@blockrun/franklin 3.29.1 → 3.29.3

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.
package/README.md CHANGED
@@ -500,6 +500,10 @@ Same wallet. Same tools. From your phone.
500
500
 
501
501
  ## Documentation
502
502
 
503
+ Full BlockRun docs: **https://blockrun.ai/docs**
504
+
505
+ - [Agent developer guide](https://blockrun.ai/docs/getting-started/agent-developers) — building agents on BlockRun
506
+ - [Models, SDKs & APIs](https://blockrun.ai/docs) — the full gateway reference
503
507
  - [Plugin SDK guide](docs/plugin-sdk.md) — build your own workflow vertical
504
508
  - [Changelog](CHANGELOG.md) — every release explained
505
509
  - [Roadmap](docs/ROADMAP.md) — what's coming next
@@ -242,7 +242,7 @@ You run on the BlockRun AI Gateway. When the user asks you to "test the BlockRun
242
242
  - \`GET /.well-known/x402\` — x402 resource list with prices
243
243
 
244
244
  **LLM (POST, x402-paid)**
245
- - \`POST /v1/chat/completions\` — OpenAI-compatible. Body: \`{ model, messages, stream?, tools?, max_tokens?, temperature? }\`. \`model\` MUST come from \`GET /v1/models\` (real frontier examples on the gateway as of 2026-06: \`anthropic/claude-sonnet-4.6\`, \`anthropic/claude-opus-4.8\`, \`deepseek/deepseek-v4-pro\`, \`zai/glm-5.1\`, \`nvidia/llama-4-maverick\`, \`openai/gpt-5-nano\`). Do NOT invent versions like \`openai/gpt-5.1\` or \`xai/grok-5\` — those don't exist; the gateway 400s with the valid list in the error body, so when in doubt fetch \`GET /v1/models\` first.
245
+ - \`POST /v1/chat/completions\` — OpenAI-compatible. Body: \`{ model, messages, stream?, tools?, max_tokens?, temperature? }\`. \`model\` MUST come from \`GET /v1/models\` (real frontier examples on the gateway as of 2026-06: \`anthropic/claude-sonnet-4.6\`, \`anthropic/claude-opus-4.8\`, \`deepseek/deepseek-v4-pro\`, \`zai/glm-5.2\`, \`nvidia/llama-4-maverick\`, \`openai/gpt-5-nano\`). Do NOT invent versions like \`openai/gpt-5.1\` or \`xai/grok-5\` — those don't exist; the gateway 400s with the valid list in the error body, so when in doubt fetch \`GET /v1/models\` first.
246
246
  - \`POST /v1/messages\` — Anthropic-compatible. Body: \`{ model, messages, max_tokens, system?, tools? }\`.
247
247
 
248
248
  **Media (POST, x402-paid; GET to poll async jobs)**
package/dist/agent/llm.js CHANGED
@@ -949,9 +949,16 @@ export class ModelClient {
949
949
  }
950
950
  if (currentText) {
951
951
  if (textEmission.mode === 'hold' && isRoleplayedJsonToolCallText(currentText)) {
952
+ // A roleplayed JSON tool call. Do NOT stream it to the UI,
953
+ // but DO keep it in `collected` so the loop's scavenge pass
954
+ // (src/agent/repair/scavenge.ts) can turn it into a real
955
+ // tool_use and execute it. If scavenge can't recover a valid
956
+ // call, the loop strips this text, restoring the old
957
+ // "non-productive → switch model" recovery.
958
+ collected.push({ type: 'text', text: currentText });
952
959
  if (this.debug) {
953
960
  console.error(`[franklin] Model ${request.model} emitted a raw JSON function-call object as text. ` +
954
- 'Treating it as non-productive output so recovery can try another model.');
961
+ 'Held from display; loop will scavenge it into a tool_use or strip it.');
955
962
  }
956
963
  }
957
964
  else if (textEmission.mode === 'hold' && isNemotronProse) {
@@ -1038,9 +1045,12 @@ export class ModelClient {
1038
1045
  }
1039
1046
  if (currentText) {
1040
1047
  if (textEmission.mode === 'hold' && isRoleplayedJsonToolCallText(currentText)) {
1048
+ // See the matching branch in content_block_stop: keep it for the
1049
+ // loop's scavenge pass instead of dropping it.
