@mastra/mcp-docs-server 1.2.13-alpha.0 → 1.2.13-alpha.11

package/.docs/reference/evals/summarization.md

@@ -0,0 +1,203 @@
+> Discover all available pages from the documentation index: https://mastra.ai/llms.txt
+
+# Summarization scorer
+
+The `createSummarizationScorer()` function creates a scorer that evaluates a summary on two axes: whether every claim it makes is supported by the source text, and whether it preserves the information the source states. The final score is the lower of the two, so a summary can't pass by being faithful but empty, or thorough but wrong.
+
+The summary is the agent's last message that carries text, and the source text defaults to the first user message of the run input. Pass `source` or `sourceExtractor` when the text being summarized lives somewhere else, such as a tool result.
+
+## Usage example
+
+Score a summary against the document it condenses.
+
+```typescript
+import { createSummarizationScorer } from '@mastra/evals/scorers/prebuilt'
+
+const scorer = createSummarizationScorer({
+  model: 'openai/gpt-5.6-sol',
+})
+
+const result = await scorer.run({
+  input: {
+    inputMessages: [{ id: '1', role: 'user', content: sourceDocument }],
+  },
+  output: [{ id: '2', role: 'assistant', content: summary }],
+})
+
+console.log(result.score)
+console.log(result.reason)
+```
+
+## Summarization evaluation
+
+Use this scorer when an agent condenses text:
+
+- Document and transcript summarization
+- Support thread and email digests
+- Any step that compresses a long input into a short output
+
+## Parameters
+
+**model** (`MastraModelConfig`): The language model to use for judging claims and coverage questions
+
+**options** (`SummarizationMetricOptions`): Configuration options for the scorer
+
+**options.source** (`string`): Text the summary is judged against. Defaults to the user message of the run input
+
+**options.sourceExtractor** (`(input, output) => string`): Function to derive the source text from the run input and output. Takes precedence over source
+
+**options.maxQuestions** (`number`): Upper bound on the coverage questions drawn from the source (default: 10)
+
+**options.scale** (`number`): Scale factor to multiply the final score (default: 1)
+
+## `.run()` returns
+
+**score** (`number`): Summarization score between 0 and scale (default 0-1), the lower of the alignment and coverage scores
+
+**reason** (`string`): Human-readable explanation naming the axis that produced the score and the claims or questions behind it. Both axis scores appear in the text
+
+**preprocessStepResult** (`object`): The alignment verdicts and the questions drawn from the source
+
+**preprocessStepResult.alignment** (`{ claim: string; supported: boolean; reason: string }[]`): One verdict per claim the summary makes
+
+**preprocessStepResult.questions** (`string[]`): The coverage questions drawn from the source text
+
+**analyzeStepResult** (`object`): The coverage verdicts
+
+**analyzeStepResult.coverage** (`{ question: string; answered: boolean; reason: string }[]`): One verdict per question, answered from the summary alone
+
+The axis scores are derived from these verdicts rather than stored: alignment is the share of `alignment` entries with `supported: true`, and coverage is the share of `questions` whose `coverage` entry has `answered: true`.
+
+## Scoring details
+
+### Two-axis evaluation
+
+The scorer runs a three-step pipeline:
+
+1. **Source judgement**: the claims the summary makes are extracted and checked against the source, and closed-ended questions are drawn from the source. Every question is written so the source answers it "yes".
+2. **Coverage**: each question is answered using the summary alone.
+3. **Scoring**: the two ratios are computed and the lower one becomes the score.
+
+The coverage step runs as a separate model call that never receives the source text. A judge that could see the source would answer questions from it rather than from the summary, which would hide the omissions this axis exists to measure.
+
+### Scoring formula
+
+```text
+Alignment    = supported_claims / total_claims
+Coverage     = answered_questions / total_questions
+Summarization = min(Alignment, Coverage) × scale
+```
+
+The score is 0 when the summary yields no claims or the source yields no questions.
+
+### Score interpretation
+
+These ranges assume the default `scale` of 1. When using a custom scale, multiply accordingly.
+
+- **0.9-1.0**: Excellent summary, faithful to the source and covering its main points
+- **0.7-0.8**: Good summary with a small omission or an unsupported detail
+- **0.4-0.6**: Moderate summary, either missing important information or drifting from the source
+- **0.1-0.3**: Poor summary, most of the source is lost or contradicted
+- **0.0**: The summary produced nothing to judge, or it failed to support any claims. A summary that answers no questions also receives this score
+
+### Reading the two axes
+
+Both axes leave their verdicts on the run result: the alignment verdicts on the preprocess step, and the coverage verdicts on the analyze step. Each verdict carries the claim or question it belongs to and the reason behind it. A low alignment score has a different meaning from a low coverage score:
+
+- A low alignment score with high coverage means the summary invents or distorts detail
+- A low coverage score with high alignment means the summary is accurate but leaves too much out
+
+The reason field names whichever axis produced the score.
+
+### What the score leaves out
+
+Length plays no part in the score. A summary that repeats the source word for word supports every claim and answers every question, so it scores 1. Add a length check of your own when compression is part of what you're testing.
+
+### Cost
+
+Each evaluation makes three model calls. `maxQuestions` bounds the coverage half of the work, which otherwise grows with source length. Raise it for long documents where ten questions can't represent the content.
+
+## Scorer configuration
+
+### Summarizing the run input
+
+```typescript
+const scorer = createSummarizationScorer({
+  model: 'openai/gpt-5.6-sol',
+})
+```
+
+### Summarizing a document from elsewhere
+
+```typescript
+import { extractToolResults } from '@mastra/evals/scorers/utils'
+
+const scorer = createSummarizationScorer({
+  model: 'openai/gpt-5.6-sol',
+  options: {
+    sourceExtractor: (input, output) => {
+      return extractToolResults(output)
+        .filter(({ toolName }) => toolName === 'fetchDocument')
+        .map(({ result }) => String(result))
+        .join('\n\n')
+    },
+    maxQuestions: 20,
+  },
+})
+```
+
+## Example
+
+Evaluate a summarization agent against a set of documents:
+
+```typescript
+import { runEvals } from '@mastra/core/evals'
+import { createSummarizationScorer } from '@mastra/evals/scorers/prebuilt'
+import { summarizerAgent } from './agent'
+
+const scorer = createSummarizationScorer({
+  model: 'openai/gpt-5.6-sol',
+  options: { maxQuestions: 10 },
+})
+
+const result = await runEvals({
+  target: summarizerAgent,
+  scorers: [scorer],
+  data: [
+    {
+      input:
+        'The company was founded in 1995 by John Smith. It started with 10 employees and grew to 500 by 2020. The company is based in Seattle.',
+    },
+  ],
+  onItemComplete: ({ scorerResults }) => {
+    console.log({
+      score: scorerResults[scorer.id].score,
+      reason: scorerResults[scorer.id].reason,
+    })
+  },
+})
+
+console.log(result.scores)
+```
+
+For more details on `runEvals`, see the [runEvals reference](https://mastra.ai/reference/evals/run-evals).
+
+To add this scorer to an agent, see the [Scorers overview](https://mastra.ai/docs/evals/overview) guide.
+
+## Comparison with faithfulness
+
+| Use case                  | Summarization                   | Faithfulness                      |
+| ------------------------- | ------------------------------- | --------------------------------- |
+| **What it measures**      | Support and coverage together   | Support only                      |
+| **Judged against**        | The source text being condensed | Retrieved context or tool results |
+| **Catches omission**      | Yes                             | No                                |
+| **Needs the full source** | Yes                             | No, context alone is enough       |
+
+Use `faithfulness` when the question is whether an answer stays grounded in retrieved context. Use `summarization` when the output is meant to stand in for a longer text.
+
+## Related
+
+- [Faithfulness Scorer](https://mastra.ai/reference/evals/faithfulness): Measures answer groundedness in context
+- [Completeness Scorer](https://mastra.ai/reference/evals/completeness): Compares element coverage without a model
+- [Content Similarity Scorer](https://mastra.ai/reference/evals/content-similarity): Compares text similarity without a model
+- [Custom Scorers](https://mastra.ai/docs/evals/custom-scorers): Creating your own evaluation metrics

