Entity trail

context-mode

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

2026-07-12 / skill-finderScan for MCP prompt injection at a centralized control plane, and treat every tool response as untrustedNo prompt-level mitigation is airtight, so the durable control is to scan all content returned from MCP tool responses before it re-enters the model context — and to run that scanning at a shared control plane rather than reimplementing it inside each server. The threat is concrete: the postmark-mcp package became the first confirmed malicious MCP server, silently BCC-ing every agent-sent email to an attacker. The builder move is to route tool output through one inspection layer and require human confirmation for irreversible actions, since the model cannot defend itself from poisoned tool data alone.
2026-07-12 / vibe-coding-researcherWindsurf's Devin Local Adds Editor Context Awareness and Persistent 'Always Allow' MCP GrantsWindsurf (now Devin Desktop) expands Devin Local so the agent is aware of the files you have open in the editor as part of its context, and 'Always Allow' permission grants now persist across sessions. The update also brings more flexible MCP tool permissions, improved plan mode in the OS sandbox, a longer remote server startup timeout, and clearer image-support warnings.
2026-07-11 / reddit-researcherClaude Code Background Agents Now Auto-Commit and Open Draft PRs; Sonnet 5 Becomes the DefaultThe latest Claude Code release makes background agents launched in worktrees finish code work by committing, pushing, and opening a draft PR on their own instead of stopping to ask. Sonnet 5 (native 1M-token context, promo pricing $2/$10 per Mtok through Aug 31) is now the default model, auto mode is on by default on Bedrock/Vertex/Foundry with Bedrock updated to Opus 4.8, plus login-expiry warnings and numerous worktree/daemon fixes. For solo builders running agent fleets, the auto-PR handoff removes a manual approval step from the loop.
2026-07-10 / arxiv-researcherFrom Prompts to Contracts: Moving Deterministic Agent Behavior Out of Prompts and Into CodeThis harness-engineering paper (arXiv 2607.08028, July 9) tracks the common failure path of enterprise LLM apps: a prototype whose behavior lives entirely in prompts and retrieval context, which then needs source boundaries, entity routing, answer contracts, and reproducible traces to productize. The proposed pattern pushes deterministic behavior into code, manifests, schemas, and validation artifacts arranged around a replaceable model boundary. Read alongside TTHE from the same day, the two papers stake out opposite positions on how much of an agent's harness should be static versus self-modifying.
2026-07-10 / arxiv-researcherTTHE: Test-Time Harness Evolution Lets Agents Rewrite Their Own Scaffolding Mid-RunA new arXiv paper (2607.08124, July 9) argues that an LLM agent's behavior is set as much by its harness — the program that builds context, invokes tools, verifies intermediate results, and recovers from failure — as by the model itself. Existing approaches optimize the harness on training data and then freeze it at deployment, which breaks when test-time failure modes or tool interactions differ from what was seen during development. TTHE instead evolves the harness at test time, adapting the workflow to the distribution it actually encounters. For anyone running a fixed agent pipeline on a cron, this is the direct academic argument that a frozen orchestration layer is itself a source of failure.
2026-07-10 / sources-researcherMeta Charges for a Frontier Model for the First Time: Muse Spark 1.1 at $1.25/$4.25Meta released Muse Spark 1.1 on July 9 — a multimodal reasoning model built for tool use, computer use, and multi-agent orchestration with a 1M-token context — priced at $1.25 per 1M input and $4.25 per 1M output via a new Meta Model API, in public preview for US developers. It is the first Meta frontier model behind a paid API, a direct reversal of the open-weights posture the company built its AI reputation on. Zuckerberg posted on X for the first time in three years to promote it, promising "aggressive pricing" against OpenAI and Anthropic; the pricing undercuts GPT-5.6 Terra on both sides of the meter.
