldedInto))} にまとめました` : `merged into #${esc(String(g.foldedInto))}`;
-    return `<div class="task revrow foldrow" data-goal="${g.id}"><div class="ic pending"></div><div class="tx"><div class="name">${esc(checkLine)} ${goalSourceBadge(g.source)}</div>
-     <div class="cond">${mergedNote}</div></div></div>`;
-  }).join('');
+  // PRD §3.5 → Masa 2026-07-21: a folded goal no longer gets its own "→ #12 にまとめました"
+  // row in the Review column (Masa: review欄にいらない・わかりにくい). It is never deleted —
+  // where its work went is now logged inside the carrier review's detail ("Also carries",
+  // see saItemHtml), so the column stays clean while the trail is still one tap away.
   const doneRows = grouped(list.filter((t) => ['done', 'skipped'].includes(t.status)).sort((a, b) => b.num - a.num), (t) => row(t));
   // Pending = deliberately shelved goals (e.g. "npm公開はPhase 2"): never
   // planned or queued until Start is pressed. Parked below To do on purpose.
@@ -4458,7 +4494,7 @@ function renderTasks(){
   const savedQAns = captureFocusedAnswerInput($('tasklist'), 'data-qansinput');
   $('tasklist').innerHTML =
     sec('Attention', attention, 'attn')
-    + sec(pickColumnLabel(cols, 'review', 'Review'), reviewDigest + review + foldedRows, 'reviewsec', reviewDefinitionFor(state.active).description, REVIEW_DEF_HINT)
+    + sec(pickColumnLabel(cols, 'review', 'Review'), reviewDigest + review, 'reviewsec', reviewDefinitionFor(state.active).description, REVIEW_DEF_HINT)
     + sec(pickColumnLabel(cols, 'doing', 'Doing'), doing)
     + sec(pickColumnLabel(cols, 'todo', 'To do'), todo, 'todosec')
     + sec('Pending', pendingRows, 'dim')
@@ -4910,9 +4946,6 @@ function renderRdEditor(){
     <label class="rdrow"><input type="checkbox" id="rdCardScreenshots"${state.rdDraft.reviewCard.screenshots ? ' checked' : ''}>
       <span><span class="rdlabel">スクショ / Screenshots</span>
       <span class="rdhint">レビューカードに証跡(スクショ/GIF)を表示する。</span></span></label>
-    <label class="rdrow"><input type="checkbox" id="rdCardBeforeAfter"${state.rdDraft.reviewCard.beforeAfter ? ' checked' : ''}>
-      <span><span class="rdlabel">Before/After</span>
-      <span class="rdhint">変更前後を並べて表示する(変更前キャプチャが無い場合は通常表示になります)。</span></span></label>
 
     <div class="rdsub">出力言語 / Output language</div>
     <label class="rdrow"><select id="rdLanguage">
@@ -4933,7 +4966,6 @@ function renderRdEditor(){
   $('rdPrScreenshots').addEventListener('change', (e) => { state.rdDraft.prSections.screenshots = e.target.checked; });
   $('rdPrRisk').addEventListener('change', (e) => { state.rdDraft.prSections.risk = e.target.checked; });
   $('rdCardScreenshots').addEventListener('change', (e) => { state.rdDraft.reviewCard.screenshots = e.target.checked; });
-  $('rdCardBeforeAfter').addEventListener('change', (e) => { state.rdDraft.reviewCard.beforeAfter = e.target.checked; });
   $('rdLanguage').addEventListener('change', (e) => { state.rdDraft.language = e.target.value; });
 }
 function renderWfEditor(){
@@ -5127,10 +5159,8 @@ function renderReviewChecklist(){
   $('reviewDoneMsg').classList.toggle('show', goal.status === 'done');
   // task 97: doneカードの『元に戻す』——押し間違いのApproveを取り消し、reviewへ戻す。
   $('reviewUndoBtn').onclick = () => revertDoneGoal(goal);
-  // reviewCard.screenshots off → the media block (and its "No capture yet"
-  // placeholder) is skipped entirely, not just emptied. reviewCard.beforeAfter
-  // on → try the before/after comparison (falls back to the single shot when
-  // there's no "before" capture — see beforeAfterMediaHtml).
+  // reviewCard.screenshots off → the media block is skipped entirely, not just
+  // emptied. (The Before/After switch is gone with the staged pair it controlled — P1.)
   const rc = reviewCardConfigFor(state.active);
   $('reviewList').innerHTML = reqs.map((t, i) => {
     const st = approvals[t.id];
@@ -5141,13 +5171,14 @@ function renderReviewChecklist(){
         <div class="reqhead"><span class="tick">✓</span><span class="reqnum">${t.num ?? i + 1}</span><span class="reqtitle">${esc(t.title ?? '')}</span><span class="apvd">Approved · tap to reopen</span></div>
       </div>`;
     }
-    const mediaInner = rc.beforeAfter ? beforeAfterMediaHtml(t.id, t.proof, esc(t.title ?? '')) : proofMediaHtml(t.id, t.proof, esc(t.title ?? ''));
-    const media = mediaInner || '<div class="reqnovideo">No capture yet</div>';
+    // P1: one shot when a promised check actually captured one, otherwise nothing at
+    // all — "No capture yet" was us apologising for a photo we no longer go and take.
+    const media = proofMediaHtml(t.id, t.proof, esc(t.title ?? ''));
     // task 156/234 + 160: show the WORKER'S REPORT (what it actually did/found,
     // incl. any test-only verification it describes) — that's what you read to
     // decide. Falls back to the instruction if the worker left no result.
     const report = String(t.result ?? '').replace(/^検証(PASS|FAIL)[^\n]*\n+/, '').replace(/\s+$/, '').trim();
-    const desc = report ? `<div class="reqdesc"><span class="reqreportlbl">Worker report</span>${mdLite(report.slice(0, 1600))}</div>`
+    const desc = report ? `<div class="reqdesc"><span class="reqreportlbl">Worker report</span>${mdLite(report)}</div>`
       : t.detail ? `<div class="reqdesc">${mdLite(t.detail.slice(0, 600))}</div>` : '';
     const usage = formatUsageCompact(t.usage);
     const usageBadge = usage ? `<span class="tbadge usage">${usage} tokens</span>` : '';
@@ -6093,6 +6124,11 @@ function renderFsBoard(){
   const paused = !!state.projects.find((p) => p.id === state.active)?.paused; // Doing列トグルの状態
   const { pgoals, list } = fsBucketTodo();
   const tt = (s, n) => esc(String(s ?? '').replace(/\s+/g, ' ').slice(0, n ?? 60));
+  // Card copy language — same rule as reviewLangFor(state.active) === 'ja', inlined
+  // (not calling reviewLangFor itself) because engine/test/board-goal-level.test.mjs
+  // evals this function body standalone off a minimal `state` stub with no
+  // state.reviewDefinitions; reviewLangFor would throw there.
+  const boardJa = state.reviewDefinitions?.[state.active]?.language === 'ja';
   const goalNo = new Map();
   for (const g of pgoals) { const n = goalReviewNumber({ goal: g, tasks: list }); if (n != null) goalNo.set(g.id, n); }
   // + New in EVERY column, matching the flagship Artifact: a hover-reveal "+" in the header
@@ -6107,6 +6143,35 @@ function renderFsBoard(){
   // status, so the card never repeats it. Stopped/failed carry NO bar (a dead bar would claim
   // the goal is moving) — a coloured word (Failed/Interrupted) says it instead.
   const goalTasks = (g) => list.filter((t) => t.goalId === g.id && !t.reply);
+  // P0b (Masa 2026-07-21): the card's breakdown is the AI's OWN TodoWrite list — never
+  // the planner's invented sub-tasks (there aren't any anymore; 1 goal = 1 task). Live
+  // list wins while it has content, else the frozen task.todosFinal (mirrors the
+  // module-level todoSourceFor() used by the chat feed's todosHtml — duplicated inline
+  // rather than called, same reason as boardJa above: this whole function is evaluated
+  // standalone by the board-goal-level test harness, which only stubs `state`/`esc`).
+  const todoListFor = (t) => {
+    const live = state.todos?.[t?.id];
+    if (Array.isArray(live) && live.length) return live;
+    return Array.isArray(t?.todosFinal) ? t.todosFinal : [];
+  };
+  // Capped + folded rather than scrolled (§1.5-8). While the task is actively running
+  // but hasn't reached its first TodoWrite yet, the list is genuinely empty — show one
+  // small "preparing" line instead of an empty container (§1.5-3), never a bare gap.