package/.docs/reference/file-based-agents/config.md

@@ -94,4 +94,42 @@ Keep `config.ts` focused on runtime options. Use sibling files for concerns that
 | Memory       | `config.memory`               | [`memory.ts`](https://mastra.ai/reference/file-based-agents/memory)             | `config.memory`                                         |
 | Workspace    | `config.workspace`            | [`workspace.ts`](https://mastra.ai/reference/file-based-agents/workspace)       | `config.workspace`                                      |
 
-Missing both `instructions.md` and `config.instructions` fails the build. Missing both `config.memory` and `memory.ts` leaves the agent without memory.
+Missing both `instructions.md` and `config.instructions` fails the build. Missing both `config.memory` and `memory.ts` leaves the agent without memory.
+
+## Discovery lifecycle
+
+File-based primitives are discovered by the Mastra bundler under `mastra dev` and `mastra build`. During discovery, Mastra reads supported files under `src/mastra/`, imports TypeScript and JavaScript modules, reads markdown instructions and skills, copies workspace seed files, and registers the assembled primitives with your Mastra app.
+
+After discovery, a file-based agent runs as a normal [`Agent`](https://mastra.ai/reference/agents/agent). Calling it from the Agent API, Studio, workflows, or your application code uses the same runtime as a code-defined agent.
+
+Discovery is source-based and conservative.
+
+It skips symlinks, test files, and directories that aren't agent directories. Workflows and singleton project files are file-routed only when they have a default export.
+
+Start your app through the Mastra CLI so discovery runs:
+
+**npm**:
+
+```bash
+npx mastra dev
+```
+
+**pnpm**:
+
+```bash
+pnpm dlx mastra dev
+```
+
+**Yarn**:
+
+```bash
+yarn dlx mastra dev
+```
+
+**Bun**:
+
+```bash
+bun x mastra dev
+```
+
+If you import your `mastra` instance directly, `agents/<name>/` directories and the other conventions aren't discovered. When you consume Mastra as a library, register those primitives in code instead.