1050
+ collected.push({ type: 'text', text: currentText });
1041
1051
  if (this.debug) {
1042
1052
  console.error(`[franklin] Model ${request.model} emitted a raw JSON function-call object as text. ` +
1043
- 'Treating it as non-productive output so recovery can try another model.');
1053
+ 'Held from display; loop will scavenge it into a tool_use or strip it.');
1044
1054
  }
1045
1055
  }
1046
1056
  else if (textEmission.mode === 'hold' && isNemotronProse) {
@@ -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 { ModelClient } from './llm.js';
5
+ import { ModelClient, isRoleplayedJsonToolCallText } from './llm.js';
6
6
  import { autoCompactIfNeeded, forceCompact, microCompact, projectCompactionSavings } from './compact.js';
7
7
  import { estimateHistoryTokens, updateActualTokens, resetTokenAnchor, getAnchoredTokenCount, getContextWindow, setEstimationModel } from './tokens.js';
8
8
  import { handleSlashCommand } from './commands.js';
@@ -1368,6 +1368,14 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
1368
1368
  responseParts = [...responseParts, ...novelCalls];
1369
1369
  logger.warn(`[franklin] scavenged ${repaired.report.scavenged} leaked tool call(s) from ${config.model}: ${repaired.report.notes.join('; ')}`);
1370
1370
  }
1371
+ // Strip pure roleplayed-JSON tool-call text parts. The LLM client
1372
+ // holds these out of the live stream (never shown) but keeps them
1373
+ // in `content` so the scavenge above can recover the call. Whether
1374
+ // or not a call was recovered, the raw JSON must not linger as the
1375
+ // assistant's "answer" — drop it. When nothing was scavenged this
1376
+ // also empties the turn, so the empty-response recovery below can
1377
+ // switch models, preserving the prior behavior.
1378
+ responseParts = responseParts.filter(p => !(p.type === 'text' && isRoleplayedJsonToolCallText(p.text)));
1371
1379
  }
1372
1380
  // ── Empty response recovery ──
1373
1381
  // If the model returns nothing, DON'T just retry the same model with the same input.
@@ -1376,7 +1384,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
1376
1384
  const hasTools = responseParts.some(p => p.type === 'tool_use');
1377
1385
  const hasThinking = responseParts.some(p => p.type === 'thinking');
1378
1386
  if (!hasText && !hasTools && !hasThinking) {
1379
- const EMPTY_FALLBACK_MODELS = ['nvidia/llama-4-maverick', 'nvidia/deepseek-v4-flash', 'zai/glm-5.1'];
1387
+ const EMPTY_FALLBACK_MODELS = ['nvidia/llama-4-maverick', 'nvidia/deepseek-v4-flash', 'zai/glm-5.2'];
1380
1388
  const nextModel = EMPTY_FALLBACK_MODELS.find(m => m !== config.model && !turnFailedModels.has(m));
1381
1389
  if (nextModel && recoveryAttempts < 2) {
1382
1390
  recoveryAttempts++;
@@ -50,8 +50,9 @@ export function scavengeToolCalls(text, opts) {
50
50
  for (const candidate of iterateJsonObjects(nonDsml)) {
51
51
  if (out.length >= max)
52
52
  break;
53
- const call = coerceToInvocation(candidate, opts.allowedNames);
54
- if (call) {
53
+ for (const call of coerceToInvocations(candidate, opts.allowedNames)) {
54
+ if (out.length >= max)
55
+ break;
55
56
  out.push(call);
56
57
  notes.push(`scavenged call: ${call.name}`);
57
58
  }
@@ -134,17 +135,41 @@ function* iterateJsonObjects(text) {
134
135
  }
135
136
  }
136
137
  }
137
- function coerceToInvocation(candidateJson, allowedNames) {
138
+ function coerceToInvocations(candidateJson, allowedNames) {
138
139
  let parsed;
139
140
  try {
140
141
  parsed = JSON.parse(candidateJson);
141
142
  }
142
143
  catch {
143
- return null;
144
+ return [];
144
145
  }
145
146
  if (!parsed || typeof parsed !== 'object')
146
- return null;
147
+ return [];
147
148
  const obj = parsed;
149
+ // Envelope A — OpenAI assistant message { tool_calls: [ {...}, ... ] }.
