@blockrun/franklin 3.49.16 → 3.49.18

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1344,7 +1344,7 @@ export async function handleSlashCommand(input, ctx) {
1344
1344
  }
1345
1345
  // Argument-based rewrite commands (prefix match)
1346
1346
  for (const { prefix, rewrite } of ARG_COMMANDS) {
1347
- if (input.startsWith(prefix)) {
1347
+ if (input.startsWith(prefix) && (prefix.endsWith(' ') || input === prefix)) {
1348
1348
  const arg = input.slice(prefix.length).trim();
1349
1349
  return { handled: false, rewritten: rewrite(arg) };
1350
1350
  }
@@ -54,6 +54,7 @@ export declare function evaluatorModel(): string;
54
54
  export declare function checkGrounding(userInput: string, history: Dialogue[], assistantText: string, client: ModelClient, opts?: {
55
55
  abortSignal?: AbortSignal;
56
56
  model?: string;
57
+ userImageCount?: number;
57
58
  }): Promise<GroundingResult>;
58
59
  /**
59
60
  * Convert a grounding result into a user-facing follow-up message. Returns
@@ -39,7 +39,10 @@ const EVALUATOR_PROMPT = `You are a GROUNDING CHECK agent. Your job is to verify
39
39
  ### A. Ungrounded claims
40
40
  Every **factual claim** in the answer must trace to ONE of:
41
41
  (a) A tool call result from this turn (model-initiated OR listed under "Pre-fetched by Franklin harness"), OR
42
- (b) Explicit acknowledgment of uncertainty ("I'm not sure", "based on older data")
42
+ (b) User-provided material, including observations of an attached image when the summary confirms an image was supplied, OR
43
+ (c) Explicit acknowledgment of uncertainty ("I'm not sure", "based on older data")
44
+
45
+ **User-provided images are evidence for visual observations.** The main assistant can see attached images directly; a Read or web tool is not required to transcribe visible text or describe shapes. When the summary confirms user-provided images, do not flag those observations merely because no tool ran. This text-only check cannot verify their visual accuracy: use PARTIAL if that is the only uncertainty, not a demand for a tool call. Images do NOT establish fresh external facts: claims about live prices, latest news, or facts beyond the supplied image still require appropriate evidence.
43
46
 
44
47
  **Harness-prefetched data is evidence.** When the turn includes a "Pre-fetched by Franklin harness" section, the data listed there was fetched live from tools on the assistant's behalf (TradingMarket, ExaAnswer, etc). Treat it identically to a model-initiated tool call — claims that reference prefetched prices, numbers, or news snippets are GROUNDED.
45
48
 
@@ -141,11 +144,16 @@ export function shouldCheckGrounding(userInput, assistantText) {
141
144
  * + tool result snippets + assistant's final answer. Bounded to keep the
142
145
  * evaluator call cheap; it doesn't need every byte of every tool output.
143
146
  */
144
- function summarizeTurn(userInput, history, assistantText) {
147
+ function summarizeTurn(userInput, history, assistantText, userImageCount = 0) {
145
148
  const lines = [];
146
149
  lines.push(`## User question`);
147
150
  lines.push(userInput.trim().slice(0, 800));
148
151
  lines.push('');
152
+ if (userImageCount > 0) {
153
+ lines.push(`## User-provided images: ${userImageCount}`);
154
+ lines.push('The assistant received these images directly in this turn. Visual observations and transcription can use them as evidence without a tool call. Image bytes are omitted from this text-only check.');
155
+ lines.push('');
156
+ }
149
157
  // ── Harness prefetch (treated as synthetic tool calls) ──
150
158
  // When intent-prefetch fires, it prepends a [FRANKLIN HARNESS PREFETCH]
151
159
  // block to the user message. The LLM answers based on that data, but
@@ -166,7 +174,7 @@ function summarizeTurn(userInput, history, assistantText) {
166
174
  const toolLines = [];
167
175
  for (let i = history.length - 1; i >= 0 && found < 40; i--) {
168
176
  const msg = history[i];
169
- if (msg.role === 'user' && typeof msg.content === 'string')
177
+ if (msg.role === 'user' && (typeof msg.content === 'string' || !msg.content.some(part => part.type === 'tool_result')))
170
178
  break;
171
179
  if (msg.role === 'assistant' && Array.isArray(msg.content)) {
172
180
  for (const part of msg.content) {
@@ -257,7 +265,7 @@ const MAX_EVAL_TOKENS = 512;
257
265
  const EVAL_TIMEOUT_MS = 15_000;
258
266
  export async function checkGrounding(userInput, history, assistantText, client, opts = {}) {
259
267
  const model = opts.model || evaluatorModel();
260
- const summary = summarizeTurn(userInput, history, assistantText);
268
+ const summary = summarizeTurn(userInput, history, assistantText, opts.userImageCount);
261
269
  // Run independently of the main agent — the evaluator gets NO tools
262
270
  // (it just reads and grades). Limit tokens so a chatty evaluator can't
263
271
  // balloon the cost of a cheap check.