package/.docs/reference/index.md

@@ -65,7 +65,6 @@ The Reference section provides documentation of Mastra's API, including paramete
 - [SlackProvider](https://mastra.ai/reference/channels/slack-provider)
 - [create-mastra](https://mastra.ai/reference/cli/create-mastra)
 - [mastra](https://mastra.ai/reference/cli/mastra)
-- [Agent Builder API](https://mastra.ai/reference/client-js/agent-builder)
 - [Agents API](https://mastra.ai/reference/client-js/agents)
 - [Conversations API](https://mastra.ai/reference/client-js/conversations)
 - [Error Handling](https://mastra.ai/reference/client-js/error-handling)
@@ -123,10 +122,10 @@ The Reference section provides documentation of Mastra's API, including paramete
 - [BlobStoreProvider](https://mastra.ai/reference/editor/blob-store-provider)
 - [MastraEditor Class](https://mastra.ai/reference/editor/mastra-editor)
 - [ProcessorProvider](https://mastra.ai/reference/editor/processor-provider)
+- [Prompt blocks](https://mastra.ai/reference/editor/prompt-blocks)
+- [Tool configuration](https://mastra.ai/reference/editor/tools)
 - [ToolProvider](https://mastra.ai/reference/editor/tool-provider)
-- [AgentBuilderOptions](https://mastra.ai/reference/editor/agent-builder/agent-builder-options)
-- [BuilderAgentDefaults](https://mastra.ai/reference/editor/agent-builder/builder-agent-defaults)
-- [Models default](https://mastra.ai/reference/editor/agent-builder/builder-models)
+- [Versioning](https://mastra.ai/reference/editor/versioning)
 - [BrowserProvider](https://mastra.ai/reference/editor/browser-provider)
 - [StorageBrowserRef](https://mastra.ai/reference/editor/storage-browser-ref)
 - [FilesystemProvider](https://mastra.ai/reference/editor/filesystem-provider)
@@ -152,6 +151,7 @@ The Reference section provides documentation of Mastra's API, including paramete
 - [Noise Sensitivity Scorer](https://mastra.ai/reference/evals/noise-sensitivity)
 - [Prompt Alignment Scorer](https://mastra.ai/reference/evals/prompt-alignment)
 - [Rubric Scorer](https://mastra.ai/reference/evals/rubric)
+- [Summarization Scorer](https://mastra.ai/reference/evals/summarization)
 - [Textual Difference Scorer](https://mastra.ai/reference/evals/textual-difference)
 - [Tone Consistency Scorer](https://mastra.ai/reference/evals/tone-consistency)
 - [Tool Call Accuracy Scorers](https://mastra.ai/reference/evals/tool-call-accuracy)

package/.docs/reference/memory/observational-memory.md

@@ -27,7 +27,7 @@ export const agent = new Agent({
 
 ## Configuration
 
-The `observationalMemory` option accepts `true`, a configuration object, or `false`. Setting `true` enables OM with `google/gemini-2.5-flash` as the default model. When passing a config object, a `model` must be explicitly set: either at the top level or on `observation.model` and/or `reflection.model`.
+The `observationalMemory` option accepts `true`, a configuration object, or `false`. Setting `true` enables OM with `google/gemini-2.5-flash` as the default model. When passing a config object, set `model` at the top level or on `observation.model` and/or `reflection.model`; when all model fields are omitted, OM falls back to `google/gemini-2.5-flash`.
 
 Observer input is multimodal-aware. OM keeps text placeholders like `[Image #1: screenshot.png]` in the transcript it builds for the Observer, and also sends the underlying image parts when possible. This applies to both single-thread observation and batched multi-thread observation. Non-image files appear as placeholders only.
 
@@ -35,7 +35,7 @@ OM performs thresholding with fast local token estimation. Text uses `tokenx`, a
 
 **enabled** (`boolean`): Enable or disable Observational Memory. When omitted from a config object, defaults to true. Only enabled: false explicitly disables it. (Default: `true`)
 
-**model** (`string | LanguageModel | DynamicModel | ModelByInputTokens | ModelWithRetries[]`): Model for both the Observer and Reflector agents. Sets the model for both at once. Cannot be used together with observation.model or reflection.model — an error will be thrown if both are set. When using observationalMemory: true, defaults to google/gemini-2.5-flash. When passing a config object, this or observation.model/reflection.model must be set. Use "default" to explicitly use the default model (google/gemini-2.5-flash). (Default: `'google/gemini-2.5-flash' (when using observationalMemory: true)`)
+**model** (`string | LanguageModel | DynamicModel | ModelByInputTokens | ModelWithRetries[]`): Model for both the Observer and Reflector agents. Sets the model for both at once. Cannot be used together with observation.model or reflection.model — an error will be thrown if both are set. When this and observation.model/reflection.model are all omitted, OM falls back to google/gemini-2.5-flash. Use "default" to explicitly use the default model (google/gemini-2.5-flash). (Default: `'google/gemini-2.5-flash'`)
 
 **scope** (`'resource' | 'thread'`): Memory scope for observations. 'thread' keeps observations per-thread. 'resource' (experimental) shares observations across all threads for a resource, enabling cross-conversation memory. (Default: `'thread'`)
 
@@ -47,7 +47,9 @@ OM performs thresholding with fast local token estimation. Text uses `tokenx`, a
 
 **temporalMarkers** (`boolean`): Insert temporal-gap reminder markers before new user messages when the previous message in the thread is at least 10 minutes older. The marker is persisted in memory, emitted as an inline reminder event so clients can render it specially, and shown to the observer so it can anchor observations to when events occurred. (Default: `false`)
 
