papilot 0.6.0 → 0.8.0

package/dist/index.html

@@ -8,6 +8,6 @@
   document.documentElement.dataset.theme = dark ? 'dark' : 'light';
   document.querySelector('meta[name="theme-color"]').content = dark ? '#26272b' : '#ffffff';
 })();
-</script><style>html{--lightningcss-light:initial;--lightningcss-dark: ;color-scheme:light;background:#fff}html[data-theme=dark]{--lightningcss-light: ;--lightningcss-dark:initial;color-scheme:dark;background:#26272b}</style>  <script type="module" crossorigin src="/assets/index-Bt9UXsGX.js"></script>
-  <link rel="stylesheet" crossorigin href="/assets/index-BNGqA0_L.css">
+</script><style>html{--lightningcss-light:initial;--lightningcss-dark: ;color-scheme:light;background:#fff}html[data-theme=dark]{--lightningcss-light: ;--lightningcss-dark:initial;color-scheme:dark;background:#26272b}</style>  <script type="module" crossorigin src="/assets/index-D_VX51no.js"></script>
+  <link rel="stylesheet" crossorigin href="/assets/index-Cc-awdNn.css">
 </head><body><div id="root"></div></body></html>

package/examples/papilot-paper/main.tex

@@ -12,7 +12,7 @@
 Papilot is a personal workspace for revising \LaTeX{} papers with an AI agent
 and reviewing every resulting change as a PDF difference before it reaches the
 manuscript. Each editing task runs in an isolated Git worktree, may be driven
-by Codex, Claude or Antigravity, and proceeds in review rounds in which the
+by Codex, Claude, Antigravity or Kimi Code, and proceeds in review rounds in which the
 author accepts, rejects or comments on individual changes. This paper describes
 the system through a sequence of small formal models. We define edits and
 review rounds as operators on a document, prove that compatible edits commute

package/examples/papilot-paper/sections/02-using.tex

@@ -91,7 +91,8 @@ remote (\cref{sec:git}).
 \subsection{Agents}
 
 Each task runs on one agent: Codex through its app server, Claude through the
-Claude Agent SDK, or Antigravity through its command line. Every agent uses its
+Claude Agent SDK, Antigravity through its command line, or Kimi Code through
+ACP. For Kimi, install the CLI and run \texttt{kimi login}. Every agent uses its
 own login on the host, and an agent that is not installed or signed in is listed
 as unavailable. When an agent reports usage, the new-task page shows each usage
 window with how much is used and when it resets (\cref{sec:usage}).
@@ -124,6 +125,7 @@ Variable & Purpose \\
 \texttt{PAPILOT\_DATA} & state and generated example, \texttt{\textasciitilde/.papilot} by default \\
 \texttt{PAPILOT\_CODEX\_BIN} & Codex executable \\
 \texttt{ANTHROPIC\_API\_KEY} & API key for Claude tasks \\
+\texttt{PAPILOT\_KIMI\_BIN} & Kimi Code executable \\
 \texttt{PAPILOT\_AGY\_BIN} & Antigravity command line \\
 \texttt{PAPILOT\_AGY\_PROXY} & HTTP proxy used only for Antigravity \\
 \texttt{PAPILOT\_REVIEW\_ENGINE} & review engine directory \\

package/examples/papilot-paper/sections/09-agents.tex

@@ -1,7 +1,7 @@
 \section{Agents and scheduling}
 \label{sec:agents}
 
-\subsection{One protocol for three agents}
+\subsection{One protocol for every agent}
 
 Every agent is driven through the same small protocol: a thread is started or
 resumed, a turn is started with the author's text, the agent streams items
@@ -40,8 +40,9 @@ task in a recoverable state with a \emph{Check status} action (not shown).}
 \end{figure}
 
