@blockrun/franklin 3.45.0 → 3.46.0
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 +13 -11
- package/dist/agent/context.js +58 -10
- package/dist/agent/llm.js +14 -6
- package/dist/agent/loop.js +34 -6
- package/dist/agent/optimize.d.ts +1 -1
- package/dist/agent/optimize.js +17 -7
- package/dist/agent/secret-redact.d.ts +26 -0
- package/dist/agent/secret-redact.js +48 -0
- package/dist/agent/streaming-executor.js +15 -0
- package/dist/commands/doctor.js +18 -0
- package/dist/commands/usage.d.ts +14 -0
- package/dist/commands/usage.js +86 -0
- package/dist/commands/wallet-adopt.d.ts +12 -0
- package/dist/commands/wallet-adopt.js +49 -0
- package/dist/index.js +16 -0
- package/dist/payments/account.d.ts +17 -0
- package/dist/payments/account.js +105 -0
- package/dist/payments/auth-mode.d.ts +5 -19
- package/dist/payments/auth-mode.js +15 -17
- package/dist/payments/billing-copy.d.ts +37 -0
- package/dist/payments/billing-copy.js +57 -0
- package/dist/payments/post-with-payment.js +8 -15
- package/dist/payments/price-catalog.d.ts +9 -0
- package/dist/payments/price-catalog.js +13 -0
- package/dist/payments/usage.d.ts +106 -0
- package/dist/payments/usage.js +168 -0
- package/dist/stats/tracker.d.ts +9 -1
- package/dist/stats/tracker.js +6 -1
- package/dist/tools/bash.js +17 -7
- package/dist/tools/blockrun.js +1 -1
- package/dist/tools/defillama.js +2 -2
- package/dist/tools/imagegen.js +5 -4
- package/dist/tools/modal.js +12 -6
- package/dist/tools/musicgen.js +4 -3
- package/dist/tools/phone.js +16 -15
- package/dist/tools/prediction.js +12 -11
- package/dist/tools/realface.js +2 -2
- package/dist/tools/rpc.js +2 -2
- package/dist/tools/sensitive-paths.d.ts +1 -1
- package/dist/tools/sensitive-paths.js +13 -4
- package/dist/tools/surf.js +2 -2
- package/dist/tools/videogen.js +6 -5
- package/dist/tools/voice.d.ts +1 -1
- package/dist/tools/voice.js +5 -4
- package/dist/wallet/reservation.d.ts +24 -0
- package/dist/wallet/reservation.js +50 -4
- package/dist/wallet/solana-migration.d.ts +48 -0
- package/dist/wallet/solana-migration.js +81 -0
- package/package.json +2 -2
package/README.md
CHANGED
|
@@ -121,7 +121,7 @@ VS Code → Extensions (Cmd+Shift+X / Ctrl+Shift+X)
|
|
|
121
121
|
→ click the Franklin icon in the Activity Bar
|
|
122
122
|
```
|
|
123
123
|
|
|
124
|
-
Free models work immediately. Paid models, image gen, and video gen
|
|
124
|
+
Free models work immediately. Paid models, image gen, and video gen use your selected payment method: funded API account credit or a USDC wallet. The CLI and the extension share the same `~/.blockrun/` config and session history, so jumping between terminal and VS Code is seamless.
|
|
125
125
|
|
|
126
126
|
### Franklin Desktop (Beta)
|
|
127
127
|
|
|
@@ -205,8 +205,10 @@ franklin balance --wallet # show the wallet balance, not the key status
|
|
|
205
205
|
```
|
|
206
206
|
|
|
207
207
|
If the key is rejected, or a tool needs an endpoint the key gateway does not serve,
|
|
208
|
-
Franklin
|
|
209
|
-
|
|
208
|
+
Franklin returns that API error without charging a wallet. Correct the key or explicitly
|
|
209
|
+
retry with `--wallet`. An empty or malformed configured key is an error; remove it to
|
|
210
|
+
return to the wallet default. `franklin logout` removes the saved key; if
|
|
211
|
+
`BLOCKRUN_API_KEY` is exported, unset it too.
|
|
210
212
|
|
|
211
213
|
### Spend tracking, and the one thing that stays an estimate
|
|
212
214
|
|
|
@@ -339,7 +341,7 @@ Every trade journals itself: thesis on open, P&L on close, written to a **wallet
|
|
|
339
341
|
Saved: generated-logo-1713052800.png (1024x1024)
|
|
340
342
|
```
|
|
341
343
|
|
|
342
|
-
Generates images via DALL-E / GPT Image directly from the CLI. Paid from your wallet — no OpenAI API key needed.
|
|
344
|
+
Generates images via DALL-E / GPT Image directly from the CLI. Paid from your selected BlockRun API account or wallet — no OpenAI API key needed.
|
|
343
345
|
|
|
344
346
|
### 📱 Remote control via Telegram
|
|
345
347
|
|
|
@@ -380,7 +382,7 @@ Every tool call is itemized. Every token is priced. When the wallet hits zero, F
|
|
|
380
382
|
|
|
381
383
|
## Smart Router
|
|
382
384
|
|
|
383
|
-
**<!-- br:models.chatVisible -->
|
|
385
|
+
**<!-- br:models.chatVisible -->78<!-- /br:models.chatVisible --> models. One decision. Zero guesswork.**
|
|
384
386
|
|
|
385
387
|
You don't pick models. Franklin picks for you.