-**retrieval** (`boolean | { vector?: boolean; scope?: 'thread' | 'resource' }`): Enable retrieval-mode observation groups as durable pointers to raw message history. true enables cross-thread browsing by default. { vector: true } also enables semantic search using Memory's vector store and embedder. { scope: 'thread' } restricts the recall tool to the current thread only. Default scope is 'resource'. (Default: `false`)
+**retrieval** (`boolean | { vector?: boolean; scope?: 'thread' | 'resource'; instructions?: string }`): Let the agent look up the raw message history behind its observations. Observation groups keep durable pointers to the original messages, and a recall tool is registered so the agent can browse them. true enables cross-thread browsing by default. { vector: true } also enables semantic search using Memory's vector store and embedder. { scope: 'thread' } restricts the recall tool to the current thread only. Default scope is 'resource'. { instructions: '...' } appends application-specific recall guidance after Mastra's built-in retrieval instructions. (Default: `false`)
+
+**hooks** (`ObserveHooks`): Lifecycle hooks fired for every observation/reflection cycle — the manual observe()/reflect() APIs, turn-driven synchronous observation, and fire-and-forget async buffering. Callbacks receive threadId/resourceId/trigger call context ('manual' | 'turn-sync' | 'async-buffer'), and the end hooks (onObservationEnd/onReflectionEnd) additionally receive the OM model call's token usage and providerMetadata (where providers such as the AI Gateway report per-call cost), so apps can account for OM model spend without wrapping the observer/reflector models in middleware. Failed async-buffered cycles never throw; they report through the end hook's error field. Errors thrown by these hooks are caught and logged — they never fail the cycle.
 
 **observation** (`ObservationalMemoryObservationConfig`): Configuration for the observation step. Controls when the Observer agent runs and how it behaves.
 
@@ -59,29 +61,33 @@ OM performs thresholding with fast local token estimation. Text uses `tokenx`, a
 
 **observation.extract** (`Extractor[]`): Custom values to extract after observation. Schema-less extractors are requested inline in the Observer output. Schema-backed extractors run as a follow-up structured output call and are stored in thread OM metadata.
 
-**observation.observeAttachments** (`boolean | string[]`): Controls which image/file attachments are forwarded to the Observer model alongside their placeholder text lines. true (default) forwards all attachments. false drops all attachments while keeping placeholders visible. An array is a case-insensitive mimeType allowlist supporting exact matches ('application/pdf'), wildcard subtypes ('image/\*'), and bare '\*' for everything. Useful when the Observer model is text-only (e.g. some DeepSeek endpoints) while the main agent uses a multimodal model. Tool-result attachments are filtered using the same rule.
+**observation.manageWorkingMemory** (`boolean`): Let the Observer manage working memory through OM extraction. Adds WorkingMemoryExtractor, defaults workingMemory.agentManaged to false, and defaults workingMemory.useStateSignals to true. See Working memory updates.
+
+**observation.observeAttachments** (`'auto' | boolean | string[]`): Controls which image/file attachments are forwarded to the Observer model alongside their placeholder text lines. true (default) forwards all attachments. false drops all attachments while keeping placeholders visible. 'auto' uses the provider capabilities registry to decide: attachments are forwarded when the Observer model supports multimodal input, dropped otherwise, and forwarded when no capability data is available for the model. An array is a case-insensitive mimeType allowlist supporting exact matches ('application/pdf'), wildcard subtypes ('image/\*'), and bare '\*' for everything. Useful when the Observer model is text-only (e.g. some DeepSeek endpoints) while the main agent uses a multimodal model. Tool-result attachments are filtered using the same rule.
 
 **observation.messageTokens** (`number`): Token count of unobserved messages that triggers observation. When unobserved message tokens exceed this threshold, the Observer agent is called. Text is estimated locally with tokenx. Image parts are included with model-aware heuristics when possible, with deterministic fallbacks when image metadata is incomplete. Image-like file parts are counted the same way when uploads are normalized as files.
 
 **observation.maxTokensPerBatch** (`number`): Maximum tokens per batch when observing multiple threads in resource scope. Threads are chunked into batches of this size and processed in parallel. Lower values mean more parallelism but more API calls.
 
-**observation.modelSettings** (`ObservationalMemoryModelSettings`): Model settings for the Observer agent.
+**observation.modelSettings** (`ObservationalMemoryModelSettings`): Model settings for the Observer agent. The maxOutputTokens: 100\_000 default is only applied with default model selection (no model set, "default", or a ModelByInputTokens selector). Custom models get no maxOutputTokens default.
 
 **observation.modelSettings.temperature** (`number`): Temperature for generation. Lower values produce more consistent output.
 
-**observation.modelSettings.maxOutputTokens** (`number`): Maximum output tokens. Set high to prevent truncation of observations.
+**observation.modelSettings.maxOutputTokens** (`number`): Maximum output tokens. Set high to prevent truncation of observations. The 100000 default is only applied with default model selection; custom models get no default.
 
-**observation.bufferTokens** (`number | false`): Token interval for async background observation buffering. Can be an absolute token count (e.g. 5000) or a fraction of messageTokens (e.g. 0.25 = buffer every 25% of threshold). When set, observations run in the background at this interval, storing results in a buffer. When the main messageTokens threshold is reached, buffered observations activate instantly without a blocking LLM call. Must resolve to less than messageTokens. Set to false to explicitly disable all async buffering (both observation and reflection).
+**observation.providerOptions** (`ProviderOptions`): Provider-specific options passed to the Observer agent, such as Google thinking configuration.
+
+**observation.bufferTokens** (`number | false`): How often background observation buffering runs. Values between 0 and 1 are fractions of messageTokens: 0.25 buffers every 25% of the threshold (7.5k tokens with the default 30k). Values of 1 or more are absolute token counts: 5000 buffers every 5k tokens. Buffered observations are stored until the messageTokens threshold is reached, then activate instantly without a blocking LLM call. Must resolve to less than messageTokens. Set to false to disable all async buffering (both observation and reflection).
 