+  const TODO_CARD_CAP = 5, TODO_MARK_BOARD = { completed: '✓', in_progress: '▸', pending: '○' };
+  const goalCardTodosHtml = (t) => {
+    if (!t) return '';
+    const list_ = todoListFor(t);
+    if (!list_.length) return t.status === 'running' ? `<div class="lbc-todos prep">${boardJa ? '準備中…' : 'Preparing…'}</div>` : '';
+    const shown = list_.slice(0, TODO_CARD_CAP);
+    const restN = list_.length - shown.length;
+    const rows = shown.map((d) => {
+      const st = ['completed', 'in_progress', 'pending'].includes(d.status) ? d.status : 'pending';
+      const label = st === 'in_progress' && d.activeForm ? d.activeForm : d.content;
+      return `<div class="todo ${st}"><span class="box">${TODO_MARK_BOARD[st]}</span><span class="tdtext">${esc(label ?? '')}</span></div>`;
+    }).join('');
+    const more = restN > 0 ? `<div class="lbc-more">+${restN} more</div>` : '';
+    return `<div class="lbc-todos">${rows}${more}</div>`;
+  };
   const goalCard = (g) => {
     const bkt = fsBoardBucket(g, list);
     // errored is status-based now (blocked/failed/partial/interrupted live in the Needs-you column,
@@ -6131,13 +6196,21 @@ function renderFsBoard(){
       + (word ? `<span class="lbc-w ${amber ? 'intr' : 'bad'}">${word}</span>` : '')
       + '</div>'
       + `<div class="lbc-t">${tt(g.text || g.plan?.[0], 90)}</div>`;
+    // The goal's one task (post-demotion there's exactly one non-reply task; [0]
+    // is just defensive against any pre-migration leftovers with more than one).
+    const task = goalTasks(g)[0];
+    // Todos: not shown for To Do / Later — nothing has actually started yet there, so
+    // there is nothing honest to fold open (quiet stays quiet). Doing/Needs/Review/Done
+    // all show the AI's own breakdown — that's what makes it survive past "running".
+    if (task && bkt !== 'todo' && bkt !== 'later') h += goalCardTodosHtml(task);
     // n/m is Doing-only (Masa 2026-07-17): a fraction on a stopped/queued/done card is
     // residue, not progress — a card showing n/m reads unambiguously as "moving right now".
-    // done/total are the goal's real tasks, so it never invents a number.
-    if (healthyRun) {
-      const gt = goalTasks(g), total = gt.length;
+    // P0b: done/total now reflect the AI's OWN todos (not the planner's invented tasks,
+    // which under 1-goal=1-task degenerate to a useless 0/1 → 1/1).
+    if (healthyRun && task) {
+      const tl = todoListFor(task), total = tl.length;
       if (total) {
-        const done = gt.filter((t) => ['done', 'skipped'].includes(t.status)).length;
+        const done = tl.filter((d) => d.status === 'completed').length;
         h += `<div class="lbc-meter"><span class="bar"><i style="width:${Math.round(done / total * 100)}%"></i></span><span class="frac">${done}/${total}</span></div>`;
       }
     }
@@ -6227,16 +6300,11 @@ function renderFsReview(){
   const pgoals = state.goals.filter((g) => g.projectId === state.active);
   const list = state.tasks.filter((t) => t.projectId === state.active);
   const rows = buildReviewDigest({ goals: pgoals, tasks: list });
-  // PRD §3.5/§6.5: folded goals still have a row — they were never deleted, their work
-  // just went somewhere else. They must be able to hold this section up ALONE: the goal
-  // they were folded into flips back to 'running' while it re-runs their request, so
-  // there is a window where `rows` is empty and only the folded ones remain. Returning
-  // early there would make #13 vanish for the length of the re-run — exactly the "where
-  // did my finished work go" this feature exists to prevent.
-  const foldedGoals = pgoals.filter((g) => g.status === 'folded');
-  if (!rows.length && !foldedGoals.length) { el.innerHTML = ''; return; } // §1.5-3: no reviews → the UI itself does not exist
+  // PRD §3.5/§6.5 → Masa 2026-07-21: folded goals no longer get their own card in the
+  // Review column (Masa: review欄にいらない). They were never deleted — where their work
+  // went is now logged inside the carrier review's detail ("Also carries", see saItemHtml).
+  if (!rows.length) { el.innerHTML = ''; return; } // §1.5-3: no reviews → the UI itself does not exist
   const green = rows.filter((r) => !r.testResult || r.testResult.ok !== false).length;
-  const proofN = rows.filter((r) => r.proof).length;
   const blockedN = pgoals.filter((g) => g.status === 'blocked').length;
   const pct = rows.length ? Math.round((green / rows.length) * 100) : 0;
   // CDO review 2026-07-08 §6: same fix as the s5sum digest above — green/rows.length
@@ -6244,7 +6312,7 @@ function renderFsReview(){
   // ratio was already correct; "all green" is the line that used to fire regardless
   // of blockedN, reading as "18/18 ready · all green · 1 blocked" (self-contradictory).
   const allClear = green === rows.length && !blockedN;
-  const sline = `${allClear ? 'all green' : green === rows.length ? 'blocked needs attention' : `${rows.length - green} failing`} · ${proofN} with proof · ${blockedN ? `${blockedN} blocked` : 'none blocked'}`;
+  const sline = `${allClear ? 'all green' : green === rows.length ? 'blocked needs attention' : `${rows.length - green} failing`} · ${blockedN ? `${blockedN} blocked` : 'none blocked'}`;
   const RCAP = 3;
   const shown = fsUI.revMore ? rows : rows.slice(0, RCAP);
   const card = (r) => {
@@ -6264,31 +6332,18 @@ function renderFsReview(){
     // retest ×3 note (v45 §5.3): the goal came back green from a Ledger Dismiss
     const g0 = pgoals.find((g) => g.id === r.goalIds[0]);
     const rnote = g0?.retest?.passed ? `<span class="rnote">✓ ${esc(g0.retest.note)}</span>` : '';
-    // PRD §6.5: goals that landed on the same files ride this card's PR instead of
-    // opening a second one that edits the same lines. Whoever opens this review has
-    // to know it answers more than one request. Only when it happened (§1.5-3).
-    const carried = pgoals.filter((x) => r.goalIds.includes(x.foldedInto));
-    const cnote = carried.length ? `<span class="fnote">${foldLangFor(g0) === 'ja'
-      ? `+ ${carried.map((x) => `#${x.id}`).join(', ')} の仕事もここに入っています`
-      : `also carries ${carried.map((x) => `#${x.id}`).join(', ')}`}</span>` : '';
+    // PRD §6.5 → Masa 2026-07-21: goals folded onto this card's PR no longer get an
+    // inline "+ #901 の仕事もここに入っています" note here — same report-in-the-list
+    // clutter the old standalone fold row was (Masa: review欄にいらない). It is logged
+    // inside the opened detail's "Also carries" section instead (see saItemHtml).
     // Title = the goal's own title (saReviewTitle → text-first), same as the See-all
     // Ledger and the To Do list, so a review card and its list row read identically
     // (Masa 2026-07-20: this card showed reviewSummary.check "…が出ない" while the list
     // showed the request text). Plan-reviews keep their "PLAN — approve to execute:"
     // checkLine, which is the actionable label there, not a title mismatch.
-    return `<div class="rcard${img ? '' : ' noimg'}" data-fsrev="${r.goalIds[0]}">${img}<div class="rc"><div class="t">${rid}${esc(r.plan ? r.checkLine : saReviewTitle(g0))}</div>${rnote}${cnote}${pr}</div>
+    return `<div class="rcard${img ? '' : ' noimg'}" data-fsrev="${r.goalIds[0]}">${img}<div class="rc"><div class="t">${rid}${esc(r.plan ? r.checkLine : saReviewTitle(g0))}</div>${rnote}${pr}</div>
       <div class="act"><button class="b" data-fsapprove="${ids}" title="Approve">✓</button><button class="b no" data-fsdismiss="${ids}" title="Send back">✕</button></div></div>`;
   };
-  // A folded goal is settled, not failed: no ✓/✕ (the thing to approve lives on the goal
-  // it was folded into), no red, no painted surface (§1.5-8) — just its own number, what
-  // it asked for, and where the work went. Tapping it opens that goal (§1.5-9 wording).