@@ -13,9 +13,10 @@
13
13
  * user input keeps waiting and is delivered at the next boundary. The
14
14
  * user's typing is never dropped.
15
15
  */
16
+ import type { AgentInput } from './types.js';
16
17
  export interface InputMultiplexer {
17
18
  /** Drop-in replacement for the driver's getUserInput. */
18
- getUserInput: () => Promise<string | null>;
19
+ getUserInput: () => Promise<AgentInput | null>;
19
20
  /** Queue a synthetic input for the next turn boundary. */
20
21
  enqueue: (text: string, opts?: {
21
22
  priority?: boolean;
@@ -23,4 +24,4 @@ export interface InputMultiplexer {
23
24
  /** Number of injected inputs currently waiting. */
24
25
  pending: () => number;
25
26
  }
26
- export declare function createInputMultiplexer(base: () => Promise<string | null>): InputMultiplexer;
27
+ export declare function createInputMultiplexer(base: () => Promise<AgentInput | null>): InputMultiplexer;
@@ -2,7 +2,7 @@
2
2
  * Franklin Agent Loop
3
3
  * The core reasoning-action cycle: prompt → model → extract capabilities → execute → repeat.
4
4
  */
5
- import type { AgentConfig, ContentPart, Dialogue, StreamEvent } from './types.js';
5
+ import type { AgentConfig, ContentPart, Dialogue, StreamEvent, AgentInput } from './types.js';
6
6
  export declare function isExternalWallFailure(toolName: string, output: string, isError?: boolean): boolean;
7
7
  /**
8
8
  * Detect when the gateway leaked an upstream rate-limit / quota error as a
@@ -60,4 +60,4 @@ export declare function isWeakModel(model: string): boolean;
60
60
  * Each user message triggers a full agent loop.
61
61
  * Returns the accumulated conversation history.
62
62
  */
63
- export declare function interactiveSession(config: AgentConfig, getUserInput: () => Promise<string | null>, onEvent: (event: StreamEvent) => void, onAbortReady?: (abort: () => void) => void): Promise<Dialogue[]>;
63
+ export declare function interactiveSession(config: AgentConfig, getUserInput: () => Promise<AgentInput | null>, onEvent: (event: StreamEvent) => void, onAbortReady?: (abort: () => void) => void): Promise<Dialogue[]>;
@@ -418,6 +418,22 @@ export function stripLargeImageData(message) {
418
418
  // in via the loop's outcome-building path. Runtime structure is what
419
419
  // matters here; we only mutate when we positively identify the shape.
420
420
  const newContent = message.content.map((part) => {
421
+ // Images the user attached directly (Desktop paste / file picker) sit at
422
+ // the top level of the user message, not inside a tool_result. They have
423
+ // no source path to re-Read, so the placeholder asks for a re-attach.
424
+ if (typeof part === 'object' &&
425
+ part !== null &&
426
+ part.type === 'image' &&
427
+ part.source?.type === 'base64' &&
428
+ (part.source.data?.length ?? 0) > SESSION_IMAGE_STRIP_THRESHOLD) {
429
+ mutated = true;
430
+ const sz = part.source.data.length;
431
+ return {
432
+ type: 'text',
433
+ text: `<user-attached image stripped from session log: ${(sz / 1024).toFixed(1)} KB base64. ` +
434
+ `Ask the user to attach it again if you need to see it.>`,
435
+ };
436
+ }
421
437
  if (typeof part === 'object' &&
422
438
  part !== null &&
423
439
  part.type === 'tool_result' &&
@@ -599,6 +615,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
599
615
  const inputMux = createInputMultiplexer(getUserInput);
600
616
  const history = [];
601
617
  let lastUserInput = ''; // For /retry
618
+ let lastUserImages = [];
602
619
  // budget-cap-usd of the skill that kicked off this turn (`/name` invocation);
603
620
  // kept across /retry, which replays the same skill prompt.
604
621
  let turnSkillBudgetUsd;
@@ -616,7 +633,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
616
633
  onEvent({ kind: 'text_delta', text: `[skills] ${err.path}: ${err.error}\n` });
617
634
  }
618
635
  }
619
- const skillRegistry = skillBoot.registry;
636
+ let skillRegistry = skillBoot.registry;
620
637
  // Track models that failed with 402 (payment required) across turns.