 Codex speaks this protocol natively. Claude runs through the Claude Agent SDK,
-and Antigravity runs the \texttt{agy} command line once per turn; both are
-translated by small adapters into the same requests and notifications. Each
+Antigravity runs the \texttt{agy} command line once per turn, and Kimi Code
+runs \texttt{kimi acp}; adapters translate their interfaces into the same
+requests and notifications. Kimi sessions resume after a restart. Each
 agent uses its own login on the host, and Papilot never reads, stores or
 forwards a credential. An agent that cannot ask for permission, such as a
 headless command line, is told in advance which tools it may not use.
@@ -102,5 +103,7 @@ answers after a mean delay $\Theta$, the expected wall-clock time of the turn is
   \E[T] = T_w \bigl(1 + \zeta \Theta\bigr),
   \label{eq:approvals}
 \end{equation}
-which is why Papilot lets agents edit and run sandboxed commands in the task's
-worktree without asking, and asks only for what leaves it.
+which motivates reducing unnecessary approval requests. Codex and Claude can
+run sandboxed commands under their permission policies; Kimi uses manual
+approvals for tools on the host, without a Papilot sandbox. Antigravity edits
+files but cannot run shell commands in Papilot.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "papilot",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "A Cursor-like workspace for your papers: annotate and review PDFs directly without ever touching LaTeX.",
   "keywords": [
     "latex",
@@ -12,17 +12,19 @@
     "codex",
     "claude",
     "antigravity",
+    "kimi",
+    "kimi-code",
     "writing",
     "academic",
     "manuscript"
   ],
-  "homepage": "https://github.com/cu1ch3n/papilot#readme",
+  "homepage": "https://papilot.ai",
   "repository": {
     "type": "git",
-    "url": "git+https://github.com/cu1ch3n/papilot.git"
+    "url": "git+https://github.com/PapilotAI/papilot.git"
   },
   "bugs": {
-    "url": "https://github.com/cu1ch3n/papilot/issues"
+    "url": "https://github.com/PapilotAI/papilot/issues"
   },
   "license": "MIT",
   "author": "Cui Chen",
@@ -34,9 +36,13 @@
     "bin",
     "src/server",
     "src/shared.ts",
+    "src/skill-icons.ts",
+    "src/skill-metadata.ts",
     "src/notes.ts",
+    "src/note-groups.ts",
     "src/conversation.ts",
     "src/commit.ts",
+    "src/task-caption.ts",
     "skills",
     "dist",
     "scripts/demo.ts",
@@ -62,9 +68,9 @@
     "doctor": "node bin/papilot.mjs doctor",
     "service:install": "tsx scripts/service.ts install",
     "service:stop": "tsx scripts/service.ts stop",
-    "test:ui": "npm run build && node scripts/check-ui.mjs && node scripts/check-review-notes.mjs && node --import tsx scripts/check-review-flow.mjs && node scripts/check-skills.mjs && node --import tsx scripts/check-accounts.mjs",
+    "test:ui": "npm run build && node scripts/check-ui.mjs && node scripts/check-appearance.mjs && node scripts/check-updates.mjs && node scripts/check-papers.mjs && node scripts/check-imports.mjs && node scripts/check-compilation.mjs && node scripts/check-build-problems.mjs && node scripts/check-review-notes.mjs && node --import tsx scripts/check-note-groups.mjs && node scripts/check-note-selection.mjs && node scripts/check-ai-review.mjs && node --import tsx scripts/check-note-results.mjs && node --import tsx scripts/check-review-flow.mjs && node scripts/check-skills.mjs && node --import tsx scripts/check-skill-library.mjs && node scripts/check-attachments.mjs && node --import tsx scripts/check-accounts.mjs",
     "site": "node scripts/build-site.mjs",
-    "site:video": "node scripts/build-site.mjs --video",
+    "test:site": "npm run site && node scripts/check-site.mjs",
     "prepack": "npm run build && npm run setup:review"
   },
   "dependencies": {
@@ -76,7 +82,8 @@
     "fastify": "^5.12.7",
     "mupdf": "^1.28.1",
     "proper-lockfile": "^4.1.2",
-    "tsx": "^4.23.15"
+    "tsx": "^4.23.15",
+    "yauzl": "^3.4.0"
   },
   "devDependencies": {
     "@anthropic-ai/claude-agent-sdk": "^0.3.311",
@@ -85,6 +92,7 @@
     "@types/node": "^22.20.5",
     "@types/react": "^19.3.0",
     "@types/react-dom": "^19.3.0",
+    "@types/yauzl": "^3.4.0",
     "@vitejs/plugin-react": "^6.1.2",
     "katex": "^0.16.47",
     "pdfjs-dist": "^6.4.299",