-  const foldCard = (g) => {
-    const line = saReviewTitle(g).slice(0, 60); // title-first, same as the review card above
-    const mergedNote = foldLangFor(g) === 'ja' ? `→ #${esc(String(g.foldedInto))} にまとめました` : `merged into #${esc(String(g.foldedInto))}`;
-    return `<div class="rcard noimg fold" data-fsfold="${g.id}">
-      <div class="rc"><div class="t"><span class="rid gno" data-n="${g.id}">#${g.id}</span>${esc(line)}</div>
-      <span class="fnote">${mergedNote}</span></div></div>`;
-  };
   // V3-C (Masa決定 2026-07-07): no collapsed strip — it duplicated the sm2 meter right
   // above. Reviews render directly: meter + Review header + cards (+N more fold, §1.5-8).
   // The meter counts what is reviewable, so it only exists when something is (§1.5-3):
@@ -6304,7 +6359,7 @@ function renderFsReview(){
         <button class="rvb${fsUI.rv === 1 ? ' on' : ''}" data-fsrv="1" title="List"><svg viewBox="0 0 24 24"><rect x="3" y="4" width="7" height="7" rx="1.5"/><path d="M13 6h8M13 9h5"/><rect x="3" y="14" width="7" height="7" rx="1.5"/><path d="M13 16h8M13 19h5"/></svg></button>
         <button class="rvb${fsUI.rv === 2 ? ' on' : ''}" data-fsrv="2" title="Compact"><svg viewBox="0 0 24 24"><path d="M3 6h18M3 12h18M3 18h18"/></svg></button>
       </span></div>
-    <div class="revlist">${shown.map(card).join('')}${foldedGoals.map(foldCard).join('')}
+    <div class="revlist">${shown.map(card).join('')}
       ${rows.length > RCAP ? `<div class="morefold" id="fsRevMore">${fsUI.revMore ? 'less' : `+${rows.length - RCAP} more`}<span class="g">›</span></div>` : ''}</div>`;
   // Mirror the ready-review count onto the mobile top-bar "Review N" opener (H23).
   const _hn = document.getElementById('fsHambN'); if (_hn) _hn.textContent = green > 0 ? String(green) : '';
@@ -6324,19 +6379,8 @@ function renderFsReview(){
     if (e.target.closest('a') || e.target.closest('.act')) return;
     openReviewChecklist(Number(c.dataset.fsrev));
   };
-  // A folded goal has nothing of its own to review — its work is in the goal it was
-  // folded into, so the tap goes there. That goal is only in Review once its re-run
-  // finishes, so resolve from its live status instead of assuming (openReviewChecklist
-  // on a running goal would present an empty checklist).
-  for (const c of el.querySelectorAll('[data-fsfold]')) c.onclick = (e) => {
-    if (e.target.closest('a')) return;
-    const g = state.goals.find((x) => x.id === Number(c.dataset.fsfold));
-    const target = state.goals.find((x) => x.id === g?.foldedInto);
-    if (!target) return;
-    goToGoalInChat(target.id);
-    if (target.status === 'review') openReviewChecklist(target.id);
-    else openGoalDetail(target.id);
-  };
+  // Folded goals no longer render their own card here (Masa 2026-07-21) — where their
+  // work went is logged inside the carrier review's detail ("Also carries"), not tapped.
   for (const b of el.querySelectorAll('[data-fsapprove]')) b.onclick = async (e) => { e.stopPropagation();
     let ok = true, mergeGoal = null;
     for (const id of b.dataset.fsapprove.split(',')) {
@@ -7413,8 +7457,14 @@ async function refresh(){
     // back the already-accumulated activity for 'running' tasks.
     for (const t of state.tasks) { if (t.activity?.length) state.act[t.id] = t.activity; }
     // Same for the live to-do list: /api/state ships t.todos for running tasks
-    // (task 108/109) so a mid-run refresh restores the checklist immediately.
-    for (const t of state.tasks) { if (t.todos?.length) state.todos[t.id] = t.todos; }
+    // (task 108/109) so a mid-run refresh restores the checklist immediately. A
+    // FINISHED task has no live t.todos (trimTaskActivityForState strips it) but
+    // keeps t.todosFinal — seed from that too (P0b), so a reload of a completed
+    // goal still shows the AI's own breakdown instead of losing it silently.
+    for (const t of state.tasks) {
+      if (t.todos?.length) state.todos[t.id] = t.todos;
+      else if (t.todosFinal?.length) state.todos[t.id] = t.todosFinal;
+    }
     state.connectCommand = s.connectCommand ?? ''; state.models = s.models ?? ['sonnet']; state.agentModels = s.agentModels ?? null; state.agentEfforts = s.agentEfforts ?? null; state.agents = s.agents ?? ['claude-code', 'codex']; state.uiVersion = s.uiVersion;
     state.billing = s.billing ?? { licensed: false, email: null };
     renderAppSelect(); // picker follows the server's installed-agent set; may correct a stale connectedApp
@@ -8053,7 +8103,7 @@ function closeSummaryPera(){ $('summaryOverlay').classList.remove('show'); }
 // 4 archived / 5 reverted — session-local; the server round-trip is what
 // settles it (see saSync's "returned to review" reset).
 // ============================================================================
-const SA = { open: false, items: [], diffs: {}, diffBoxHtml: {}, detOpen: new Set(), focusGid: null, focusScrolled: false, focusTimer: null, cur: null };
+const SA = { open: false, items: [], diffs: {}, actLog: {}, shotBox: {}, detOpen: new Set(), focusGid: null, focusScrolled: false, focusTimer: null, cur: null };
 window.SA = SA; // top-level const isn't a window property; external checks probe window.SA (same convention as window.state above)
 const SA_SVG = {
   ok: '<svg viewBox="0 0 24 24"><path d="M4 12l5 5L20 6"/></svg>',
@@ -8086,8 +8136,13 @@ function saRows(){
           return { no: goalReviewNumber({ goal: sg, tasks: ts }), title: saReviewTitle(sg), cur: gid === r.goalIds[0] };
         })
       : null;
+    // Masa 2026-07-21: goals folded onto this review no longer take their own Review-column
+    // row ("→ #12 にまとめました" read as clutter). Instead this review's detail logs what
+    // it carries. Mirrors engine/lib.mjs foldedIntoReview() (browser can't import it).
+    const folded = gs.filter((x) => x.status === 'folded' && x.foldedInto != null && r.goalIds.includes(x.foldedInto))
+      .map((x) => ({ id: x.id, title: saReviewTitle(x) }));
     return {
-      ids: r.goalIds, goalId: r.goalIds[0], goal: g,
+      ids: r.goalIds, goalId: r.goalIds[0], goal: g, folded,
       no: r.reviewNumber,
       title: saReviewTitle(g),
       check: String(g?.reviewSummary?.check || r.checkLine || '').replace(/\s+/g, ' ').trim(),
@@ -8095,6 +8150,47 @@ function saRows(){
       did, siblings, risk: g?.risk ?? null,
       note: g?.retest?.passed ? g.retest.note : null,
       pr: r.pr,
+      // P0b (Masa 2026-07-21): the AI's own TodoWrite breakdown for the "Steps" section
+      // below — the primary goal's own non-reply task, frozen at completion. Distinct
+      // from `did` above (task.result prose) and from Activity (raw tool-call log).
+      todosFinal: (() => {
+        const t = ts.find((x) => x.goalId === g?.id && !x.reply);
+        return Array.isArray(t?.todosFinal) ? t.todosFinal : [];
+      })(),
+      // Screenshots the worker took of what it built (#229). Measured 2026-07-22 on real
+      // runs: it takes 1-3 depending on the ask — three variants as three shots, or four
+      // variants as ONE comparison sheet. A single shot is not a lazy shot, so none of
+      // these get thumbnailed down.
+      //
+      // Folded goals count: a goal folded into this one keeps its own tasks, and those
+      // tasks carry shots. Reading only the primary goal's tasks makes real deliverables
+      // vanish from the screen (measured: #4 folded into #3, its images disappeared).
+      shots: (() => {
+        const ids = [...r.goalIds, ...folded.map((f) => f.id)];
+        return ts.filter((t) => ids.includes(t.goalId)).flatMap((t) => t.shots ?? []);
+      })(),
+      // The task whose implementation log the engine froze at completion (P1.5). The
+      // lines are NOT in state — 1000 of them per task would bloat every page load — so
+      // this is just the id to ask for when the reviewer opens Activity.
+      logTaskId: (() => {
+        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.activityLog);
+        return t?.id ?? null;
+      })(),
+      // P1 (Masa 2026-07-21): the AI's own final report, verbatim and untrimmed — the
+      // card renders it as Markdown and interprets nothing. This is the deliverable for
+      // every ask that has no running thing to open ("research this", "explain that"),
+      // and the explanation for every ask that does.