621
638
  // These persist until the session ends — unlike transient errors, payment failures
622
639
  // will keep failing until the user adds funds. Map stores failure timestamp for future TTL.
@@ -795,11 +812,15 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
795
812
  process.once('SIGINT', exitFlush);
796
813
  process.once('SIGTERM', exitFlush);
797
814
  while (true) {
798
- let input = await inputMux.getUserInput();
799
- if (input === null)
815
+ const incoming = await inputMux.getUserInput();
816
+ if (incoming === null)
800
817
  break; // User wants to exit
818
+ let input = typeof incoming === 'string' ? incoming : incoming.text;
819
+ let images = typeof incoming === 'string' ? [] : incoming.images;
801
820
  if (input === '')
802
821
  continue; // Empty input → re-prompt
822
+ // Desktop imports and toggles take effect without restarting this session.
823
+ skillRegistry = loadAllSkills(workDir).registry;
803
824
  // ── Slash command dispatch ──
804
825
  if (!input.startsWith('/'))
805
826
  turnSkillBudgetUsd = undefined;
@@ -816,6 +837,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
816
837
  continue;
817
838
  }
818
839
  input = lastUserInput;
840
+ images = lastUserImages;
819
841
  }
820
842
  else {
821
843
  const cmdResult = await handleSlashCommand(input, {
@@ -847,6 +869,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
847
869
  input = redactedText;
848
870
  }
849
871
  lastUserInput = input;
872
+ lastUserImages = images;
850
873
  fireLifecycleHook('UserPromptSubmit');
851
874
  // ── Skill trigger auto-invoke ──
852
875
  // Match the user message against every skill's `triggers:` list. Strong
@@ -891,11 +914,14 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
891
914
  // (pushback SYSTEM NOTE, prefetch context block) are applied AFTER
892
915
  // the turn analyzer runs so they get driven by model-decided flags
893
916
  // instead of keyword regex.
894
- history.push({ role: 'user', content: input });
917
+ const userContent = images.length
918
+ ? [{ type: 'text', text: input }, ...images]
919
+ : input;
920
+ history.push({ role: 'user', content: userContent });
895
921
  turnCount++;
896
922
  toolGuard.startTurn();
897
923
  publishLiveState('working');
898
- persistSessionMessage({ role: 'user', content: input });
924
+ persistSessionMessage({ role: 'user', content: userContent });
899
925
  // ── Model recovery: try original model at the start of each new turn ──
900
926
  // If we fell back to a free model last turn due to a transient error, try original again.
901
927
  // But DON'T reset if the original model had a payment failure — it will just fail again.
@@ -1397,7 +1423,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
1397
1423
  // images, Read's tool_result image blocks get tokenized as base64 text
1398
1424
  // by the gateway (verified 2026-05-09) and the model hallucinates from
1399
1425
  // the "Image file: <path>" stub. Detect upfront, route accordingly.
1400
- const turnNeedsVision = loopCount === 1 && messageNeedsVision(lastUserInput);
1426
+ const turnNeedsVision = loopCount === 1 && (images.length > 0 || messageNeedsVision(lastUserInput));
1401
1427
  // ── Router: resolve routing profiles to concrete models ──
1402
1428
  // The shared Router is local/deterministic and runs once per user turn.
1403
1429
  // The turn analyzer remains useful for planning and pushback, but no
@@ -2413,6 +2439,7 @@ async function runInteractiveSession(config, getUserInput, onEvent, onAbortReady
2413
2439
  if (!config.disableGroundingRetry && shouldCheckGrounding(lastUserInput || '', assistantText)) {
2414
2440
  const gResult = await checkGrounding(lastUserInput, history, assistantText, client, {
2415
2441
  abortSignal: abort.signal,
2442
+ userImageCount: images.length,
2416
2443
  });
2417
2444
  if (gResult.verdict === 'UNGROUNDED' && groundingRetryCount < MAX_GROUNDING_RETRIES) {
