Conversation
Route every AI feature through the BTL Runtime (an OpenAI-compatible gateway) and add a tool-calling agent that simulates real transactions. - Repoint both LLM proxies (api/llm-recommend.ts, vite dev plugin) from Gemini to BTL: inject the Bearer key server-side and forward the per-request cost headers to the browser. - Shared BTL client (src/lib/btl): OpenAI-shape request/response + cost-meta parsing, a non-streaming tool-calling loop (runBtlAgent), and a freeform "explain" hook. - Concierge (LI.FI Earn): intent parsing + vault ranking on BTL, a live per-call cost chip, and a one-click OpenAI/DeepSeek provider toggle. - Deposit preflight: a tool-calling agent runs the real REVM simulator and narrates the deposit. - Four explainer weaves: revert explainer, decoded-calldata explainer, contract-upgrade auditor (verified-source diff to a risk table), and a storage-slot annotator (structured JSON to hoverable chips). - Cost-transparency UI: AiCostChip, BtlRuntimePanel, a foldable BtlExplanation panel, and a "Powered by BTL" badge across surfaces. The rules-based fallback and the LLM_MODE=off kill switch keep the app crash-proof.
The latest updates on your projects. Learn more about Vercel for GitHub.
|
There was a problem hiding this comment.
Pull request overview
This PR switches HexKit’s AI backbone from Gemini to the BTL Runtime (OpenAI-compatible gateway), adds a shared BTL client + tool-calling agent loop, and surfaces cost/route transparency + “AI explain” panels across several UX surfaces (concierge, simulation revert, call decode, storage slots, and contract upgrade diff).
Changes:
- Repoint
/api/llm-recommend(Vercel + Vite dev proxy) to BTL chat completions, forwarding per-call cost/routing headers to the browser. - Introduce shared BTL utilities (
src/lib/btl/*) including a generic “explain” hook and a tool-calling agent loop used to preflight deposits via real REVM simulation. - Add UI components for explanations and cost transparency (
BtlExplanation,AiCostChip,BtlRuntimePanel,BtlBadge) and integrate them into multiple features.
Reviewed changes
Copilot reviewed 29 out of 30 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| vite.config.ts | Dev-server proxy now forwards BTL chat completions + cost headers; expands watch ignores. |
| src/lib/btl/useBtlExplain.ts | Generic “explain” hook wrapping BTL calls + meta. |
| src/lib/btl/models.ts | BTL model registry + defaults and AB toggle list. |
| src/lib/btl/client.ts | Shared request builder, text extraction, meta parsing, JSON parsing helpers. |
| src/lib/btl/agent.ts | Tool-calling loop (runBtlAgent) for non-streaming chat/tool execution. |
| src/components/smart-decoder/SmartDecoder.tsx | Adds “What does this call do?” AI explanation panel for decoded calldata. |
| src/components/simulation-results/SummaryTab.tsx | Adds “Explain this revert” AI explanation panel for reverted simulations. |
| src/components/integrations/lifi-earn/earnApi.ts | postLlmRecommend now returns { data, meta } and parses BTL headers. |
| src/components/integrations/lifi-earn/DepositFlow.tsx | Adds AI preflight that calls REVM simulation via tool-calling agent; extracts tx builder. |
| src/components/integrations/lifi-earn/concierge/VaultRecommendations.tsx | Shows per-recommendation cost chip for AI-sourced recommendations. |
| src/components/integrations/lifi-earn/concierge/types.ts | Adds optional meta to recommendations for cost/routing. |
| src/components/integrations/lifi-earn/concierge/LlmErrorAlert.tsx | Updates error copy from Gemini to BTL. |
| src/components/integrations/lifi-earn/concierge/intent/IntentPanel.tsx | Adds per-parse cost chip + model A/B toggle + cache key fingerprinting. |
| src/components/integrations/lifi-earn/concierge/intent/hooks/useIntentRecommendation.ts | Migrates intent recommendation request/parse to OpenAI-style responses + stores meta. |