|
|
386
388
|
|
|
@@ -461,7 +463,7 @@ No email. No phone. No KYC. Your Solana or Base address is your account — port
|
|
|
461
463
|
| Writes code | ✅ | ✅ | ⚠️ | ✅ |
|
|
462
464
|
| **Spends money for you** | ❌ | ❌ | ❌ | ✅ **USDC wallet, x402** |
|
|
463
465
|
| **Buys data + APIs + images + search** | ❌ | ❌ | ❌ | ✅ **55+ APIs, one wallet** |
|
|
464
|
-
| Picks best model per task | ❌ single-vendor | ❌ plan-tied | ❌ | ✅ **Smart Router, <!-- br:models.chatVisible -->
|
|
466
|
+
| Picks best model per task | ❌ single-vendor | ❌ plan-tied | ❌ | ✅ **Smart Router, <!-- br:models.chatVisible -->78<!-- /br:models.chatVisible --> models** |
|
|
465
467
|
| Pricing model | Subscription | Subscription | Subscription | **YOPO** — per outcome, USDC |
|
|
466
468
|
| Monthly fee | $20–$200 | $20–$40 | $20+ | **$0** |
|
|
467
469
|
| Rate-limited | Yes | Yes | Yes | No — limited only by wallet |
|
|
@@ -489,14 +491,14 @@ Ask "what's BTC looking like?" — Franklin fetches live price data, computes RS
|
|
|
489
491
|
**🎨 AI image generation**
|
|
490
492
|
Ask "generate a logo" — Franklin calls DALL-E / GPT Image, saves the result locally, paid from your wallet.
|
|
491
493
|
|
|
492
|
-
**🧠 <!-- br:models.chatVisible -->
|
|
494
|
+
**🧠 <!-- br:models.chatVisible -->78<!-- /br:models.chatVisible --> models via one wallet**
|
|
493
495
|
Anthropic, OpenAI, Google, xAI, DeepSeek, GLM, Kimi, Minimax, NVIDIA free tier. One wallet, one interface, automatic fallback.
|
|
494
496
|
|
|
495
497
|
**💳 x402 micropayments (YOPO)**
|
|
496
498
|
HTTP 402 native. Every paid action is a signed USDC micropayment via EIP-712 — non-custodial, your keys never leave your machine. YOPO: you pay only for outcomes.
|
|
497
499
|
|
|
498
500
|
**🧠 Learned model router**
|
|
499
|
-
Trained on 2M+ real requests. Classifies your task and picks the best model from <!-- br:models.chatVisible -->
|
|
501
|
+
Trained on 2M+ real requests. Classifies your task and picks the best model from <!-- br:models.chatVisible -->78<!-- /br:models.chatVisible --> LLMs. Four profiles (auto/eco/premium/free). Adapts to your usage over time.
|
|
500
502
|
|
|
501
503
|
</td>
|
|
502
504
|
<td width="50%" valign="top">
|
|
@@ -552,13 +554,13 @@ Core is workflow-agnostic. Add new verticals without touching the loop. Discover
|
|
|
552
554
|
│ Intent → Smart Router → Tool Use → Spend Control → Result │
|
|
553
555
|
├──────────────────────────────────────────────────────────────┤
|
|
554
556
|
│ Learned Router │
|
|
555
|
-
│ 2M+ requests · <!-- br:models.chatVisible@live -->
|
|
557
|
+
│ 2M+ requests · <!-- br:models.chatVisible@live -->78<!-- /br:models.chatVisible@live --> models · category detection · Elo scores │
|
|
556
558
|
├──────────────────────────────────────────────────────────────┤
|
|
557
559
|
│ Agent Loop │
|
|
558
560
|
│ 16 tools · Sessions · Compaction · Pricing · Plugin SDK │
|
|
559
561
|
├──────────────────────────────────────────────────────────────┤
|
|
560
562
|
│ BlockRun Gateway │
|
|
561
|
-
│ <!-- br:models.chatVisible@live -->
|
|
563
|
+
│ <!-- br:models.chatVisible@live -->78<!-- /br:models.chatVisible@live --> LLMs · CoinGecko · Search · Image APIs · paid services │
|
|
562
564
|
├──────────────────────────────────────────────────────────────┤
|
|
563
565
|
│ x402 Micropayment Protocol │
|
|
564
566
|
│ HTTP 402 · USDC on Solana & Base · signed payment payloads │
|
|
@@ -602,7 +604,7 @@ src/
|
|
|
602
604
|
├── stats/ Usage tracking + insights engine
|
|
603
605
|
├── ui/ Ink-based terminal UI
|
|
604
606
|
├── proxy/ Payment proxy for external tools
|
|
605
|
-
├── router/ Learned model router (<!-- br:models.chatVisible@live -->
|
|
607
|
+
├── router/ Learned model router (<!-- br:models.chatVisible@live -->78<!-- /br:models.chatVisible@live --> models, Elo scoring)
|
|
606
608
|
├── wallet/ Wallet management (Base + Solana)
|
|
607
609
|
├── mcp/ MCP server auto-discovery
|
|
608
610
|
└── commands/ CLI subcommands
|
package/dist/agent/context.js
CHANGED
|
@@ -10,21 +10,41 @@ import { getWalletAddress as getBaseWalletAddress } from '@blockrun/llm';
|
|
|
10
10
|
import { Keypair } from '@solana/web3.js';
|
|
11
11
|
import bs58 from 'bs58';
|
|
12
12
|
import { loadLearnings, decayLearnings, saveLearnings, formatForPrompt } from '../learnings/store.js';
|
|
13
|
+
import { isKeyMode, resolvePayMode } from '../payments/auth-mode.js';
|
|
13
14
|
// ─── System Instructions Assembly ──────────────────────────────────────────
|
|
14
15
|
// Composable prompt sections — each independently maintainable and conditionally includable.