2418
2445
  groundingRetryCount++;
@@ -37,7 +37,12 @@ export interface CapabilityOutcome {
37
37
  is_error?: boolean;
38
38
  }
39
39
  export type ContentPart = TextSegment | CapabilityInvocation | ThinkingSegment;
40
- export type UserContentPart = TextSegment | CapabilityOutcome;
40
+ export type UserContentPart = TextSegment | ImageSegment | CapabilityOutcome;
41
+ /** Structured driver input; CLI and scheduled prompts can remain plain text. */
42
+ export type AgentInput = string | {
43
+ text: string;
44
+ images: ImageSegment[];
45
+ };
41
46
  export interface Dialogue {
42
47
  role: Role;
43
48
  content: ContentPart[] | UserContentPart[] | string;
@@ -52,6 +52,8 @@ export interface GatewayModel {
52
52
  name: string;
53
53
  description?: string;
54
54
  owned_by?: string;
55
+ available?: boolean;
56
+ release_date?: string;
55
57
  billing_mode: BillingMode;
56
58
  categories: string[];
57
59
  context_window?: number;
@@ -0,0 +1,4 @@
1
+ import type { AgentInput } from '../agent/types.js';
2
+ /** Accept only inline image bytes. Never turn a renderer-supplied path into a
3
+ * server-side file read or URL fetch. Keep image data out of the text prompt. */
4
+ export declare function desktopInput(text: unknown, attachments: unknown): AgentInput;
@@ -0,0 +1,33 @@
1
+ const MAX_IMAGE_BYTES = 4_000_000;
2
+ const MAX_DATA_URL_LENGTH = Math.ceil(MAX_IMAGE_BYTES / 3) * 4 + 64;
3
+ /** Accept only inline image bytes. Never turn a renderer-supplied path into a
4
+ * server-side file read or URL fetch. Keep image data out of the text prompt. */
5
+ export function desktopInput(text, attachments) {
6
+ const prompt = typeof text === 'string' ? text.trim() : '';
7
+ if (attachments === undefined || (Array.isArray(attachments) && attachments.length === 0)) {
8
+ if (!prompt)
9
+ throw new Error('empty input');
10
+ return prompt;
11
+ }
12
+ if (!Array.isArray(attachments) || attachments.length !== 1) {
13
+ throw new Error('Attach one image per message.');
14
+ }
15
+ const value = attachments[0]?.path;
16
+ if (typeof value !== 'string' || value.length > MAX_DATA_URL_LENGTH)
17
+ throw new Error('Image attachment exceeds the 4 MB limit.');
18
+ const match = /^data:(image\/(?:png|jpeg|webp|gif));base64,([A-Za-z0-9+/]+={0,2})$/.exec(value);
19
+ if (!match)
20
+ throw new Error('Attach a PNG, JPEG, WebP or GIF image.');
21
+ const [, mediaType, data] = match;
22
+ const bytes = Buffer.from(data, 'base64');
23
+ if (bytes.length > MAX_IMAGE_BYTES || bytes.toString('base64') !== data)
24
+ throw new Error('Invalid or oversized image attachment.');
25
+ const valid = mediaType === 'image/png' ? bytes.subarray(0, 8).equals(Buffer.from([137, 80, 78, 71, 13, 10, 26, 10]))
26
+ : mediaType === 'image/jpeg' ? bytes[0] === 255 && bytes[1] === 216 && bytes[2] === 255
27
+ : mediaType === 'image/webp' ? bytes.toString('ascii', 0, 4) === 'RIFF' && bytes.toString('ascii', 8, 12) === 'WEBP'
28
+ : /^GIF8[79]a$/.test(bytes.toString('ascii', 0, 6));
29
+ if (!valid)
30
+ throw new Error('Image attachment does not match its declared format.');
31
+ const image = { type: 'image', source: { type: 'base64', media_type: mediaType, data } };
32
+ return { text: prompt || 'Describe the attached image.', images: [image] };
33
+ }
@@ -0,0 +1,10 @@
1
+ /** Free-text questions have their own protocol; answers never grant tool permissions. */
2
+ export declare class DesktopQuestions<Owner> {
3
+ private timeoutMs;
4
+ private sequence;
5
+ private pending;
6
+ constructor(timeoutMs?: number);
7
+ request(owner: Owner, question: string, options: string[] | undefined, emit: (kind: 'agent.questionAsk' | 'agent.questionClosed', payload: object) => void): Promise<string>;
8
+ respond(owner: Owner, askId: unknown, answer: unknown): boolean;
9
+ cancel(owner: Owner): void;
10
+ }
@@ -0,0 +1,40 @@
1
