One developer's AI stack: how Cursor, Kiro, and ChatGPT share a single architecture

·1 views

A software developer has detailed a personal AI workflow that connects Cursor, Kiro, and ChatGPT through a shared architecture rather than treating each tool as a standalone product. The system uses a Logseq vault as persistent memory and a set of shared 'skills' — reusable instruction files — that any of the three IDEs can load depending on the task at hand. A Cursor Pro+ feature called Grok Bot acts as an orchestrator, routing work to specialized agents with persistent context about team members, coding standards, and project history. In practice, the setup handles tasks ranging from PR code reviews and production incident analysis to drafting technical articles, with each agent drawing on the same shared knowledge base. The developer argues the architecture solves a common pain point: paying for multiple AI tools that operate in isolation with no shared memory of previous work.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Diffusion Models Could Replace Autoregressive AI for Smarter Code Generation

Dominant AI coding tools like GitHub Copilot and ChatGPT rely on autoregressive models that generate code sequentially, token by token, making them ill-suited for tasks like mid-file insertion or large-scale codebase editing. Diffusion models, originally popularized in image generation, offer an alternative by starting with random token sequences and iteratively refining them into functional code. This non-autoregressive approach allows simultaneous edits across an entire sequence, enabling more holistic structural reasoning. Adapting diffusion models to code requires solving discrete-domain challenges, since unlike pixels, tokens cannot be partially corrupted with continuous noise. Emerging tools such as DiffCoder and CodeDiff are beginning to explore how these models can be integrated into production systems to address limitations that traditional large language models struggle to overcome.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Mismatched ARTF Envelopes Silently Mislead Auction Agents Despite Clean Parsing

The IAB Tech Lab's Agentic Real Time Framework (ARTF) places a containerized agent alongside ad exchange components, using a structured envelope that specifies the auction lifecycle stage and the relevant OpenRTB object. A common production error occurs when a bid_response from a prior auction hop is left on an envelope labeled LIFECYCLE_PUBLISHER_BID_REQUEST, a stage at which no DSP should have yet responded. Although gRPC accepts the malformed bytes without error, the agent receives a distorted view of its position in the auction, potentially acting on data that does not belong at that stage. The open-source linter RTBlint flags this as artf.lifecycle.payload_unexpected and catches the inverse case — a response-stage envelope missing a bid_response — as artf.lifecycle.payload_mismatch, which can block bid-shading logic. Both issues typically stem from teams reusing a single JSON blob across auction hops without stripping stage-inappropriate fields before forwarding.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Anthropic Open-Sources Bloom and Petri Frameworks for AI Behavioral Auditing

Anthropic released Bloom and Petri on December 19, 2025, two open-source tools designed to make behavioral evaluation of frontier AI models more systematic and reproducible. Bloom automates the creation of evaluation suites through a four-stage pipeline — Understanding, Ideation, Rollout, and Judgment — starting from a seed configuration. Petri complements Bloom by enabling parallel exploration of risk interactions during auditing. The tools were tested across 16 frontier models and four behaviors, including sycophancy, sabotage, self-preservation, and self-preferential bias. Both frameworks are available under the MIT license, though Anthropic notes that benchmark results do not substitute for real-world safety testing in production environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Quota Catchup Script to Recover Skipped launchd Jobs After API Limits Reset

A developer extended their Claude API quota circuit breaker, claude-quota-guard.py, which protects 15 launchd jobs on macOS, to address a gap that emerged after quota recovery. When the circuit breaker halts jobs during quota exhaustion, launchd does not reschedule them until their next calendar slot, potentially causing a full day of missed executions. To fix this, a companion script called quota-catchup.py was built to identify and re-run only the jobs that were skipped that day. The script applies three conditions to filter candidates: the job must be guarded by claude-quota-guard.py, must show a skipped marker from today, and must have had at least one scheduled time slot already pass. Jobs on repeating intervals are excluded since they self-recover, while calendar-scheduled jobs with missed slots are prioritised for manual re-triggering.

0 comments Read more at DEV Community