 **observation.bufferOnIdle** (`boolean`): Run background observation buffering when an agent turn ends and the agent becomes idle. This is separate from bufferTokens, which controls step-time async buffering. Set this to true to buffer short idle turns without waiting for the next turn or the messageTokens threshold.
 
-**observation.bufferActivation** (`number`): Controls how much of the message window to retain after activation. Accepts a ratio (0-1) or an absolute token count (≥ 1000). For example, 0.8 means: activate enough buffers to remove 80% of messageTokens and leave 20% as active message history. An absolute token count like 4000 targets a goal of keeping \~4k message tokens remaining after activation. Higher values remove more message history per activation when using a ratio. Higher values keep more message history when using a token count.
+**observation.bufferActivation** (`number`): How much of the message window to clear when buffered observations activate. Values between 0 and 1 are the fraction of messageTokens to remove: 0.8 removes \~80% of the message history and keeps \~20% (6k tokens with the default 30k). Values of 1000 or more are the token count to keep: 4000 keeps \~4k message tokens after activation. Note the direction flips: a higher ratio removes more history, while a higher token count keeps more.
 
-**observation.activateAfterIdle** (`number | string | false | "auto"`): Time before buffered observations are forced to activate after inactivity. Accepts milliseconds, a duration string, "auto" for a provider-aware prompt cache TTL, or false. If unset, the top-level activateAfterIdle value is used for observations. Set false to disable the top-level idle setting for observations.
+**observation.activateAfterIdle** (`number | string | false | "auto"`): Time before buffered observations are forced to activate after inactivity. Accepts milliseconds, a duration string, "auto" for a provider-aware prompt cache TTL, or false. If unset, the top-level activateAfterIdle value is used for observations. Set false to disable the top-level idle setting for observations. Currently only applied when using the standalone ObservationalMemory class; new Memory(...) applies the top-level activateAfterIdle only.
 
-**observation.activateOnProviderChange** (`boolean`): Force buffered observations to activate when the actor provider or model changes. If unset, the top-level activateOnProviderChange value is used for observations.
+**observation.activateOnProviderChange** (`boolean`): Force buffered observations to activate when the actor provider or model changes. If unset, the top-level activateOnProviderChange value is used for observations. Currently only applied when using the standalone ObservationalMemory class; new Memory(...) applies the top-level activateOnProviderChange only.
 
-**observation.blockAfter** (`number`): Token threshold above which synchronous (blocking) observation is forced. Between messageTokens and blockAfter, only async buffering/activation is used. Above blockAfter, a synchronous observation runs as a last resort, while buffered activation still preserves a minimum remaining context (min(1000, retention floor)). Accepts a multiplier (1 < value < 2, multiplied by messageTokens) or an absolute token count (≥ 2, must be greater than messageTokens). Only relevant when bufferTokens is set. Defaults to 1.2 when async buffering is enabled.
+**observation.blockAfter** (`number`): Safety net that forces a synchronous (blocking) observation when background buffering can't keep up. Values from 1 up to (but not including) 100 are multipliers of messageTokens: 1.2 forces a blocking observation at 120% of the threshold (36k tokens with the default 30k). Values of 100 or more are absolute token counts and must be greater than messageTokens. Between messageTokens and blockAfter, only async buffering and activation run; buffered activation still preserves a minimum remaining context (the smaller of 1000 tokens or the retention floor). Only relevant when bufferTokens is set. Defaults to 1.2 when async buffering is enabled.
 
 **observation.previousObserverTokens** (`number | false`): Optional token budget for the observer's previous-observations context. When set to a number, the observations passed to the Observer agent are tail-truncated to fit within this budget while keeping the newest observations and preserving highlighted 🔴 items when possible. When a buffered reflection is pending, the already-reflected observation lines are automatically replaced with the reflection summary before truncation. Set to 0 to omit previous observations entirely, or false to disable truncation explicitly.
 
@@ -95,19 +101,21 @@ OM performs thresholding with fast local token estimation. Text uses `tokenx`, a
 
 **reflection.observationTokens** (`number`): Token count of observations that triggers reflection. When observation tokens exceed this threshold, the Reflector agent is called to condense them.
 
-**reflection.modelSettings** (`ObservationalMemoryModelSettings`): Model settings for the Reflector agent.
+**reflection.modelSettings** (`ObservationalMemoryModelSettings`): Model settings for the Reflector agent. The maxOutputTokens: 100\_000 default is only applied with default model selection (no model set, "default", or a ModelByInputTokens selector). Custom models get no maxOutputTokens default.
 
 **reflection.modelSettings.temperature** (`number`): Temperature for generation. Lower values produce more consistent output.
 
-**reflection.modelSettings.maxOutputTokens** (`number`): Maximum output tokens. Set high to prevent truncation of observations.
+**reflection.modelSettings.maxOutputTokens** (`number`): Maximum output tokens. Set high to prevent truncation of observations. The 100000 default is only applied with default model selection; custom models get no default.
+
+**reflection.providerOptions** (`ProviderOptions`): Provider-specific options passed to the Reflector agent, such as Google thinking configuration.
 
