# elizaOS — User Feedback Analysis (2026-04-14)

## Data context & limits
- Primary signals available for this date are **GitHub issues + PR review findings** (including automated review notes). Several daily Discord artifacts referenced in the index were not provided here (multiple “File not found”), so **community chat feedback is underrepresented**.
- Sample used for quantification: **10 high-signal threads** (5 top issues + 5 PRs with substantial review feedback). Percentages below are relative to these 10 threads unless otherwise noted.

---

## 1) Pain Point Categorization (Top recurring friction areas)

### A. Technical Functionality (Most severe / highest risk)
**Recurring problems**
1) **Core memory correctness & persistence regressions**  
   - PR #6562 review flags:
     - IGNORE memory persistence bug when `ALLOW_MEMORY_SOURCE_IDS` is set (hardcoded `"agent_response"` check causes silent drops).
     - Embedding failure persists a **zero vector**, making stored memories effectively unretrievable (silent data loss).
   - PR #6701 (MnemoPay) review flags: “economic memory” is **in-memory only** (no persistence), undermining stated purpose.

**Frequency / severity**
- Mentioned in **3/10 threads (30%)**, but severity is high because it affects correctness, safety, and user trust (silent data loss, misleading “stored” state).

**Users affected most**
- Developers building agents with long-lived memory, evaluation loops, or financial workflows (where auditability and retention matter).

---

### B. Developer Experience / Onboarding (CLI & local dev workflow)
**Recurring problems**
1) **`elizaos create` fails on macOS due to Bun postinstall behavior** (Issue #6704)  
   - Failure: “Bun’s postinstall script was not run”  
   - CLI cleans up directory on failure, leaving users with nothing to inspect or salvage.
2) **Local dev harness/submodule workflow inconsistencies can break fresh clones** (PR #6702 review)  
   - Root workspace entries for submodules committed in a way that can cause missing-workspace install failures.
   - `workspace:*` vs lockfile resolution mismatch.
3) **Windows-specific friction** (weekly summary notes fix attempt in `plugin-openrouter` for Windows git checkout blocks).

**Frequency / severity**
- Mentioned in **3/10 threads (30%)**. Severity is medium-to-high because it blocks initial adoption.

**Users affected most**
- Newcomers, macOS Apple Silicon users, Windows developers, and anyone cloning repo fresh without submodules initialized.

_Category label_: Documentation + Technical Functionality (onboarding) + Integration (toolchain)

---

### C. Integration (Plugins & external ecosystems)
**Recurring problems**
1) **Demand for standardized, secure integrations for agent economies**
   - Marketplace plugin proposal (Issue #6700): unified AI-to-AI marketplace + swaps + GPU rental.
   - Token safety plugin request (Issue #6706): multi-chain pre-trade safety checks.
   - Capability token enforcement proposal (Issue #6707): cryptographic authorization for tool calls (including irreversible actions).
2) **Connector-specific behavior gaps**  
   - PR #6709: TOON schema missing `params` caused required action params not to flow through (notably in Discord/Milady connectors).

**Frequency / severity**
- Mentioned in **4/10 threads (40%)**. Severity varies: connector param loss is high friction; new integrations are strategic requests.

**Users affected most**
- Builders deploying agents into Discord/community rooms, crypto trading/DeFi agents, and agent-to-agent commerce experiments.

---

### D. UX/UI (Conversation behavior in group rooms)
**Recurring problems**
1) **Agents responding in the wrong contexts or creating loops in group chats**  
   - PR #6712 targets addressee routing and anti-loop guidance, but review flags a P1 ambiguity bug (`aliasEntity` causing failure to resolve addressees when `entityId ≠ agentId`).

**Frequency / severity**
- Mentioned in **1/10 threads (10%)**, but severe for real deployments in group channels (spam/loop risk; incorrect silence).