|
|
15
16
|
function getCoreInstructions() {
|
|
16
|
-
|
|
17
|
+
// Key mode has no wallet. getWalletKnowledgeSection() below already branches
|
|
18
|
+
// on this; these lines did not, so every key-mode turn briefed the model on
|
|
19
|
+
// a wallet it cannot read and a USDC balance it cannot spend — then pointed
|
|
20
|
+
// the Wallet tool at it, which returns a portal link instead. Keep the
|
|
21
|
+
// spend-without-hesitating posture either way; only the funding source
|
|
22
|
+
// changes.
|
|
23
|
+
const keyMode = isKeyMode();
|
|
24
|
+
const identity = keyMode
|
|
25
|
+
? 'an autonomous AI agent that pays its own way with prepaid BlockRun account credits'
|
|
26
|
+
: 'an autonomous AI agent with a wallet';
|
|
27
|
+
const walletTool = keyMode
|
|
28
|
+
? '- **Wallet** — reports that this session bills prepaid account credits and where to manage them. Use it for any "what\'s my balance / how much money / billing status" question instead of running \`franklin balance\` via Bash. Free, one call.'
|
|
29
|
+
: '- **Wallet** — read your own chain, address, and USDC balance. Use this for any "what\'s my balance / how much money / wallet status" question instead of running \`franklin balance\` via Bash. Free, one call, never costs USDC.';
|
|
30
|
+
const tradingMarket = keyMode
|
|
31
|
+
? '- **TradingMarket** — current stock / FX / crypto / commodity prices (BlockRun Gateway / Pyth; billed to account credits, $0.001/stock call, free for everything else).'
|
|
32
|
+
: '- **TradingMarket** — current stock / FX / crypto / commodity prices (BlockRun Gateway / Pyth; wallet pays automatically, $0.001/stock call, free for everything else).';
|
|
33
|
+
const spendPosture = keyMode
|
|
34
|
+
? 'Your brand is spending real money to get real answers; $0.001 for a stock quote is exactly what the account credits are for. Don\'t hesitate on cents.'
|
|
35
|
+
: 'Your brand is spending USDC to get real answers; $0.001 for a stock quote is exactly what the wallet is for. Don\'t hesitate on cents.';
|
|
36
|
+
return `You are Franklin, ${identity}. You help users with software engineering, marketing campaigns, trading signals, and any task that benefits from an agent that can reason, act, and spend.
|
|
17
37
|
|
|
18
38
|
You are an interactive agent — not a chatbot. Use the tools available to you to accomplish tasks. Your job is to be a highly capable collaborator who takes initiative, makes progress, and delivers results.
|
|
19
39
|
|
|
20
40
|
# Franklin has hands
|
|
21
41
|
You run with live tools by default:
|
|
22
|
-
|
|
23
|
-
|
|
42
|
+
${walletTool}
|
|
43
|
+
${tradingMarket}
|
|
24
44
|
- **ExaAnswer / ExaSearch / ExaReadUrls** — cited current-events answers, semantic web search, clean URL content.
|
|
25
45
|
- **WebSearch / WebFetch** — live web.
|
|
26
46
|
|
|
27
|
-
When a user asks for a current price, today's news, or any live-world state, **call the tool**. Refusal phrases like "I can't provide real-time data" or "check Yahoo Finance" are a bug — they belong to systems without tools.
|
|
47
|
+
When a user asks for a current price, today's news, or any live-world state, **call the tool**. Refusal phrases like "I can't provide real-time data" or "check Yahoo Finance" are a bug — they belong to systems without tools. ${spendPosture}
|
|
28
48
|
|
|
29
49
|
# System
|
|
30
50
|
- All text you output outside of tool use is displayed to the user. Use markdown for formatting.
|
|
@@ -230,22 +250,39 @@ Hard rules:
|
|
|
230
250
|
- If \`open\` fails (e.g. no GUI on a remote box), fall back to giving them the URL as plain text and tell them to paste it into a browser.`;
|
|
231
251
|
}
|
|
232
252
|
function getBlockRunApiSection() {
|
|
253
|
+
// Which host answers depends on how this session pays, and the three are not
|
|
254
|
+
// interchangeable: api.blockrun.ai authenticates a bearer key and 401s
|
|
255
|
+
// without one, while blockrun.ai / sol.blockrun.ai answer a 402 challenge and
|
|
256
|
+
// ignore a bearer key entirely. This section used to call api.blockrun.ai an
|
|
257
|
+
// "alias" of the Base gateway, which invited the model to send a wallet-signed
|
|
258
|
+
// request to a host that cannot settle it, or a key to a host that cannot read
|
|
259
|
+
// it. Name the one host this session actually uses.
|
|
260
|
+
//
|
|
261
|
+
// assembleInstructions() keys its cache on the pay mode, so this is safe to
|
|
262
|
+
// vary and cannot be served stale after a switch.
|
|
263
|
+
const mode = resolvePayMode();
|
|
264
|
+
const chargedBy = mode.kind === 'key' ? 'account credits' : 'x402 wallet payment';
|
|
265
|
+
const hosts = mode.kind === 'key'
|
|
266
|
+
? `- **Your host: \`https://api.blockrun.ai\`** — the account API. Requests carry your configured key; the built-in tools attach it. It is a SEPARATE service, not an alias of the Base gateway.
|
|
267
|
+
- Not yours: \`https://blockrun.ai/api\` (Base x402 wallet) and \`https://sol.blockrun.ai/api\` (Solana x402 wallet). They ignore an account key and answer 402. Never send one there, and never attach a wallet payment proof to an account request.`
|
|
268
|
+
: `- **Your host: \`${mode.apiBase}\`** — the ${mode.chain === 'solana' ? 'Solana' : 'Base'} x402 wallet gateway. Paid calls answer 402 and the tools sign from your ${mode.chain === 'solana' ? 'Solana' : 'Base'} wallet.