-**reflection.bufferActivation** (`number`): Ratio (0-1) controlling when async reflection buffering starts. When observation tokens reach observationTokens \* bufferActivation, reflection runs in the background. On activation at the full threshold, the buffered reflection replaces the observations it covers, preserving any new observations appended after that range.
+**reflection.bufferActivation** (`number`): When background reflection starts, as a ratio (0-1) of observationTokens: 0.5 starts reflecting in the background once observations reach 50% of the threshold (20k tokens with the default 40k). When the full threshold is reached, the buffered reflection replaces the observations it covers, preserving any new observations appended after that range.
 
-**reflection.activateAfterIdle** (`number | string | false | "auto"`): Time before buffered reflections are forced to activate after inactivity. Accepts milliseconds, a duration string, "auto" for a provider-aware prompt cache TTL, or false. Reflections do not inherit top-level activateAfterIdle; set this explicitly to opt reflections into idle activation.
+**reflection.activateAfterIdle** (`number | string | false | "auto"`): Time before buffered reflections are forced to activate after inactivity. Accepts milliseconds, a duration string, "auto" for a provider-aware prompt cache TTL, or false. Reflections do not inherit top-level activateAfterIdle; set this explicitly to opt reflections into idle activation. Currently only applied when using the standalone ObservationalMemory class; this setting has no effect through new Memory(...).
 
-**reflection.activateOnProviderChange** (`boolean`): Force buffered reflections to activate when the actor provider or model changes. Reflections do not inherit top-level activateOnProviderChange; set this explicitly to opt reflections into provider-change activation.
+**reflection.activateOnProviderChange** (`boolean`): Force buffered reflections to activate when the actor provider or model changes. Reflections do not inherit top-level activateOnProviderChange; set this explicitly to opt reflections into provider-change activation. Currently only applied when using the standalone ObservationalMemory class; this setting has no effect through new Memory(...).
 
-**reflection.blockAfter** (`number`): Token threshold above which synchronous (blocking) reflection is forced. Between observationTokens and blockAfter, only async buffering/activation is used. Above blockAfter, a synchronous reflection runs as a last resort. Accepts a multiplier (1 < value < 2, multiplied by observationTokens) or an absolute token count (≥ 2, must be greater than observationTokens). Only relevant when bufferActivation is set. Defaults to 1.2 when async reflection is enabled.
+**reflection.blockAfter** (`number`): Safety net that forces a synchronous (blocking) reflection when background reflection can't keep up. Values from 1 up to (but not including) 100 are multipliers of observationTokens: 1.2 forces a blocking reflection at 120% of the threshold (48k tokens with the default 40k). Values of 100 or more are absolute token counts and must be greater than observationTokens. Between observationTokens and blockAfter, only async buffering and activation run. Only relevant when bufferActivation is set. Defaults to 1.2 when async reflection is enabled.
 
 ### Token estimate metadata cache
 
@@ -150,7 +158,7 @@ const memory = new Memory({
 
 **name** (`string`): Human-readable extractor name. OM slugifies this value into the extractor slug. Names must be unique after slug generation.
 
-**slug** (`string`): Generated stable identifier for persisted values and XML tags. Slugs use lowercase letters, numbers, and hyphens. Built-in slugs and reserved XML tags cannot be used by custom extractors.
+**slug** (`string`): Read-only property derived from name — not a constructor option. Generated stable identifier for persisted values and XML tags. Slugs use lowercase letters, numbers, and hyphens. Built-in slugs and reserved XML tags cannot be used by custom extractors.
 
 **instructions** (`string | (context) => string`): Instructions for what to extract and when to update the value. Use a function to derive instructions from runtime context.
 
@@ -158,6 +166,8 @@ const memory = new Memory({
 
 **includePreviousExtraction** (`boolean`): Controls whether the previous extraction is shown to the extractor on future OM runs. Set to false for values that should only come from the current OM run. (Default: `true`)
 
+**metadataKeyPath** (`string | false`): Dot-separated OM metadata path used to persist the extracted value. Set to false to skip OM metadata persistence entirely. (Default: `'extracted.<slug>'`)
+
 **onExtracted** (`(context) => T | void | Promise<T | void>`): Optional hook called after a custom extractor returns a value and before metadata is persisted. Returning a value replaces the extracted value. Throwing records an extraction failure.
 
 ### Extraction behavior
@@ -663,18 +673,45 @@ Emitted when buffered observations or reflections are activated (moved into the
 
 **observations** (`string`): The activated observations text.
 
+**triggeredBy** (`'threshold' | 'ttl' | 'provider_change'`): Whether activation was triggered by threshold crossing, activateAfterIdle expiry, or a model/provider change.
+
+**lastActivityAt** (`number`): Unix-ms timestamp of the last assistant message part used for TTL checks.
+
+**ttlExpiredMs** (`number`): How long activateAfterIdle had been exceeded when activation fired.
+
+**previousModel** (`string`): Previous assistant model identifier that triggered activation (e.g. openai/gpt-4o).
+
+**currentModel** (`string`): Current actor model identifier that triggered activation.
+
 **recordId** (`string`): The OM record ID.
 
 **threadId** (`string`): This thread's ID.
 
 **config** (`ObservationMarkerConfig`): Snapshot of config at activation time.
 
+### `data-om-thread-update`
+
+Emitted when the Observer updates the thread title. Only emitted when `observation.threadTitle` is enabled.
+
+**cycleId** (`string`): Unique ID for this observation cycle — shared with observation markers.
+
+**threadId** (`string`): The thread ID that was updated.
+
+**oldTitle** (`string`): The previous thread title. Undefined if the thread had no title.
+
+**newTitle** (`string`): The new thread title.
+
+**timestamp** (`string`): When this update occurred.
+
 ## Standalone usage
 
 Most users should use the `Memory` class above. Using `ObservationalMemory` directly is mainly useful for benchmarking, experimentation, or when you need to control processor ordering with other processors (like [guardrails](https://mastra.ai/docs/agents/guardrails)).
 