**Users affected most**
- Operators running agents in multi-party Discord/Telegram-style rooms.

---

### E. Performance / Reliability (Latency + timeouts)
**Recurring problems**
1) **Timeout defaults and provider behavior can shift latency unexpectedly**
   - PR #6562 changes provider timeout behavior; review notes `PROVIDERS_TOTAL_TIMEOUT_MS` default increased (1s → 5s), potentially impacting P99 latency for existing deployments.

**Frequency / severity**
- Mentioned in **1/10 threads (10%)**; severity depends on deployment sensitivity.

---

### F. Documentation (Expectations vs reality)
**Recurring problems**
1) **Mismatch between “plugin promises” and operational reality**
   - MnemoPay PR claims durable “economic memory,” but implementation is ephemeral and unbounded; no tests included.
2) **Tooling assumptions not spelled out**
   - Bun runtime dependency vs system-installed bun mismatch (Issue #6704) suggests docs/packaging expectations are unclear.
3) **Strategic pivot communication**
   - Weekly summary indicates deprioritization of standalone messaging protocols (e.g., XMTP closure in registry) in favor of AgentID—users may still expect messaging-first priorities.

**Frequency / severity**
- Mentioned in **3/10 threads (30%)**, typically as “surprise” failures or overpromised capabilities.

---

## 2) Usage Pattern Analysis (Actual vs intended usage)

### Observed “actual usage”
1) **Agents as economic actors** (dominant emerging pattern)
   - Cross-chain wallet operations (agent wallet plugin, swaps/bridges).
   - Pre-trade safety checks (SafeAgent).
   - Marketplace access + GPU rental + escrow settlement (MAXIA proposal).
   - Reward economies for contributions (AIGEN protocol issue #6708).

2) **Agents operating in group chat environments**
   - Discord connector testing surfaced critical schema/param issues (PR #6709).
   - Addressee routing + anti-loop guidance is being treated as core (PR #6712).

3) **Local “agent harness” workflows for devs**
   - The repo is being used as a runnable “agent runtime laboratory,” not only as a library (PR #6702).

### Divergence from intended usage (implied)
- The framework appears to be shifting from “chat-first agents” to **trust/identity + capability authorization + economic primitives**, but many developer workflows still start with “`elizaos create` → run a chat agent,” making setup failures disproportionately damaging.

### Feature requests aligned with real usage
- **Capability-based authorization / cryptographic approvals** for tool calls (#6707) aligns with agents doing high-stakes actions (trades, posts, physical actions).
- **Token safety scoring and pre-trade blocking** (#6706) aligns with increased trading automation.
- **Unified marketplace plugin** (#6700) aligns with agents needing composable service procurement (compute + AI services + escrow).

---

## 3) Implementation Opportunities (Concrete solutions per major pain point)

### Pain Point 1: Memory correctness & persistence trust (Core)
**Opportunities (prioritized)**
1) **Fix correctness bugs + add regression tests (High impact / Low–Medium effort)**
   - Patch IGNORE persistence logic so allowlists don’t silently drop agent responses (match non-IGNORE path behavior).
   - Replace “zero-vector fallback” with:
     - “do not persist embedding-dependent memory until embedding succeeds,” or
     - “persist with `embeddingStatus=failed` and exclude from similarity search until re-embedded.”
   - Add tests covering:
     - `ALLOW_MEMORY_SOURCE_IDS` + IGNORE path
     - embedding failure behavior + subsequent retrieval

2) **Introduce a durable storage adapter for “economic memory” plugins (High impact / Medium effort)**
   - For MnemoPay-like plugins: ship a minimal persistence interface backed by existing SQL adapter patterns (plugin-sql / sqlite / pglite).
   - Add eviction/retention policy primitives (cap by count/time/importance).

3) **Adopt an “event log” model for high-stakes actions (Medium–High impact / Medium effort)**
   - Similar to approaches in: Temporal (event history), Kafka-based audit logs, OpenZeppelin Defender (action logs).
   - Provide an optional append-only evidence ledger service for financial actions and approvals.

