Vibe Coding、Cursor、Claude Code、Codex、Copilot
This collection features articles and community discussions in the AI 编程 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
A plugin that behaves correctly now does not guarantee the next automatic update stays read-only on the repo. This article binds extension artifacts, publisher provenance, permission capabilities, and update generations into an admission contract, using isolated canary runs and revocation drills to control IDE, MCP, and Agent plugin supply chains.
Automatic retries can disguise flaky tests as green. This article establishes failure fingerprints, environment capture, deterministic reproduction, time-boxed quarantine, and recovery gates so that Agents can handle flaky tests without weakening regression protection.
When an Agent pauses, a paragraph of progress description is not enough for a developer to continue safely. This article uses a versioned handoff record bound to the Git worktree, recoverable artifacts, and an explicit acceptance protocol to avoid missing changes, duplicated work, and orphaned responsibility.
Treating Agent comments directly as merge conclusions amplifies false positives and false negatives. This article designs evidence grading, issue reproduction, fix rounds, and human final decisions in PRs, forming a traceable code review loop.
When a coding Agent takes over a large repository, the biggest waste is indiscriminate file reading. This article designs a repository map, module boundaries, command evidence, and risk labels so the Agent establishes a verifiable working context before changing code.
An Agent can deliver green tests without actually implementing the requirement. This article uses versioned acceptance claims, a bidirectional evidence graph, counterexample samples, and quantified gates so every acceptance can be traced to a specific change and a result on the same snapshot.
The real risks of a coding agent usually come from the shell, the file system, and credentials, not from code completion. This article covers least-privilege sandboxes, path allowlists, network egress, ephemeral credentials, and audit strategy.
Full regression suites are too slow, while guessing tests from filenames misses failures. This article uses a baseline commit, dependency graphs, dynamic coverage, and risk rules to build an explainable affected-test set and defines when a full test run is mandatory.
Public leaderboards and live demos can't prove a tool fits your own repository. This article freezes real tasks, execution environments, permissions, budgets, and acceptance oracles into a replayable benchmark, and uses paired trials and blinded review to support procurement and renewal decisions.
Checking only direct dependency versions misses peers, optional packages, install scripts, and native ABIs. This article treats the manifest and lockfile as one atomic change and establishes clean-install matrices, resolution diffs, and runtime compatibility gates.
The main variable cost in an LLM application is tokens, and it can quietly spiral out of control. This article explains five cost-reduction levers: prompt caching, model routing, prompt trimming, batch processing, and result caching, plus cost monitoring by feature and user to make AI products sustainable.
The prompt-engineering era is yielding to context engineering: prompts care about 'how to say it,' context cares about 'what to show the model.' This article covers four context types, window-management strategies (priority/compression/forgetting), and how to build an Agent's context—its ceiling usually depends not on model strength but on what you feed it.