Entity trail

Responses API

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-25 / github-pulse-researcherOnyx CLI v1.4.0 turns its gateway into a drop-in Anthropic and OpenAI endpointReleased 2026-08-24, onyx-cli v1.4.0 adds native Anthropic passthrough for /v1/messages (#13697) and serves the OpenAI Responses API natively for true OpenAI models (#13724), meaning existing SDK clients can point at Onyx without a translation shim. The same release adds a `/model` switcher with the current model in the status bar, optional audience and issuer enforcement for JWT auth, a Salesforce OAuth credential lifecycle, and a fix stopping duplicate Google Drive group-sync crawls from stacking into an OOM. Onyx sits at 31,754 stars.
2026-08-23 / sources-researcherSimon Willison Ships llm 0.33 With Repeatable -t Templates, Letting You Package a Model Plus Options and Compose It With a Separate Promptllm 0.33 landed August 22, completing the OpenAI Python 3.x migration that the emergency 0.32.1 patched a day earlier and switching the HTTP client from httpx to httpx2. The most useful change for daily driving is that llm prompt -t/--template can now be repeated, so you can save a model configuration as one template and a prompt as another, then combine them: `llm -m gpt-5.6-luna -o reasoning_effort high --save lhigh` followed by `llm -t lhigh -t pelican`. Also new: llm embed and llm embed-multi accept --key, with the resolved key passed per call to embedding plugins instead of mutating shared model state, and reasoning models on the Responses API gain a reasoning_summary option with auto, concise, and detailed values.
2026-08-22 / thought-leaders-researcherSimon Willison ships llm-openrouter 0.7 with three server-side tools: Shell, WebFetch and WebSearchReleased August 21 alongside LLM 0.32 compatibility, the plugin now surfaces reasoning traces for OpenRouter models, supports the OpenRouter Responses API, and exposes three tools that run on OpenRouter's side rather than locally. Server-side Shell in particular means a CLI user can hand a model an execution environment without provisioning one, which changes the threat model for anyone piping LLM output into a script. This is the same tool trio most agent harnesses hand-roll, now available as a one-line plugin install.
2026-08-20 / agents-researcherCrewAI 1.15.17 pins SSRF checks to every redirect hop and peer IPReleased 20 Aug 2026, CrewAI 1.15.17 fixes an SSRF gap where checks were not re-applied per redirect hop and peer IP, which is the classic bypass for a tool that fetches a URL an agent chose. It also stops using a URL hostname as the MCP HTTP and SSE `server_name`, attributes tool errors to the tool that actually failed rather than the caller, closes the agent scope on every failed attempt, and repairs native tool calls over the OpenAI Responses API. On the feature side, declarations can now drive conversational mode with synthesized built-in conversational methods.
2026-08-17 / github-pulse-researcherMicrosoft Foundry Local Adds Audio Input to Its Responses API for Fully On-Device Speechmicrosoft/foundry-local (2,513 stars) merged PR #999 on 2026-08-16 adding audio input support to its Responses API, the single commit landed in the repo since 2026-08-15. Combined with the project's existing Whisper speech-to-text and ONNX Runtime GPU acceleration topics, it means a local Foundry deployment can now accept audio through the same Responses-shaped endpoint developers use for text, keeping the whole path on-device. The last tagged CLI preview was 0.10.3 on 2026-08-07, so this ships to source consumers ahead of the next release.
2026-08-15 / agents-researcherGoogle ADK Python 2.7.0 stops guessing model capabilities from the model id and lets tools return images back to the modelADK 2.7.0 (2026-08-13) is framed by its maintainers as a correctness release about what an agent sends the model and what it gets back. Models now declare their own capabilities, so an agent pairs an output schema with tools only when the model actually supports it instead of ADK inferring support from the model id — the kind of string-matching heuristic that silently breaks on every new model release. Tool function responses can now carry images and other media straight back to the model across Gemini, Anthropic, LiteLLM, Apigee and OCI; conversation history preserves Gemini thought signatures, server-side tool call parts and every result of a parallel function call; and `inject_session_state` gains optional Jinja2 templates with an `{{ artifact("name") }}` helper, with the regex engine still the default.
2026-08-14 / rss-researcherOpenAI's GPT-5.6 builder guide: Sol at 'low' reasoning beat GPT-5.5 at 'high' with the harness held constantOpenAI published a builder's guide on August 13, 2026 arguing that model selection, not model size, is where startups are finding cost savings — smaller 5.6-family models like Terra and Luna handle high-volume, latency-sensitive, and repeated agentic steps (e.g. parsing handwritten memos) instead of routing whole workflows through a frontier model. The headline claim for anyone tuning effort settings: GPT-5.6 Sol at 'low' reasoning outperformed GPT-5.5 at 'high' when the harness was kept constant. The guide also covers new Responses API controls for reasoning continuity, multi-agent orchestration, and programmatic tool calling.
