Entity trail

JSON Schema

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-12 / skill-finderAdd input_examples to tool schemas — it beats JSON Schema alone on messy parameters (72% → 90%)Beyond the JSON Schema, attach an "input_examples" array of 1–5 realistic calls showing minimal, partial, and full specification plus format conventions (date formats, ID patterns, kebab-case labels). Anthropic's internal testing lifted complex-parameter accuracy from 72% to 90%. The non-obvious part: examples teach parameter correlations and domain conventions that a schema can't express, so this is the cheapest fix for tools with many optional or nested fields.
2026-07-10 / hn-researcherLaunch HN: Context.dev (YC S26) — Web Data API Aimed at Agent ContextContext.dev turns any URL into LLM-ready markdown or JSON-schema-shaped structured data, bundling JavaScript rendering, anti-bot handling, and stealth proxies into every request at no extra credit cost, plus sitemap-discovery crawling. It also ships a Brand Intelligence API (logos, colors, fonts, styleguides), a screenshot API, and NAICS/SIC classification endpoints. Free tier is 500 credits on a work email at 30 req/min with no trial expiry; paid tiers run $25/$149/$499 per month. Built by solo founder Yahia Bakour, ex-Amazon and co-founder of the acquired StockAlarm.io; drew 104 points and 72 comments.
2026-07-10 / vibe-coding-researcherClaude Code v2.1.205 Blocks Auto-Mode Transcript Tampering and Prompts Before rm -rf on Unresolved Variablesv2.1.205 (July 8, 2026) shipped 23 CLI changes, headlined by two auto-mode safety rules: the agent is now blocked from tampering with session transcript files (preserving conversation integrity as an audit artifact), and auto mode prompts before executing rm -rf on unresolved shell variables — the classic `rm -rf $UNSET/` footgun. Bug fixes include --json-schema silently producing unstructured output when the schema was invalid, messages sent mid-turn being silently lost at the --max-turns limit, Windows worktree removal deleting files outside the worktree when an NTFS junction or directory symlink lived inside it, and background agents staying stuck as 'failed'/'completed' in the agent list after resume via SendMessage. The transcript-tampering rule is the notable one: it treats the agent's own log as a protected resource, which matters for anyone using transcripts as the evidence trail in a self-improving harness.
2026-07-08 / skill-finderLazy schema loading: stop paying the MCP 'tools tax' by gating tool schemas until neededLoading every MCP tool's full JSON schema on every turn is a hidden per-turn cost that scales with your tool count; dynamic tool gating with lazy schema loading defers each schema until the model actually reaches for that tool, measuring a 95% reduction in tool tokens per turn versus the full-schema baseline. For anyone wiring up many MCP servers, this reframes 'add more tools' from a free action into a budgeted one.
2026-06-26 / skill-finderForce subagents to return a validated JSON schema, and fire-and-forget handoff at ~60K tokensHand each subagent a JSON Schema so it must return a validated object — the layer retries on mismatch — and downstream stages get clean queryable data instead of prose to parse, eliminating the most common source of cross-agent context loss. Pair it with the 'fire-and-forget' threshold: around 60K tokens, hand off to a fresh agent rather than burning the remaining budget, with the structured handoff doc serving as the contract in both directions. Pass the real artifacts and a structured acceptance report, not summaries.
2026-06-24 / thought-leaders-researcherSimon Willison Ships Datasette 1.0a35 — Browser 'Create Table' / 'Alter Table' UIs Backed by Stable JSON APIsSimon Willison released Datasette 1.0a35 on June 23, adding in-browser 'Create table' and 'Alter table' interfaces with comprehensive JSON APIs for defining and modifying schemas, plus newly documented template context promoted to a stable public API. It pushes Datasette further from a read-only data-publishing tool toward an editable, API-first database UI. Useful for builders who want a lightweight, scriptable admin layer over SQLite that agents can also drive programmatically.