|
|
269
|
+
- Not yours: \`${mode.chain === 'solana' ? 'https://blockrun.ai/api' : 'https://sol.blockrun.ai/api'}\` (the other chain's wallet gateway) and \`https://api.blockrun.ai\` (the account API, which needs a key and 401s without one — it is not an alias of either wallet host).`;
|
|
233
270
|
return `# BlockRun Gateway API (the network you live on)
|
|
234
271
|
You run on the BlockRun AI Gateway. When the user asks you to "test the BlockRun API", "check all endpoints", or call the gateway directly, use ONLY the paths below. **Never invent, pluralize, or singularize an endpoint** — \`/v1/image/generate\` (singular) is wrong, \`/v1/images/generations\` (plural) is correct. If a path you have in mind isn't in this list, fetch the canonical discovery endpoints before calling it.
|
|
235
272
|
|
|
236
273
|
**Base URLs**
|
|
237
|
-
|
|
238
|
-
-
|
|
274
|
+
${hosts}
|
|
275
|
+
Use the built-in tools so the right host and credential are applied. Never switch billing modes to work around an API error.
|
|
239
276
|
|
|
240
277
|
**Discovery (always free, GET) — fetch these BEFORE guessing a path**
|
|
241
278
|
- \`GET /openapi.json\` (or \`/.well-known/openapi.json\`) — full OpenAPI 3.1 contract, every route + request schema
|
|
242
279
|
- \`GET /.well-known/x402\` — x402 resource list with prices
|
|
243
280
|
|
|
244
|
-
**LLM (POST,
|
|
281
|
+
**LLM (POST, billed to ${chargedBy})**
|
|
245
282
|
- \`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, verified live 2026-08-30: \`anthropic/claude-sonnet-5\`, \`anthropic/claude-opus-5\`, \`openai/gpt-5.6-sol\`, \`deepseek/deepseek-v4-pro\`, \`zai/glm-5.3\`, \`zai/glm-5.3-flash\`, \`xai/grok-4.5\`, \`qwen/qwen3.7-flash\`, \`nvidia/nemotron-3-nano-omni-30b-a3b-reasoning\` (free, the only free id on BOTH the Base and Solana gateways — the other free models are Base-only)). 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
283
|
- \`POST /v1/messages\` — Anthropic-compatible. Body: \`{ model, messages, max_tokens, system?, tools? }\`.
|
|
247
284
|
|
|
248
|
-
**Media (POST,
|
|
285
|
+
**Media (POST, billed to ${chargedBy}; GET to poll async jobs)**
|
|
249
286
|
- \`POST /v1/images/generations\` — text-to-image. Body: \`{ model, prompt, size?, n?, response_format? }\`.
|
|
250
287
|
- \`POST /v1/images/image2image\` — image-to-image. Body: \`{ model, prompt, image, ... }\`.
|
|
251
288
|
- \`GET /v1/images/generations/{id}\` — fetch a generated image by id.
|
|
@@ -253,7 +290,7 @@ You run on the BlockRun AI Gateway. When the user asks you to "test the BlockRun
|
|
|
253
290
|
- \`GET /v1/videos/generations/{id}\` — poll video job (settles payment when complete).
|
|
254
291
|
- \`POST /v1/audio/generations\` — music/audio. Body: \`{ model, prompt, ... }\`. Default \`model\`: \`minimax/music-2.5+\`.
|
|
255
292
|
|
|
256
|
-
**Search (POST,
|
|
293
|
+
**Search (POST, billed to ${chargedBy})**
|
|
257
294
|
- \`POST /v1/search\` — Exa-backed web search. Body: \`{ query }\` (1–1000 chars).
|
|
258
295
|
- \`/v1/exa/{...path}\` — Exa passthrough (answer / search / contents).
|
|
259
296
|
|
|
@@ -452,7 +489,18 @@ const _instructionCache = new Map();
|
|
|
452
489
|
* Result is memoized per workingDir for the process lifetime.
|
|
453
490
|
*/
|
|
454
491
|
export function assembleInstructions(workingDir, model) {
|
|
455
|
-
|
|
492
|
+
// The pay mode is part of the key because getCoreInstructions() and
|
|
493
|
+
// getWalletKnowledgeSection() both branch on it. invalidateKey() can demote
|
|
494
|
+
// a session from key mode to wallet mode mid-process after a 401, and
|
|
495
|
+
// without this the model would keep being told it bills account credits
|
|
496
|
+
// long after Franklin went back to signing from the wallet.
|
|
497
|
+
// Keyed on the full billing identity, not just key-vs-wallet: the gateway
|
|
498
|
+
// host and the wallet guidance both name the active CHAIN, so a
|
|
499
|
+
// base -> solana switch inside one process must not be served the previous
|
|
500
|
+
// chain's instructions. resolvePayMode() is memoised, so this is cheap.
|
|
501
|
+
const payMode = resolvePayMode();
|
|
502
|
+
const billing = payMode.kind === 'key' ? 'key' : `wallet:${payMode.chain}`;
|
|
503
|
+
const cacheKey = model ? `${workingDir}::${model}::${billing}` : `${workingDir}::${billing}`;
|
|
456
504
|
const cached = _instructionCache.get(cacheKey);
|
|
457
505
|
if (cached)
|
|
458
506
|
return cached;
|
package/dist/agent/llm.js
CHANGED
|
@@ -658,7 +658,9 @@ export class ModelClient {
|
|
|
658
658
|
signal: requestController.signal,
|
|
659
659
|
}), requestController, createModelTimeoutError('request', request.model, requestTimeoutMs), requestTimeoutMs);
|
|
660
660
|
// Handle x402 payment
|
|
661
|
-
|
|
661
|
+
// Account 402 is a credit refusal. Only a request that started on
|
|
662
|
+
// the wallet rail may initiate a signing handshake.
