Thoughts on how AI is changing how we build, inspect, and fix software.
InvestorsDear investors — memo & calendarPR Review v4.1 deepens the OpenCode sandbox: exploratory testing, granular GitHub inline comments, live OpenCode activity streaming to the dashboard, and AES-256-GCM encrypted per-repository secrets for real integration runs.
Critique Builder is a browser-first cloud agent surface for starting repo-backed coding jobs, watching live execution, and inspecting results without turning the product into a browser IDE. Managed sandboxes, OpenCode, subagents, and a stronger bias toward shipping real software.
Kimi K2.6 replaces Kimi K2.5 across Critique’s PR review stack and Remedy chat models: architecture, benchmarks vs K2.5 and peers, integration notes, and April 21–27 2026 introductory pricing at 50% off former Kimi K2.5 credits.
Arcee AI’s Trinity-Large-Thinking is now available in Critique as `arcee-ai/trinity-large-thinking` at a 1-credit floor. We break down the Apache 2.0 release, MoE architecture, reasoning-trace constraints, benchmark profile, and why this model matters for long-horizon agentic review and Remedy.
Critique PR Review v4 is live in beta: nearly 4,000 PRs reviewed, sandbox-native OpenCode execution getting tighter, and Critique now showing up as the #2 public app this month on OpenRouter’s Embed V1 4B activity view just behind AnythingLLM.
Gemma-4-31B and MiMo-V2-Flash join critique.sh at a permanent 0.5cr floor. We compare SWE-bench Verified scores and critique.sh credit costs across Gemini 3 Flash, MiniMax-M2.5, Qwen3.5-27B, GPT-5.4 mini, Claude Haiku 4.5, GPT-5.1 Codex Max, and Claude Sonnet 4.6 — with a price-to-performance ratio chart.
Critique PR Review v3.1 moves final review synthesis into the sandbox. The sandbox now authors the structured review artifact with summary, findings, command timeline, runtime checks, and sub-agent output, while the app stays the control plane.
Critique PR Review v3 is live one day after the V2 release. Sandbox-backed AI code review, dedicated review-analysis workers, canonical review-run pages, GitHub App publishing, Remedy handoff, fix prompts, and an owner-only review assistant are now part of the product.
We shipped ten deep improvements to the Critique review engine: a semantic index, subsystem clustering for large PRs, intrinsic-risk-based drill-down targeting, cross-file relationship analysis, a CODE_QUALITY specialist, two-phase reasoning, domain heuristic packs, contract-diff analyzers, better evidence packaging, and expanded model support across GPT-5.4, Grok-4.2, GLM-5V-Turbo, MiMo v2.5, and more.
Critique now supports interactive PR chat via @critique mentions in GitHub pull request comments. Ask questions, run slash commands for review, explain, fix, security, tests, and more. Powered by Qwen 3.6 Plus — a frontier-level model with a 1M-token context window that beats GPT-5.1-Codex, GPT-5.2, and GPT-5.3-Codex on long-horizon benchmarks.
AI code review is now mainstream, but developer trust still lags dangerously behind adoption. Here is why engineering leaders should evaluate AI review tooling as governance infrastructure — and what the real buying criteria look like in 2026.
Remedy — Critique's autonomous code-fixing engine — is now available directly inside Critique Chat. Start from a review, choose a model, and watch fixes land on your PR branch in real time. Your machine never runs a line of this code.
A technical deep dive into hybrid retrieval for codebases: exact string search, semantic search, warmed snapshots, reciprocal rank fusion, and the operational details that make code search trustworthy.
Staff engineer or platform lead evaluating AI PR review automation? These five questions — permissions, token model, data egress, check behavior, and uninstall path — separate a trustworthy GitHub App from a risk review waiting to happen.
Critique Chat is the dashboard assistant grounded in your real GitHub codebase. Connect a repo, ask questions, get answers backed by live code search. Free for all signed-in users with frontier models like KAT Coder Pro V2, MiniMax M2.7, GLM-5V-Turbo, and MiMo v2.5.
MCP and CLI are often framed as rivals for AI agent tool invocation. Primary benchmarks show CLI can be 9–32× cheaper in tokens for well-known tools, while MCP wins on discoverability, OAuth, and tenant isolation. This article synthesizes the evidence and ends with a decision framework.
A detailed critique.sh essay on open-source and open-weight AI models, MiniMax M2.7 vs Claude Opus 4.6, why Chinese labs are leading the cost-capability curve, and the March 2026 credit promos for MiniMax, GLM-5, and Kimi K2.5.
Deep dive on Relace Search (relace/relace-search): 256K context, parallel tool calls, report_back handoff to oracle agents, OpenRouter pricing and telemetry, benchmarks vs naive retrieval, and how Critique Chat uses it today.
An honest map of cloud coding agents across GitHub, Cursor, Jules, Devin, OpenCode, E2B, and Vercel Sandbox, plus where Critique Remedy fits as a review-attached execution layer with BYOA support.
Xiaomi’s MiMo-V2-Flash and MiMo-V2-Pro: costs, coding benchmarks, hybrid attention, 1M context, and how critique.sh uses them for critique, fixes, and agentic review at every tier.
Inside the fastest adoption curve in developer tooling history — and why the bottleneck has moved from writing code to reviewing it. Claude Code, PR review crisis, and the infrastructure critique.sh was built to close.
For the last two years, most of the conversation around AI coding has focused on the wrong layer. In 2026, the more important reality is much bigger: AI is increasingly helping maintain, inspect, debug, and repair software too.