+The `ObservationalMemory` class is the engine; to attach it to an agent, wrap it in an `ObservationalMemoryProcessor`, which needs a `Memory` instance for loading and persisting messages. Note that `stores.memory` is typed as optional on storage adapters, so a non-null assertion (or a runtime check) is needed:
+
 ```typescript
-import { ObservationalMemory } from '@mastra/memory/processors'
+import { ObservationalMemory, ObservationalMemoryProcessor } from '@mastra/memory/processors'
+import { Memory } from '@mastra/memory'
 import { Agent } from '@mastra/core/agent'
 import { LibSQLStore } from '@mastra/libsql'
 
@@ -683,8 +720,11 @@ const storage = new LibSQLStore({
   url: 'file:./memory.db',
 })
 
+const memory = new Memory({ storage })
+
 const om = new ObservationalMemory({
-  storage: storage.stores.memory,
+  storage: storage.stores.memory!,
+  memory,
   model: 'google/gemini-2.5-flash',
   scope: 'resource',
   observation: {
@@ -695,13 +735,15 @@ const om = new ObservationalMemory({
   },
 })
 
+const omProcessor = new ObservationalMemoryProcessor(om, memory)
+
 export const agent = new Agent({
   id: 'my-agent',
   name: 'my-agent',
   instructions: 'You are a helpful assistant.',
   model: 'openai/gpt-5-mini',
-  inputProcessors: [om],
-  outputProcessors: [om],
+  inputProcessors: [omProcessor],
+  outputProcessors: [omProcessor],
 })
 ```
 
@@ -719,15 +761,19 @@ The standalone `ObservationalMemory` class accepts all the same options as the `
 
 When `retrieval` is set (any truthy value), a `recall` tool is registered so the agent can page through raw messages behind observation group ranges. By default (scope `'resource'`), the tool supports listing threads (`mode: "threads"`), browsing other threads (`threadId`), and cross-thread search. With `retrieval: { vector: true }`, semantic search is available (`mode: "search"`). Set `scope: 'thread'` to restrict the tool to the current thread only. The tool is automatically added to the agent's tool list.
 
+Mastra also injects scope-aware usage instructions into the agent's context. For resource scope with `vector: true`, these cover routing between `search`, `threads`, and `messages`, including fallback to thread discovery when search results are unsuitable. Without `vector: true`, the instructions only cover `threads` and `messages` browsing, so the agent isn't steered toward a search mode that isn't configured. Resource-scoped instructions are injected even before any observation group exists, so the agent can browse other threads from the first message. Use `retrieval: { instructions: '...' }` to append application-specific guidance after the built-in instructions.
+
 ### Parameters
 
 **mode** (`'messages' | 'threads' | 'search'`): What to retrieve. "messages" (default) pages through message history. "threads" lists all threads for the current user. "search" finds messages by semantic similarity across all threads (requires vector store and embedder). (Default: `'messages'`)
 
 **query** (`string`): Search query for mode: "search". Finds messages semantically similar to this text across all threads for the current user.
 
-**cursor** (`string`): A message ID to anchor the recall query. Required for mode: "messages" when browsing the current thread. Extract the start or end ID from an observation group range (e.g. from \_range: \startId:endId\\\_, use either startId or endId). If a range string is passed directly, the tool returns a hint explaining how to extract the correct ID. Can be omitted when threadId is provided to start reading from the beginning of that thread.
+**cursor** (`string`): A message ID to anchor the recall query. Extract the start or end ID from an observation group range (e.g. from \_range: \startId:endId\\\_, use either startId or endId). If a range string is passed directly, the tool returns a hint explaining how to extract the correct ID. When both cursor and threadId are omitted for mode: "messages", the tool browses the current thread from the position set by anchor.
+
+**threadId** (`string`): Browse a different thread by its ID, or pass "current" for the active thread. Use mode: "threads" first to discover thread IDs. When provided without a cursor, reading starts from the beginning of the thread.
 
-**threadId** (`string`): Browse a different thread by its ID. Use mode: "threads" first to discover thread IDs. When provided without a cursor, reading starts from the beginning of the thread.
+**anchor** (`'start' | 'end'`): For mode: "messages" without a cursor, page from the start (oldest-first) or end (newest-first) of the thread. (Default: `'start'`)
 
 **page** (`number`): Pagination offset. For messages: positive values page forward from cursor, negative values page backward. For threads: page number (0-indexed). 0 is treated as 1 for messages. (Default: `1`)
 
@@ -735,6 +781,10 @@ When `retrieval` is set (any truthy value), a `recall` tool is registered so the
 
 **detail** (`'low' | 'high'`): Controls how much content is shown per message part. 'low' shows truncated text and tool names with positional indices (\[p0], \[p1]). 'high' shows full content including tool arguments and results, clamped to one part per call with continuation hints. (Default: `'low'`)
 
+**partType** (`'text' | 'tool-call' | 'tool-result' | 'reasoning' | 'image' | 'file'`): Filter results to only include message parts of this type. Only applies to mode: "messages".
+
+**toolName** (`string`): Filter results to only include tool-call and tool-result parts matching this tool name. Only applies to mode: "messages".
+
 **partIndex** (`number`): Fetch a single message part at full detail by its positional index. Use this when a low-detail recall shows an interesting part at \[p1] — call again with partIndex: 1 to see the full content without loading every part.
 