+      report: (() => {
+        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.result);
+        return String(t?.result ?? '').replace(/^検証(PASS|FAIL)[^\n]*\n+/, '').trim();
+      })(),
+      // …and the worker's own "here's how you see it running" declaration (task.run,
+      // written by the worker into .manager-run.json). Absent for asks with nothing to
+      // run — then no button, which is the honest state, not a missing feature.
+      run: (() => {
+        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.run);
+        return t ? { taskId: t.id, ...t.run } : null;
+      })(),
       // Explicit review-summary display fields (spec §11: the card's shape follows the
       // task). A goal's reviewSummary may carry them directly (e.g. a conversational task
       // that reports an Answer + its Activity, or a ticket whose test count / diff totals
@@ -8102,19 +8198,13 @@ function saRows(){
       testResult: (typeof (g?.reviewSummary?.tests) === 'string')
         ? { ran: true, ok: true, passed: parseInt(g.reviewSummary.tests, 10) || 0, failed: 0, label: g.reviewSummary.tests }
         : r.testResult,
-      answer: g?.reviewSummary?.answer ?? null,
       activity: Array.isArray(g?.reviewSummary?.activity) ? g.reviewSummary.activity : null,
-      discussion: Array.isArray(g?.reviewSummary?.discussion) ? g.reviewSummary.discussion : null,
       changedDesc: g?.reviewSummary?.changed ?? null,
       metaFiles: Array.isArray(g?.reviewSummary?.files) ? g.reviewSummary.files : null,
       metaAdd: g?.reviewSummary?.a ?? null, metaDel: g?.reviewSummary?.d ?? null,
       ghRows: Array.isArray(g?.reviewSummary?.ghdiff) ? g.reviewSummary.ghdiff : null,
       diffTotal: g?.reviewSummary?.diffTotal ?? null,
       proofTaskId: pt?.id ?? null, proofImg: pt?.proof?.png ?? pt?.proof?.gif ?? null,
-      // Ledger before/after (docs/design/HANDOFF-review-card-evidence-first-2026-07-18.md
-      // §3 kind==='image'): most tasks never have a "before" capture, so this is almost
-      // always null and beforeAfterMediaHtml() falls back to the single-shot view.
-      proofBefore: pt?.proof?.beforePng ?? pt?.proof?.beforeGif ?? null,
       // Recording (spec §5/§6): a proof task may carry a screen recording (proof.mp4)
       // or just a recording:true flag; the goal's reviewSummary can also name one.
       // Drives the neutral "Verify recording ↗" meta chip + the Proof video player.
@@ -8123,7 +8213,6 @@ function saRows(){
       recordingUrl: pt?.proof?.mp4 ? withKey(`/proof/${pt.id}/${pt.proof.mp4}`) : (typeof g?.reviewSummary?.recording === 'string' ? g.reviewSummary.recording : null),
       // Local URL (spec §5): a running localhost the reviewer can open ("Local localhost:4809 ↗").
       // Falls back to the existing "Local run ↗" (runlocal) button when absent.
-      localUrl: g?.reviewSummary?.localUrl ?? null,
       // Client-side diff fallback: the git-backed /diff endpoint is null for seeded/
       // offline goals, so "What changed" never renders. A unified-diff string stored on
       // the goal (or any of its tasks) lets saGhDiffHtml render locally with no server.
@@ -8146,85 +8235,33 @@ function saRows(){
     };
   });
 }
-// Meta row: PR link / token savings (real tap-to-open toggle) / Run locally — under
-// the title (H23 §1). Run-locally has no real per-task command yet — this is the
-// CDO's answer to "let me test this locally", left for CTO/engine to decide what it
-// actually triggers (docs/design/HANDOFF-review-card-evidence-first-2026-07-18.md
-// §5). Shown only for tasks with a real automated check to plausibly re-run
-// (it.testResult.ran), and tapping it is honest about not being wired yet rather
-// than pretending to execute anything — never a fake "running…" state.
-function saMetaRowHtml(it, i){
-  const canRunLocal = Boolean(it.testResult?.ran);
-  if (!it.pr && !it.tokenOpt && !canRunLocal) {
-    return '<div class="lg-meta"><span class="none">No PR link / token-savings data on this task.</span></div>';
-  }
-  const chips = [];
-  if (it.pr) chips.push(`<a class="sa-prlink sm" href="${esc(it.pr)}" target="_blank" title="Open pull request">PR #${esc((it.pr.match(/pull\/(\d+)/) ?? [, ''])[1])} ↗</a>`);
-  if (it.tokenOpt) {
-    const pct = Number(it.tokenOpt.percent);
-    const label = pct > 0 ? `saved ${pct}% tokens` : 'token usage';
-    chips.push(`<button class="tokchip" type="button" data-saact="tok:${i}">${esc(label)} <i class="tokchevron">›</i></button>`);
-  }
-  if (canRunLocal) chips.push(`<button class="sa-runlocal" type="button" data-saact="runlocal:${i}" title="Run this locally"><svg viewBox="0 0 24 24"><path d="M6 4l14 8-14 8V4z"/></svg>Run locally</button>`);
-  const tokCmp = it.tokenOpt?.basis === 'baseline' ? 'vs your recent average' : 'from prompt caching';
-  const tokTech = (it.tokenOpt?.techniques ?? []).slice(0, 4).join(' / ');
-  const tokDetail = it.tokenOpt ? `<div class="tokdetail" id="tokdetail${i}"><span class="k">How this run used tokens</span><span class="v">${esc(tokCmp)} · ${esc(tokTech)}</span></div>` : '';
-  const runNote = canRunLocal ? `<div class="tokdetail" id="runlocal${i}"><span class="v">No local run command is wired for this task yet.</span></div>` : '';
-  return `<div class="lg-meta">${chips.join('')}</div>${tokDetail}${runNote}`;
-}
-// Context: closed by default, one-line verbatim-quote summary, chevron large and at
-// the far right (H23 §2). Quotes goal.text — the user's own words — never
-// task.detail/title (AI-authored; an earlier round of this design mistakenly quoted
-// the AI's own read on the request as "your ask" and it wasn't recognizable).
-function saContextHtml(it, gk){
-  const goalTxt = it.goalText;
-  const summary = goalTxt
-    ? `"${esc(goalTxt.length > 60 ? `${goalTxt.slice(0, 60)}…` : goalTxt)}"`
-    : 'No verbatim ask on record for this goal.';
-  const split = it.siblings
-    ? `<div class="lg-splitnote"><div class="hd">Split into ${it.siblings.length} goals — nothing else was dropped</div><ul>${it.siblings.map((s) => `<li class="${s.cur ? 'cur' : ''}">${s.no != null ? `#${esc(s.no)} ` : ''}${esc(s.title)}${s.cur ? ' <b>(this one)</b>' : ''}</li>`).join('')}</ul></div>`
-    : '';
-  const did = it.did?.length ? `<div class="lg-bullethd">What we did</div><ul class="lg-bullets">${it.did.map((s) => `<li>${esc(s)}</li>`).join('')}</ul>` : '';
-  const look = it.look?.length ? `<div class="lg-bullethd">What to look at</div><ul class="lg-bullets">${it.look.map((s) => `<li>${esc(s)}</li>`).join('')}</ul>` : '';
-  const inner = (split || did || look) ? `${split}${did}${look}` : '<div class="lg-notfound">Nothing else on record for this goal.</div>';
-  const open = SA.detOpen.has(`${gk}:ctx`) ? ' open' : '';
-  return `<details class="lg-goal" data-detkey="${gk}:ctx"${open}><summary><span class="k">Context</span><span class="v">${summary}</span><i class="gtoggle">›</i></summary><div class="lg-context">${inner}</div></details>`;
-}
-// Evidence hero: shape follows task kind, not one image-shaped template (H23 §3).
-// Real data check: most finished tasks have no proof image at all (logic-only /
-// docs-only / research changes verify without a screenshot) — kind is derived, see
-// saEvidenceKind() above.