150
+ // Gateway models that lack native tool-calling sometimes dump the whole
151
+ // assistant-message JSON into the text channel.
152
+ if (Array.isArray(obj.tool_calls)) {
153
+ const calls = [];
154
+ for (const tc of obj.tool_calls) {
155
+ const c = coerceObj(tc, allowedNames);
156
+ if (c)
157
+ calls.push(c);
158
+ }
159
+ return calls;
160
+ }
161
+ // Envelope B — legacy single { function_call: { name, arguments } }.
162
+ if (obj.function_call && typeof obj.function_call === 'object') {
163
+ const c = coerceObj({ type: 'function', function: obj.function_call }, allowedNames);
164
+ return c ? [c] : [];
165
+ }
166
+ const c = coerceObj(obj, allowedNames);
167
+ return c ? [c] : [];
168
+ }
169
+ function coerceObj(value, allowedNames) {
170
+ if (!value || typeof value !== 'object')
171
+ return null;
172
+ const obj = value;
148
173
  // Pattern 1 — { name, arguments } (Anthropic-ish flat form) AND the flat
149
174
  // OpenAI leak { type: "function", name, parameters } emitted by the free
150
175
  // DeepSeek model — both carry the name at the top level. Args land under
@@ -188,10 +213,18 @@ function resolveAllowedName(name, allowedNames) {
188
213
  return null;
189
214
  if (allowedNames.has(name))
190
215
  return name;
191
- const target = normalizeName(name);
192
- for (const allowed of allowedNames) {
193
- if (normalizeName(allowed) === target)
194
- return allowed;
216
+ // Try the name as-is, then with any provider namespace stripped
217
+ // (`functions.web_search` / `tools.web_search` → `web_search`), which
218
+ // Gemini- and OpenAI-style function declarations sometimes prepend.
219
+ const candidates = [name];
220
+ if (name.includes('.'))
221
+ candidates.push(name.slice(name.lastIndexOf('.') + 1));
222
+ for (const candidate of candidates) {
223
+ const target = normalizeName(candidate);
224
+ for (const allowed of allowedNames) {
225
+ if (normalizeName(allowed) === target)
226
+ return allowed;
227
+ }
195
228
  }
196
229
  return null;
197
230
  }
@@ -232,6 +232,7 @@ const MODEL_CONTEXT_WINDOWS = {
232
232
  'xai/grok-4-0709': 131_072,
233
233
  'xai/grok-4-1-fast-reasoning': 131_072,
234
234
  // Others
235
+ 'zai/glm-5.2': 1_000_000, // flagship bump 2026-06 — context jumped 200K → 1M
235
236
  'zai/glm-5.1': 200_000,
236
237
  'moonshot/kimi-k2.7': 256_000,
237
238
  'moonshot/kimi-k2.6': 256_000,
package/dist/pricing.js CHANGED
@@ -92,6 +92,7 @@ export const MODEL_PRICING = {
92
92
  // PROMOTION (active ~2026-04): flat $0.001/call for all GLM models
93
93
  'zai/glm-5': { input: 0.6, output: 1.92 }, // flat promo ended 2026-06-06 — per-token now
94
94
  'zai/glm-5.1': { input: 1.40, output: 4.40 }, // launch promo ended 2026-06-05 — per-token now
95
+ 'zai/glm-5.2': { input: 1.40, output: 4.40 }, // new flagship 2026-06 — 1M context, same per-token price as 5.1
95
96
  'zai/glm-5-turbo': { input: 1.2, output: 4.0 }, // flat promo ended 2026-06-06 — per-token now
96
97
  'zai/glm-5.1-turbo': { input: 1.2, output: 4.0 }, // client alias for zai/glm-5-turbo
97
98
  };
@@ -158,9 +158,12 @@ const MODEL_SHORTCUTS = {
158
158
  'm3': 'minimax/minimax-m3',
159
159
  'm2.7': 'minimax/minimax-m2.7',
160
160
  // Others
161
- glm: 'zai/glm-5.1',
161
+ glm: 'zai/glm-5.2',
162
+ 'glm-5.2': 'zai/glm-5.2',
163
+ // glm-5.1 demoted to back-compat pin 2026-06 (flagship is 5.2) — still routes.
