- banner, init hook, reply optimization, roles warnedUnnamedEntities, runtime (embedding cache, callerInfo, safeReplacer, onlyInclude, logPrompt/logResponse gating), logger file output - DISABLE_MEMORY_CREATION/ALLOW_MEMORY_SOURCE_IDS in message service; JSON5, parseBooleanFromText try/catch, formatPosts; add DISABLE_MEMORY_CREATION guard to runtime.evaluate() (was missed) - keepExistingResponses/BOOTSTRAP_KEEP_RESP, ANXIETY provider, null service guard, provider timeout, HandlerCallback actionName, hasRequestedInState in reply - Docs: CHANGELOG, ROADMAP, docs/DESIGN.md, README env/design section, WHY comments; fix ALLOW_MEMORY_SOURCE_IDS comment in message.ts - New: bootstrap banner (preview script), anxiety provider Made-with: Cursor # Relates to # Risks # Background ## What does this PR do? ## What kind of change is this? # Documentation changes needed? # Testing ## Where should a reviewer start? ## Detailed testing steps --- > [!NOTE] > **Medium Risk** > Touches core runtime/message pipeline behavior (provider timeouts, action-chain state recomposition, memory persistence gating, callback signature) and adds new logging side effects, so regressions could impact response latency, persistence, or integrations. > > **Overview** > **Core runtime hardening & observability.** `composeState` now enforces a per-provider 30s timeout (continuing with empty results on failure) and profiles provider timings; action chains recompose state using `onlyInclude` to refresh just `RECENT_MESSAGES`/`ACTION_STATE` while preserving other cached provider data; action-result caching uses `safeReplacer()`. > > **Logging & callbacks.** Adds opt-in file logging via `LOG_FILE` to write `output.log`, `prompts.log`, and `chat.log` (with ANSI stripping and prompt/response correlation), and extends `HandlerCallback` to accept an optional `actionName` (runtime and bootstrap `reply` now pass it). > > **Memory/pipeline controls & bootstrap behavior.** Implements `DISABLE_MEMORY_CREATION` with `ALLOW_MEMORY_SOURCE_IDS` allowlisting and updates race handling to use resolved `keepExistingResponses`; skips memory-dependent evaluators when memory creation is disabled. Adds a bootstrap startup banner (plugin `init` hook + preview script), a new `ANXIETY` provider (registered by default), budget-based truncation of action history/results via shared `sliceToFitBudget`, and reduces roles-provider warning spam with per-runtime bounded tracking. Documentation and examples are updated accordingly (README/CHANGELOG/ROADMAP/DESIGN, Telegram model selection, JSON5 dependency). > > ^{Written by [Cursor Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit d4e24be84caf81f419a4a5dd4b50652daf168a0e. This will update automatically on new commits. Configure [here](https://cursor.com/dashboard?tab=bugbot).} ## Summary by CodeRabbit * **New Features** * Memory persistence controls (DISABLE_MEMORY_CREATION, ALLOW_MEMORY_SOURCE_IDS), shared config helpers, new anxiety provider, startup banner, and prompt/response file logging. * **Improvements** * Per-provider timeouts and a 30s provider limit, first-action reply optimization, budgeted truncation of action/history context, safer JSON parsing from text, and handler callbacks include action names. * **Bug Fixes** * Explicit timer cleanup on provider completion and hardened plugin registration. * **Tests & Docs** * Expanded tests and new design/roadmap/README documentation.

Greptile Summary

This PR is a broad hardening and observability pass on the core `@elizaos/typescript` runtime. Key changes include: per-provider 30 s timeouts in `composeState`, `onlyInclude` optimization for action-chain state recomposition, `DISABLE_MEMORY_CREATION` / `ALLOW_MEMORY_SOURCE_IDS` controls in both the message service and `evaluate()`, optional file logging (`LOG_FILE` → `output.log` / `prompts.log` / `chat.log`), an `ANXIETY` verbosity-guidance provider registered by default, a bootstrap ANSI banner on plugin init, JSON5-tolerant LLM output parsing, and `HandlerCallback` extended with an optional `actionName` parameter. - **Bug — IGNORE response persistence bypassed by `ALLOW_MEMORY_SOURCE_IDS`** (`message.ts:948–952`): the ignore path checks `allowedSources.includes("agent_response")` — a hardcoded string users will never configure — so all IGNORE decisions are silently dropped when any source allowlist is active. The regular response path (line ~773) correctly skips this check; the ignore path should match that behavior. - **Bug — zero-vector fallback corrupts semantic memory** (`runtime.ts:4861`): on embedding failure the memory is persisted with a zero vector, making it unretrievable by cosine-similarity search while appearing normally stored. Silent data loss. - **`PROVIDERS_TOTAL_TIMEOUT_MS` default raised from 1 000 ms → 5 000 ms** (`message.ts:~2019`): existing deployments relying on the 1 s default will now allow providers up to 5 s before bailing, which may noticeably increase P99 response latency. - **Variable shadowing in ignore block** (`message.ts:949`): `allowedSources` is re-resolved inside `if (!disableMemoryCreation)`, shadowing the binding already computed on line 461 and adding a redundant runtime call. - **Dead `catch` block in `parseBooleanFromText`** (`utils.ts:941–950`): `String(value).trim().toUpperCase()` cannot throw after the early-return guards; the `logger.warn` is unreachable.