-function saEvidenceHtml(kind, factRow, img, filePaths, diffText, gtr, it){
-  if (kind === 'logic') {
-    const passed = gtr?.passed ?? 0, failed = gtr?.failed ?? 0, ok = gtr?.ok !== false;
-    const n = ok ? String(passed) : `${passed}/${passed + failed}`;
-    return `<div class="lg-ev"><div class="lg-evhd">Evidence — tests, no UI to show</div>
-      <div class="lg-herostat${ok ? '' : ' fail'}"><span class="n">${esc(n)}</span><span class="l">tests passed<br>${esc(filePaths.join(', '))}</span></div></div>`;
-  }
-  if (kind === 'docs') {
-    const lines = diffText ? diffText.split('\n').filter((l) => l.length) : [];
-    const diffHtml = lines.map((l) => {
-      const isAdd = l.startsWith('+');
-      return `<div class="ln${isAdd ? ' a' : ''}"><span class="gutter">+</span><span class="txt">${esc(l.slice(1) || l)}</span></div>`;
-    }).join('');
-    return `<div class="lg-ev"><div class="lg-evhd">Evidence — the diff, no test to run</div>
-      <div class="row"><span class="fls">${esc(filePaths.join(', '))} · docs-only, no test target</span></div>
-      ${diffHtml ? `<div class="lg-diffblock">${diffHtml}</div>` : ''}</div>`;
-  }
-  if (kind === 'missing') {
-    const note = it.check || 'Please verify this change manually — no automated capture exists for it.';
-    return `<div class="lg-ev"><div class="lg-missinghd"><svg viewBox="0 0 24 24"><path d="M12 8v5M12 16.5v.5"/><path d="M10.3 3.9 2.4 18a1.9 1.9 0 0 0 1.7 2.9h15.8a1.9 1.9 0 0 0 1.7-2.9L13.7 3.9a1.9 1.9 0 0 0-3.4 0z"/></svg>No capture — this touched the UI and should have one</div>
-      <div class="lg-manualnote">${esc(note)}</div></div>`;
-  }
-  // kind === 'image' (default)
-  return `<div class="lg-ev"><div class="lg-evhd">Evidence</div><div class="row">${factRow}</div>${img}</div>`;
-}
+// (P1, Masa 2026-07-21) saMetaRowHtml() was deleted here — the pre-redesign meta row,
+// uncalled since the review-card redesign. Its "Run locally" chip only ever opened a
+// note saying no run command was wired; the card's Open it chip now starts the real one.
+// (saContextHtml() was deleted here — Masa 2026-07-22. The pre-redesign Context fold:
+// verbatim quote, 'What we did' / 'What to look at' bullets, sibling split note. Nothing
+// has called it since the review-card redesign; Request and Result carry that ground now.)
 // Unified-diff → GitHub two-gutter model (REVIEW-CARD-SPEC §7 "GitHub diff = Bright"):
 // walk the @@ hunk headers to recover the old/new line numbers, emit one row per line
 // {o:oldNo|null, n:newNo|null, t:'add'|'del'|'ctx', c:text}. Word-level highlight is
 // NOT fabricated from a raw diff (the product never invents evidence) — the Bright
 // line+gutter tint carries the change and .gh-hi simply stays absent. Pure, so it is
 // extracted and unit-tested in engine/test/review-card-diff.test.mjs.
+// How many body rows the diff really has, ignoring the display cap. A big change used
+// to be cut at 80 rows with nothing said (SPEC §7 asks for "…+N 行 · GitHub で開く ↗"
+// precisely so a reviewer is never shown a slice that looks like the whole thing).
+// Same skip rules as saUnifiedToGh below — headers and @@ markers are not rows.
+function saDiffRowCount(diff){
+  let n = 0;
+  for (const raw of String(diff || '').split('\n')) {
+    if (raw === '') continue;
+    if (/^(diff --git|index |--- |\+\+\+ |new file|deleted file|old mode|new mode|similarity|rename |Binary )/.test(raw)) continue;
+    if (/^@@ -\d+(?:,\d+)? \+\d+(?:,\d+)? @@/.test(raw)) continue;
+    if (raw.startsWith('\\')) continue;
+    n++;
+  }
+  return n;
+}
 function saUnifiedToGh(diff, max){
   const cap = max || 80; const out = []; let o = 0, n = 0;
   for (const raw of String(diff || '').split('\n')) {
@@ -8240,10 +8277,13 @@ function saUnifiedToGh(diff, max){
   }
   return out;
 }
-function saGhDiffHtml(diff){
+function saGhDiffHtml(diff, pr){
   const rows = saUnifiedToGh(diff);
   if (!rows.length) return '';
-  return saGhDiffRowsHtml(rows);
+  // Say what was left out. Before this, a 5000-line change rendered its first 80 rows
+  // and stopped — a quiet lie, and the one thing SPEC §7 names as forbidden.
+  const more = Math.max(0, saDiffRowCount(diff) - rows.length);
+  return saGhDiffRowsHtml(rows, more, pr || '');
 }
 // One Activity transcript row (spec §3): You / tool (mono, bold name) / res (green) /
 // Claude. Mirrors the flagship artifact's tx mapping so the pulldown pixel-matches.
@@ -8253,6 +8293,18 @@ function saTxRowHtml(e){
   if (e[0] === 'user') return `<div class="txr you"><span class="w">You</span><span class="m">${esc(e[1])}</span></div>`;
   return `<div class="txr cl"><span class="w">Claude</span><span class="m">${esc(e[1])}</span></div>`;
 }
+// One line of the frozen implementation log. The worker's own tool calls come through
+// as "Read app/index.html" / "Bash npm test"; the Manager's own progress notes are
+// prefixed with ✎ by sendProcessAct() and read as commentary, so they take the quieter
+// row. Nothing is parsed beyond that split — the line is shown as it was written.
+function saLogRowHtml(line){
+  const s = String(line ?? '');
+  if (s.startsWith('✎')) return `<div class="txr note"><span class="w"></span><span class="m">${esc(s.replace(/^✎\s*/, ''))}</span></div>`;
+  const m = s.match(/^([A-Z][A-Za-z]*)\s+([\s\S]*)$/);
+  return m
+    ? `<div class="txr tool"><span class="w">tool</span><span class="m"><b>${esc(m[1])}</b> ${esc(m[2])}</span></div>`
+    : `<div class="txr cl"><span class="w"></span><span class="m">${esc(s)}</span></div>`;
+}
 // Structured GitHub diff (REVIEW-CARD-SPEC §7): renders {o,n,t,c,h} rows with a
 // word-level highlight (h) and, for a huge change, a "…+N 行 · GitHub で開く ↗"
 // footer (only linking to GitHub when a PR exists). Mirrors the flagship artifact's
@@ -8289,8 +8341,18 @@ function saVideoFigHtml(url){
 // No fabricated pipeline: an untested goal shows □ Tested, not a green ✓.
 function saWorkflowHtml(it){
   const ran = Boolean(it.testResult?.ran);
-  const hasProof = Boolean(it.proofImg && it.proofTaskId != null);
-  const steps = [['Implemented', 'ok'], ['Tested', ran ? 'ok' : 'todo'], ['Proof captured', hasProof ? 'ok' : 'todo'], ['In review', 'now'], ['Merge', 'todo'], ['Deploy', 'todo']];
+  // P1 (Masa 2026-07-21): "Proof captured □" is gone. We no longer go take photos of
+  // the work, so an empty box there was scoring the goal against a step that isn't
+  // part of the process any more — a permanent unchecked box nobody could ever check.
+  // "Deploy" is gone too (Masa 2026-07-22): nothing in the product observes a deploy,
+  // so that box could never be ticked — the same permanent-□ problem as Proof captured.
+  // Merge stays because it IS observed: the PR poller flips the goal to done on merge.
+  //
+  // (PRD §5.5-A/15 asks for the USER'S OWN pipeline — PRD→design→implement→check→PR→
+  // merge→deploy, varying per person. There is no place to record such a pipeline yet,
+  // so inventing one here would be this layer guessing someone's process. Left as a
+  // decision, not faked.)
+  const steps = [['Implemented', 'ok'], ['Tested', ran ? 'ok' : 'todo'], ['In review', 'now'], ['Merge', 'todo']];
   return '<div class="rwf">' + steps.map(([s, st]) => `<div class="wf-s ${st}"><span class="mk">${st === 'ok' ? '✓' : st === 'now' ? '→' : '□'}</span>${s}</div>`).join('') + '</div>';
 }
 function saItemHtml(it, i){
@@ -8304,15 +8366,9 @@ function saItemHtml(it, i){
   const d = SA.diffs[it.goalId];
   const gtr = it.testResult;
   const hasProofImg = it.proofImg && it.proofTaskId != null;
-  // Before/after (spec §6): reuse the same beforeafter/ba-col/balabel renderer as the
-  // per-task checklist (beforeAfterMediaHtml). it.proofBefore is null for the ~all
-  // tasks with no "before" capture, so it falls back to the single <img class="proofimg">.
-  // Diff-box + zoom inset is computed async after first paint (saEnhanceDiffBoxes);
-  // SA.diffBoxHtml caches it so the box survives every later saRender().