---

### Pain Point 2: `elizaos create` / Bun install failures & destructive cleanup
**Opportunities (prioritized)**
1) **Packaging fix: remove `bun` runtime dependency from CLI/bootstrap (High impact / Low effort)**
   - Implement the proposed change in Issue #6704: depend on `@types/bun` in dev only; avoid pulling `bun` npm package into user projects.

2) **Change CLI failure mode: keep generated project on error (High impact / Low effort)**
   - Instead of cleaning up, leave the directory with:
     - a clear `BUILD_FAILED.md` (or console instructions),
     - captured logs, and
     - a suggested manual recovery path.

3) **Add a “doctor” command (Medium impact / Medium effort)**
   - `elizaos doctor` checks:
     - bun version, node version, postinstall script behavior,
     - OS-specific constraints (macOS aarch64),
     - workspace integrity.
   - Similar to: `flutter doctor`, `rustup check`, `expo doctor`.

---

### Pain Point 3: Plugin & workspace/dev workflow fragility (submodules, lockfiles, Windows)
**Opportunities (prioritized)**
1) **Make submodules truly optional with CI validation (High impact / Medium effort)**
   - Enforce in CI that a fresh clone **without submodules** can `bun install` successfully.
   - Provide a single documented path:
     - either “submodules are dev-only; never committed as workspaces,” or
     - “submodules are first-class; provide bootstrap script and ensure lockfile consistency.”

2) **Replace submodules with a supported local override mechanism (Medium impact / Medium effort)**
   - Use workspace protocol + `patch:`/`overrides` (pnpm) equivalents where possible, or a documented `bun link` flow.
   - Similar patterns: Next.js monorepo examples, Rust path dependencies for local dev.

3) **Windows friction checklist + preflight (Medium impact / Low effort)**
   - Codify known Windows git/path constraints in docs and `doctor` output.
   - Mirror the approach used by many CLIs: detect platform and print targeted instructions.

---

### Pain Point 4: Group chat misrouting / loops / addressee confusion
**Opportunities (prioritized)**
1) **Fix alias ambiguity bug + add missing tests (High impact / Low effort)**
   - Add coverage for `entityId ≠ agentId` in `NameVariationRegistry` tests (explicitly called out as missing).
2) **Add connector-level metadata contract (High impact / Medium effort)**
   - Standardize fields like `replyToEntityId`, `inReplyTo`, thread IDs across adapters.
   - Similar to: Matrix event relations, Slack thread_ts, Discord message references.
3) **Operational guardrails: rate limiting + loop detection (Medium impact / Medium effort)**
   - Add a simple “anti-ping-pong” heuristic at runtime level: detect alternating agents or repeated near-duplicate replies and force cooldown.

---

### Pain Point 5: Authorization expectations for irreversible tool calls
**Opportunities (prioritized)**
1) **First-class “capability/approval” middleware interface (High impact / Medium effort)**
   - Provide a stable interception point for tool calls (before execution) with:
     - approval tiering,
     - human-in-the-loop hooks,
     - cryptographic signing support.
   - Similar to: Kubernetes admission controllers + RBAC; GitHub protected branches.

2) **Reference implementation plugin (Medium impact / Medium effort)**
   - Accept #6707 directionally by defining an official plugin API contract for auth gates, then allow SINT (or others) as implementation.

3) **UI/Docs: explicit “risk tiers” for actions (Medium impact / Low effort)**
   - Document recommended defaults:
     - read-only: auto
     - social posting: approve
     - funds transfer: approve + signed intent

---

## 4) Communication Gaps (Expectations vs reality)

### Recurring expectation mismatches
1) **“Memory stored” does not guarantee “memory retrievable”**
   - Zero-vector embedding fallback creates the illusion of persistence without utility.
   - Action: document memory lifecycle states (queued → embedded → searchable) and expose metrics.