|
|
663
|
+
if (response.status === 402 && !headers.Authorization) {
|
|
662
664
|
if (this.debug)
|
|
663
665
|
console.error('[franklin] Payment required — signing...');
|
|
664
666
|
const signedPayment = await this.signPayment(response, request.model);
|
|
@@ -1012,11 +1014,17 @@ export class ModelClient {
|
|
|
1012
1014
|
currentToolInput = '';
|
|
1013
1015
|
}
|
|
1014
1016
|
else if (currentThinking) {
|
|
1015
|
-
|
|
1016
|
-
|
|
1017
|
-
|
|
1018
|
-
|
|
1019
|
-
|
|
1017
|
+
if (currentThinkingSignature) {
|
|
1018
|
+
collected.push({
|
|
1019
|
+
type: 'thinking',
|
|
1020
|
+
thinking: currentThinking,
|
|
1021
|
+
signature: currentThinkingSignature,
|
|
1022
|
+
});
|
|
1023
|
+
}
|
|
1024
|
+
else if (this.debug) {
|
|
1025
|
+
console.error(`[franklin] Dropped unsigned native thinking block from ${request.model}; ` +
|
|
1026
|
+
'provider requires thinking.signature on replay.');
|
|
1027
|
+
}
|
|
1020
1028
|
currentThinking = '';
|
|
1021
1029
|
currentThinkingSignature = '';
|
|
1022
1030
|
}
|
package/dist/agent/loop.js
CHANGED
|
@@ -1494,6 +1494,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1494
1494
|
let callToolDefs = buildCallToolDefs();
|
|
1495
1495
|
let callMaxTokens = maxTokens;
|
|
1496
1496
|
let callSystemPrompt = systemPrompt;
|
|
1497
|
+
const toolCapFinalRound = toolCapWarned && turnToolCalls >= MAX_TOOL_CALLS_PER_TURN;
|
|
1497
1498
|
if (planActive && loopCount === 1) {
|
|
1498
1499
|
callToolDefs = []; // No tools during planning
|
|
1499
1500
|
callMaxTokens = 2048; // Short plan output
|
|
@@ -1526,6 +1527,13 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1526
1527
|
'\n\nThe image is already included above. Answer the question about it ' +
|
|
1527
1528
|
'directly. Do not call any tools.';
|
|
1528
1529
|
}
|
|
1530
|
+
if (toolCapFinalRound && callToolDefs.length > 0) {
|
|
1531
|
+
callToolDefs = [];
|
|
1532
|
+
callSystemPrompt = callSystemPrompt + '\n\n# Tool budget exhausted\n' +
|
|
1533
|
+
`You already used ${turnToolCalls} tool calls this turn and received a system stop signal. ` +
|
|
1534
|
+
'No more tools are available. Based only on the results already in the conversation, ' +
|
|
1535
|
+
'give the user a concise status/update now. If the work is incomplete, say exactly what remains.';
|
|
1536
|
+
}
|
|
1529
1537
|
const onFinalTurn = config.forceAnswerOnFinalTurn && loopCount === maxTurns;
|
|
1530
1538
|
const toolBudgetSpent = config.maxToolCalls != null && turnToolCalls >= config.maxToolCalls;
|
|
1531
1539
|
if ((onFinalTurn || toolBudgetSpent) && callToolDefs.length > 0) {
|
|
@@ -1568,7 +1576,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1568
1576
|
// Consume any pending forced tool_choice from the previous round's
|
|
1569
1577
|
// grounding-retry decision. `tool_choice` is dropped automatically in
|
|
1570
1578
|
// llm.ts if `tools` ended up empty, so it's safe to attach here.
|
|
1571
|
-
const callToolChoice = forceToolChoiceNextRound;
|
|
1579
|
+
const callToolChoice = toolCapFinalRound ? null : forceToolChoiceNextRound;
|
|
1572
1580
|
forceToolChoiceNextRound = null;
|
|
1573
1581
|
// Wall-clock start of the model call. Used by the recordUsage call
|
|
1574
1582
|
// a few hundred lines below so franklin-stats.json captures real
|
|
@@ -1627,11 +1635,16 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1627
1635
|
.filter((p) => p.type === 'text')
|
|
1628
1636
|
.map(p => p.text)
|
|
1629
1637
|
.join('\n');
|
|
1630
|
-
|
|
1631
|
-
|
|
1632
|
-
|
|
1633
|
-
|
|
1634
|
-
|
|
1638
|
+
if (callToolDefs.length > 0) {
|
|
1639
|
+
const repaired = callRepair.process(declaredCalls, reasoningText || null, contentText || null);
|
|
1640
|
+
if (repaired.report.scavenged > 0) {
|
|
1641
|
+
const novelCalls = repaired.calls.slice(declaredCalls.length);
|
|
1642
|
+
responseParts = [...responseParts, ...novelCalls];
|
|
1643
|
+
logger.warn(`[franklin] scavenged ${repaired.report.scavenged} leaked tool call(s) from ${config.model}: ${repaired.report.notes.join('; ')}`);
|
|
1644
|
+
}
|
|
1645
|
+
}
|
|
1646
|
+
else if (declaredCalls.length > 0) {
|
|
1647
|
+
logger.warn(`[franklin] Dropping ${declaredCalls.length} tool call(s) because tools are disabled for this round`);
|
|
1635
1648
|
}
|
|
1636
1649
|
// Strip pure roleplayed-JSON tool-call text parts. The LLM client
|
|
1637
1650
|
// holds these out of the live stream (never shown) but keeps them
|
|
@@ -1641,6 +1654,9 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1641
1654
|
// also empties the turn, so the empty-response recovery below can
|
|
1642
1655
|
// switch models, preserving the prior behavior.
