Files
openmaic/OpenMAIC/lib/agent/tools/read-scene-content.ts
2026-08-16 14:58:47 +08:00

414 lines
17 KiB
TypeScript

/**
* `read_scene_content` agent tool
*
* Read-only. Surfaces the current scene's outline + content to the model so it
* can reason about the slide, answer questions about it, and distil a precise
* instruction for `regenerate_scene`. This is the "read" half of read-then-act:
* the model SEES the slide here (instead of regenerating blind), while the
* trusted content used for execution still flows from the injected context.
*
* The content is pulled from the same client-injected `SceneContext`
* (`getSceneContext`) the other tools use — no new data plumbing, and strictly
* more token-efficient than pre-stuffing every scene into the system prompt.
*/
import { Type, type Static } from 'typebox';
import type { AgentTool } from '@earendil-works/pi-agent-core';
import type { PPTElement } from '@openmaic/dsl';
import type { SceneContext } from './regenerate-scene-actions';
import { ALLOWED_EDIT_PROPS } from './edit-elements-gate';
import { EXACT_CONTENT_RAW_CAP, isExactContentEditable } from './edit-elements-content-contract';
// ── Deps ─────────────────────────────────────────────────────────────────────
export interface ReadSceneContentDeps {
/** Returns the trusted scene/stage context for a scene id (client-sourced). */
getSceneContext: (sceneId: string) => SceneContext | undefined;
/** The active scene id, used when the model omits sceneId. */
activeSceneId?: string;
/** Active canvas selection ids, used to resolve "this" and "these". */
getSelection?: () => readonly string[];
}
// ── Content projection (model-visible text) ──────────────────────────────────
// The model only sees `content[].text`, NOT `details`. Serialize a compact,
// human-readable projection of the actual content so it can reason about what
// is on the slide.
const PROJECTION_CAP = 2000;
const ELEMENT_TEXT_CAP = 80;
const SLIDE_PATCH_PROJECTION_CAP = 30000;
const MANIFEST_PAGE_CAP = 10000;
function stripHtml(html: string): string {
return html
.replace(/<[^>]*>/g, ' ')
.replace(/\s+/g, ' ')
.trim();
}
function truncate(s: string, max: number): string {
return s.length > max ? `${s.slice(0, max)}…` : s;
}
/**
* Pull the human-readable text out of a single slide element. Mirrors the
* @openmaic/dsl `PPTElement` union — each variant stashes its text differently:
* - text: `content` (HTML string)
* - shape: `text.content` (ShapeText, HTML string)
* - table: `data` (TableCell[][], each cell carries `text`)
* - code: `lines` (CodeLine[], each carries `content`)
* - latex: `latex` (source string)
* Returns '' for elements with no meaningful text (image / line / chart / …).
*/
function extractElementText(el: unknown): string {
const e = el as {
type?: string;
content?: unknown;
text?: { content?: unknown };
data?: unknown;
lines?: unknown;
latex?: unknown;
};
switch (e.type) {
case 'text':
return typeof e.content === 'string' ? stripHtml(e.content) : '';
case 'shape':
return typeof e.text?.content === 'string' ? stripHtml(e.text.content) : '';
case 'table': {
const rows = Array.isArray(e.data) ? (e.data as unknown[]) : [];
return rows
.flatMap((row) => (Array.isArray(row) ? (row as unknown[]) : []))
.map((cell) => {
const c = cell as { text?: unknown };
return typeof c.text === 'string' ? stripHtml(c.text) : '';
})
.filter(Boolean)
.join(' | ');
}
case 'code': {
const lines = Array.isArray(e.lines) ? (e.lines as unknown[]) : [];
return lines
.map((line) => {
const l = line as { content?: unknown };
return typeof l.content === 'string' ? l.content : '';
})
.join(' ')
.trim();
}
case 'latex':
return typeof e.latex === 'string' ? e.latex.trim() : '';
default:
return '';
}
}
// Interactive pages are edited with `edit_interactive_html`, which needs the
// model to author EXACT `oldText` anchors — so it must see the full, raw,
// un-escaped HTML (not a truncated JSON projection). Cap only as a pathological
// safety net; real pages (after eliding base64) are well under this.
