feat(example): Week 2 agent flow - build/typecheck unblock + deploy-pending UX + graceful errors#25
Merged
Merged
Conversation
TypeScript 6 (merged in #13) regressed the example app typecheck, which CI does not cover (CI builds packages only). Two errors blocked next build: - TS2882: side-effect imports of @txkit/themes/base, @txkit/themes/dark and globals.css need ambient module declarations. Added globals.d.ts. - TS2769: the as-const tool definitions typed required as a readonly tuple, not assignable to the Anthropic SDK Tool input_schema required string[]. Switched to satisfies Anthropic.Tool, which keeps literal types without readonly.
Before deploy, deployed.json holds placeholder addresses and preparing an envelope returns a 503; the flow only surfaced that reactively after a visitor typed a prompt. DeployPendingBanner reads the checkIsAgentPolicyGateDeployed / checkIsMockPendleRouterDeployed predicates and shows a preview-mode banner up front, rendering nothing once both contracts hold real addresses.
Bump the agent model from claude-opus-4-5 to claude-opus-4-8 (current default; same request surface, none of the removed params are in use). The system prompt told the model to summarise after calling the tool, but the route is single-shot - it never feeds a tool result back, so that turn never happens. Reworded to state intent in one sentence before the tool call, which lands in the same response and shows up as the assistant reply.
A blank ANTHROPIC_API_KEY= (from clearing a copied .env.example value, or a host that injects empty vars) passed through as an empty string and failed z.string().min(1), surfacing a raw 500 zod dump instead of the route intended friendly not-set 503. Coerce blank values to undefined before parsing so optional fields stay optional and defaulted fields fall back to their default.
The post-deploy sync step pointed at packages/tx-decoder/src/registry/data/agent-policy-gate.json, which no longer carries AgentPolicyGate after the alpha.4 cleanup. The address lives in examples/arbitrum-london/decoder-data/agent-policy-gate.json, matching DEPLOY.md, the deploy script comment, and the decode route.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Two changes in /api/agent, both reducing Claude cost without touching demo quality: - Model claude-opus-4-8 -> claude-haiku-4-5. The route does one thing: turn a natural-language intent into a single validated tool call against known token addresses. That is simple structured extraction, not reasoning - Haiku 4.5 handles it at ~1/5 the price ($1/$5 vs $5/$25 per 1M tokens), roughly $0.0015 per interaction. - Skip the model call entirely when the scenario A contracts are still placeholder addresses. Previously a visitor typing in preview mode would spend a Claude call that then 503s because the envelope cannot be built. Now the deploy check short-circuits to 503 before the API call, so preview mode costs nothing.
|
Deployment failed with the following error: Learn More: https://vercel.com/mikediamonds-projects?upgradeToPro=build-rate-limit |
…ain tx hashes The AI Agentic category requires real tx hashes, not preview-only. The UI sign button already lands a real executeEnvelope tx, but Robinhood Chain has no UI flow (scenario C is a placeholder), so a script is the only way to get a verifiable tx there. SmokeExecuteEnvelope reproduces exactly what the dApp does - build the inner MockPendleRouter swap, agent-sign the EIP-712 envelope, execute through the gate - and is chain-agnostic, so the same command lands a real tx on Arbitrum Sepolia and on Robinhood (sponsor bonus). DEPLOY.md gains the router deploy + allow-list on Robinhood and a step-7 capture flow. The full deploy -> allow-list -> smoke tx path was rehearsed on a local anvil (executeEnvelope succeeded, real tx hash produced).
Judge-facing overview plus a Live on-chain section with placeholder tables for both chains contract addresses and example executeEnvelope tx hashes, filled in after deploy. States what the demo proves (agent prepares, human verifies, gate enforces) and the honest scope (deterministic mock router; scenario C UI still roadmap).
…actions Split the 348-line chat component (now 237). Extracted the pure formatters plus a new resolveExplorerLabel to utils/formatters.ts, the /api/decode call and its DecodedCall type to utils/fetchDecoded.ts, and the sign/reject/tx-link block to SignEnvelopeActions.tsx. Also from review: split the double ternary in the error path into named consts, stable crypto.randomUUID keys instead of array index, aria-live on the thinking status, role=alert on errors, aria-label plus focus-visible on the input, and the explorer link label now matches the chain (Arbiscan vs Robinhood explorer) instead of a hardcoded Arbiscan. Color classes use the Tailwind theme shorthands.