|
|
1643
1656
|
responseParts = responseParts.filter(p => !(p.type === 'text' && isRoleplayedJsonToolCallText(p.text)));
|
|
1657
|
+
if (callToolDefs.length === 0) {
|
|
1658
|
+
responseParts = responseParts.filter(p => p.type !== 'tool_use');
|
|
1659
|
+
}
|
|
1644
1660
|
}
|
|
1645
1661
|
// ── Empty response recovery ──
|
|
1646
1662
|
// If the model returns nothing, DON'T just retry the same model with the same input.
|
|
@@ -1648,6 +1664,18 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
|
|
|
1648
1664
|
const hasText = responseParts.some(p => p.type === 'text' && p.text?.trim());
|
|
1649
1665
|
const hasTools = responseParts.some(p => p.type === 'tool_use');
|
|
1650
1666
|
const hasThinking = responseParts.some(p => p.type === 'thinking');
|
|
1667
|
+
if (toolCapFinalRound && !hasText) {
|
|
1668
|
+
const spendNote = turnCostUsd > 0
|
|
1669
|
+
? `${turnToolCalls} tool calls, $${turnCostUsd.toFixed(4)} spent this turn`
|
|
1670
|
+
: `${turnToolCalls} tool calls this turn`;
|
|
1671
|
+
logger.error(`[franklin] Tool cap final-answer round produced no text after tools were disabled — ending turn`);
|
|
1672
|
+
onEvent({
|
|
1673
|
+
kind: 'text_delta',
|
|
1674
|
+
text: `\n\n⚠️ Tool loop stopped: ${spendNote}. Franklin disabled tools after the cap warning, but the model did not produce a final status. Try rephrasing or use \`/model\` to switch.\n`,
|
|
1675
|
+
});
|
|
1676
|
+
onEvent({ kind: 'turn_done', reason: 'cap_exceeded' });
|
|
1677
|
+
break;
|
|
1678
|
+
}
|
|
1651
1679
|
if (!hasText && !hasTools && !hasThinking) {
|
|
1652
1680
|
// Free-only recovery chain — a free/empty-response session must NEVER
|
|
1653
1681
|
// fall back to a paid model (would silently charge the wallet). Both
|
package/dist/agent/optimize.d.ts
CHANGED
|
@@ -3,7 +3,7 @@
|
|
|
3
3
|
*
|
|
4
4
|
* Five layers of optimization to minimize token usage:
|
|
5
5
|
* 1. Tool result size budgeting — cap large outputs, keep preview
|
|
6
|
-
* 2. Thinking block stripping — remove old thinking from history
|
|
6
|
+
* 2. Thinking block stripping — remove old/invalid thinking from history
|
|
7
7
|
* 3. Time-based cleanup — clear stale tool results after idle gap
|
|
8
8
|
* 4. Adaptive max_tokens — start low (8K), escalate on hit
|
|
9
9
|
* 5. Pre-compact stripping — remove images/docs before summarization
|
package/dist/agent/optimize.js
CHANGED
|
@@ -3,7 +3,7 @@
|
|
|
3
3
|
*
|
|
4
4
|
* Five layers of optimization to minimize token usage:
|
|
5
5
|
* 1. Tool result size budgeting — cap large outputs, keep preview
|
|
6
|
-
* 2. Thinking block stripping — remove old thinking from history
|
|
6
|
+
* 2. Thinking block stripping — remove old/invalid thinking from history
|
|
7
7
|
* 3. Time-based cleanup — clear stale tool results after idle gap
|
|
8
8
|
* 4. Adaptive max_tokens — start low (8K), escalate on hit
|
|
9
9
|
* 5. Pre-compact stripping — remove images/docs before summarization
|
|
@@ -204,9 +204,14 @@ export function budgetToolResults(history) {
|
|
|
204
204
|
}
|
|
205
205
|
// ─── 2. Thinking Block Stripping ───────────────────────────────────────────
|
|
206
206
|
/**
|
|
207
|
-
* Remove thinking blocks from older assistant messages
|
|
208
|
-
*
|
|
209
|
-
*
|
|
207
|
+
* Remove thinking blocks from older assistant messages and any unsigned native
|
|
208
|
+
* thinking blocks. Anthropic-compatible APIs require `thinking.signature` when
|
|
209
|
+
* a prior assistant thinking block is replayed; an unsigned block poisons the
|
|
210
|
+
* next request with a 400, so it must never survive into outbound history.
|
|
211
|
+
*
|
|
212
|
+
* Keeps only signed thinking in the most recent N assistant messages (default:
|
|
213
|
+
* last 2 turns). Older thinking blocks are large and not needed after the
|
|
214
|
+
* decision is made.
|
|
210
215
|
*/
|
|
211
216
|
const KEEP_THINKING_TURNS = 2;
|
|
212
217
|
export function stripOldThinking(history) {
|
|
@@ -226,9 +231,14 @@ export function stripOldThinking(history) {
|
|
|
226
231
|
let modified = false;
|
|
227
232
|
for (let i = 0; i < history.length; i++) {
|
|
228
233
|
const msg = history[i];
|
|
229
|
-
// Strip thinking
|
|
230
|
-
if (msg.role === 'assistant' &&
|
|
231
|
-
const filtered = msg.content.filter((part) =>
|
|
234
|
+
// Strip old thinking, plus unsigned thinking even on recent turns.