+ /** Free-text questions have their own protocol; answers never grant tool permissions. */
2
+ export class DesktopQuestions {
3
+ timeoutMs;
4
+ sequence = 0;
5
+ pending = new Map();
6
+ constructor(timeoutMs = 15 * 60 * 1000) {
7
+ this.timeoutMs = timeoutMs;
8
+ }
9
+ request(owner, question, options, emit) {
10
+ const askId = `question_${++this.sequence}`;
11
+ return new Promise(resolve => {
12
+ const finish = (answer) => {
13
+ if (!this.pending.delete(askId))
14
+ return;
15
+ clearTimeout(timer);
16
+ emit('agent.questionClosed', { askId });
17
+ resolve(answer);
18
+ };
19
+ const timer = setTimeout(() => finish(''), this.timeoutMs);
20
+ timer.unref?.();
21
+ this.pending.set(askId, { owner, finish });
22
+ emit('agent.questionAsk', { askId, question, options: options ?? [] });
23
+ });
24
+ }
25
+ respond(owner, askId, answer) {
26
+ if (typeof askId !== 'string' || typeof answer !== 'string' || answer.length > 16_000)
27
+ return false;
28
+ const request = this.pending.get(askId);
29
+ if (!request || request.owner !== owner)
30
+ return false;
31
+ request.finish(answer);
32
+ return true;
33
+ }
34
+ cancel(owner) {
35
+ for (const request of [...this.pending.values()]) {
36
+ if (request.owner === owner)
37
+ request.finish('');
38
+ }
39
+ }
40
+ }
@@ -0,0 +1,10 @@
1
+ import type { GatewayModel } from '../gateway-models.js';
2
+ /** The models.list payload used by both Desktop and the browser UI. */
3
+ export declare function buildModelList(models: GatewayModel[]): {
4
+ id: string;
5
+ label: string;
6
+ free: boolean;
7
+ group: string;
8
+ provider: string;
9
+ contextWindow: number | undefined;
10
+ }[];
@@ -0,0 +1,40 @@
1
+ import { sortModelsNewestFirst } from '@blockrun/model-catalog';
2
+ // Keep company sections in the established picker order. Recency within each
3
+ // section belongs to Model Core, shared with the expanded terminal picker.
4
+ const PROVIDER_LABEL = {
5
+ anthropic: 'Anthropic', openai: 'OpenAI', azure: 'OpenAI', google: 'Google', 'google-vertex': 'Google',
6
+ xai: 'xAI', deepseek: 'DeepSeek', meta: 'Meta', 'meta-llama': 'Meta', nvidia: 'NVIDIA',
7
+ moonshot: 'Moonshot', moonshotai: 'Moonshot', qwen: 'Qwen', alibaba: 'Qwen', mistral: 'Mistral',
8
+ mistralai: 'Mistral', minimax: 'MiniMax', zhipu: 'Zhipu', bytedance: 'ByteDance', cohere: 'Cohere',
9
+ perplexity: 'Perplexity', amazon: 'Amazon', microsoft: 'Microsoft', '01-ai': 'Yi', ai21: 'AI21',
10
+ };
11
+ const PROVIDER_ORDER = ['Anthropic', 'OpenAI', 'Google', 'xAI', 'DeepSeek', 'Qwen', 'Moonshot', 'Meta', 'Mistral', 'MiniMax', 'NVIDIA'];
12
+ function providerLabel(id, ownedBy) {
13
+ const p = (id.split('/')[0] || ownedBy || '').toLowerCase();
14
+ return PROVIDER_LABEL[p] || (p ? p.charAt(0).toUpperCase() + p.slice(1) : 'Other');
15
+ }
16
+ /** The models.list payload used by both Desktop and the browser UI. */
17
+ export function buildModelList(models) {
18
+ const groups = new Map();
19
+ for (const model of models) {
20
+ if (model.available === false || !model.categories.includes('chat'))
21
+ continue;
22
+ const provider = providerLabel(model.id, model.owned_by);
23
+ const group = groups.get(provider) ?? [];
24
+ group.push(model);
25
+ groups.set(provider, group);
26
+ }
27
+ const rank = (provider) => {
28
+ const index = PROVIDER_ORDER.indexOf(provider);
29
+ return index < 0 ? PROVIDER_ORDER.length : index;
30
+ };
31
+ return [...groups].sort(([a], [b]) => rank(a) - rank(b) || a.localeCompare(b))
32
+ .flatMap(([provider, rows]) => sortModelsNewestFirst(rows).map(model => ({
33
+ id: model.id,
34
+ label: model.name,
35
+ free: model.billing_mode === 'free',
36
+ group: provider,
37
+ provider,
38
+ contextWindow: model.context_window,
39
+ })));
40
+ }
@@ -16,6 +16,8 @@
16
16
  * fed by a getUserInput queue. Good enough for the desktop app; multi-session
17
17
  * fan-out can come later.
