{
"interval": {
"intervalStart": "2026-04-04T00:00:00.000Z",
"intervalEnd": "2026-04-05T00:00:00.000Z",
"intervalType": "day"
},
"repository": "elizaos/eliza",
"overview": "From 2026-04-04 to 2026-04-05, elizaos/eliza had 0 new PRs (0 merged), 2 new issues, and 4 active contributors.",
"topIssues": [
{
"id": "I_kwDOMT5cIs76q52a",
"title": "Plugin proposal: @sint/eliza-plugin — capability token enforcement for Eliza agent tool calls",
"author": "pshkv",
"number": 6707,
"repository": "elizaos/eliza",
"body": "## Context\n\nEliza agents execute tool calls — blockchain transactions, Twitter posts, file operations, and increasingly physical actions (robots, drones, IoT). Today there's no formal authorization layer between \"the LLM decided to call this tool\" and \"the tool executes.\"\n\nThe existing `GuardrailsService` and plugin system handle some of this, but without:\n- Cryptographic proof of authorization (who issued permission for this specific action)\n- Irreversibility-aware approval flows (on-chain transaction ≠ camera read)\n- Behavioral drift detection (agent acting anomalously → escalate before next action)\n- Physical constraint enforcement (for Eliza agents controlling robots)\n\n## Proposed: `@sint/eliza-plugin`\n\nA thin wrapper around [SINT Protocol](https://github.com/pshkv/sint-protocol) that intercepts Eliza tool calls:\n\n```typescript\n// eliza.config.ts\nimport { SintPlugin } from \"@sint/eliza-plugin\";\n\nexport default {\n plugins: [\n new SintPlugin({\n resolveToken: (toolName, agentId) => tokenStore.get(`${agentId}:${toolName}`),\n tierRules: [\n // Blockchain transactions require human review\n { resourcePattern: \"solana://*\", actions: [\"transfer\", \"swap\"], baseTier: \"T3_commit\" },\n // Social posts need operator approval\n { resourcePattern: \"twitter://*\", actions: [\"post\", \"retweet\"], baseTier: \"T2_act\" },\n // Read-only ops auto-allow\n { resourcePattern: \"twitter://*\", actions: [\"search\", \"read\"], baseTier: \"T0_observe\" },\n ],\n emitLedgerEvent: (event) => evidenceLedger.write(event),\n })\n ]\n}\n```\n\n## What this adds to Eliza\n\n| Feature | Benefit |\n|---|---|\n| Ed25519 capability tokens | Cryptographic proof of who authorized each tool call |\n| Tier-based approval (`T2_act`, `T3_commit`) | On-chain transactions pause for human sign-off |\n| CSML behavioral drift | Agent acting anomalously → auto-escalate before next action |\n| Rate limiting | Per-tool sliding window enforcement |\n| Evidence ledger | Hash-chained audit trail, TEE-attested |\n| Physical constraint enforcement | For Eliza agents controlling robots/drones |\n\n## Connection to #6688 (AgentID)\n\n@haroldmalikfrimpong-ops is proposing behavioral fingerprinting in #6688 — SINT's `AvatarRegistry` already does this via the CSML metric. If AgentID and SINT use compatible schemas, agents could port behavioral history across both systems.\n\n## Implementation\n\nThis would be a thin bridge — ~200 lines wrapping `@sint/gate-policy-gateway`. The hard part (policy engine, evidence ledger, approval flow) is already in SINT Protocol (950 tests).\n\nFull repo: https://github.com/pshkv/sint-protocol \nWould the Eliza core team be open to this as a community plugin? Happy to build it.",
"createdAt": "2026-04-04T18:32:12Z",
"closedAt": null,
"state": "OPEN",
"commentCount": 0
},
{
"id": "I_kwDOMT5cIs76qj66",
"title": "Plugin: SafeAgent — Token safety checks before trading (honeypot, scam, rug pull detection)",
"author": "CryptoGenesisSecurity",
"number": 6706,
"repository": "elizaos/eliza",
"body": "## Plugin: @elizaos/plugin-safeagent\n\n**Repo**: https://github.com/CryptoGenesisSecurity/plugin-safeagent\n\n### Problem\nElizaOS agents trading crypto have limited pre-trade safety checks. `plugin-base-signals` only covers Base L2. No comprehensive multi-chain token safety exists in the ElizaOS ecosystem.\n\n### Solution\nSafeAgent plugin provides two actions:\n\n**1. CHECK_TOKEN_SAFETY** — Full safety scan on demand\n```\nUser: \"Is 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913 safe on base?\"\nAgent: 🟢 Token Safety Report — USDC (USDC) — Score: 90/100 — SAFE\n```\n\n**2. PRE_TRADE_SAFETY_CHECK** — Auto-blocks dangerous trades\n- Score < 40: **TRADE BLOCKED** (likely scam)\n- Score 40-70: **CAUTION** (reduce position)\n- Score > 70: proceed normally\n\n### Detection\n- Honeypot simulation (real DEX swap test on UniV2, V3, Aerodrome)\n- 17 scam pattern checks on source code\n- LP lock verification (Unicrypt, TeamFinance, PinkSale)\n- Owner analysis (renounced, privileged functions)\n- 6 EVM chains: Base, Ethereum, Arbitrum, Optimism, Polygon, BSC\n- Sub-second response (<1s cached, <1.5s fresh)\n\n### Standard\nImplements the **ERC Token Safety Score** — an open standard for on-chain token safety scoring.\n- Live oracle on Optimism: `0x3B8A6D696f2104A9aC617bB91e6811f489498047`\n- MCP SSE: `https://cryptogenesis.duckdns.org/mcp/sse`\n\n### Installation\n```bash\nnpm install CryptoGenesisSecurity/plugin-safeagent\n```\n\n```typescript\nimport safeAgentPlugin from \"@elizaos/plugin-safeagent\";\nconst agent = new AgentRuntime({ plugins: [safeAgentPlugin] });\n```\n\n### vs plugin-base-signals\n| Feature | plugin-base-signals | plugin-safeagent |\n|---------|-------------------|-----------------|\n| Chains | Base only | 6 EVM chains |\n| Honeypot sim | GoPlus API | Direct DEX swap simulation |\n| Scam patterns | Basic | 17 patterns |\n| LP lock check | No | Yes (5 lockers) |\n| On-chain oracle | No | Yes (Optimism) |\n| Response time | Variable | <1s (cached) |\n\nHappy to contribute this to the official elizaos-plugins registry.",
"createdAt": "2026-04-04T17:50:59Z",
"closedAt": null,
"state": "OPEN",
"commentCount": 0
}
],
"topPRs": [
{
"id": "PR_kwDOMT5cIs7Pdfjt",
"title": "feat: add plugin-mnemopay — economic memory for AI agents",
"author": "t49qnsx7qt-kpanks",
"number": 6701,
"body": "## Summary\n\nAdds **plugin-mnemopay**, a new plugin that gives Eliza agents economic memory. Agents can remember payment outcomes, learn from settlements and refunds, and build reputation over time — making them smarter about financial interactions.\n\nThis is powered by [MnemoPay](https://github.com/t49qnsx7qt-kpanks/mnemopay-sdk) (`@mnemopay/sdk`), a TypeScript SDK for AI agent economic memory.\n\n### Why this matters\n\nStandard AI agents treat every financial interaction as a blank slate. With MnemoPay, agents:\n- **Remember** which providers delivered quality work and which didn't\n- **Learn** from payment disputes and successful settlements\n- **Build reputation** through consistent positive outcomes (capped at 2.0)\n- **Make informed decisions** by recalling past financial experiences before acting\n\n### Plugin components\n\n| Component | Name | Purpose |\n|-----------|------|---------|\n| **Service** | `MnemoPayService` | Manages the MnemoPayLite engine lifecycle |\n| **Actions** | `REMEMBER_OUTCOME` | Store a payment/interaction outcome in economic memory |\n| | `CHARGE_PAYMENT` | Create an escrow payment (wallet debit) |\n| | `SETTLE_PAYMENT` | Confirm delivery, reinforce reputation (+delta) |\n| | `REFUND_PAYMENT` | Reverse payment, dock reputation (-delta) |\n| | `RECALL_MEMORIES` | Query past financial experiences |\n| **Provider** | `MnemoPayProvider` | Injects wallet balance, reputation, recent transactions, and relevant memories into conversation context |\n| **Evaluator** | `MnemoPayEvaluator` | Auto-tracks financial outcomes after every agent response (passive learning loop) |\n\n### Architecture decisions\n\n- Follows the exact same patterns as `advanced-memory` and `basic-capabilities` plugins\n- Service extends `Service` base class with static `start()` factory\n- Actions return `ActionResult` with `success` field\n- Provider returns `ProviderResult` with `text`, `values`, and `data`\n- Evaluator uses `alwaysRun: true` for passive financial outcome detection\n- Built-in lightweight engine included — no external dependency required at runtime\n- Configurable via `MNEMOPAY_AGENT_ID` and `MNEMOPAY_REPUTATION_DELTA` env vars\n\n### Usage\n\n```typescript\nimport { createMnemoPayPlugin } from \"./plugin-mnemopay\";\n\nconst agent: ProjectAgent = {\n character: myCharacter,\n plugins: [createMnemoPayPlugin()],\n};\n```\n\n## Test plan\n\n- [ ] Verify `MnemoPayService` initializes correctly with default and custom config\n- [ ] Test each action (REMEMBER_OUTCOME, CHARGE_PAYMENT, SETTLE_PAYMENT, REFUND_PAYMENT, RECALL_MEMORIES) with valid and invalid inputs\n- [ ] Verify provider injects correct context (wallet, reputation, recent txs, relevant memories)\n- [ ] Verify evaluator auto-tracks financial keywords and stores with correct importance/tags\n- [ ] Confirm plugin registers correctly via `createMnemoPayPlugin()` factory\n- [ ] Test settle/refund reputation bounds (0.0 to 2.0)\n- [ ] Verify graceful degradation when service is not available\n\n🤖 Generated with [Claude Code](https://claude.com/claude-code)\n\n\n\nGreptile Summary
\n\nThis PR adds `plugin-mnemopay`, a new Eliza plugin that gives agents \\\"economic memory\\\" — tracking payment charges, settlements, refunds, reputation, and past financial interactions. The plugin follows the standard Eliza plugin shape (Service + Actions + Provider + Evaluator) and the code structure is clean and readable.\n\nHowever, there are several significant issues that should be resolved before merging:\n\n- **No state persistence** — the core `MnemoPayLiteEngine` stores all memories, transactions, and reputation in plain in-memory JavaScript structures. Every agent restart wipes the slate clean, which directly defeats the plugin's stated purpose of building reputation and memory over time.\n- **NaN reputation corruption** — if `MNEMOPAY_REPUTATION_DELTA` is set to any non-numeric string, `Number.parseFloat()` silently returns `NaN`, permanently corrupting the reputation score.\n- **Unsafe null casts in action handlers** — every `validate()` checks `if (!service) return false`, but every `handler()` casts the same `getService()` call without a null guard, creating a latent null dereference across all 5 actions and the evaluator.\n- **Ambiguous action validation** — both `RECALL_MEMORIES` and `REMEMBER_OUTCOME` match the keyword \\\"remember\\\", causing both actions to fire on messages like \\\"Remember that Provider X is excellent.\\\"\n- **Unbounded memory growth** — `this.memories.push(entry)` has no eviction policy; the evaluator fires on very common financial keywords after every agent response.\n- **Dead interface field** — `MnemoPayConfig.initialBalance` is declared in `types.ts` but never read by the engine or service initializer.\n- **No tests** — the test plan in the PR description is entirely unchecked; no test files are included in the changeset.\n\nConfidence Score: 1/5
\n\nNot safe to merge — the plugin's core value proposition (persistent economic memory) is unimplemented, and there are several logic bugs that will silently corrupt state or cause null-dereference crashes.\n\nThe fundamental design flaw (all state is ephemeral in-memory) means the plugin cannot fulfill its stated purpose in any real deployment. Combined with the NaN-corruption bug for misconfigured reputation delta, unsafe null dereferences in every action handler, and absence of any tests, the PR needs substantial rework.\n\nmnemopay-service.ts requires the most attention (persistence, NaN guard, memory eviction). All action handler files need null-safety fixes. recall-memories.ts needs its validate keyword set de-conflicted from remember-outcome.ts.\n\nImportant Files Changed