|
|
235
|
+
if (msg.role === 'assistant' && Array.isArray(msg.content)) {
|
|
236
|
+
const filtered = msg.content.filter((part) => {
|
|
237
|
+
if (part.type !== 'thinking')
|
|
238
|
+
return true;
|
|
239
|
+
const signature = part.signature;
|
|
240
|
+
return keepSet.has(i) && typeof signature === 'string' && signature.trim().length > 0;
|
|
241
|
+
});
|
|
232
242
|
if (filtered.length < msg.content.length) {
|
|
233
243
|
modified = true;
|
|
234
244
|
result.push({
|
|
@@ -62,6 +62,32 @@ export declare function redactSecrets(input: string): RedactionResult;
|
|
|
62
62
|
* Returns the names of env vars that were set, deduped, in stash order.
|
|
63
63
|
* Safe to call on an empty match list (no-op).
|
|
64
64
|
*/
|
|
65
|
+
/**
|
|
66
|
+
* Redact secrets from text Franklin did not author — tool output, not user
|
|
67
|
+
* input.
|
|
68
|
+
*
|
|
69
|
+
* `redactSecrets` above is the INPUT path: it scrubs what a user pastes into
|
|
70
|
+
* chat and hands the caller the raw values to stash on process.env. That is
|
|
71
|
+
* the wrong contract here. Tool output flows straight into the model prompt
|
|
72
|
+
* (over the network, to whichever gateway model serves the turn) and into the
|
|
73
|
+
* persisted session transcript on disk, so the value must not survive at all
|
|
74
|
+
* and there is nothing to stash.
|
|
75
|
+
*
|
|
76
|
+
* The case that forced this: `printenv` and `echo` are auto-approved by
|
|
77
|
+
* bash-guard as safe commands, so `printenv BLOCKRUN_API_KEY` needed no
|
|
78
|
+
* confirmation and returned the account bearer key as tool output. Stripping
|
|
79
|
+
* the variable from the Bash subprocess env (tools/bash.ts) closes that exact
|
|
80
|
+
* command; this closes the class — a key can still reach output via a config
|
|
81
|
+
* file, a `curl -v` trace, an MCP result, or a shell history dump.
|
|
82
|
+
*
|
|
83
|
+
* `extraLiterals` carries values known only at runtime, e.g. the configured
|
|
84
|
+
* API key when it does not match a published prefix shape. Empty and very
|
|
85
|
+
* short entries are ignored so a blank env var cannot blank out the text.
|
|
86
|
+
*/
|
|
87
|
+
export declare function redactSecretsInOutput(input: string, extraLiterals?: readonly string[]): {
|
|
88
|
+
text: string;
|
|
89
|
+
labels: string[];
|
|
90
|
+
};
|
|
65
91
|
export declare function stashSecretsToEnv(matches: RedactionMatch[]): string[];
|
|
66
92
|
/**
|
|
67
93
|
* Build a one-paragraph warning + usage hint to surface to the user when
|
|
@@ -216,6 +216,54 @@ export function redactSecrets(input) {
|
|
|
216
216
|
* Returns the names of env vars that were set, deduped, in stash order.
|
|
217
217
|
* Safe to call on an empty match list (no-op).
|
|
218
218
|
*/
|
|
219
|
+
/**
|
|
220
|
+
* Redact secrets from text Franklin did not author — tool output, not user
|
|
221
|
+
* input.
|
|
222
|
+
*
|
|
223
|
+
* `redactSecrets` above is the INPUT path: it scrubs what a user pastes into
|
|
224
|
+
* chat and hands the caller the raw values to stash on process.env. That is
|
|
225
|
+
* the wrong contract here. Tool output flows straight into the model prompt
|
|
226
|
+
* (over the network, to whichever gateway model serves the turn) and into the
|
|
227
|
+
* persisted session transcript on disk, so the value must not survive at all
|
|
228
|
+
* and there is nothing to stash.
|
|
229
|
+
*
|
|
230
|
+
* The case that forced this: `printenv` and `echo` are auto-approved by
|
|
231
|
+
* bash-guard as safe commands, so `printenv BLOCKRUN_API_KEY` needed no
|
|
232
|
+
* confirmation and returned the account bearer key as tool output. Stripping
|
|
233
|
+
* the variable from the Bash subprocess env (tools/bash.ts) closes that exact
|
|
234
|
+
* command; this closes the class — a key can still reach output via a config
|
|
235
|
+
* file, a `curl -v` trace, an MCP result, or a shell history dump.
|
|
236
|
+
*
|
|
237
|
+
* `extraLiterals` carries values known only at runtime, e.g. the configured
|
|
238
|
+
* API key when it does not match a published prefix shape. Empty and very
|
|
239
|
+
* short entries are ignored so a blank env var cannot blank out the text.
|
|
240
|
+
*/
|
|
241
|
+
export function redactSecretsInOutput(input, extraLiterals = []) {
|
|
242
|
+
if (!input)
|
|
243
|
+
return { text: input, labels: [] };
|
|
244
|
+
let text = input;
|
|
245
|
+
const labels = [];
|
|
246
|
+
for (const literal of extraLiterals) {
|
|
247
|
+
const value = literal?.trim();
|
|
248
|
+
// 8 is below any credential this guards and above the length at which a
|
|
249
|
+
// stray value would collide with ordinary output.