const INTERACTIVE_HTML_CAP = 120000;
// Generated pages often inline vendor libs / media as huge base64 data-URIs
// (a single page can be ~760KB, ~95% base64). That payload is noise the model
// can't usefully edit and it pushes the actual code past any cap, so the model
// never sees the real JS/markup it needs to fix. Elide the base64 body to a tiny
// marker for the MODEL'S VIEW only — the edit tool still applies against the full
// stored HTML, and the model authors oldText from the real (non-base64) code.
function elideDataUris(html: string): string {
return html.replace(
/(data:[^;,]*;base64,)([A-Za-z0-9+/=]+)/g,
(_m, prefix: string, payload: string) => `${prefix}…[${payload.length} base64 chars elided]`,
);
}
function projectElementManifest(element: PPTElement, index: number): Record<string, unknown> {
const manifest: Record<string, unknown> = {
_index: index,
_path: `/elements/${index}`,
id: element.id,
type: element.type,
};
if (element.name !== undefined) manifest.name = element.name;
if (element.type === 'text' && element.textType !== undefined) {
manifest.textType = element.textType;
} else if (element.type === 'image' && element.imageType !== undefined) {
manifest.imageType = element.imageType;
} else if (element.type === 'chart') {
manifest.chartType = element.chartType;
}
if (element.lock !== undefined) manifest.lock = element.lock;
if (element.groupId !== undefined) manifest.groupId = element.groupId;
const summary = truncate(extractElementText(element), ELEMENT_TEXT_CAP);
if (summary) manifest._textSummary = summary;
return manifest;
}
function projectElementForPatch(element: PPTElement, index: number): Record<string, unknown> {
const source = element as unknown as Record<string, unknown>;
const projected: Record<string, unknown> = {
_index: index,
_path: `/elements/${index}`,
id: element.id,
type: element.type,
};
if (element.name !== undefined) projected.name = element.name;
for (const key of ['lock', 'groupId', 'left', 'top', 'width', 'height', 'rotate']) {
if (source[key] !== undefined) projected[key] = source[key];
}
for (const key of ALLOWED_EDIT_PROPS) {
if (source[key] !== undefined) projected[key] = source[key];
}
if (element.type === 'text') {
if (isExactContentEditable(element.content)) {
projected.content = element.content;
} else {
projected._contentEdit = `content editing unavailable: HTML exceeds the exact ${EXACT_CONTENT_RAW_CAP}-character/serialized-record limit`;
}
} else if (element.type === 'shape' && element.text) {
projected.text = {
...(isExactContentEditable(element.text.content)
? { content: element.text.content }
: {
_contentEdit: `content editing unavailable: HTML exceeds the exact ${EXACT_CONTENT_RAW_CAP}-character/serialized-record limit`,
}),
defaultFontName: element.text.defaultFontName,
defaultColor: element.text.defaultColor,
lineHeight: element.text.lineHeight,
wordSpace: element.text.wordSpace,
paragraphSpace: element.text.paragraphSpace,
align: element.text.align,
};
} else {
const summary = truncate(extractElementText(element), ELEMENT_TEXT_CAP);
if (summary) projected._textSummary = summary;
}
return projected;
}
function projectContent(
content: SceneContext['content'],
selectionIds: readonly string[] = [],
requestedElementIds: readonly string[] = [],
manifestOffset = 0,
): string {
const c = content as
| { type?: string; canvas?: { elements?: unknown[] }; html?: unknown }
| undefined;
// Interactive: return the full page HTML verbatim so edits can anchor exactly.
if (c?.type === 'interactive') {
const raw = typeof c.html === 'string' ? c.html : '';
if (!raw) return '(this interactive scene has no embedded HTML)';
// Elide base64 payloads first so the real code is visible within the cap.
const html = elideDataUris(raw);
const capped =
html.length > INTERACTIVE_HTML_CAP
? `${html.slice(0, INTERACTIVE_HTML_CAP)}…(truncated)`
: html;
return `Interactive page HTML (to fix a bug, call edit_interactive_html with exact oldText snippets copied verbatim from below; base64 payloads are shown elided — never use them as oldText):\n${capped}`;
}
let projection: string;
if (c?.type === 'slide') {
const elements = Array.isArray(c.canvas?.elements) ? (c.canvas!.elements as PPTElement[]) : [];
const indexById = new Map<string, number>();
const duplicateIds = new Set<string>();
elements.forEach((element, index) => {
if (indexById.has(element.id)) duplicateIds.add(element.id);
else indexById.set(element.id, index);
});
const selected = selectionIds.filter((id) => indexById.has(id));
const requested = [...new Set(requestedElementIds)].filter((id) => indexById.has(id));
const allManifestEntries = elements.map((element, index) =>
projectElementManifest(element, index),
);
const manifestEntries: Record<string, unknown>[] = [];
let nextManifestOffset: number | null = null;
if (requested.length > 0) {
for (const id of requested) manifestEntries.push(allManifestEntries[indexById.get(id)!]);
} else {
let manifestSize = 2;
for (let index = Math.max(0, manifestOffset); index < allManifestEntries.length; index++) {
const entry = allManifestEntries[index];
const serialized = JSON.stringify(entry);
if (manifestEntries.length > 0 && manifestSize + serialized.length > MANIFEST_PAGE_CAP) {
nextManifestOffset = index;
break;
}
manifestEntries.push(entry);
manifestSize += serialized.length + 1;
}
}
const manifest = JSON.stringify(manifestEntries, null, 2);
const allDetails = elements.map((element, index) => projectElementForPatch(element, index));
const allDetailsJson = JSON.stringify(allDetails, null, 2);
const detailIds = requested.length > 0 ? requested : selected;
const candidateDetails =
allDetailsJson.length <= SLIDE_PATCH_PROJECTION_CAP && requested.length === 0
? allDetails
: detailIds.length > 0
? detailIds.map((id) => {
const index = indexById.get(id)!;
return projectElementForPatch(elements[index], index);
})
: [];
const details: Record<string, unknown>[] = [];
const omittedDetailIds: string[] = [];
let detailSize = 2;
const detailBudget =
requested.length === 1 ? Number.POSITIVE_INFINITY : SLIDE_PATCH_PROJECTION_CAP;
for (const detail of candidateDetails) {
const serialized = JSON.stringify(detail, null, 2);
if (detailSize + serialized.length <= detailBudget) {
details.push(detail);
detailSize += serialized.length + 2;
} else if (typeof detail.id === 'string') {
omittedDetailIds.push(detail.id);
}
}
const detailsText =
details.length > 0
? JSON.stringify(details, null, 2)
: '(omitted because the exact records are large; call read_scene_content again with elementIds for the elements you intend to patch)';
projection =
'Slide element manifest page. Identity fields (`id`, `type`, `name`) and underscore-prefixed projection metadata are read-only; patch editable real paths under `/elements` and guard every index with an id test.\n' +
manifest +
(nextManifestOffset !== null
? `\nManifest continues. Next manifestOffset: ${nextManifestOffset}.`
: '') +
'\nExact editable element records:\n' +
detailsText +
(omittedDetailIds.length > 0
? `\nExact records omitted to stay within the read budget: ${omittedDetailIds.join(', ')}. Request fewer elementIds.`
: '') +
(duplicateIds.size > 0
? `\nWarning: duplicate element ids make JSON Patch editing unavailable: ${[...duplicateIds].join(', ')}.`
: '') +
`\nSelected element ids on this slide: ${selected.length > 0 ? selected.join(', ') : '(none)'}`;
return projection;
} else {
projection = JSON.stringify(content ?? {});
}
return projection.length > PROJECTION_CAP
? `${projection.slice(0, PROJECTION_CAP)}…(truncated)`
: projection;
}
// ── Params ───────────────────────────────────────────────────────────────────
// The model only needs to say WHICH scene to read; defaults to the active one.