package/scripts/service.ts

@@ -22,6 +22,7 @@ if (process.argv[2] === 'stop') {
     } catch { /* An older or incomplete plist may not contain environment settings. */ }
   }
   const env = serviceEnvironment(ROOT, os.homedir(), process.execPath, process.env, previous);
+  env.PAPILOT_SERVICE_MANAGER = 'launchd';
   const args = ['/usr/bin/caffeinate', '-i', process.execPath, '--import', 'tsx', path.join(ROOT, 'src/server/main.ts')];
   const plist = `<?xml version="1.0" encoding="UTF-8"?><!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"><plist version="1.0"><dict>
 <key>Label</key><string>${label}</string><key>ProgramArguments</key><array>${args.map(a => `<string>${xml(a)}</string>`).join('')}</array>

package/skills/ai-review.md

@@ -0,0 +1,18 @@
+---
+name: AI review
+description: Read the paper and leave PDF comments without editing the manuscript
+hint: What should the reviewer focus on? Leave empty for the whole paper
+order: 8
+output: annotations
+---
+Review the paper as a constructive academic reviewer. Read it and leave suggestions; do not edit, rewrite, commit, or apply changes to the manuscript.
+
+- Follow the author's requested focus, or review the whole paper when none is given.
+- Look for unsupported claims, gaps in reasoning, unclear definitions, inconsistent terminology, and concrete writing issues.
+- Anchor each comment to the smallest relevant passage in the supplied PDF: the specific sentence, claim, or equation at issue, usually 1–3 lines. Avoid whole-paragraph or page-level highlights. If the issue concerns a symbol within a line, name that symbol in the comment.
+- Write one issue per comment, usually in one short sentence: aim for 10–30 English words or 20–60 Chinese characters. State the problem and a concrete next step; add a second short sentence only when needed for accuracy. Example: “This bound is undefined at p = 1; state that case separately.”
+- Omit praise, preambles, repeated quotations, background explanations, full derivations, and rewritten paragraphs. For a complex issue, identify the failing step and needed clarification; leave extended reasoning for follow-up.
+- Prefer a few high-impact, actionable comments to minor or speculative criticism. Flag a repeated issue once at a representative location. Do not invent sources, results, or assumptions, or claim a citation is wrong without evidence.
+- Use the author's requested language, otherwise the manuscript's language.
+- The supplied PDF extraction contains text, equations, and captions, not figure images. Do not claim to have inspected page appearance.
+- Return PDF annotations in the format Papilot supplies, with a single brief sentence summarizing the main priority. Do not repeat the findings in the summary or add a review essay. The author decides which suggestions to use in a separate editing task.

package/skills/latex-warnings.md

@@ -6,7 +6,7 @@ order: 6
 ---
 Fix the problems LaTeX reports when it builds the paper.
 