|
|
250
|
+
if (!value || value.length < 8)
|
|
251
|
+
continue;
|
|
252
|
+
if (!text.includes(value))
|
|
253
|
+
continue;
|
|
254
|
+
text = text.split(value).join('[REDACTED:configured_credential]');
|
|
255
|
+
labels.push('configured_credential');
|
|
256
|
+
}
|
|
257
|
+
for (const { label, pattern } of SECRET_PATTERNS) {
|
|
258
|
+
pattern.lastIndex = 0;
|
|
259
|
+
if (!pattern.test(text))
|
|
260
|
+
continue;
|
|
261
|
+
pattern.lastIndex = 0;
|
|
262
|
+
text = text.replace(pattern, `[REDACTED:${label}]`);
|
|
263
|
+
labels.push(label);
|
|
264
|
+
}
|
|
265
|
+
return { text, labels: [...new Set(labels)] };
|
|
266
|
+
}
|
|
219
267
|
export function stashSecretsToEnv(matches) {
|
|
220
268
|
const set = [];
|
|
221
269
|
for (const m of matches) {
|
|
@@ -8,6 +8,8 @@ import { join } from 'node:path';
|
|
|
8
8
|
import { recordFailure } from '../stats/failures.js';
|
|
9
9
|
import { estimateSpendUsd, isSpendTool } from '../tools/spend-tools.js';
|
|
10
10
|
import { BLOCKRUN_DIR } from '../config.js';
|
|
11
|
+
import { redactSecretsInOutput } from './secret-redact.js';
|
|
12
|
+
import { resolvePayMode } from '../payments/auth-mode.js';
|
|
11
13
|
import { logger } from '../logger.js';
|
|
12
14
|
/** Persist a large tool result to disk and return a preview string. */
|
|
13
15
|
const PERSIST_THRESHOLD = 50_000;
|
|
@@ -287,6 +289,19 @@ export class StreamingExecutor {
|
|
|
287
289
|
const preview = (this.inputPreview(invocation) || '').replace(/[\r\n]+/g, ' ');
|
|
288
290
|
logger.info(`[franklin] Slow tool: ${invocation.name} ${status} after ${(execElapsed / 1000).toFixed(1)}s${preview ? ` — ${preview.slice(0, 80)}` : ''}`);
|
|
289
291
|
}
|
|
292
|
+
// Scrub credentials before the output goes anywhere. This is the one
|
|
293
|
+
// funnel every tool result passes through, and it runs BEFORE the
|
|
294
|
+
// persist-to-disk branch below so a redacted body is what reaches the
|
|
295
|
+
// model, the transcript, and the on-disk overflow file alike.
|
|
296
|
+
//
|
|
297
|
+
// resolvePayMode() is memoised per process, so reading the configured
|
|
298
|
+
// key here costs nothing per tool call.
|
|
299
|
+
const payMode = resolvePayMode();
|
|
300
|
+
const redacted = redactSecretsInOutput(result.output, payMode.kind === 'key' ? [payMode.key] : []);
|
|
301
|
+
if (redacted.labels.length > 0) {
|
|
302
|
+
logger.warn(`[franklin] Redacted ${redacted.labels.join(', ')} from ${invocation.name} output`);
|
|
303
|
+
result = { ...result, output: redacted.text };
|
|
304
|
+
}
|
|
290
305
|
// Persist large results to disk with preview.
|
|
291
306
|
// Instead of just truncating, save the full result to disk so it can be re-read later.
|
|
292
307
|
if (result.output.length > PERSIST_THRESHOLD) {
|
package/dist/commands/doctor.js
CHANGED
|
@@ -287,6 +287,24 @@ async function runChecks() {
|
|
|
287
287
|
remedy: 'Add /opt/homebrew/bin to PATH in ~/.zshrc',
|
|
288
288
|
});
|
|
289
289
|
}
|
|
290
|
+
// ── Solana wallet divergence (blockrun#119) ───────────────────────
|
|
291
|
+
// A tightened SDK selection can leave a user spending from a different
|
|
292
|
+
// address than the one holding their USDC, with nothing printed. Report it
|
|
293
|
+
// where they already look when money seems missing.
|
|
294
|
+
try {
|
|
295
|
+
const { detectSolanaWalletDivergence } = await import('../wallet/solana-migration.js');
|
|
296
|
+
const divergence = await detectSolanaWalletDivergence();
|
|
297
|
+
if (divergence) {
|
|
298
|
+
out.push({
|
|
299
|
+
name: 'Solana wallet',
|
|
300
|
+
status: 'warn',
|
|
301
|
+
detail: `active ${divergence.active}; ${divergence.alternatives.length} other wallet(s) found: ` +
|
|
302
|
+
divergence.alternatives.map((w) => w.address).join(', '),
|
|
303
|
+
remedy: 'If your USDC is on one of those: franklin wallet-adopt <address>',
|
|
304
|
+
});
|
|
305
|
+
}
|
|
306
|
+
}
|
|
307
|
+
catch { /* a diagnostic must never take down `franklin doctor` */ }
|
|
290
308
|
return out;
|
|
291
309
|
}
|
|
292
310
|
function printHuman(checks) {
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* `franklin usage` — what the account actually charged, joined to what
|
|
3
|
+
* Franklin recorded.
|
|
4
|
+
*
|
|
5
|
+
* `franklin stats` reports Franklin's own tally, which in key mode is built
|
|
6
|
+
* from catalog prices because the account gateway settles without returning a
|
|
7
|
+
* charge. This command reports the gateway's ledger, which is authoritative,
|
|
8
|
+
* and lines the two up per request rather than comparing totals — a total that
|
|
9
|
+
* matches can still be two errors cancelling out.
|
|
10
|
+
*/
|
|
11
|
+
export declare function usageCommand(opts?: {
|
|
12
|
+
days?: string;
|
|
13
|
+
json?: boolean;
|
|
14
|
+
}): Promise<void>;
|