2) **“Plugin provides X” vs missing operational requirements**
   - MnemoPay’s stated goal (long-term economic memory) conflicts with in-memory-only implementation.
   - Action: require plugin submissions to declare:
     - persistence behavior,
     - growth/eviction policy,
     - tests/status badge.

3) **Strategic pivot away from messaging protocols**
   - Registry closed XMTP messaging issue to focus on AgentID; users may still expect messaging connectors to be a core roadmap item.
   - Action: publish a short “What we are/aren’t building” page: messaging connectors supported, but not new messaging protocols as core.

### Recurring questions implied by issues
- “Why does `elizaos create` fail and then delete everything?”
- “Why do my actions not receive required params in Discord connectors?”
- “Why is my agent replying in group chats when it shouldn’t (or staying silent)?”
- “How do I safely let agents trade or move funds?”

---

## 5) Community Engagement Insights

### Power users & maintainers (and what they need)
- **odilitime** (largest contributor this period): pushing core runtime changes (logging, routing, harness). Needs:
  - stronger CI gates for behavior changes (latency defaults, routing invariants),
  - clear backwards-compat policy for runtime options.
- **NubsCarson**: connector/schema correctness (TOON params, async action continuation). Needs:
  - connector conformance tests,
  - clearer schema contracts between runtime ↔ adapters.
- **dirtybits** (Issue #6704): newcomer experience + packaging correctness. Needs:
  - fast turnaround on install blockers,
  - “doctor” tooling and non-destructive scaffolding.

### Newcomer friction signals
- Install/bootstrap failures on macOS and Windows are classic “first hour” drop-off points.
- Plugin proposals indicate enthusiasm, but also that contributors need clearer contribution acceptance criteria (tests, persistence, security model).

### Converting passive users into contributors
- Create a “Plugin submission checklist” PR template (persistence/tests/security/limits).
- Label and curate “onboarding-blocker” issues (like #6704) with high visibility and quick response SLAs.
- Provide small, well-scoped tasks:
  - add missing tests for known bugs (e.g., aliasEntity case),
  - docs improvements (“Bun + elizaOS” guide),
  - connector metadata conformance matrix.

---

## 6) Feedback Collection Improvements

### Current channel effectiveness
- GitHub issues/PR reviews are producing actionable, technical feedback.
- Discord/community feedback is missing from this dataset for the day due to unavailable files; even when available, it tends to be less structured.

### Improvements to gather more structured feedback
1) **Add an “Installation Report” issue form**
   - Auto-captures OS/arch, bun/node versions, CLI version, logs, and whether project directory was deleted.
2) **Add a “Connector Behavior” issue form**
   - Captures connector type, encapsulation format (TOON/XML), expected vs actual action params, sample payload.
3) **Opt-in anonymous telemetry for CLI scaffolding**
   - Count failure points (without collecting user code), similar to many CLIs that report install step failures.

### Underrepresented segments whose feedback is missing
- Non-crypto agent builders (customer support, RPA, knowledge work) — current visible feedback skew is strongly “agent economy” oriented.
- Python/Rust runtime users — PRs mention regenerated prompts across languages, but direct user feedback from those ecosystems is absent.
- Teams deploying at scale (latency SLOs, multi-agent operations) — only hinted via timeout concerns.

---

## Prioritized High-Impact Actions (next 1–2 weeks)
1) **Fix and test core memory correctness issues** (IGNORE allowlist bug + embedding failure handling) and publish a brief “memory lifecycle” doc.  
2) **Unblock onboarding: patch CLI Bun dependency + make `elizaos create` non-destructive on failure**, plus add `elizaos doctor` (even a minimal version).  
3) **Stabilize dev workflow: enforce “fresh clone without submodules installs” in CI** and resolve workspace/lockfile consistency rules.  
4) **Ship connector conformance tests for action param propagation (TOON/XML) and group addressee routing**, including the `entityId ≠ agentId` case.  
5) **Define a first-class tool-call authorization interception API** (capability tiers + approvals) to support the dominant “agents as economic actors” usage pattern.