Confidence Score: 3/5

- Not safe to merge without addressing the IGNORE memory persistence bug and zero-vector embedding fallback. - Two confirmed logic bugs affect correctness in production: IGNORE responses are silently dropped from memory when ALLOW_MEMORY_SOURCE_IDS is set (contradicts stated design and the regular response path), and failed embeddings produce unretrievable zero-vector memories. The PROVIDERS_TOTAL_TIMEOUT_MS default change is a silent breaking change for latency-sensitive deployments. The rest of the PR is well-structured, well-commented, and tested. - packages/typescript/src/services/message.ts (IGNORE persistence logic at lines 948–974) and packages/typescript/src/runtime.ts (zero-vector embedding fallback at lines 4858–4865)

Important Files Changed

| Filename | Overview | |----------|----------| | packages/typescript/src/services/message.ts | Core message processing pipeline — adds DISABLE_MEMORY_CREATION / ALLOW_MEMORY_SOURCE_IDS filtering and keepExistingResponses race-check; the IGNORE path incorrectly applies the source allowlist to agent-generated responses, causing all IGNORE memories to be silently dropped when ALLOW_MEMORY_SOURCE_IDS is configured (contradicts stated design). Also contains variable shadowing and redundant comments. | | packages/typescript/src/runtime.ts | Major runtime hardening: per-provider 30s timeouts, onlyInclude optimization for action-chain state recomposition, DISABLE_MEMORY_CREATION guard for evaluate(), file-based prompt/response logging, caller-info stack traces. Zero-vector embedding fallback silently corrupts memory search for affected records. | | packages/typescript/src/logger.ts | Adds optional file logging (output.log, prompts.log, chat.log) gated by LOG_FILE env-var; prompt/response correlation now uses the same slug passed through metadata; all log files are placed in the same directory as LOG_FILE. Implementation is sound. | | packages/typescript/src/bootstrap/providers/anxiety.ts | New ANXIETY provider returning channel-specific verbosity guidance via random selection of 3 examples per turn; uses Math.random()-0.5 sort (biased shuffle) but non-critical for this use case. | | packages/typescript/src/bootstrap/providers/roles.ts | Switches from a module-level Set to a WeakMap> for per-runtime warn-once tracking of unnamed entities; correctly addressed from previous review. | | packages/typescript/src/utils/json-llm.ts | New JSON5-based tolerant extraction helper for LLM output; handles code-fenced blocks and single-quote / trailing-comma JSON; throws on failure so callers (parseJSONObjectFromText) can catch and return null. | | packages/typescript/src/utils.ts | parseBooleanFromText accepts broader types, formatPosts adds metadata fallbacks and text delimiters, parseJSONObjectFromText delegates to JSON5-based json-llm helper. The try-catch in parseBooleanFromText is dead code but harmless. | | packages/typescript/src/types/components.ts | HandlerCallback extended with optional actionName parameter (backward compatible); allows callers to attribute responses to the producing action without content parsing. |

Flowchart

```mermaid %%{init: {'theme': 'neutral'}}%% flowchart TD A[Incoming Message] --> B{DISABLE_MEMORY_CREATION?} B -- No --> C{ALLOW_MEMORY_SOURCE_IDS set?} C -- No allowlist --> D[Persist message memory + queue embedding] C -- Allowlist set --> E{sourceId in allowlist?} E -- Yes --> D E -- No --> F[Skip persistence, assign UUID] B -- Yes --> F D --> G[shouldRespond check] F --> G G -- IGNORE --> H[callback ignoreContent] H --> I{DISABLE_MEMORY_CREATION?} I -- No --> J{allowedSources includes 'agent_response'?} J -- Yes --> K[Persist IGNORE memory] J -- No → bug --> L["❌ IGNORE memory silently dropped"] I -- Yes --> L2[Skip IGNORE memory] G -- Respond --> M[runSingleShotCore / LLM] M --> N{canPersistMemory = !disableMemory} N -- true --> O[Persist response memories] N -- false --> P[Skip response memories] O --> Q[run evaluators] P --> Q Q --> R{DISABLE_MEMORY_CREATION?} R -- Yes --> S[Skip evaluators entirely] R -- No --> T[Run evaluators with timeout] style L fill:#ff4444,color:#fff style L2 fill:#aaa,color:#fff ``` _{Last reviewed commit: ["docs: add review dis..."](https://github.com/elizaos/eliza/commit/d7d1ad5199b8fe12b302326d321ce50fede6912a)} > Greptile also left **2 inline comments** on this PR.

elizaOS