18
18
  */
19
+ import { desktopInput } from './desktop-input.js';
20
+ import { listManagedSkills, importSkill, toggleSkill, removeSkill, resolveManagedSkill } from '../skills/management.js';
19
21
  import http from 'node:http';
20
22
  import crypto from 'node:crypto';
21
23
  import fs from 'node:fs';
@@ -30,6 +32,7 @@ import { assembleInstructions } from '../agent/context.js';
30
32
  import { interactiveSession } from '../agent/loop.js';
31
33
  import { allCapabilities, createSubAgentCapability } from '../tools/index.js';
32
34
  import { getModelsByCategory } from '../gateway-models.js';
35
+ import { buildModelList } from './model-list.js';
33
36
  import { listSessions, loadSessionHistory } from '../session/storage.js';
34
37
  import { loadSdkSettlements } from '../stats/cost-log.js';
35
38
  import { readSwaps } from '../stats/swap-log.js';
@@ -37,6 +40,7 @@ import { isCloudSyncEnabled, cloudList, cloudSync } from './cloud-sync.js';
37
40
  import { setupAgentWallet, setupAgentSolanaWallet } from '@blockrun/llm';
38
41
  import { retryFetchBalance } from '../commands/balance-retry.js';
39
42
  import { AgentHost } from './agent-host.js';
43
+ import { DesktopQuestions } from './desktop-questions.js';
40
44
  import { bus } from '../events/bus.js';
41
45
  import { FREE_DEFAULT_MODEL } from '../free-models.js';
42
46
  import { readEvmBalance } from '../wallet/manager.js';
@@ -144,26 +148,11 @@ const TOOL_LABELS = {
144
148
  function labelFor(name) {
145
149
  return TOOL_LABELS[name] ?? name;
146
150
  }
147
- // Model list grouping — by provider (company), like OpenRouter/Together. The
148
- // provider is the id's vendor prefix (e.g. "anthropic/claude-…"); PROVIDER_ORDER
149
- // puts the most-wanted vendors first, the rest fall in alphabetically.
150
- const PROVIDER_LABEL = {
151
- anthropic: 'Anthropic', openai: 'OpenAI', azure: 'OpenAI', google: 'Google', 'google-vertex': 'Google',
152
- xai: 'xAI', deepseek: 'DeepSeek', meta: 'Meta', 'meta-llama': 'Meta', nvidia: 'NVIDIA',
153
- moonshot: 'Moonshot', moonshotai: 'Moonshot', qwen: 'Qwen', alibaba: 'Qwen', mistral: 'Mistral',
154
- mistralai: 'Mistral', minimax: 'MiniMax', zhipu: 'Zhipu', bytedance: 'ByteDance', cohere: 'Cohere',
155
- perplexity: 'Perplexity', amazon: 'Amazon', microsoft: 'Microsoft', '01-ai': 'Yi', ai21: 'AI21',
156
- };
157
- const PROVIDER_ORDER = ['Anthropic', 'OpenAI', 'Google', 'xAI', 'DeepSeek', 'Qwen', 'Moonshot', 'Meta', 'Mistral', 'MiniMax', 'NVIDIA'];
158
151
  const MEDIA_MIME = {
159
152
  mp3: 'audio/mpeg', wav: 'audio/wav', m4a: 'audio/mp4', ogg: 'audio/ogg', flac: 'audio/flac',
160
153
  mp4: 'video/mp4', webm: 'video/webm',
161
154
  png: 'image/png', jpg: 'image/jpeg', jpeg: 'image/jpeg', webp: 'image/webp', gif: 'image/gif',
162
155
  };
163
- function providerLabel(id, ownedBy) {
164
- const p = (id.split('/')[0] || ownedBy || '').toLowerCase();
165
- return PROVIDER_LABEL[p] || (p ? p.charAt(0).toUpperCase() + p.slice(1) : 'Other');
166
- }
167
156
  // ─── Browser-attack surface gate ────────────────────────────────────────────
168
157
  // Loopback binding alone is NOT an auth boundary: any web page the user has
169
158
  // open can reach 127.0.0.1 (the browser attaches an Origin header but happily
@@ -453,6 +442,7 @@ export async function startServer(opts) {
453
442
  // Pending approval asks for the legacy desktop session (tools + trade plans).
454
443
  // Keyed by askId — the wire contract the desktop UI already declares
455
444
  // (agent.permissionAsk / agent.permissionResponse with askId + decision).
445
+ const questions = new DesktopQuestions();
456
446
  let askSeq = 0;
457
447
  const legacyAsks = new Map();
458
448
  const requestDesktopPermission = (toolName, description, timeoutMs = 15 * 60 * 1000) => new Promise((resolve) => {
@@ -615,6 +605,8 @@ export async function startServer(opts) {
615
605
  break;
616
606
  }
617
607
  case 'turn_done':
608
+ if (activeWs)
609
+ questions.cancel(activeWs);
618
610
  if (event.reason === 'completed') {
619
611
  emit('agent.done', { sessionId: '', costUsd: 0 });
620
612
  }
@@ -667,13 +659,12 @@ export async function startServer(opts) {
667
659
  allowPaidPrefetch: !desktopBoundary,
668
660
  ...(desktopBoundary ? { maxSpendUsd: 5 } : {}),
669
661
  costSaver,
670
- // Tools such as ImageGen perform their own exact-price confirmation.