| src/components/integrations/lifi-earn/concierge/intent/hooks/useIntentParser.ts | Migrates NL intent parsing to OpenAI-style responses + returns meta. |
| src/components/integrations/lifi-earn/concierge/IdleSweepPanel.tsx | Adds session-level BTL receipt panel and “Powered by BTL” badge gating. |
| src/components/integrations/lifi-earn/concierge/hooks/useVaultRecommendations.ts | Migrates idle-sweep recommendations to OpenAI-style responses + stores meta. |
| src/components/integrations/lifi-earn/concierge/BtlRuntimePanel.tsx | New aggregated “receipt” UI for total BTL cost across calls. |
| src/components/integrations/lifi-earn/buildDepositTx.ts | Extracts quote→tx mapping used by both manual sim and AI tool. |
| src/components/explorer/StorageLayoutViewer.tsx | Adds AI storage-slot annotation flow returning structured JSON rendered as chips. |
| src/components/explorer/ContractDiff.tsx | Adds AI “upgrade audit” panel; includes verified-source snippets when available. |
| src/components/BtlBadge.tsx | New “Powered by BTL” badge component + external link. |
| src/components/BtlBadge.css | Styling for the badge. |
| src/components/btl/ThinkingIndicator.tsx | New animated “thinking” indicator shared by explanation panels. |
| src/components/btl/SlotAnnotationChips.tsx | New tooltip chip UI for per-slot annotations + anomaly highlighting. |
| src/components/btl/BtlExplanation.tsx | New shared collapsible explanation panel + lightweight markdown renderer + footer badge. |
| src/components/btl/AiCostChip.tsx | New per-call cost chip UI rendered from forwarded BTL headers. |
| README.md | Documentation updated from Gemini → BTL Runtime env/config. |
| public/logos/btl-runtime.svg | Adds BTL Runtime logo asset used by the badge. |
| api/llm-recommend.ts | Serverless proxy now targets BTL chat completions and forwards cost/routing headers. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+72
to
79
| if (!Array.isArray((body as any).messages)) { | ||
| return res.status(400).json({ error: "Body must include `messages` array" }); | ||
| } | ||
| const serialized = JSON.stringify(body); | ||
| if (serialized.length > MAX_BODY_BYTES) { | ||
| return res.status(413).json({ error: "Request body too large" }); | ||
| } |
Comment on lines
+26
to
+32
| function buildSourceSnippet(ctx: ContractContext): string { | ||
| if (!ctx.metadata?.sources) return ''; | ||
| return Object.entries(ctx.metadata.sources) | ||
| .map(([path, content]) => `// ${path}\n${content}`) | ||
| .join('\n\n') | ||
| .slice(0, SOURCE_CHAR_CAP); | ||
| } |
Comment on lines
+340
to
346
| const userText = | ||
| "INPUT:\n```json\n" + | ||
| JSON.stringify(userPayload, null, 2) + | ||
| "\n```\n\nReturn ONLY the JSON object matching required_output_shape. No prose, no code fences."; | ||
| function extractGeminiText(raw: unknown): string | null { | ||
| // Gemini 3 Pro can return multi-part content with `thought: true` parts | ||
| // before the answer — concatenate every non-thought text part. | ||
| try { | ||
| const r = raw as { | ||
| candidates?: Array<{ | ||
| content?: { | ||
| parts?: Array<{ text?: string; thought?: boolean }>; | ||
| }; | ||
| }>; | ||
| }; | ||
| const parts = r.candidates?.[0]?.content?.parts ?? []; | ||
| const joined = parts | ||
| .filter((p) => !p.thought && typeof p.text === "string") | ||
| .map((p) => p.text ?? "") | ||
| .join("") | ||
| .trim(); | ||
| return joined.length > 0 ? joined : null; | ||
| } catch { | ||
| return null; | ||
| } | ||
| } | ||
| function safeParseJson(text: string): unknown { | ||
| try { | ||
| return JSON.parse(text); | ||
| } catch { | ||
| // Strip common noise: code fences, leading commentary | ||
| const stripped = text | ||
| .replace(/^```(?:json)?/i, "") | ||
| .replace(/```$/i, "") | ||
| .trim(); | ||
| try { | ||
| return JSON.parse(stripped); | ||
| } catch { | ||
| // Thinking models sometimes prepend prose before the JSON object. | ||
| // Find the first `{` and last `}` and try parsing that substring. | ||
| const first = stripped.indexOf("{"); | ||
| const last = stripped.lastIndexOf("}"); | ||
| if (first >= 0 && last > first) { | ||
| try { | ||
| return JSON.parse(stripped.slice(first, last + 1)); | ||
| } catch { | ||
| /* fall through */ | ||
| } | ||
| } | ||
| return null; | ||
| } | ||
| } | ||
| return buildBtlChatRequest(system, userText, { temperature: 0.2 }); | ||
| } |
Comment on lines
+154
to
160
| const userText = | ||
| "INPUT:\n```json\n" + | ||
| JSON.stringify(payload, null, 2) + | ||
| "\n```\n\nReturn ONLY the JSON object."; | ||
| function safeParseJson(text: string): unknown { | ||
| try { | ||
| return JSON.parse(text); | ||
| } catch { | ||
| const stripped = text | ||
| .replace(/^```(?:json)?/i, "") | ||
| .replace(/```$/i, "") | ||
| .trim(); | ||
| try { | ||
| return JSON.parse(stripped); | ||
| } catch { | ||
| const first = stripped.indexOf("{"); | ||
| const last = stripped.lastIndexOf("}"); | ||
| if (first >= 0 && last > first) { | ||
| try { | ||
| return JSON.parse(stripped.slice(first, last + 1)); | ||
| } catch { | ||
| /* fall through */ | ||
| } | ||
| } | ||
| return null; | ||
| } | ||
| } | ||
| return buildBtlChatRequest(system, userText, { temperature: 0.1 }); | ||
| } |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit:b4a05a8069
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const toolRuns: Array<{ name: string; args: unknown; result: unknown }> = []; | ||
| for (let i = 0; i < maxIters; i++) { | ||
| const { data, meta } = await callBtl({ model, messages, tools: toolSpecs, tool_choice: "auto", temperature: 0.2 }); |
There was a problem hiding this comment.
When the model decides to answer without tools, which is allowed by tool_choice: "auto", runBtlAgent returns that text immediately and DepositFlow.handlePreflight displays it even if toolRuns is empty. In that scenario the “Preflight with AI” path can present a deposit assessment that never ran simulate_deposit/REVM, despite the feature relying on simulation-backed numbers; force the required tool for this preflight or reject final text until the tool has run.
Useful? React with 👍 / 👎.
| opts: { model?: string; temperature?: number; jsonMode?: boolean; tools?: unknown[]; toolChoice?: unknown; maxTokens?: number } = {}, | ||
| ): Record<string, unknown> { | ||
| const body: Record<string, unknown> = { | ||
| model: opts.model ?? BTL_DEFAULT_MODEL, |
There was a problem hiding this comment.
For normal app calls opts.model is absent, so this hard-codes deepseek-v4-flash into the browser request body; the server proxy then forwards that body unchanged, meaning the documented BTL_MODEL environment variable cannot change the deployed default model. Deployments that set BTL_MODEL to an approved or fallback model will still send the bundled default unless the client omits model or the proxy overwrites/allowlists it server-side.
Useful? React with 👍 / 👎.
Comment on lines
+42
to
+55
| // TEMP debug probe — reports env visibility without leaking the key. Remove after diagnosis. | ||
| if (req.query?.debug === "1") { | ||
| const k = process.env.BTL_API_KEY || ""; | ||
| return res.status(200).json({ | ||
| vercelEnv: process.env.VERCEL_ENV || null, | ||
| gitBranch: process.env.VERCEL_GIT_COMMIT_REF || null, | ||
| commit: process.env.VERCEL_GIT_COMMIT_SHA?.slice(0, 7) || null, | ||
| hasKey: k.length > 0, | ||
| keyLen: k.length, | ||
| keyPrefix: k ? k.slice(0, 3) : null, | ||
| baseUrl: process.env.BTL_BASE_URL || "(default)", | ||
| btlEnvNames: Object.keys(process.env).filter((n) => n.toUpperCase().includes("BTL")), | ||
| }); | ||
| } |
Comment on lines
+677
to
+679
| setPreflightText(finalText); | ||
| setPreflightMeta(metas.at(-1) ?? null); | ||
| setPreflightRan(toolRuns.some((r) => r.name === "simulate_deposit")); |
Comment on lines
+51
to
+58
| const slots = json.slots | ||
| .filter((s): s is Record<string, unknown> => !!s && typeof s === 'object' && typeof (s as { note?: unknown }).note === 'string') | ||
