Post-M9 Review Flashcards

Vocabulary flashcards covering key Post-M9 concepts including hooks, MCP primitives, prerequisite gates, error handling, subagents, and structured output.

PreToolUse

A hook that can block a tool call before execution using permissionDecision: "deny" and rewrite the input via updatedInput.

PostToolUse

A hook that receives additionalContext as commentary, but cannot block or rewrite tool calls.

Hook Session State

The state behavior of hooks where each invocation runs as a fresh subprocess with no memory of past calls, requiring state to live in the session rather than Python variables.

Tools (MCP Primitive)

Model-initiated actions within the Model Context Protocol.

Resources (MCP Primitive)

Browsable content in the Model Context Protocol that the model knows exists without needing a tool call, requiring explicit human UI attachment to read.

Prompts (MCP Primitive)

Reusable templates within the Model Context Protocol that are invocable by name.

MCP Tool Naming Convention

The standard structure mcp____ assigned to an MCP tool upon loading by a client.

Prerequisite Gate

A validation check keyed on the facility_id of a gated call matched against a prior successful get_facility_users result in the transcript.

Transient Error

An error category where isRetryable is true, causing the system to retry locally.

Validation Error

An error category where isRetryable is false, instructing the system to explain the issue to the user without retrying.

Permission Error

An error category where isRetryable is false, triggering escalation via a structured handoff.

A-2 Escalation Triggers

The three specific conditions that trigger escalation: an ambiguous facility with no live user to ask, a permission-category error, and an exhausted retry budget.

Standalone Escalation Handoff

A handoff format required because human recipients lack conversation access, requiring fields such as facility, question, attempted, reason, and recommendation to stand alone.

Subagent Context Inheritance

The structural default where subagents inherit nothing from the parent conversation, requiring all needed facts to be explicitly included in their prompt.

Parallel Dispatch

The execution model where multiple dispatch-tool calls are emitted in a single coordinator response.

Sequential Dispatch

The execution model where the same tool is invoked across separate conversation turns.

Coverage Gaps Section

A coordinator report section used when a subagent fails, explicitly distinguishing unmeasured facilities from those measured as zero.

R-3 "other" + metric_detail Pattern

A routing structure for real Mongo metrics lacking fixture coverage (kscore, bg, temperature, sleep, steps, weight) that maps them to metric:"other" and metric_detail:" ".

Schema Validation vs. Semantic Validation

The principle that type systems like Pydantic verify structure and shape (e.g., $n = 0 \rightarrow \text{null}$) but cannot detect factual fabrication or invalid ranges (e.g., $n = 10$, $\text{mean} = -97$).