[TheCodingChef-eth]

Closes #64, closes #65, closes #70, closes #71

Base repair (first commit)

develop does not compile: 3ce175e (PR #172) was committed with a bad conflict
resolution (schema syntax error, missing Prisma models and config keys,
completeLLM imported but not exported, syntax errors in the cron files). The
first commit redoes the merge of 7a8d096 and ced038e keeping both sides, plus
the few type fixes needed to compile. Please review it on its own.

• closes Run the contracts x402 trace API in an end-to-end test against backend-published traces #64 x402 end-to-end

• tests/x402-e2e.test.ts runs the submodule's real createTraceApi against
  traces from generateReasoningTrace + pinAndPublishDocument: 402 challenge,
  paid 200 (content equals the canonical payload), tampered 502, unknown 404,
  and no content when settlement fails.

• In-memory IPFS, in-process stub facilitator, and a fetch guard that fails the
  test on any non-loopback request.

• tests/canonical-parity.test.ts compares canonicalJson with the contracts
  canonicalize over an edge-case corpus; pinJson now rejects non-plain
  objects (IPFS_NON_PLAIN_DOCUMENT) because Date diverges (ISO vs {}).

• npm run test:x402:prepare and npm run test:e2e:x402; CI installs and
  caches the x402 deps. Skips locally with guidance, fails when CI=true.

• closes Use provider-native structured outputs generated from the zod proposal schemas #65 Structured outputs

• src/lib/llm-schema.ts is the single schema registry (zod, Anthropic tool
  schema, Gemini responseSchema, postValidate).

• callLLMStructured(): Anthropic forced tool use, Gemini responseSchema, one
  repair round per provider, LLM_STRUCTURED_OUTPUT_FAILED. Fence stripping is
  gone from llm.ts, gemini.ts and market.service.ts.

• generateMarketQuestion / generateHedgeConditions migrated; seams
  setAnthropicClient / setGeminiModelFactory; fixtures in tests/fixtures/llm/.

• closes Defend market generation against prompt injection in ingested signals #70 Prompt injection

• llm-guard.service.ts: sanitiser (NFKC, control/zero-width/bidi, caps, URLs
  moved to url), quarantine, nonce-delimited untrusted blocks, checkProposal
  (grounding, probability, expiry, oracle allowlist, content policy).

• All fetchers sanitise before caching; GDELT URL is stored in url.

• Rejections/quarantines go to agent_logs with reason codes and hashes only.

• Corpus: 28 attack + 18 benign cases in tests/fixtures/injection/.

• closes Enforce LLM token budgets and a shared circuit breaker across instances #71 Budgets and shared breaker

• llm-budget.ts + atomic Lua: per-provider token bucket, circuit breaker with
  half-open probe, daily token/cost budgets with an admin reserve, reservations
  released on error/missing usage, crashed leases reclaimed.

• Usage read from Anthropic usage and Gemini usageMetadata; cron stops at
  budget minus reserve, admin runs to the full budget; LLM_BUDGET_EXCEEDED
  logged once per day. LLM_REDIS_FAIL_MODE=closed|open.

• Same scenario table runs against the JS reference store and real Redis.

Notes

• New env vars are in .env.example and docs/architecture.md.
• npm test has 15 failures that also fail on the repaired base (request-id
  tests, copy-trade/reconcile, some resilience tests, DB-dependent tests).
• LLM fixtures are hand-built in the real API shapes, not live captures.

[drips-wave]

@TheCodingChef-eth Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits