# ElizaOS Developer Update (2026-04-23)  
**Week:** 2026-04-19 → 2026-04-23

This update summarizes core framework and ecosystem changes across `elizaos/eliza`, `elizaos/cloud`, and `elizaos-plugins/registry`, plus notable developer Discord discussions.

---

## 1) Core Framework

### Runtime stability + consistency work
- **Milady CI regressions repaired (restores Telegram export + website blocker; fixes LifeOps signal client collisions)**  
  The Milady pipeline regressions were resolved, restoring downstream functionality that depends on Telegram export and website-blocker plumbing, and fixing a method collision in the LifeOps signal client.  
  PR: https://github.com/elizaos/eliza/pull/7008

- **Typecheck alignment across LifeOps / agent / app-core / shared**  
  Continued work to keep the monorepo type surface coherent across apps and packages (reducing “works in app A, fails in app B” drift).  
  PR: https://github.com/elizaos/eliza/pull/6984

### Cloud auth architecture: Steward migration
- **Authorization flows migrated to Steward** to decouple identity/auth from legacy paths and enable more reliable, independently evolvable login flows.  
  PR: https://github.com/elizaos/cloud/pull/466

- **Native Ethereum + Solana login buttons** added to improve wallet-based identity UX.  
  PR: https://github.com/elizaos/cloud/pull/462

### Work in progress to watch
- **Local inference hub with provider switcher + local probes** is actively being developed (useful for builders targeting Ollama/local LLMs and wanting runtime-level failover). Mentioned in the 2026-04-22 dev digest.  
  Daily digest context: https://elizaos.github.io/api/summaries/overall/day/2026-04-22.json

---

## 2) New Features

### LemonCake payment plugin (agent-safe paid API access)
A new open-source plugin, **`eliza-plugin-lemoncake`** (MIT licensed), was introduced to solve a common agent deployment problem: *letting autonomous agents pay for third-party APIs without embedding credit cards or long-lived API keys.*

**Key capabilities (as announced):**
- Spend-capped, time-boxed, revocable **Pay Tokens**
- Uses **USDC / JPYC on Polygon**
- Integrations for **32 paid APIs** (search, scraping, LLM, image, etc.)
- Auto idempotency keys, structured receipts
- Accounting exports (QuickBooks / Xero / Zoho / freee)
- Sandbox mode for testing

Developer discussion (Discord coders channel):  
https://discord.com/channels/1253563208833433701/1300025221834739744

**Example (minimal integration pattern)**  
(Exact API may vary by plugin release; conceptually you add the plugin and provide credentials / environment config.)

```ts
import { AgentRuntime } from "@elizaos/core";
import { lemoncakePlugin } from "eliza-plugin-lemoncake";

const runtime = new AgentRuntime({
  plugins: [
    lemoncakePlugin({
      // Typical config: issuer / policy / sandbox flags
      // sandbox: true,
      // defaultSpendCapUsdCents: 500, // $5.00
    }),
  ],
});
```

**Operational guidance**
- Treat Pay Tokens like short-lived delegated credentials:
  - set **tight caps** for unattended agents
  - rotate/revoke on anomaly detection
  - record receipts in your own datastore even if you export to accounting

---

## 3) Bug Fixes

### Billing misclassification for image-generation models (Cloud)
A pricing issue where image-generation models were billed under an incorrect classification was fixed. If you rely on model-type based billing rules, re-validate your internal assumptions about SKU mapping and invoice line items.  
PR: https://github.com/elizaos/cloud/pull/470

### Telegram export functionality restored (Milady regression)
Telegram export flows were restored as part of the Milady CI regression repair. If your deployments depend on Telegram history export/import, re-run your export job and verify message continuity.  
PR: https://github.com/elizaos/eliza/pull/7008

### Open issue: Telegram read receipts performance
A follow-up is needed to optimize Telegram read receipt handling (reduce nested lookups / improve message ID fetching).  
Issue: https://github.com/elizaos/eliza/issues/7009

---

## 4) API Changes (Developer-Facing)

### Runtime composition surface is evolving (watch for V2→V3 alignment)
Recent work (notably the “starter agent workspace” effort) introduced changes around how characters are loaded and composed (e.g., `loadCharacters` accepting file paths/options in some flows). If you have custom character loaders or wrappers, expect minor churn until the V2 alpha → stable line hardens.

Reference (starter workspace PR): https://github.com/elizaos/eliza/pull/6702

### Prompt output formatting: configurable to support non-Claude models
A new environment variable was added to control output encapsulation used by `dynamicPromptExecFromState`, improving reliability on models that don’t emit TOON consistently (e.g., Gemini, Llama/Ollama).  
PR: https://github.com/elizaos/eliza/pull/6978

**Example**
```bash
# Use a model-friendly output format when TOON is unreliable
export PROMPT_OUTPUT_FORMAT=json
```

---

## 5) Social Media Integrations (Twitter / Telegram / Discord / Farcaster)

### Discord
- **LifeOps Discord routing fixes** landed as part of ongoing stability work (reducing misroutes and improving parity with other connector behavior).  
  PR (clean-pick / superseding prior branch): https://github.com/elizaos/eliza/pull/6971

- **Community incident:** multiple users reported Discord messages being auto-deleted with no audit log entries (likely Discord-side automation/anti-spam behavior). This is not currently tracked as a GitHub issue, but it impacts developer support workflows and link sharing. Track internally if you rely on Discord for operational links.

### Telegram
- Export functionality restored (see Bug Fixes).
- Read receipt optimization remains open (see API/Issues).

### Twitter/X
- A builder reported a high-traction X agent (9M impressions) later hit repeated bans. Treat this as a reminder that X automation risk remains high; for production agents consider:
  - rate limiting + “human-like” pacing
  - alternate channels (Discord/Telegram) for durable distribution
  - auditable action logs for ban appeals / debugging

(Discussion context)  
https://discord.com/channels/1253563208833433701/1253563209462448241

---

## 6) Model Provider Updates

### Local inference + provider switching (in progress)
Work is underway on a **local inference hub** and **provider switcher** (see Core Framework). This will matter if you:
- want automatic fallback between OpenAI/Anthropic/local providers
- need consistent tracing across providers
- run “offline-first” agents with local probes

### Dependency hardening across providers + SDKs
Broad dependency updates landed this week across Rust/Python/TS packages (security/perf posture). Notable provider-adjacent movements referenced in the repo activity include:
- updates to AI/provider SDK packages (e.g., Anthropic SDK bumps observed in dependency maintenance streams)
- continued effort to reduce prompt bloat and improve model output determinism (ties directly into `PROMPT_OUTPUT_FORMAT`)

For the full dependency maintenance context (day digest):  
https://elizaos.github.io/api/summaries/overall/day/2026-04-22.json

---

## 7) Breaking Changes / Migration Warnings (V1 → V2 → V3)

### Version naming is currently confusing (community call-out)
In Discord, the team clarified the project’s versioning convention being used in some areas:
- `0.x` = “first version”
- `1.x` = “second version”
- `2.x` = “third version” (often referred to as “v3” conceptually)

If you ship SDK bindings or templates, **pin exact versions** and avoid semantic assumptions based on the major digit alone until the naming settles.

Discussion:  
https://discord.com/channels/1253563208833433701/1253563209462448241

### V2 is usable but still alpha
Core devs indicated **v2.x is ready for active agent development**, but **v2 stable is not yet announced**. Plan accordingly:
- pin to a known-good alpha tag
- vendor critical prompt templates/specs if you need reproducibility
- run CI against `develop` if you’re building infra plugins

### Plugin ecosystem migration: Elisym + registry integration
- `@elisym/plugin-elizaos` was added to the plugin registry (enables v1 agents to integrate with the Elisym marketplace via Nostr NIP-89/NIP-90 + Solana payments).  
  Registry PR: https://github.com/elizaos-plugins/registry/pull/346

- Next step called out by maintainers: **port Elisym to the V3 runtime line** after the ecosystem merge.

### CI/CD breaking point to address (registry)
A workflow failure was identified in the plugin registry: `claude-code-action` can’t fetch an OIDC token due to missing permissions (`id-token: write` / token config). This blocks some PR automation.

Reference discussion (PR #346 context):  
https://github.com/elizaos-plugins/registry/pull/346  
Discord summary (2026-04-21): https://discord.com/channels/1253563208833433701/1253563209462448241

**Typical fix pattern**
```yaml
permissions:
  contents: read
  id-token: write
```

---

### Security Note (ecosystem safety)
A community member reported losing tokens to a wallet drainer at `bulkdao.co/allocation`. Treat this as an ops reminder: **never ask users to connect wallets to unverified links**, and consider adding link-scanning / allowlist policies in any agent that posts URLs.

Discord warning context (coders channel):  
https://discord.com/channels/1253563208833433701/1300025221834739744