| .map((s) => ({ | ||
| slot: (s.slot as string) ?? null, | ||
| label: (s.label as string) ?? null, | ||
| note: s.note as string, | ||
| unusual: !!s.unusual, | ||
| })); |
| argsList.map((a) => | ||
| a | ||
| ? `${a.synthChainId}:${a.synthTokenAddress}:${a.sourceTokenSymbol ?? ""}:${a.rankedVaults.slice(0, 8).map((v) => v.slug).join(",")}` | ||
| ? `${a.synthChainId}:${a.synthTokenAddress}:${a.sourceTokenSymbol ?? ""}:${a.model ?? ""}:${intentFingerprint(a.intent)}:${a.rankedVaults.map((v) => v.slug).join(",")}` |
Comment on lines
+1
to
+4
| import { useCallback, useState } from "react"; | ||
| import { buildBtlChatRequest, extractOpenAiText, type BtlRuntimeMeta } from "./client"; | ||
| import { postLlmRecommend } from "@/components/integrations/lifi-earn/earnApi"; | ||
Comment on lines
+150
to
151
| const request = buildGeminiIntentRequest(intent, candidates, args.walletAssets, args.sourceTokenSymbol, args.sourceChainId, args.model); | ||
| let lastError: string | null = null; |
Comment on lines
76
to
79
| const serialized = JSON.stringify(body); | ||
| if (serialized.length > MAX_BODY_BYTES) { | ||
| return res.status(413).json({ error: "Request body too large" }); | ||
| } |
Comment on lines
+30
to
+34
| try { result = tool ? await tool.run(JSON.parse(c.function.arguments || "{}")) : { error: "unknown tool" }; } | ||
| catch (e: any) { result = { error: e?.message ?? "tool failed" }; } | ||
| toolRuns.push({ name: c.function.name, args: c.function.arguments, result }); | ||
| messages.push({ role: "tool", tool_call_id: c.id, content: JSON.stringify(result) }); | ||
| } |
Comment on lines
+173
to
+175
| | `BTL_API_KEY` | BTL Runtime API key for the yield concierge LLM | | ||
| | `BTL_MODEL` | BTL model (default: `deepseek-v4-flash`) | | ||
| | `BTL_BASE_URL` | BTL Runtime base URL (default: `https://api.badtheorylabs.com`) | |
Comment on lines
262
to
265
| argsList.map((a) => | ||
| a | ||
| ? `${a.synthChainId}:${a.synthTokenAddress}:${a.sourceTokenSymbol ?? ""}:${a.rankedVaults.slice(0, 8).map((v) => v.slug).join(",")}` | ||
| ? `${a.synthChainId}:${a.synthTokenAddress}:${a.sourceTokenSymbol ?? ""}:${a.model ?? ""}:${intentFingerprint(a.intent)}:${a.rankedVaults.map((v) => v.slug).join(",")}` | ||
| : "null", |
Comment on lines
67
to
71
| function buildParseRequest( | ||
| userText: string, | ||
| rawUserText: string, | ||
| chains: EarnChainInfo[], | ||
| protocols: EarnProtocolInfo[] | ||
| ) { |
Comment on lines
+184
to
187
| export async function postLlmRecommend(body: unknown): Promise<{ data: unknown; meta: BtlRuntimeMeta }> { | ||
| const res = await fetch(LLM_PROXY, { | ||
| method: "POST", | ||
| headers: { "Content-Type": "application/json", ...proxyHeaders() }, |
| } | ||
| const request = buildGeminiIntentRequest(intent, candidates, args.walletAssets, args.sourceTokenSymbol, args.sourceChainId); | ||
| const request = buildGeminiIntentRequest(intent, candidates, args.walletAssets, args.sourceTokenSymbol, args.sourceChainId, args.model); |
| @@ -204,11 +206,12 @@ async function fetchRecommendationForAsset( | |||
| const request = buildGeminiRequest([source], candidateMap); | |||
…lls 502/empty-response flakiness
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integrates the BTL Runtime (an OpenAI-compatible gateway) as HexKit's AI backbone, and adds a tool-calling agent that simulates real transactions.
What's in it
src/lib/btl) + a non-streaming tool-calling loop (runBtlAgent) that runs the real REVM simulator on a deposit and narrates it.AiCostChip,BtlRuntimePanel, a foldableBtlExplanationpanel, and a "Powered by BTL" badge across surfaces.The rules-based fallback and the
LLM_MODE=offkill switch keep the app crash-proof. Build is green and the BTL client is unit-tested.To deploy: set
BTL_API_KEY(Preview scope); leavePROXY_SECRETunset. Existing LI.FI / Etherscan / EDB env covers the surfaces the AI runs on.