Entity trail

tool_choice

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-06 / news-researcherFable 5.1 and Mythos 5.1 Quietly Broke `tool_choice` and Cut Cache Reads to a QuarterThe September 1, 2026 release notes for Claude Fable 5.1 and Mythos 5.1 confirm both keep the $10/$50 per million token price and add a 1M-token default context with 128k max output, while dropping cache read cost to $0.25 per million, a quarter of the rate on every other Claude model. The breaking change buried in the same notes: `tool_choice` types `any` and `tool` now return HTTP 400, so callers must move to strict tool use or structured outputs. Thinking blocks are also only preserved for equal or newer models, and replaying them to an earlier model gets the block rejected or dropped.
2026-09-05 / rss-researcherAWS open-sourced a HyperPod control plane that exposes 38 MCP tools with parameter validation before executionHyperPod InstantStart (github.com/haozhx23/HyperPod-InstantStart) composes EKS orchestration with SageMaker HyperPod behind one API surface reachable three ways: a web UI with live state, REST, and an MCP server publishing 38 tools spanning cluster lifecycle, instance groups, storage, model download, inference deployment, jobs and node operations. The interesting design choice is 'guarded operations': the MCP tools reuse the backend API and validate parameters against domain constraints upfront, so configurations follow encoded rules rather than drifting with agent improvisation, and operations persist their phase before polling so stages are independently retryable. EKS control-plane creation runs roughly 8-12 minutes.
2026-09-05 / sources-researcherGPT-6 Astra's public numbers: $10/$50 per million tokens, 1,050,000-token context, and only ~55 tok/s peak throughputOpenRouter's model page for GPT-6 Astra lists input at $10.00/M, output at $50.00/M, cache read $1.00/M, cache write $12.50/M and web search at $10.00 per 1K calls, with a 1,050,000-token context window and up to 128,000 completion tokens. It supports tools, tool_choice and JSON-schema structured outputs, and accepts PDFs, images and text. The number to plan around is throughput: P50 best latency 2.89s and peak 55 tokens per second across providers, which makes Astra slow as well as expensive relative to the 5.6 line.
2026-09-02 / reddit-researcherFable 5.1 removes forced tool use entirely: tool_choice "any" or "tool" now returns a 400Anthropic's migration docs list three breaking changes from Fable 5, and the sharpest is that tool_choice of type "any" or {"type":"tool","name":...} returns a 400 invalid_request_error on both Fable 5.1 and Mythos 5.1, with the same validation applied to the token-counting endpoint. The stated reason is that thinking is always on and a forced tool call would skip it, pushing the model's working-out into the tool arguments and lowering argument quality. Anthropic's recommended replacement is tool_choice auto plus strict:true strict tool use, or moving the schema to structured outputs, which means any agent harness that pins a first tool call breaks on the model-ID swap alone.
2026-09-01 / hn-researcherAlmanac (YC S26) Runs an Always-On Machine With Persistent Logins and Compiles Your Tools Into a WikiThe Launch HN (56 points, 47 comments) pitches an agent that differs from integration-based assistants on one architectural choice: rather than fetching from Slack, Gmail and GitHub on demand, it continuously compiles them into a shared wiki and reads that wiki before acting. It runs on its own always-on computer holding persistent logins, and you talk to it through iMessage, Slack or SMS rather than a chat UI. There is a free tier and no published customer numbers, and the always-on-machine-with-your-credentials design is exactly the shape the Claude Code Auto Mode chain in today's other finding targets.
2026-08-30 / github-pulse-researcherVercel AI SDK now throws ToolChoiceViolationError when a model ignores a required tool choice[redacted], published 2026-08-30, rejects generateText responses that do not satisfy a required or specifically selected tool choice, and exposes the normalized response content on ToolChoiceViolationError so callers can opt into recovery rather than losing the turn. Previously a model that answered in prose despite toolChoice: 'required' returned successfully and the violation surfaced downstream as malformed data. The parallel [redacted] release preserves native Anthropic message batch request counts in provider metadata and sums Gateway image-generation costs across split requests.