671
- // Map the Desktop yes/no approval bridge back to the tool's choices.
672
- onAskUser: async (question, options) => {
673
- const decision = await requestDesktopPermission('Agent confirmation', question);
674
- if (decision === 'yes' || decision === 'always')
675
- return options?.[0] ?? 'yes';
676
- return options?.[options.length - 1] ?? 'no';
662
+ onAskUser: (question, options) => {
663
+ const ws = activeWs;
664
+ const turnId = activeTurnId;
665
+ if (!ws || !turnId)
666
+ return Promise.resolve('');
667
+ return questions.request(ws, question, options, (kind, payload) => send(ws, turnId, kind, payload));
677
668
  },
678
669
  // Structured approvals (trade plans): surfaced to the desktop over the
679
670
  // in-flight turn stream as agent.permissionAsk; the UI answers with
@@ -701,6 +692,27 @@ export async function startServer(opts) {
701
692
  const { id, kind, payload } = msg;
702
693
  const p = (payload ?? {});
703
694
  switch (kind) {
695
+ case 'skills.list':
696
+ case 'skills.upload':
697
+ case 'skills.toggle':
698
+ case 'skills.remove':
699
+ case 'skills.resolve': {
700
+ try {
701
+ const result = kind === 'skills.list' ? listManagedSkills(workDir)
702
+ : kind === 'skills.upload' ? importSkill(workDir, p.content)
703
+ : kind === 'skills.toggle' ? toggleSkill(workDir, p.name, p.enabled)
704
+ : kind === 'skills.remove' ? removeSkill(workDir, p.name)
705
+ : resolveManagedSkill(workDir, p.input);
706
+ send(ws, id, 'response', result);
707
+ }
708
+ catch (err) {
709
+ send(ws, id, 'error', { message: err instanceof Error ? err.message : 'Skill management failed' });
710
+ }
711
+ break;
712
+ }
713
+ case 'skills.create':
714
+ send(ws, id, 'error', { message: 'Automatic skill generation is not available. Import a SKILL.md from Skills instead.' });
715
+ break;
704
716
  case 'session.list': {
705
717
  const metas = listSessions();
706
718
  send(ws, id, 'response', {
@@ -863,26 +875,7 @@ export async function startServer(opts) {
863
875
  case 'models.list': {
864
876
  try {
865
877
  const models = await getModelsByCategory('chat');
866
- const mapped = models.map((m) => {
867
- const provider = providerLabel(m.id, m.owned_by);
868
- return {
869
- id: m.id,
870
- label: m.name,
871
- free: m.billing_mode === 'free',
872
- group: provider,
873
- provider,
874
- contextWindow: m.context_window,
875
- };
876
- });
877
- // Group by provider (PROVIDER_ORDER first, then alphabetical); free
878
- // models surface first within each provider. The picker renders
879
- // consecutive same-`group` items as one section, so sort accordingly.
880
- const rank = (g) => { const i = PROVIDER_ORDER.indexOf(g); return i < 0 ? 999 : i; };
881
- mapped.sort((a, b) => rank(a.group) - rank(b.group) ||
882
- a.group.localeCompare(b.group) ||
883
- (a.free === b.free ? 0 : a.free ? -1 : 1) ||
884
- a.label.localeCompare(b.label));
885
- send(ws, id, 'response', { models: mapped });
878
+ send(ws, id, 'response', { models: buildModelList(models) });
886
879
  }
887
880
  catch (err) {
888
881
  send(ws, id, 'error', { message: err instanceof Error ? err.message : 'models error' });
@@ -890,9 +883,12 @@ export async function startServer(opts) {
890
883
  break;
891
884
  }
892
885
  case 'agent.send': {
893
- const text = String(p.text ?? '').trim();
894
- if (!text) {
895
- send(ws, id, 'agent.error', { sessionId: '', message: 'empty input' });
886
+ let input;
887
+ try {
888
+ input = desktopInput(p.text, p.attachments);
889
+ }
890
+ catch (error) {
891
+ send(ws, id, 'agent.error', { sessionId: '', message: error instanceof Error ? error.message : 'Invalid attachment' });
896
892
  break;
897
893
  }
898
894
  activeWs = ws;
@@ -919,7 +915,7 @@ export async function startServer(opts) {
919
915
  pushInput(`/model ${desiredModel}`);
920
916
  currentModel = desiredModel;
921
917
  }
922
- pushInput(text);
918
+ pushInput(input);
923
919
  break;
924
920
  }
925
921
  case 'settings.get':
@@ -936,8 +932,14 @@ export async function startServer(opts) {
936
932
  break;
937
933
  }
938
934
  case 'agent.cancel':
939
- if (abortFn)
940
- abortFn();
935
+ if (ws === activeWs) {
936
+ if (abortFn)
937
+ abortFn();
938
+ questions.cancel(ws);
939
+ }
940
+ break;
941
+ case 'agent.questionResponse':
942
+ send(ws, id, 'response', { ok: questions.respond(ws, p.askId, p.answer) });
941
943
  break;
942
944
  case 'agent.permissionResponse': {
943
945
  // Two resolution paths:
@@ -1115,6 +1117,7 @@ export async function startServer(opts) {
1115
1117
  });
1116
1118
  });
1117
1119
  ws.on('close', () => {
1120
+ questions.cancel(ws);
1118
1121
  // Tear down this socket's agent subscriptions; the agents keep running.