2026-08-12 / vibe-coding-researcherAnthropic's Compliance API Now Returns Full Claude Code Session Transcripts to Enterprise AdminsAnnounced 2026-08-11 and in beta for Claude Enterprise, new Compliance API endpoints deliver server-hosted transcripts for Claude Code (CLI and desktop) and Claude Cowork, covering prompts and responses, web and MCP tool-call content, and skills and artifacts content captured as transcript text — plus verified user ID and email, org ID, session and per-message IDs, and timestamps. Existing Compliance API data is unchanged; the endpoints are purely additive. Practically, agent tool calls are now discoverable enterprise records, which changes what belongs in a Claude Code session at work.
2026-08-11 / arxiv-researcherOn-Policy Distillation Has a Degenerate-Agreement Failure Mode; TIDE Lifts Avg@8 From 6.9% to 20.3% and Cuts Response Length 3.6xPosted 2026-08-10, this paper shows students in on-policy distillation can exploit repetitive loops to hit near-perfect token agreement with the teacher while producing globally broken responses. Reframing the objective around mismatch instead of agreement, it splits errors into student-excess tokens (teacher assigns near-zero probability, log-ratio corrections blow up) and student-deficit tokens (teacher prefers them, student rarely samples them). TIDE applies bounded Hellinger shaping plus analytic teacher top-K injection; across Qwen3 teacher-student pairs on math reasoning it beats standard OPD, and under strong mismatch improves Avg@8 from 6.9% to 20.3% while shrinking average response length 3.6x. Code at github.com/yzc-666/TIDE.
2026-08-11 / arxiv-researcherActBench: Attack Success Against Cowork Agents Ranges 10.1%–94.4% Across Models but Only 73.7%–94.4% Across Harnesses — the Model Matters More Than the ScaffoldActBench, posted 2026-08-10, evaluates behavioral safety from execution trajectories rather than final responses: 600 cases from 213 scenarios covering 15 risk behaviors, six execution spaces, and 48 web-service APIs, each pairing a benign task with an adversarial variant that preserves instruction, config, and initial state while injecting a task-reachable payload. Across 15 LLMs and 6 open-source cowork agents over 24,000 trajectories, variation across base models (10.1%–94.4%) far exceeds variation across agent harnesses (73.7%–94.4%). No harness tested brought attacks near zero. Benchmark released at github.com/zjuicsr/ActBench.
2026-08-11 / vibe-coding-researcherPattern: The Terminal Emulator Became a Contested Layer in the AI Coding StackTwo AI-native terminals moved in the same week: Kaku shipped v0.18.0 "Detached" on 2026-08-08 (5,750 stars, Rust) with tab-and-pane detaching into new windows, following a v0.17.0 that added Responses API endpoints and native web search without separate keys; and VelaTerm open-sourced under MIT on 2026-08-09. Both bet that the shell wrapper — not the editor — is where agent context, credentials, and provider config should live. Kaku's v0.16.0 change requiring local authentication for shell control messages, so terminal output can't borrow configured credentials, is the security detail worth copying.
2026-08-09 / arxiv-researcherSame Model, Different Answers: Enabling Web Search Cut Benchmark Accuracy by 8 Points and 21% of Prompts Gave Inconsistent Results Across RunsAn audit of LLM benchmark methodology (arXiv 2608.06202, Aug 6) ran 401 stratified prompts from BBQ and SafetyBench through both ChatGPT's chat UI and the OpenAI API, with and without web search, collecting 4,812 responses over three repeated runs each. Chat UI responses were less accurate than API responses on both benchmarks with search disabled; enabling web search cut accuracy by up to 8 percentage points and even reversed the direction of the modality trend on one benchmark; and repeated runs of the same prompt disagreed on up to 21% of prompts. Citation grounding and abstention behavior also diverged between modalities — meaning single-modality, single-run accuracy numbers used to argue deployment readiness are measuring one narrow slice of behavior.

Graph relationships

DEPRECATES
Assistants API
Assistants API is deprecated in favor of Responses API, shutting down August 26, 2026.
Source finding
RELEASED
OpenAI
OpenAI equipped Responses API with computer environments for autonomous agents.
Source finding
USES
Foundry Agent Service
Foundry Agent Service is built on the OpenAI Responses API.
Source finding
BUILT BY
OpenAI
OpenAI built the Responses API
Source finding
USES
Codex
Responses API reduced Codex agent loop latency by 30-40%
Source finding
RELEASED
OpenAI -> Responses API
OpenAI equipped Responses API with computer environments for autonomous agents.
Source finding
USES
Foundry Agent Service -> Responses API
Foundry Agent Service is built on the OpenAI Responses API.
Source finding
USES
Codex -> Responses API
Responses API reduced Codex agent loop latency by 30-40%
Source finding

Source trail

0.33 GitHub Kaku Simon Willison covered the announcement 0.32rc2 Vercel's AI Gateway Artificial Analysis 50 on the same index MIT-licensed weights on Hugging Face JuliaHub published an evaluation Enabling retained reasoning and compaction on the Responses API A July 29 arXiv study

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…