 **before** (`string`): For mode: "threads" only. Filter to threads created before this date. Accepts ISO 8601 format (e.g. "2026-03-15", "2026-03-10T00:00:00Z").
@@ -753,6 +803,8 @@ When `retrieval` is set (any truthy value), a `recall` tool is registered so the
 
 **limit** (`number`): The limit used for this query.
 
+**detail** (`'low' | 'high'`): The detail level used for this query.
+
 **hasNextPage** (`boolean`): Whether more messages exist after this page.
 
 **hasPrevPage** (`boolean`): Whether more messages exist before this page.

package/.docs/reference/memory/serialized-memory-config.md

@@ -4,7 +4,7 @@
 
 `SerializedMemoryConfig` is the JSON-serializable subset of [`Memory`](https://mastra.ai/reference/memory/memory-class) configuration that lives on a stored agent record. The runtime hydrates it back into a `Memory` instance by resolving the vector and embedder IDs against the configured `Mastra` instance.
 
-It's the type used by [`BuilderAgentDefaults.memory`](https://mastra.ai/reference/editor/agent-builder/builder-agent-defaults) and by `EditorAgentNamespace.create({ memory })`.
+It's the type used by [`BuilderAgentDefaults.memory`](https://agent-builder.mastra.ai/reference/builder-agent-defaults) and by `EditorAgentNamespace.create({ memory })`.
 
 ## Usage example
 
@@ -70,5 +70,5 @@ new MastraEditor({
 
 - [Memory class](https://mastra.ai/reference/memory/memory-class): The runtime type this config hydrates into.
 - [Observational memory](https://mastra.ai/reference/memory/observational-memory): Full observational memory reference.
-- [BuilderAgentDefaults](https://mastra.ai/reference/editor/agent-builder/builder-agent-defaults): Where this type is pinned as the Builder default.
-- [Agent Builder: Memory](https://mastra.ai/docs/agent-builder/memory): Concept and worked examples.
+- [BuilderAgentDefaults](https://agent-builder.mastra.ai/reference/builder-agent-defaults): Where this type is pinned as the Builder default.
+- [Agent Builder: Memory](https://agent-builder.mastra.ai/memory): Concept and worked examples.

package/.docs/reference/observability/tracing/interfaces.md

@@ -319,7 +319,9 @@ enum SpanType {
    * Provider-executed (server-side) tool span. Reconstructed from
    * tool-call and tool-result stream chunks for tools the model
    * provider executes (e.g. Anthropic code execution, server-side
-   * web search). Opened on tool-call chunk, closed on tool-result.
+   * web search). Created on the tool-result chunk under the model
+   * step that delivered it, with the start time backdated to the
+   * tool-call chunk.
    */
   PROVIDER_TOOL_CALL = 'provider_tool_call',
 
@@ -473,6 +475,7 @@ interface ToolCallAttributes {
   toolId?: string
   toolType?: string
   toolDescription?: string
+  toolCallId?: string
   success?: boolean
 }
 ```
@@ -495,6 +498,8 @@ interface MCPToolCallAttributes {
   /** Tool description */
   toolDescription?: string
 
+  toolCallId?: string
+
   /** Whether tool execution was successful */
   success?: boolean
 }

package/.docs/reference/processors/processor-interface.md

@@ -399,7 +399,7 @@ processLLMRequest?(
 
 **tracingContext** (`TracingContext`): Tracing context for observability.
 
-**writer** (`ProcessorStreamWriter`): Stream writer for emitting custom data chunks during streaming. Use writer.custom() to send transient UI signals.
+**writer** (`ProcessorStreamWriter`): Stream writer for emitting custom data chunks during streaming. Call writer.custom() to emit a data-\* chunk.
 
 **abortSignal** (`AbortSignal`): Signal for cancelling the operation.
 
@@ -504,7 +504,7 @@ processAPIError?(args: ProcessAPIErrorArgs): Promise<ProcessAPIErrorResult | voi
 
 **abort** (`(reason?: string, options?: { retry?: boolean; metadata?: unknown }) => never`): Function to abort processing.
 
-**writer** (`ProcessorStreamWriter`): Stream writer for emitting custom data chunks during streaming. Use writer.custom() to send transient UI signals.
+**writer** (`ProcessorStreamWriter`): Stream writer for emitting custom data chunks during streaming. Call writer.custom() to emit a data-\* chunk.
 
 **requestContext** (`RequestContext`): Request context passed through from the agent call.
 
@@ -961,6 +961,18 @@ await writer.custom({
 })
 ```
 
+When memory is configured, custom `data-*` chunks emitted from `processOutputStream` or `processOutputResult` are saved as parts of the assistant message. Set `transient: true` on the chunk object to stream it without saving it to memory:
+
+```typescript
+await writer.custom({
+  type: 'data-progress',
+  data: { status: 'Processing' },
+  transient: true,
+})
+```
+
+Pass `transient` as a property of the chunk, not as the second argument to `writer.custom()`. The second argument contains writer options such as `messageId`.
+
 By default, processors **don't** see `data-*` chunks in `processOutputStream` so they don't accidentally process tool telemetry or their own output. Opt in by setting `processDataParts: true` on the processor:
 
 ```typescript