164
+ 'glm-5.1': 'zai/glm-5.1',
162
165
  'glm-turbo': 'zai/glm-5-turbo',
163
- 'glm5': 'zai/glm-5.1',
166
+ 'glm5': 'zai/glm-5.2',
164
167
  kimi: 'moonshot/kimi-k2.7',
165
168
  'k2.7': 'moonshot/kimi-k2.7',
166
169
  // K2.6 demoted 2026-06 (flagship is K2.7) but still routes — `k2.6` pins it.
@@ -97,10 +97,14 @@ export const MODEL_SHORTCUTS = {
97
97
  minimax: 'minimax/minimax-m3',
98
98
  'm3': 'minimax/minimax-m3',
99
99
  'm2.7': 'minimax/minimax-m2.7',
100
- glm: 'zai/glm-5.1',
100
+ glm: 'zai/glm-5.2',
101
101
  'glm-5': 'zai/glm-5',
102
+ 'glm-5.2': 'zai/glm-5.2',
103
+ // GLM-5.1 demoted to a back-compat pin 2026-06 (flagship is 5.2) — still
104
+ // routes for anyone who wants the 200K-context build explicitly.
105
+ 'glm-5.1': 'zai/glm-5.1',
102
106
  'glm-turbo': 'zai/glm-5-turbo',
103
- 'glm5': 'zai/glm-5.1',
107
+ 'glm5': 'zai/glm-5.2',
104
108
  kimi: 'moonshot/kimi-k2.7',
105
109
  'k2.7': 'moonshot/kimi-k2.7',
106
110
  // K2.6 demoted 2026-06 (gateway flagship is K2.7) but still routes — the
@@ -169,6 +173,9 @@ export const PICKER_CATEGORIES = [
169
173
  { id: 'deepseek/deepseek-v4-pro', shortcut: 'deepseek-v4-pro', label: 'DeepSeek V4 Pro', price: '$0.435/$0.87', highlight: true },
170
174
  { id: 'deepseek/deepseek-reasoner', shortcut: 'r1', label: 'DeepSeek V4 Flash R.', price: '$0.2/$0.4' },
171
175
  { id: 'xai/grok-4-1-fast-reasoning', shortcut: 'grok-fast', label: 'Grok 4.1 Fast R.', price: '$0.2/$0.5' },
176
+ // GLM-5.2: Z.AI's new flagship — 1M context, top open-source on
177
+ // long-horizon coding. `glm`/`glm5` shortcuts pin it.
178
+ { id: 'zai/glm-5.2', shortcut: 'glm', label: 'GLM-5.2', price: '$1.4/$4.4' },
172
179
  ],
173
180
  },
174
181
  {
@@ -181,7 +188,8 @@ export const PICKER_CATEGORIES = [
181
188
  { id: 'deepseek/deepseek-chat', shortcut: 'deepseek', label: 'DeepSeek V4 Flash Chat', price: '$0.2/$0.4' },
182
189
  { id: 'moonshot/kimi-k2.7', shortcut: 'kimi', label: 'Kimi K2.7', price: '$0.95/$4' },
183
190
  // GLM flat-rate promos fully ended 2026-06-06 — whole family per-token
184
- // now (glm-5 $0.60/$1.92; `glm` shortcut still pins flagship glm-5.1).
191
+ // now (glm-5 $0.60/$1.92; `glm` shortcut pins flagship glm-5.2, listed
192
+ // in Reasoning above).
185
193
  { id: 'zai/glm-5', shortcut: 'glm-5', label: 'GLM-5', price: '$0.6/$1.92' },
186
194
  // Minimax M2.7 hidden to make room for V4 Pro in Reasoning + V4 Flash
187
195
  // (free) without exceeding the picker's 24-entry cap. Shortcut `minimax`
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@blockrun/franklin",
3
- "version": "3.29.1",
3
+ "version": "3.29.3",
4
4
  "description": "Franklin Agent — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
5
5
  "type": "module",
6
6
  "exports": {