1119
1122
  const subs = wsSubs.get(ws);
1120
1123
  if (subs)
@@ -40,7 +40,9 @@ export declare function loadBundledSkills(): BundledLoad;
40
40
  * unioned; the Registry resolves name collisions by source precedence
41
41
  * (project > user > learned > bundled).
42
42
  */
43
- export declare function loadAllSkills(workDir: string): AllSkillsLoad;
43
+ export declare function loadAllSkills(workDir: string, options?: {
44
+ includeDisabled?: boolean;
45
+ }): AllSkillsLoad;
44
46
  /** Ensure the user-global skills root exists so `~/.blockrun/skills/<name>/`
45
47
  * edits don't fail with ENOENT. Idempotent. */
46
48
  export declare function ensureUserSkillsDir(): string;
@@ -25,6 +25,7 @@ import { fileURLToPath } from 'node:url';
25
25
  import { dirname, join } from 'node:path';
26
26
  import { existsSync, mkdirSync } from 'node:fs';
27
27
  import { loadSkillsFromDir } from './loader.js';
28
+ import { disabledSkills } from './state.js';
28
29
  import { Registry } from './registry.js';
29
30
  import { BLOCKRUN_DIR } from '../config.js';
30
31
  import { migrateLegacyLearnedSkills } from '../learnings/store.js';
@@ -51,7 +52,7 @@ export function loadBundledSkills() {
51
52
  * unioned; the Registry resolves name collisions by source precedence
52
53
  * (project > user > learned > bundled).
53
54
  */
54
- export function loadAllSkills(workDir) {
55
+ export function loadAllSkills(workDir, options = {}) {
55
56
  const errors = [];
56
57
  const all = [];
57
58
  const counts = {};
@@ -85,7 +86,8 @@ export function loadAllSkills(workDir) {
85
86
  errors.push(...result.errors);
86
87
  counts[source] = (counts[source] ?? 0) + result.skills.length;
87
88
  }
88
- return { registry: Registry.fromLoaded(all), errors, counts };
89
+ const disabled = options.includeDisabled ? new Set() : disabledSkills();
90
+ return { registry: Registry.fromLoaded(all.filter((s) => !disabled.has(s.skill.name))), errors, counts };
89
91
  }
90
92
  /** Ensure the user-global skills root exists so `~/.blockrun/skills/<name>/`
91
93
  * edits don't fail with ENOENT. Idempotent. */
@@ -33,6 +33,9 @@ export function matchSkill(input, registry, vars) {
33
33
  if (!skill)
34
34
  return null;
35
35
  const args = space < 0 ? '' : input.slice(space + 1).trim();
36
- const rewritten = substituteVariables(skill.skill.body, vars, args);
36
+ const rendered = substituteVariables(skill.skill.body, vars, args);
37
+ // External SKILL.md files often omit Franklin's argument placeholder.
38
+ const rewritten = args && !skill.skill.body.includes('$ARGUMENTS')
39
+ ? `${rendered}\n\nUser request:\n${args}` : rendered;
37
40
  return { rewritten, skill: skill.skill };
38
41
  }
@@ -1,12 +1,4 @@
1
- /**
2
- * SKILL.md loader — parses Anthropic-spec frontmatter + markdown body.
3
- *
4
- * The frontmatter parser is intentionally minimal: flat `key: value` lines
5
- * with scalar values (string, boolean, number). Quoted strings, nested
6
- * objects, and arrays are out of scope for MVP — Anthropic's SKILL.md spec
7
- * never needs them and adding a YAML dependency would be heavier than the
8
- * spec warrants.
9
- */
1
+ /** SKILL.md frontmatter and Markdown loader. Metadata is data, never executable. */
10
2
  import type { LoadResult, ParseResult, SkillSource } from './types.js';
11
3
  export declare function parseSkill(content: string): ParseResult;
12
4
  /**