-  const boxedImg = SA.diffBoxHtml[it.goalId];
-  const img = hasProofImg
-    ? (typeof boxedImg === 'string' ? boxedImg : beforeAfterMediaHtml(it.proofTaskId, { png: it.proofImg, beforePng: it.proofBefore }, esc(it.title ?? '')))
-    : '';
+  // Proof image (P1): the single capture from a check the AI itself promised and ran.
+  // No pair, no red box — those compared against a "before" shot we used to stage.
+  const img = hasProofImg ? proofMediaHtml(it.proofTaskId, { png: it.proofImg }, esc(it.title ?? '')) : '';
   // Changed-file paths: d from the diff endpoint ({add,del,files:[{path,add,del}],diff}),
   // falling back to task.changedFiles when the server has no diff for this goal.
   // A usable git diff has actual files; the endpoint returns an empty {add:0,del:0,
@@ -8339,9 +8395,9 @@ function saItemHtml(it, i){
   const kind = saEvidenceKind(filePaths, hasProofImg, gtr);
   const gk = it.goalId;
   const rs = it.goal?.reviewSummary || {};
-  // Conversational task (no code change): drops Proof + What changed, shows an Answer
-  // instead (spec §11); its Activity IS the conversation.
-  const isConv = Boolean(it.answer);
+  // A conversational task (no code change) has no Proof / What changed to show. Its
+  // Result IS the answer (P1) — the old separate Answer section is gone.
+  const isConv = !it.report && !filePaths.length;
   // ── meta (spec §5) — fully data-driven / hide-empty: each chip renders only when its
   // data exists (no PR → no PR chip · no tests → no Tests · no file totals → no Files ·
   // no recording → no Verify). PR = bold green, no underline (a real <a href>). ──
@@ -8358,11 +8414,13 @@ function saItemHtml(it, i){
   if (testsLabel) chips.push(`<span class="mx"><span class="mk">Tests</span> <b class="ok">${esc(testsLabel)}</b></span>`);
   // Files: hide-empty — only with changed files AND totals (a bare-count chip reads odd).
   if (nf && fAdd) chips.push(`<span class="mx"><span class="mk">Files</span> <b>${nf}</b> <span class="mmut">· <i class="a">${esc(fAdd)}</i> <i class="d">${esc(fDel)}</i></span></span>`);
-  // Local: an open localhost URL (neutral, not green). A bare host:port isn't navigable in
-  // the demo, so it renders as a non-navigating chip; a real URL opens in a new tab.
-  if (it.localUrl) { const lu = String(it.localUrl); const isUrl = /^https?:/.test(lu);
-    chips.push(`<a class="mx mlink"${isUrl ? ` href="${esc(lu)}" target="_blank" rel="noopener"` : ' style="cursor:pointer"'} title="Open the running app"><span class="mk">Local</span> <b>${esc(lu.replace(/^https?:\/\//, ''))}</b>${arw}</a>`); }
-  else if (gtr?.ran) chips.push(`<button type="button" class="mx mlink" data-saact="runlocal:${i}" title="Run this locally"><span class="mk">Local</span> <b>run</b>${arw}</button>`);
+  // (The Local chip was deleted here — Masa 2026-07-22. It read reviewSummary.localUrl,
+  // a key parseReviewSummary never emits, so it only ever appeared in demo data; and the
+  // real thing it promised is the Open it section.)
+  // (Open it is NOT a meta chip — Masa 2026-07-22. It was saying the same thing as the
+  // Open it section below, in the PR chip's vocabulary, so the card spoke two languages
+  // about one thing. It lives in the body now, in the same typography as every other
+  // section.)
   // Verify = a screen recording of the check (spec §5): NEUTRAL. Opens the file when there
   // is a real URL, otherwise scrolls to the in-card player.
   if (it.recording) chips.push(it.recordingUrl
@@ -8371,23 +8429,69 @@ function saItemHtml(it, i){
   const tokCmp = it.tokenOpt?.basis === 'baseline' ? 'vs your recent average' : 'from prompt caching';
   const tokTech = (it.tokenOpt?.techniques ?? []).slice(0, 4).join(' / ');
   const metaDetail = (it.tokenOpt && tokPct > 0) ? `<div class="rmeta-detail" id="tokdetail${i}"><span class="k">How this run used tokens</span><span class="v">${esc(tokCmp)}${tokTech ? ` · ${esc(tokTech)}` : ''}</span></div>` : '';
-  const runNote = gtr?.ran ? `<div class="rmeta-detail" id="runlocal${i}"><span class="v">No local run command is wired for this task yet.</span></div>` : '';
-  const meta = `<div class="rmeta">${chips.join('')}</div>${metaDetail}${runNote}`;
+  // §1.5-3: with Open it moved into the body, a card with no PR / tokens / tests has no
+  // chips at all — render no meta row rather than an empty one holding open a gap.
+  const meta = chips.length ? `<div class="rmeta">${chips.join('')}</div>${metaDetail}` : '';
   // ── Request (spec §4): a PLAIN section — the user's own short words (goal.text). Never
   // an accordion, never the implementation spec (that lives in Activity below). ──
   const reqSec = it.goalText
     ? `<section class="rsec rreq"><div class="rsec-h">Request</div><p class="rbody rask">${esc(it.goalText)}</p></section>`
     : '';
+  // ── Open it (Masa 2026-07-22): a plain section in the card's own typography — the same
+  // heading + body shape as Request and Result, not a chip in another vocabulary. The
+  // address the worker declared IS the control (press it → the process starts on demand →
+  // it opens); its own note sits under it, and that note is where the status and any
+  // failure reason land after the press. Nothing declared → no section (§1.5-3). ──
+  const openBody = it.run
+    ? `<div class="ropen" id="runlocal${i}"><button type="button" class="olink" data-saact="preview:${i}" id="prevlink${i}">${esc(it.run.url || 'Start it')}<span class="ar">↗</span></button><p class="rbody v">${esc(it.run.note || it.run.cmd || 'Starts when you press it.')}</p></div>`
+    : '';
+  // ── Also carries (Masa 2026-07-21): goals folded onto this review are logged HERE, inside
+  // the review's own content — not as loose "→ #12 にまとめました" rows in the Review column.
+  // Chrome label stays English (§377-7); the request titles keep whatever language they were
+  // written in. Only rendered when something was actually folded in (§1.5-3 no empty box). ──
+  const carriesBody = it.folded?.length
+    ? `<ul class="rcheck">${it.folded.map((f) => `<li>#${esc(String(f.id))} ${esc(f.title)}</li>`).join('')}</ul>`
+    : '';
   // ── What to check (spec §2.5): a bulleted list when several, a line when one. ──
   // Only an explicit review-summary check renders here — never the goal text echoed back
   // (it.check falls back to goal.text via checkLine, which would fabricate a "What to
   // check" for a conversational task that has none). §0.5 no placeholder.
-  const checkArr = Array.isArray(rs.checkList) ? rs.checkList.filter(Boolean)
+  // The working AI's own sentence wins (P1, Masa 2026-07-22): it declares what to look
+  // at in the same file it declares how to run the thing. reviewSummary.check below is
+  // the intake AI's invention and only fills in for goals that predate the declaration.
+  const checkArr = it.run?.check ? [it.run.check]
+    : Array.isArray(rs.checkList) ? rs.checkList.filter(Boolean)
     : Array.isArray(rs.check) ? rs.check.filter(Boolean)
     : (typeof rs.check === 'string' && rs.check.trim() ? [rs.check.trim()] : []);
   const checkBody = checkArr.length > 1
     ? `<ul class="rcheck">${checkArr.map((c) => `<li>${esc(c)}</li>`).join('')}</ul>`
     : (checkArr.length ? `<p class="rbody rcheck1">${esc(checkArr[0])}</p>` : '');
+  // ── Steps (P0b, Masa 2026-07-21): the AI's own TodoWrite breakdown — distinct from
+  // "What to check" (the human's pass condition) and Activity (the raw tool-call log).
+  // Reuses the same .todos/.todo vocabulary as the chat feed / board card so the ✓/▸/○
+  // language is consistent everywhere it appears. Empty → no section (§1.5-3).
+  const stepsBody = it.todosFinal?.length ? `<div class="todos">${todoRowsHtml(it.todosFinal)}</div>` : '';
+  // ── Result (P1, Masa 2026-07-21): the AI's own report, rendered as written and read
+  // whole (no 160-char squeeze, no summariser of ours in the way). This is the one
+  // section that carries every kind of ask — a研究 report reads here as prose, a UI
+  // change explains itself here, and an image the AI pasted shows up here as an image.
+  // We interpret nothing; mdLite only turns Markdown into the shapes it already means. ──
+  const reportBody = it.report ? `<div class="rreport">${mdLite(it.report, { images: false })}</div>` : '';
+  // ── Screenshots (P2, Masa 2026-07-22): the pictures the AI took of its own result.
+  // Full width, at their own aspect — a four-variant comparison sheet is unreadable
+  // shrunk into a thumbnail, and that is the shape the AI actually produces. The file
+  // name it chose is the caption (it names them variant-1-linear.png / compare.png), so
+  // it labels its own work; we add nothing. ──
+  // A before/after pair is two shots the worker named as such — it decides whether the
+  // change had a "before" worth shooting, we only recognise the naming. Anything else is
+  // shown as the plain list.
+  const shotPair = saShotPair(it.shots, it.run?.compare);
+  const boxed = shotPair ? SA.shotBox[it.goalId] : null;
+  const shotsBody = shotPair
+    ? `<div class="rshotpair" data-shotpair="${it.goalId}">${typeof boxed === 'string' ? boxed : saPlainPairHtml(shotPair)}</div>`
+    : it.shots?.length
+      ? it.shots.map((sh) => `<figure class="rshot"><img src="${esc(sh.url)}" alt="${esc(sh.name || 'screenshot')}" loading="lazy"><figcaption>${esc(String(sh.name || '').replace(/\.[a-z]+$/i, ''))}</figcaption></figure>`).join('')
+      : '';
   // ── Proof (spec §6/§11): before/after (identical size) when a real image exists; a
   // run-log for logic-only tests; dropped entirely for a conversational task. ──
   let proofInner = '';
@@ -8400,14 +8504,17 @@ function saItemHtml(it, i){
   // peeks then "…+N 行 · GitHub で開く ↗" for a huge change (no GitHub link when there's no
   // PR). Structured recorded rows (ghRows) win over a git/client unified diff. Dropped
   // entirely for a conversational task. ──
-  const changedDescHtml = (it.changedDesc || rs.changed) ? `<p class="rbody rdesc">${esc(String(it.changedDesc || rs.changed).slice(0, 600))}</p>` : '';
+  // No prose here (P1, Masa 2026-07-22): reviewSummary.changed was the intake AI writing
+  // its own summary of the change, and on a real goal it came out as the request echoed
+  // back. What changed is now only what actually changed — files and the diff.
+  const changedDescHtml = '';
   const filesListHtml = it.metaFiles
     ? `<div class="rfiles">${it.metaFiles.map((f) => esc(f)).join('<br>')}</div>`
     : (d && d.files.length) ? `<div class="rfiles">${d.files.map((f) => `${esc(f.path)}  <i class="a">+${f.add}</i> <i class="d">−${f.del}</i>`).join('<br>')}</div>`
       : (clientDiff && filePaths.length) ? `<div class="rfiles">${filePaths.map((p) => esc(p)).join('<br>')}</div>` : '';
   const diffHtml = it.ghRows
     ? saGhDiffRowsHtml(it.ghRows, it.diffTotal || 0, it.pr || '')
-    : hasGitDiff ? saGhDiffHtml(d.diff) : clientDiff ? saGhDiffHtml(clientDiff) : '';
+    : hasGitDiff ? saGhDiffHtml(d.diff, it.pr) : clientDiff ? saGhDiffHtml(clientDiff, it.pr) : '';
   const changedBody = isConv ? '' : (changedDescHtml + filesListHtml + diffHtml);
   // ── sensitive-area note (real: classifyChangeRisk forced the pause) ──
   const risk = it.risk;
@@ -8419,31 +8526,50 @@ function saItemHtml(it, i){
   // below, offset-indented. The review Discussion folds in at the bottom of the body. The
   // transcript is the recorded build (activity); real goals fall back to "what we did".
   // Persisted open-state via data-detkey so SSE re-renders don't collapse it. ──
+  // The "what we did" fallback is the same report as Result above, cut to 160 chars —
+  // showing both makes the card repeat itself in a worse form (P1). Only fall back when
+  // there is no Result to read.
   const txArr = (it.activity && it.activity.length)
     ? it.activity.map((e) => saTxRowHtml(e))
-    : (it.did?.length ? it.did.map((s) => saTxRowHtml(['claude', s])) : []);
-  const discArr = it.discussion || it.thread || null;
-  const threadHtml = (discArr && discArr.length)
-    ? discArr.map((m) => `<div class="rmsg${m[0] === 'you' ? ' you' : ''}"><div class="who">${m[0] === 'you' ? 'You' : 'Manager'}</div><div class="msg">${esc(m[1])}</div></div>`).join('')
-    : '';
+    : ((!reportBody && it.did?.length) ? it.did.map((s) => saTxRowHtml(['claude', s])) : []);
+  // (Discussion was deleted here — Masa 2026-07-22. It read reviewSummary.discussion,
+  // which parseReviewSummary never emits, so it never rendered on a real goal; replies
+  // already live in the card's own reply box.)
+  const threadHtml = '';
+  // ── Activity = the implementation log (P1.5, Masa 2026-07-22): what the AI actually
+  // did, read AFTER it finished. Fetched only when the section is opened — a task's log
+  // runs to ~500KB and most cards are never expanded — so the closed state is one line
+  // and the lines arrive on the first open (same "only when asked" shape as Open it).
+  // Once fetched they live in SA.actLog so an SSE rebuild doesn't wipe them.
+  const logLines = it.logTaskId != null ? SA.actLog[it.logTaskId] : null;
+  const logHtml = Array.isArray(logLines)
+    ? (logLines.length
+      ? logLines.map((l) => saLogRowHtml(l)).join('')
+      : '<div class="txr res"><span class="w"></span><span class="m">Nothing was logged for this one.</span></div>')
+    : '<div class="txr res"><span class="w"></span><span class="m">Loading…</span></div>';
   const actOpen = SA.detOpen.has(`${gk}:act`) ? ' open' : '';
   const txFirst = txArr[0] || '<div class="txr"><span class="w"></span><span class="m">Activity</span></div>';
   const txRest = txArr.slice(1).join('');
   const discInner = threadHtml ? `<div class="rsub">Discussion</div><div class="rthread">${threadHtml}</div>` : '';
   const activitySec = (txArr.length || threadHtml)
     ? `<section class="rsec ract"><div class="rsec-h">Activity</div><details class="lacc" data-detkey="${gk}:act"${actOpen}><summary><i class="ltog">›</i><div class="rtx lfirst">${txFirst}</div></summary><div class="lbody">${txRest ? `<div class="rtx">${txRest}</div>` : ''}${discInner}</div></details></section>`
-    : '';
+    : it.logTaskId != null
+      ? `<section class="rsec ract"><div class="rsec-h">Activity</div><details class="lacc" data-detkey="${gk}:act" data-actlog="${it.logTaskId}"${actOpen}><summary><i class="ltog">›</i><div class="rtx lfirst">${saTxRowHtml(['user', it.goalText || it.title])}</div></summary><div class="lbody"><div class="rtx" data-actbody="${it.logTaskId}">${logHtml}</div>${discInner}</div></details></section>`
+      : '';
   const S = (label, body) => body ? `<section class="rsec"><div class="rsec-h">${label}</div>${body}</section>` : '';
   return `<div class="sa-item lg redesign${it.goalId === SA.cur ? ' cur' : ''}" data-gid="${it.goalId}">
     <div class="rhead"><span class="rtag">${it.no != null ? `#${it.no}` : ''}</span><span class="rst"><span class="rdot"></span>In review</span></div>
     <h2 class="rtitle">${esc(it.title)}</h2>
     ${meta}
     ${reqSec}
+    ${S('Open it', openBody)}
+    ${S('Also carries', carriesBody)}
     ${S('What to check', checkBody)}
-    ${S('Answer', it.answer ? `<p class="rbody">${esc(it.answer)}</p>` : '')}
+    ${S('Result', reportBody)}
+    ${S('Screenshots', shotsBody)}
+    ${S('Steps', stepsBody)}
     ${S('Proof', proofInner)}
     ${S('What changed', changedBody)}
-    ${S('Not done', it.risk?.summary ? `<p class="rbody rrisk">${esc(it.risk.summary)}</p>` : '')}
     ${riskHtml}
     ${activitySec}
     ${S('Workflow', saWorkflowHtml(it))}
@@ -8458,12 +8584,45 @@ function saScrollCardToTop(el){
   if (!body || !el) return;
   body.scrollTop += el.getBoundingClientRect().top - body.getBoundingClientRect().top;
 }
+// Fetch one task's frozen implementation log and paint it into the open card. Cached in
+// SA.actLog (including an empty result) so a card can be opened and closed without
+// asking twice. A failure writes an honest line rather than an empty box — the log
+// either exists or it doesn't, and we say which.