export const ReadSceneContentParams = Type.Object({
sceneId: Type.Optional(
Type.String({
description:
'The id of the scene to read. Defaults to the current scene shown in the system prompt.',
}),
),
elementIds: Type.Optional(
Type.Array(Type.String(), {
maxItems: 50,
uniqueItems: true,
description:
'Existing slide element ids whose exact editable records should be returned. Use this after reading a large slide manifest.',
}),
),
manifestOffset: Type.Optional(
Type.Integer({
minimum: 0,
description: 'Offset returned by a previous manifest page for very large slides.',
}),
),
});
export type ReadSceneContentParams = Static<typeof ReadSceneContentParams>;
// ── Details returned to the client ───────────────────────────────────────────
// Compact metadata only. The model gets the full content via `content[].text`;
// the client apply path never reads this tool's `details.content`, so echoing
// the raw scene content here is dead weight (and megabytes for base64 images).
export interface ReadSceneContentDetails {
sceneId: string;
title: string;
type: string;
/** Short text-only outline (title / description / keyPoints) — safe to keep. */
outline: SceneContext['outline'];
}
// ── Factory ──────────────────────────────────────────────────────────────────
export function makeReadSceneContentTool(
deps: ReadSceneContentDeps,
): AgentTool<typeof ReadSceneContentParams, ReadSceneContentDetails> {
return {
name: 'read_scene_content',
label: 'Read scene content',
description:
'Reads the current scene to understand what is on it — its outline (title, ' +
'description, key points) and its content (slide elements / quiz questions / etc). ' +
'Use this BEFORE answering questions about the slide or regenerating it, so your ' +
'reply and any regeneration instruction reflect what is actually on the slide. ' +
'For large slides, follow manifestOffset pages, then pass elementIds from the manifest ' +
'to retrieve exact target records.',
parameters: ReadSceneContentParams,
execute: async (_toolCallId, params) => {
const sceneId = params.sceneId || deps.activeSceneId || '';
const ctx = deps.getSceneContext(sceneId);
if (!ctx) {
return {
content: [
{
type: 'text',
text: `Error: scene context not found for sceneId ${JSON.stringify(String(sceneId).slice(0, 200))}. Cannot read the scene.`,
},
],
details: {
sceneId,
title: '',
type: '',
outline: undefined as unknown as SceneContext['outline'],
},
isError: true,
};
}
const { outline, content, runtimeErrors } = ctx;
const keyPoints = (outline.keyPoints ?? []).join('; ');
const projection = projectContent(
content,
deps.getSelection?.() ?? [],
params.elementIds ?? [],
params.manifestOffset ?? 0,
);
// Surface any runtime errors the page threw when it rendered — for an
// interactive scene these are usually the real reason it's blank/broken, so
// the agent fixes the root cause instead of guessing from the static HTML.
const errorsSection =
runtimeErrors && runtimeErrors.length > 0
? `\n\nRuntime errors this page reported when it rendered (these are the likely reason it is blank/broken — fix the ROOT CAUSE shown here, do not guess):\n` +
runtimeErrors.map((e) => `- ${e}`).join('\n')
: '';
return {
content: [
{
type: 'text',
text:
`Scene "${outline.title}" (type: ${outline.type}). ` +
`Description: ${outline.description || '(none)'}. ` +
`Key points: ${keyPoints || '(none)'}.\n` +
`Scene content:\n${projection}${errorsSection}`,
},
],
details: {
sceneId,
title: outline.title,
type: outline.type,
outline,
},
};
},
};
}