\n\n\n\n\n| Filename | Overview |\n|----------|----------|\n| packages/typescript/src/plugin-mnemopay/services/mnemopay-service.ts | Core engine stores all memory, transactions, and reputation in plain JS objects — no persistence; missing NaN guard for reputationDelta and no memory eviction cap. |\n| packages/typescript/src/plugin-mnemopay/actions/charge-payment.ts | Unsafe null cast of service in handler body, and overly broad amount extraction regex that can match non-payment numbers. |\n| packages/typescript/src/plugin-mnemopay/actions/recall-memories.ts | Validate keyword \"remember\" conflicts with REMEMBER_OUTCOME action; unsafe service cast in handler. |\n| packages/typescript/src/plugin-mnemopay/evaluators/mnemopay-evaluator.ts | Passive financial keyword detection is broadly correct; unsafe service cast in handler; combined with unbounded memory growth, every turn with words like \"cost\" or \"fee\" writes a memory entry. |\n| packages/typescript/src/plugin-mnemopay/providers/mnemopay-provider.ts | Cleanly injects wallet/reputation/recent-tx context; graceful degradation when service is absent; non-critical recall failures are swallowed appropriately. |\n| packages/typescript/src/plugin-mnemopay/types.ts | MnemoPayConfig.initialBalance is declared but never consumed by the engine or service initializer — dead interface field. |\n| packages/typescript/src/plugin-mnemopay/index.ts | Plugin registration and re-exports are clean and follow the expected Plugin interface pattern. |\n\n\n\n\n\nSequence Diagram
\n\n```mermaid\nsequenceDiagram\n participant User\n participant Action\n participant MnemoPayService\n participant MnemoPayLiteEngine\n participant Provider\n participant Evaluator\n\n Note over MnemoPayLiteEngine: In-memory only — no persistence\n\n User->>Action: \"Charge $50 for design task\"\n Action->>MnemoPayService: getEngine()\n MnemoPayService->>MnemoPayLiteEngine: charge(50, \"design task\")\n MnemoPayLiteEngine-->>Action: txId = \"tx_agent_1_1_...\"\n Action-->>User: \"Payment charged. TX: tx_agent_1_1_...\"\n\n User->>Action: \"Settle payment tx_agent_1_1_...\"\n Action->>MnemoPayLiteEngine: settle(txId)\n MnemoPayLiteEngine->>MnemoPayLiteEngine: reputation = min(2.0, rep + delta)\n MnemoPayLiteEngine-->>Action: settled tx\n Action-->>User: \"Settled. Reputation: 1.05\"\n\n Note over Evaluator: alwaysRun — fires after every response\n Evaluator->>MnemoPayLiteEngine: remember(\"[Auto-tracked] ...\", {importance, tags})\n MnemoPayLiteEngine->>MnemoPayLiteEngine: memories.push(entry) — unbounded!\n\n User->>Provider: (next conversation turn)\n Provider->>MnemoPayLiteEngine: balance() + getRecentTransactions(5)\n Provider->>MnemoPayLiteEngine: recall(messageText, 3)\n MnemoPayLiteEngine-->>Provider: memories + balance\n Provider-->>User: Economic memory context injected into prompt\n```\n\n\nComments Outside Diff (1)
\n\n1. `packages/typescript/src/plugin-mnemopay/actions/charge-payment.ts`, line 162-165 ([link](https://github.com/elizaos/eliza/blob/fad58e46328b453c6bfa5bc21dd58c5c7726c938/packages/typescript/src/plugin-mnemopay/actions/charge-payment.ts#L162-L165)) \n\n ![\"P1\"](\"https://greptile-static-assets.s3.amazonaws.com/badges/p1.svg?v=7\")
**Unsafe cast without null check in handlers — potential null dereference**\n\n Every action's `validate()` defensively checks `if (!service) return false`, but the corresponding `handler()` immediately casts the result without any null guard. If the service is not registered, `runtime.getService(\"mnemopay\")` returns `null`, and `service.getEngine()` throws a `TypeError`. The same pattern appears in `settle-payment.ts`, `refund-payment.ts`, `remember-outcome.ts`, `recall-memories.ts`, and `mnemopay-evaluator.ts`.\n\n Each handler should guard against this:\n ```typescript\n const service = runtime.getService(\"mnemopay\") as MnemoPayService | null;\n if (!service) {\n return { success: false, text: \"MnemoPayService is not available\" };\n }\n const engine = service.getEngine();\n ```\n\n