-- Compile in your worktree (for example with `latexmk`) and read the log. If you cannot run commands, look for the same problems in the source instead.
+- When the request lists problems from the last compilation, with their files and lines, fix those. Otherwise compile in your worktree (for example with `latexmk`) and read the log. If you cannot run commands, look for the same problems in the source instead.
 - Fix undefined or duplicate labels, references and citations; overfull boxes, by rewording slightly or allowing a break where the line overflows; and package warnings that point to real mistakes.
 - Do not silence warnings globally: no document-wide `\sloppy`, badness or tolerance changes, and no new packages, unless the request asks for them.
 - In your summary, list the warnings you fixed and any you left, with the reason.

package/src/conversation.ts

@@ -6,6 +6,12 @@ export const reviewInstructions = 'Continue this task using the decisions and co
 // Earlier versions stored the prompt itself as the author's message.
 export const isFeedbackPrompt = (text: string) => text === reviewInstructions || text === 'Address the comments from this review round. Preserve accepted changes and do not reintroduce rejected edits.';
 
+// Older AI reviews appended a generated copy of every PDF suggestion to the
+// chat summary. Keep those records intact, but display only their summary.
+export function aiReviewSummary(text: string) {
+  return text.split(/\n\n1\. \*\*Page \d+\*\* — /, 1)[0].trimEnd();
+}
+
 // What the agent receives when its own edits stop the review build. The chat
 // shows the error, never this prompt.
 export function buildFixPrompt(failure: BuildFailure) {

package/src/note-groups.ts

@@ -0,0 +1,30 @@
+export const NOTE_GROUPS = ['yellow', 'green', 'purple', 'orange', 'red'] as const;
+// Older saved annotations and review drafts keep their identity; blue is now
+// displayed and described as yellow, and is no longer offered in the palette.
+export type NoteGroup = typeof NOTE_GROUPS[number] | 'blue';
+export type GroupInstructions = Partial<Record<typeof NOTE_GROUPS[number], string>>;
+export const canonicalGroup = (group: NoteGroup): typeof NOTE_GROUPS[number] => group === 'blue' ? 'yellow' : group;
+export const isNoteGroup = (value: unknown): value is NoteGroup => value === 'blue' || NOTE_GROUPS.includes(value as typeof NOTE_GROUPS[number]);
+export function parseGroupInstructions(value: unknown): GroupInstructions {
+  if (value === undefined) return {};
+  if (!value || typeof value !== 'object' || Array.isArray(value)) throw new Error('Use valid instructions for each color group.');
+  const result: GroupInstructions = {};
+  for (const [key, text] of Object.entries(value)) {
+    if (!isNoteGroup(key) || typeof text !== 'string' || text.length > 4000) throw new Error('Use valid instructions for each color group (up to 4000 characters).');
+    const group = canonicalGroup(key);
+    if (result[group] && result[group] !== text.trim()) throw new Error('Use one instruction for each color group.');
+    if (text.trim()) result[group] = text.trim();
+  }
+  return result;
+}
+export const groupsWithoutInstructions = (notes: { group?: NoteGroup; comment: string }[], instructions: GroupInstructions) =>
+  NOTE_GROUPS.filter(group => !instructions[group]?.trim() && notes.some(note => note.group && canonicalGroup(note.group) === group && !note.comment.trim()));
+export const GROUP_COLORS: Record<NoteGroup, string> = { yellow: '#eab308', blue: '#eab308', green: '#16a34a', purple: '#a855f7', orange: '#ea860c', red: '#e05260' };
+export const GROUP_NAMES: Record<NoteGroup, string> = { yellow: 'Yellow', blue: 'Yellow', green: 'Green', purple: 'Purple', orange: 'Orange', red: 'Red' };
+export function noteGroups<T extends { group?: NoteGroup }>(notes: T[]) {
+  return NOTE_GROUPS.flatMap(group => {
+    const entries = notes.flatMap((note, index) => note.group && canonicalGroup(note.group) === group ? [{ note, index }] : []);
+    return entries.length ? [{ group, entries }] : [];
+  });
+}
+export const GROUP_INSTRUCTIONS = 'Color groups are labels only, with no built-in meaning. Apply the author\'s instructions for this message to every marked location in each named color group. Never infer an editing action from a color. If a group\'s intended treatment is unclear, ask the author.';

package/src/notes.ts

@@ -1,24 +1,45 @@
-import type { Annotation } from './shared.ts';
+import { canonicalGroup, GROUP_INSTRUCTIONS, GROUP_NAMES, NOTE_GROUPS, type GroupInstructions } from './note-groups.ts';
+import type { Annotation, NoteAnchor, NoteSummary } from './shared.ts';
+export type { NoteSummary } from './shared.ts';
 