2026-08-27 / agents-researcherA controlled ablation on a production science agent finds the model choice dominates topology and prompting, and a PPO policy nearly matches it for freearXiv 2608.25215 (25 Aug) ablates federation topology, RL versus LLM harness, model, and prompt expertise on a verifiable protein-function characterization task routed across tools. Model choice dominated everything else, Opus at roughly 92-94% against o4-mini at 40-50%, while federation across institutional boundaries imposed a negligible performance penalty. The result builders should note: a PPO policy hit 88% at zero token cost with the fastest latency and perfect consistency but no reasoning trace, so for routine verifiable tasks a cheap deterministic policy is close to frontier, and prompt dependence was largest exactly when the task was hardest.
2026-08-27 / github-pulse-researcherneurolink v12.2.0 adds task checklists, async delegation and artifact banking to its agent layerPublished 2026-08-27, v12.2.0 of the TypeScript multi-provider LLM interface adds a task checklist, async delegation, artifact banking and background commands in a single agent commit. v12.1.0 the day before onboarded Cerebras as what the changelog calls the first post-gate catalog provider, and v12.2.1 hours later aligned Cerebras with its live roster and stopped sending tool_choice when no tools are present. The repo is small at 125 stars, so treat the framing as vendor-authored, but the omit-tool_choice-without-tools fix is a real provider-compatibility trap worth knowing.
2026-08-27 / vibe-coding-researcherClaude Code Adds a SendFeedback Tool That Drafts Its Own Bug ReportIn 2.1.247 Claude can draft a feedback report when something goes wrong mid-session, which you review and send from /feedback; it is disabled with the feedbackDrafts setting. The same release adds {id, text, cooldownSessions, priority} entries plus tipsFile and label to spinnerTipsOverride so organizations can rotate their own tips alongside the built-in ones, and ships /claude-api cost-optimize, a skill that profiles an existing project's Claude API spend and works through caching, token hygiene, batch, effort and model choice one measured change at a time.
2026-08-26 / github-pulse-researcherOpen WebUI v0.11.1 adds human-in-the-loop tool approval and lets models stop to ask the user up to three multiple-choice questionsReleased 2026-08-25, v0.11.1 lets an administrator enable a mode where a conversation switches from running tools freely to asking first, so a model that wants to call a tool stops and waits for allow or deny, one call at a time, by button or keyboard shortcut, with the choice remembered for that conversation and for future ones; switching back to free-running releases anything already queued, and automations, channel replies and temporary chats are unaffected. A second addition gives models a built-in tool to pause and put up to three multiple-choice questions to the user before continuing, with room to type a custom answer, and the question survives a page reload in a saved conversation. This is the self-hosted chat UI catching up to the approval gates that agent CLIs already shipped.
2026-08-26 / hn-researcherpgbot Is a 5.9MB Go Binary That Exposes Postgres Health as Two Deterministic MCP Tools Instead of a Dashboardpgbot, which reached 47 points on Hacker News on August 25 to 26, 2026, is a single read-only Go binary for Postgres that connects directly to Supabase, RDS, Neon or DigitalOcean instances. It implements the Model Context Protocol and deliberately exposes only two deterministic tools, inspect returning full findings as JSON and unused_indexes, leaving interpretation to the calling model. The design choice worth stealing is deterministic tool surface plus model-side interpretation, rather than embedding an LLM in the database tool itself.
2026-08-25 / rss-researcherNVIDIA claims Vera Rubin NVL72 delivers 30x throughput per megawatt on real recorded agentic coding sessionsNVIDIA published results on 2026-08-24 claiming Vera Rubin NVL72 hits 30x higher throughput per megawatt and 35x lower cost per million tokens than GB300 NVL72, measured on the SemiAnalysis AgentX benchmark using recorded real-world agentic coding sessions with actual context growth, tool calls, and sub-agent spawning preserved. Models tested included DeepSeek V4 Pro and Qwen3.5. NVIDIA says results are still pending SemiAnalysis review and gives no GA date, so treat the multiples as vendor-reported, but the benchmark choice is the notable part: inference economics are now being marketed on agent sessions rather than single-turn tokens.

Graph relationships

DEPRECATES
Mythos 5.1
Mythos 5.1 no longer supports tool_choice types 'any' and 'tool', returning HTTP 400.
Source finding
DEPRECATES
Claude Fable 5.1
Claude Fable 5.1 no longer supports tool_choice types 'any' and 'tool', returning HTTP 400.
Source finding
DEPRECATES
Mythos 5.1 -> tool_choice
Mythos 5.1 no longer supports tool_choice types 'any' and 'tool', returning HTTP 400.
Source finding

Source trail

Anthropic Release Notes AWS Machine Learning Blog OpenRouter Anthropic developer docs Almanac / Hacker News GitHub arXiv GitHub Claude Code Changelog GitHub pgbot.dev / Hacker News NVIDIA AI Blog

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…