+async function saLoadActivityLog(taskId){
+  if (!Number.isFinite(taskId) || SA.actLog[taskId] !== undefined) return;
+  SA.actLog[taskId] = null; // in flight — keeps a second toggle from re-fetching
+  let lines = [];
+  try {
+    const r = await fetch(withKey(`/api/tasks/${taskId}/activity`));
+    lines = r.ok ? ((await r.json()).lines ?? []) : [];
+  } catch { lines = []; }
+  SA.actLog[taskId] = lines;
+  const body = document.querySelector(`#seeall [data-actbody="${taskId}"]`);
+  if (!body) return;
+  body.innerHTML = lines.length
+    ? lines.map((l) => saLogRowHtml(l)).join('')
+    : '<div class="txr res"><span class="w"></span><span class="m">Nothing was logged for this one.</span></div>';
+}
+// After first paint, compute the red box for any card showing a before/after pair and
+// patch that card's node directly — a full re-render here would close every open
+// pulldown just to draw a box. Cached per goal (null included) so it runs once.
+async function saEnhanceShotBoxes(){
+  const targets = SA.items.filter((it) => saShotPair(it.shots, it.run?.compare) && SA.shotBox[it.goalId] === undefined);
+  await Promise.all(targets.map(async (it) => {
+    const pair = saShotPair(it.shots, it.run?.compare);
+    let html = null;
+    try { html = await shotsBoxedHtml(pair.before, pair.after); } catch { html = null; }
+    SA.shotBox[it.goalId] = html;
+    if (!html || !SA.open) return;
+    const host = document.querySelector(`#seeall [data-shotpair="${it.goalId}"]`);
+    if (host) host.innerHTML = html;
+  }));
+}
 function saRender(){
   const S = $('seeall'); if (!S) return;
   const pending = SA.items.filter((x) => x.st === 0);
   const green = pending.filter((x) => !x.testResult || x.testResult.ok !== false).length;
-  const proofN = pending.filter((x) => x.proofImg).length;
-  const c = `${pending.length} ready · ${green === pending.length ? 'all green' : `${pending.length - green} failing`} · ${proofN} with proof`;
+  const c = `${pending.length} ready · ${green === pending.length ? 'all green' : `${pending.length - green} failing`}`;
   const curIdx = pending.findIndex((x) => x.goalId === SA.cur);
   const posLabel = pending.length && curIdx >= 0 ? `${curIdx + 1} / ${pending.length}` : '';
   S.innerHTML = `<div class="sa-panel"><div class="sa-hd"><span class="t">Review</span><span class="c">${c}</span><span class="sa-count">${posLabel}</span><span class="keys">↑↓ MOVE · A APPROVE · D DISMISS · ESC CLOSE</span><button class="sa-x" data-saact="close:0">✕</button></div>
@@ -8475,7 +8634,14 @@ function saRender(){
     det.addEventListener('toggle', () => {
       const k = det.dataset.detkey;
       if (det.open) SA.detOpen.add(k); else SA.detOpen.delete(k);
+      // Opening Activity is what fetches the implementation log (Masa 2026-07-22:
+      // "開いた時だけ取りに行く"). Closed cards cost nothing; a reopened one is served
+      // from SA.actLog.
+      if (det.open && det.dataset.actlog) saLoadActivityLog(Number(det.dataset.actlog));
     });
+    // A card re-rendered while its Activity was already open (SSE tick) keeps the log
+    // it had — the fetch below no-ops once SA.actLog holds the answer.
+    if (det.open && det.dataset.actlog) saLoadActivityLog(Number(det.dataset.actlog));
   }
   // "ないものは載せない" (Masa): a Proof <img> whose file is missing/broken must never
   // show a broken-image placeholder. proofImg is only a stored filename, so a file that
@@ -8528,25 +8694,6 @@ async function saFetchDiffs(){
   }));
   if (SA.open) saSync();
 }
-// Diff-box + zoom inset (HANDOFF-review-card-diffbox-2026-07-18): unlike
-// saFetchDiffs above (whose fetch settling triggers a full saSync()/saRender()), this
-// patches each card's already-rendered .beforeafter node directly — a full re-render
-// here would reset scroll, close every open <details>, and drop an in-progress reply
-// keystroke, all just to draw a box around one screenshot. SA.diffBoxHtml is still
-// cached per goal, so if a later SSE-driven saSync() does rebuild the card anyway,
-// saItemHtml() picks the cached result back up with no flicker to the plain pair.
-async function saEnhanceDiffBoxes(){
-  const targets = SA.items.filter((it) => it.proofImg && it.proofTaskId != null && it.proofBefore && SA.diffBoxHtml[it.goalId] === undefined);
-  await Promise.all(targets.map(async (it) => {
-    let html = null;
-    try { html = await beforeAfterBoxedHtml(it.proofTaskId, { png: it.proofImg, beforePng: it.proofBefore }, esc(it.title ?? '')); }
-    catch { html = null; } // a decode/pixel-diff failure is a fallback, never a crash
-    SA.diffBoxHtml[it.goalId] = html; // cache even on null — never re-attempt an already-answered "no box"
-    if (!html || !SA.open) return;
-    const host = document.querySelector(`.sa-item[data-gid="${it.goalId}"] .beforeafter`);
-    if (host) host.outerHTML = html; // html itself opens with the replacement .beforeafter
-  }));
-}
 function saOpen(focusGoalId){
   SA.open = true;
   SA.items = saRows().map((r) => ({ ...r, st: 0 }));
@@ -8576,7 +8723,7 @@ function saOpen(focusGoalId){
     if (el) saScrollCardToTop(el);
   });
   saFetchDiffs();
-  saEnhanceDiffBoxes();
+  saEnhanceShotBoxes();
 }
 function saClose(){ SA.open = false; SA.focusGid = null; if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; } $('seeall').classList.remove('open'); }
 // Continuous review (Masa 2026-07-08): move a persistent "current" card with
@@ -8635,10 +8782,32 @@ $('seeall').addEventListener('click', async (e) => {
   const it = SA.items[i];
   if (k === 'close') return saClose();
   if (!it) return;
-  // meta-row toggles (H23): tap the token chip to reveal the technique breakdown, tap
-  // Run-locally to reveal the honest "not wired yet" note — neither performs an action.
+  // meta-row toggle (H23): tap the token chip to reveal the technique breakdown.
   if (k === 'tok') { b.classList.toggle('open'); $('seeall').querySelector(`#tokdetail${i}`)?.classList.toggle('open'); return; }
-  if (k === 'runlocal') { $('seeall').querySelector(`#runlocal${i}`)?.classList.toggle('open'); return; }
+  // Open it (P1, Masa 2026-07-21): start the thing the worker said to run, then open it.
+  // Booting a real process takes seconds, so the chip says "Starting…" while it waits —
+  // and a failure lands in the detail line as a plain reason. It never turns the goal
+  // red: failing to boot a preview says nothing about whether the work itself is good.
+  if (k === 'preview') {
+    if (!it.run?.taskId) return;
+    const sec = $('seeall').querySelector(`#runlocal${i}`);
+    const link = $('seeall').querySelector(`#prevlink${i}`);
+    const v = sec?.querySelector('.v');
+    if (link) link.disabled = true;
+    if (v) v.textContent = 'Starting it…';
+    let r = null, body = {};
+    try { r = await fetch(`/api/tasks/${it.run.taskId}/preview`, { method: 'POST' }); body = await r.json().catch(() => ({})); }
+    catch (e) { body = { error: String(e.message ?? e) }; }
+    if (link) link.disabled = false;
+    const noteText = it.run.note || it.run.cmd || 'Starts when you press it.';
+    // A tab opened: that IS the feedback. Put the worker's own note back rather than
+    // narrating what the reviewer just watched happen.
+    if (r?.ok && body.url) { window.open(body.url, '_blank', 'noopener'); if (v) v.textContent = noteText; return; }
+    // Started, but the worker declared no URL: its own note is where it said to look.
+    if (r?.ok) { if (v) v.textContent = it.run.note || it.run.cmd || 'Started — nothing to open.'; return; }
+    if (v) v.textContent = `Could not start it: ${body.error || `HTTP ${r?.status ?? 'error'}`}`;
+    return;
+  }
   // Verify chip (no external recording URL) → scroll the in-card video player into view.
   if (k === 'rec') { $('seeall').querySelector(`.sa-item[data-gid="${it.goalId}"] .rvid`)?.scrollIntoView({ behavior: 'smooth', block: 'center' }); return; }
   // every verdict auto-advances the focus to the next pending card (tempo).