 // PDF notes reach the agent as one plain-text message. The chat shows that
 // message as the notes again, so the format and its reader live together.
 const PREAMBLE = 'Revise the manuscript according to these notes the author marked on the compiled PDF.';
 const QUOTE = '   PDF text: "', REGION = '   Marked region: a figure, equation or other non-text area', NOTE = '\n   Note: ';
 
-export interface NoteSummary { where: string; quote: string; comment: string }
+const noteWhere = (note: Annotation) => [`Page ${note.page + 1}`, note.source ? `${note.source.file} ${note.source.lines[0] === note.source.lines[1] ? `line ${note.source.lines[0]}` : `lines ${note.source.lines[0]}–${note.source.lines[1]}`}` : '', note.author ? `from ${note.author.name}` : ''].filter(Boolean).join(' · ');
+export const summarizeNotes = (notes: Annotation[]): NoteSummary[] => notes.map(note => ({ ...(note.group ? { group: note.group } : {}), id: note.id, where: noteWhere(note), quote: note.quote, comment: note.comment, anchor: { sources: note.source ? [{ file: note.source.file, lines: note.source.lines }] : [] } }));
 
-export function reviewNotesInstructions(notes: NoteSummary[]) {
-  const items = notes.map((note, index) => `${index + 1}. ${note.where}\n${note.quote ? `${QUOTE}${note.quote}"` : REGION}${NOTE}${note.comment.trim().replace(/\n/g, '\n   ')}`);
-  return `${PREAMBLE} These notes refer to the displayed review round and its Before, After or Unchanged text. Region coordinates are percentages of the displayed PDF strip (left, top, right, bottom); any quoted text for a region is nearby context. Read the current source before editing. Preserve accepted changes and do not reintroduce rejected edits unless a note explicitly asks for it.\n\n${items.join('\n\n')}`;
+// Also recover anchors from older messages, whose locations were only text.
+export function noteAnchor(note: NoteSummary): NoteAnchor {
+  if (note.anchor) return note.anchor;
+  const round = note.where.match(/^Round (\d+)\b/), sources: NoteAnchor['sources'] = [];
+  for (const place of note.where.split(/ · |; /)) {
+    const match = place.match(/^(.+?)(?::| lines? )(\d+(?:[–-]\d+)?(?:,\s*\d+(?:[–-]\d+)?)*)$/);
+    if (!match) continue;
+    for (const range of match[2].split(/,\s*/)) {
+      const [start, end] = range.split(/[–-]/).map(Number);
+      sources.push({ file: match[1], lines: [start, end ?? start] });
+    }
+  }
+  return { sources, ...(round ? { round: Number(round[1]) } : {}) };
+}
+
+export function reviewNotesInstructions(notes: NoteSummary[], request = '') {
+  const items = notes.map((note, index) => `${index + 1}. ${note.where}${note.group ? ` · Color group: ${GROUP_NAMES[note.group]}` : ''}\n${note.quote ? `${QUOTE}${note.quote}"` : REGION}${NOTE}${note.comment.trim().replace(/\n/g, '\n   ')}`);
+  return `${PREAMBLE}${notes.some(note => note.group) ? ` ${GROUP_INSTRUCTIONS}` : ''}${request.trim() ? ` Overall request: ${request.trim()}\n\n` : ''} These notes refer to the displayed review round and its Before, After or Unchanged text. Region coordinates are percentages of the displayed PDF strip (left, top, right, bottom); any quoted text for a region is nearby context. Read the current source before editing. Preserve accepted changes and do not reintroduce rejected edits unless a note explicitly asks for it.\n\n${items.join('\n\n')}`;
 }
 