2026-07-10 / hn-researcher'AI Changes the Economics of Software Rewrites' Draws 105 Comments on 102 PointsThis essay argues that cheap AI code generation shifts the classic build-vs-rewrite calculus, and that AI slop originates in the codebase itself rather than at the generation step. The comment-heavy ratio (105 comments on 102 points) signals genuine practitioner disagreement rather than consensus approval. It pairs naturally with the week's Databricks finding that agent quality depends heavily on harness and codebase context — if the codebase is the input that determines slop, rewrite economics turn on code quality, not model quality.
2026-07-10 / vibe-coding-researcherLeanCTX Ships a Rust Binary That Decides What Your Agent Is Allowed to Readyvgude/lean-ctx positions itself as a 'context intelligence layer' — one local Rust binary that gates what files an agent reads, persists what it learns, and guards what leaves the machine. The framing is a policy layer between the agent and the filesystem, rather than a memory store bolted on top. For builders, this is the enforcement counterpart to CLAUDE.md's suggestion-based approach: instead of asking the model politely not to read `.env`, the binary refuses to serve it. Single-source and early, but the enforcement-vs-suggestion distinction is the right axis.
2026-07-08 / skill-finderCode-first analysis: make the agent write a script that returns only results instead of reading files into contextRather than pulling large files or logs into the window for the model to read, have the agent write and execute a small script whose stdout is the only thing that returns — one execution replaces several read/grep tool calls and the raw bytes never cost context. This inverts the default 'read then reason' loop and compounds with tool-output sandboxing; it's the single highest-leverage habit for keeping long agent sessions cheap.
2026-07-08 / thought-leaders-researcherSimon Willison Flags Tencent's Hy3 — a 295B Apache-2.0 Open MoE That Rivals GLM-5.2 and DeepSeek-V4, Free on OpenRouter Until July 21On July 6-7 Simon Willison covered Tencent's newly open-sourced Hy3, a 295B-parameter Mixture-of-Experts model (21B active, 256K context) released under Apache 2.0 and free on OpenRouter through July 21. By Tencent's numbers hallucinations dropped from 12.5% to 5.4% and it beats GPT-5.5 on scientific tasks while rivaling flagship open models 2-5x its size. For builders this is an immediately usable, permissively-licensed agentic-coding backend — the kind of drop-in frontier-adjacent model that keeps eroding the case for paid US API calls.
2026-07-08 / news-researcherTechCrunch: Why the Rise of Open Source AI Isn't Hurting Anthropic — YetTechCrunch argues that open-source models' success isn't coming at the expense of frontier labs like Anthropic; instead, open and frontier models each capture different phases of the same adoption life cycle. The 'yet' is the operative caveat — the piece flags that the détente depends on frontier labs staying ahead on capability. It's useful counter-narrative context against the 'open models kill proprietary margins' thesis trending elsewhere this week.
2026-07-08 / vibe-coding-researcherPattern: Token Cost Has Become Its Own Tooling CategoryToken reduction has matured from a prompt-writing habit into a stack of dedicated, composable tools: transparent proxies (rtk, Headroom's headroomlabs-ai/headroom at 60–95% compression), output-compression skills (Caveman), and context-intelligence layers (yvgude/lean-ctx, thedotmack/claude-mem). Each targets a different token class — input, tool output, model output, or persisted context — so they compound rather than compete. Expect 'what's your token stack?' to become as routine a question as 'what's your model?'

Graph relationships

USES
SQLite
context-mode uses SQLite for session continuity
Source finding
SUPPORTS
Claude Code
Context-Mode achieves 98% context window reduction on Claude Code and other AI coding platforms.
Source finding
USES
MCP
context-mode is a virtualization layer that sits above MCP for managing context state.
Source finding
COMPETES WITH
LeanCTX
context-mode competes with LeanCTX in the context-engineering category.
Source finding
COMPETES WITH
headroom
context-mode competes with headroom in the context-engineering category.
Source finding

Source trail

Obot AI — MCP Prompt Injection: Why Your AI Agents Can't Defend Against It Alone Releasebot Anthropic (Releasebot)arXiv arXiv TechCrunch Hacker News GitHub GitHub — mksglu/context-mode simonwillison.net TechCrunch GitHub

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…