… and summary The copy button, the explorer link, and the decoded-arguments summary had hover styles but no keyboard focus indicator. Added a focus-visible ring matching the project standard so keyboard and switch-access users can see focus.
Replace 104 verbose bg-[color:var(--color-*)] arbitrary values with the Tailwind v4 theme shorthands (bg-accent, text-muted, border-border, ...) already registered in @theme inline - no config change, identical output. Review cleanup folded in: rename isDeployed -> checkIsDeployed (predicate naming rule), drop the milestone sprint-diary strings and the past-dated Phase 2 Day 10 text from /api/agent responses and the RWA builder, neutralize past-deadline footer/comment text on flow-c, extract alignClass in ChatMessage, single blank line before return where missing, remove the void unused-import hack in envelope-builder, and trim a redundant decoder comment.
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 free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
What
Week 2 of the Arbitrum London Buildathon agent demo (scenario A - Pendle yield swap on Arbitrum Sepolia). Reviews the agent loop, clears the blockers that stopped the demo building/running end to end, trims API spend, adds the reproducible real-tx-hash path the AI Agentic category requires, and a hackathon code-review polish pass.
Changes
packages/**only). Addedglobals.d.tsambient declarations (TS2882) and switched the tool defs tosatisfies Anthropic.Tool(TS2769).DeployPendingBannerreads thecheckIs*Deployedpredicates and shows a preview-mode banner on/flow-awhiledeployed.jsonis placeholder. Renders nothing once contracts are live.claude-opus-4-5->claude-haiku-4-5(simple structured extraction, ~1/5 cost, no quality loss); intent stated before the single-shot tool call; cost guard skips the model call entirely in preview mode (spends nothing pre-deploy).SmokeExecuteEnvelope.s.solreproduces the dApp's executeEnvelope (agent-sign + execute through the gate) as a reproducible forge broadcast, chain-agnostic so one command lands a real tx on Arbitrum Sepolia and on Robinhood (sponsor bonus - the only way to get a Robinhood tx while scenario C has no UI). DEPLOY.md gains the Robinhood router deploy + allow-list and a step-7 capture flow; a new demoREADME.mdcarries the judge-facing "Live on-chain" tables. Rehearsed end to end on a local anvil.ANTHROPIC_API_KEY=returns the intended 503 "not set" instead of a raw 500 zod dump.PendleAgentChat348 -> 237 lines: extractedutils/formatters.ts,utils/fetchDecoded.ts, andSignEnvelopeActions.tsx.aria-liveon the thinking status;role=alerton errors;aria-labelon the input.bg-[color:var(--color-*)]arbitrary values -> theme shorthands (bg-accent,text-muted, ...) already registered in@theme inline; no config change, identical output.key={index}-> stablecrypto.randomUUID;isDeployed->checkIsDeployed; removed sprint-diarymilestonestrings + past-dated "Phase 2 Day 10" text; blank lines beforereturn; removed thevoidunused-import hack.Verification
forge test: 20/20 ·forge build: clean (incl. the smoke script)SmokeExecuteEnvelope-> ONCHAIN EXECUTION SUCCESSFUL, realexecuteEnvelopetx hash producedtsc --noEmit: 0 errors ·eslint app src: 0 errors ·next build: clean, 9 pagesbg-accent{background-color:var(--tx-color-primary,...)},bg-card/40->color-mix)Deploy runbook for Mike (needs a funded testnet key you hold)
Full copy-paste in
examples/arbitrum-london/DEPLOY.md. After deploy, run the smoke script (step 7) per chain and paste the addresses + tx hashes intoREADME.md. Only the deployer address needs gas (~0.05 ETH per chain on Arbitrum Sepolia + Robinhood); the agent signer signs off-chain (no gas).Left for Week 3
pnpm lintlints vendoredcontracts/lib/**(noise) - flagged separately.Do not merge - review first.
🤖 Generated with Claude Code