-export function notesInstructions(notes: Annotation[]) {
+export function notesInstructions(notes: Annotation[], request = '', groupInstructions: GroupInstructions = {}) {
   const items = notes.map((note, index) => {
-    const where = [`Page ${note.page + 1}`, note.source ? `${note.source.file} ${note.source.lines[0] === note.source.lines[1] ? `line ${note.source.lines[0]}` : `lines ${note.source.lines[0]}–${note.source.lines[1]}`}` : '', note.author ? `from ${note.author.name}` : ''].filter(Boolean).join(' · ');
+    const where = noteWhere(note);
     const source = note.source?.excerpt ? ['   LaTeX source of the marked region:', '   ```latex', ...note.source.excerpt.split('\n').map(line => '   ' + line), '   ```'] : [];
-    return [`${index + 1}. ${where}`, ...source, note.quote ? `${QUOTE}${note.quote}"` : REGION, `${NOTE.slice(1)}${note.comment.trim().replace(/\n/g, '\n   ')}`].join('\n');
+    return [`${index + 1}. ${where}${note.group ? ` · Color group: ${GROUP_NAMES[note.group]}` : ''}`, ...source, note.quote ? `${QUOTE}${note.quote}"` : REGION, `${NOTE.slice(1)}${note.comment.trim().replace(/\n/g, '\n   ')}`].join('\n');
   });
-  return `${PREAMBLE} SyncTeX mapped each marked region to the LaTeX file and lines that produced it; the PDF text is what the author saw. Read the current source before editing, and change only what the notes ask for.\n\n${items.join('\n\n')}`;
+  const scope = request.trim() ? `follow both the overall request and the notes.\n\nOverall request:\n${request.trim()}` : 'change only what the notes ask for.';
+  const groups = NOTE_GROUPS.filter(group => groupInstructions[group]?.trim() && notes.some(note => note.group && canonicalGroup(note.group) === group));
+  const directions = groups.length ? `\n\nInstructions for color groups (apply only to the attached locations in that group):\n${groups.map(group => `${GROUP_NAMES[group]}:\n   ${groupInstructions[group]!.trim().replace(/\n/g, '\n   ')}`).join('\n')}` : '';
+  return `${PREAMBLE}${notes.some(note => note.group) ? ` ${GROUP_INSTRUCTIONS}` : ''} SyncTeX mapped each marked region to the LaTeX file and lines that produced it; the PDF text is what the author saw. Read the current source before editing, and ${scope}${directions}\n\n${items.join('\n\n')}`;
 }
 
 // The notes in a message written by notesInstructions, or null for any other message.
@@ -39,7 +60,10 @@ export function readNotesInstructions(text: string): NoteSummary[] | null {
     if (note < 0) return null;
     const marked = rest.slice(0, note);
     if (marked !== REGION && !(marked.startsWith(QUOTE) && marked.endsWith('"'))) return null;
-    notes.push({ where: heading[1], quote: marked === REGION ? '' : marked.slice(QUOTE.length, -1), comment: rest.slice(note + NOTE.length).replace(/\n {3}/g, '\n') });
+    const legacy = heading[1].endsWith(' · Color group: Blue');
+    const group = legacy ? 'blue' : NOTE_GROUPS.find(group => heading[1].endsWith(` · Color group: ${GROUP_NAMES[group]}`));
+    const where = group ? heading[1].slice(0, -` · Color group: ${legacy ? 'Blue' : GROUP_NAMES[group]}`.length) : heading[1];
+    notes.push({ ...(group ? { group } : {}), where, quote: marked === REGION ? '' : marked.slice(QUOTE.length, -1), comment: rest.slice(note + NOTE.length).replace(/\n {3}/g, '\n') });
   }
   return notes.length ? notes : null;
 }