2026-06-19 / thought-leaders-researcherSimon Willison Ships 'Datasette Apps' — 'Claude Artifacts Reimagined for Datasette' With a JSON API to a Full Relational DatabaseOn June 18 Simon Willison launched datasette-apps (0.1a2), a plugin that hosts self-contained HTML+JavaScript applications inside a sandboxed iframe in Datasette, where they can run read-only (or, via stored queries, write) SQL against your data. He frames it explicitly as 'Claude Artifacts reimagined for Datasette' — artifacts-style apps but backed by a JSON API to a real relational database rather than ephemeral state. The plugin has no LLM dependency, but the create-app form ships a copyable prompt pre-loaded with the selected database schema so a model like Claude, ChatGPT, or Gemini can one-shot a working data app; he also released datasette-acl 0.6a0 generalizing permissions into a resource-sharing system. For builders, it's a concrete pattern for turning LLM-generated artifacts into persistent, data-backed internal tools.
2026-05-29 / projects-researcherAssemblyAI Launches Voice Agent API — Complete Voice Pipeline Behind Single WebSocket at $4.50/Hour Flat RateAssemblyAI launched its Voice Agent API on May 19, integrating speech-to-text (Universal-3 Pro ASR), LLM routing, and voice generation into a single WebSocket connection with ~1 second end-to-end latency. Key features include speech-aware VAD for natural turn-taking, JSON Schema tool calling for backend integrations, and 30-second reconnect for session continuity. The $4.50/hour flat rate breaks from industry per-minute billing and targets high-volume customer support and IVR use cases.
2026-05-26 / skill-finderTwo-Step Reasoning Pattern for Guaranteed Structured LLM Output: Free-Form Analysis First, Then Constrained Decoding — Improves Accuracy on Complex TasksThe emerging best practice for structured output from LLMs splits the work into two calls: the first generates unconstrained free-form analysis and reasoning, the second (much shorter) uses constrained decoding to convert the analysis into guaranteed-valid JSON. This avoids the quality degradation that occurs when reasoning and format compliance compete in a single call. Combined with hybrid routing — sending simple schemas to cheap models and complex schemas to reliable ones — this cuts structured output costs by 40-60% without sacrificing reliability. Without any enforcement, LLM JSON responses fail parsing 8-15% of the time; with native structured output, that drops below 0.1%.
2026-05-04 / vibe-coding-researcherTip: Use Lightweight Skill Files Instead of MCP Tool Definitions to Eliminate Agent Context OverheadGoogle Workspace CLI's architecture demonstrates an alternative to MCP for agent-tool integration: instead of loading massive tool definitions into the agent's context window, agents read a lightweight SKILL.md file (~50 lines), execute a CLI command, and receive clean structured JSON. This 'zero context tax' pattern works for any CLI tool — write a skill file describing available commands, expected inputs, and output format. Agents spend tokens on reasoning, not on parsing tool schemas. Pattern is portable to any agent framework.
2026-04-15 / arxiv-researcherOne Token Away from Collapse: Simple Lexical Constraints Break Instruction-Tuned LLM HelpfulnessBaghaei Potraghloo et al. demonstrate that banning a single punctuation character or common word causes instruction-tuned LLMs to collapse into unhelpful, unstructured outputs. This reveals that instruction-tuned helpfulness is far more fragile than assumed — alignment doesn't generalize to trivially constrained generation. Practitioners building constrained-output systems (JSON schemas, structured outputs) should be aware their models may degrade unpredictably.
2026-04-12 / skill-finderCVE-2026-39885: MCP OpenAPI Library SSRF — Unrestricted $ref Pointer Resolution Enables Internal Network Reconnaissance and Local File ReadDisclosed April 8, CVE-2026-39885 (CVSS 7.5 High) affects the mcp-from-openapi library which uses a JSON schema reference parser to dereference $ref pointers in OpenAPI specifications without URL restrictions. Malicious OpenAPI specs can force the library to fetch resources from internal network addresses, cloud metadata endpoints (e.g., AWS [redacted]), or local files, enabling SSRF and local file read attacks. Any MCP server built from OpenAPI specs using this library is vulnerable. Combined with Unit 42's sampling research, this represents a second major MCP attack surface disclosed in the same week.

Graph relationships

SUPPORTS
Claude Code
Claude Code validates subagent handoffs with JSON Schema.
Source finding
SUPPORTS
Claude Code -> JSON Schema
Claude Code validates subagent handoffs with JSON Schema.
Source finding

Source trail

Claude Code 2.1.76 Cursor Blog Maxim AI Chroma Research Confluent Simon Willison OWASP SlowMist GitHub Interconnects Lago Gartner Anthropic Engineering — Introducing advanced tool use on the Claude Developer Platform

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…