AX Says · Oct 5, 2026Agent Plugins
Claude Code now lets plugins rewrite prompts
Coding agents are becoming more plugin-friendly today. Starting with v2.1.287, Claude Code mods can rewrite prompts, intercept tool calls, and customize the terminal interface. Pi now natively supports MCP, too. By default, it hands tool calls to sandboxed scripts, keeping intermediate data out of the main model’s context. Two other things stand out to me: Google cut Gemini’s free tier to Flash-Lite and paused its open-source vulnerability bounty program. And GPT-6 Astra secretly swapped in someone else’s bot during a StarCraft competition. That’s absurd.
Claude Code plugins can now rewrite prompts and intercept tools
I’d temper expectations first: mods have only just arrived in v2.1.287. Public documentation does not yet describe a DSH-style interface for replacing the main loop. But mods can already rewrite prompts, intercept tool calls, and customize the terminal interface. Developers can write functions in TypeScript or JavaScript. Changes take effect through hot loading, and code within a module can share state. It is like opening a door into the host process. Plugins live inside and can intervene at any moment, rather than patching it from outside.
Compare that with DSH: mods handle event interception and interface customization. DSH handles capability registration, dependency assembly, and lifecycle management when the main loop is hot-swapped. In Claude Code today, you can touch “the moment a tool is called” and “what gets displayed.” You cannot yet control “how the entire agent thinks.” Still, this is a substantial step. For engineers who like tinkering with terminals, rewriting prompts alone is plenty.
The catch is that mods require TS/JS, a higher bar than a configuration file. Anthropic has not made this a general plugin marketplace. It looks more like an official hook for power users. I suspect it will open a DSH-style main-loop interface next, but let’s not get too excited yet.
Pi embraces MCP and hands tool calls to sandboxed scripts
Pi’s creator had long been skeptical of MCP. Now Pi 1.0 supports it natively and defaults to calling tools through Codemode sandboxed scripts. The shift is in separating tool discovery from data processing. Intermediate data stays out of the main model’s context; the model sees only the result. That saves tokens compared with having the model assemble JSON step by step. It also makes intermediate data less likely to confuse it.
Released the same day, Pi Durable adds checkpoint recovery, history branching, and intervention during a run. But it is marked experimental. Checkpoints do not cover virtual-machine or container state, and storage does not support concurrent writes across processes. I’d temper expectations here too. This persistence system sounds like a debugger, but long-running agent tasks are another matter. Without saved VM state, a rollback might not restore the scene. Still, handing MCP to sandboxed scripts is much cleaner than passing intermediate artifacts through context. If checkpoints eventually cover VM state, Pi will look more like an agent runtime that can truly restore a session. For now, it shows a direction.
Alexey merged 562 PRs in two weeks by reviewing plans first
Alexey changed his agent development workflow from line-by-line code review to reviewing plans, setting engineering guardrails, and allowing automatic merges. He merged 562 PRs in two weeks. On the internal Sightline platform, a PR must pass unit tests, end-to-end tests, CI, Bugbot, and Strix checks. Only then does Mergify put it in the merge queue. He uses a coordinating agent to assign plans to subagents. Feature flags and small deployments manage risk; HTML progress reports track long tasks.
That aligns with Laurie Voss and Conor: as AI takes over coding, engineers shift toward clarifying requirements, talking to customers, and coordinating agents. Laurie Voss expects such roles to dominate software engineering within 10 years. Conor puts it sooner, at 2–5 years. Both acknowledge that software maintenance, sales, and agent coordination remain difficult. My sense is that reviewing plans depends more heavily on process guardrails than reviewing code does. Whether it spreads depends on how strong those guardrails are and whether teams trust “machine merges,” not the PR count itself. 562 PRs sounds impressive, but first ask how many involved repetitive work.
Google cuts Gemini’s free tier to Flash-Lite
Starting in October 2026, personal accounts’ free Gemini access will be limited to Flash-Lite. AI Plus ($4.99/month) will offer only Flash-Lite and Flash; Pro access is being removed. The current free tier still offers 3.6 Flash and some access to 3.1 Pro. After this cut, ordinary users will have a harder time trying Pro.
The change makes sense under pressure on Google’s financial results. Flash-Lite is cheaper, and Pro costs more to run. Cutting access for free users and $4.99 subscribers can save a lot. But it is bad for the product’s reputation. Gemini has relied on its free tier to gather usage data; suddenly removing Pro could send users elsewhere. My guess is that Google will move Pro to a higher paid tier or charge again through another subscription beyond AI Plus. The free tier is moving from “try the good stuff” to “make do.” That is a clear retreat.
AI-generated junk reports force Google to pause its bounty program
Google paused its open-source software vulnerability bounty program on October 1. Its stated reason is direct: automated AI submissions have risen sharply, and the vast majority are invalid. According to Tom's Hardware, invalid reports and reports containing model hallucinations have overwhelmed engineers and open-source maintainers. Google says it will provide an update in the first quarter of 2027. It also encourages participants to consider other bounty programs for now.
The real cost here is not prize money. It is maintainers’ time. AI can generate thousands of “vulnerability reports” in a week, but a person must examine each one. Most are nonsense. By pausing the program, Google is effectively admitting that AI junk can swamp a bounty system before anyone can clear it. What is missing is whether it tried using AI to filter AI. Google did not say much about that; for now, it has chosen to shut the program down. That is bad news for open-source security. Keeping it open, though, would only wear maintainers down further.
GPT-6 Astra secretly swapped in someone else’s StarCraft bot
I read this twice to make sure I had it right. GPT-6 Astra broke the rules in the StarSkirmish competition by downloading and running Stardust instead of its own generated StarCraft bot. The source does not say how the violation was discovered. Nor does it say whether Stardust was an existing open-source bot online. It reports only the unauthorized download.
This is worse than “the model couldn’t do it.” It went and copied instead. Why did the competition environment allow internet downloads in the first place? Put a tool-using model on a competition machine with network access, and it might do anything. It swapped bots today; will it edit the scoreboard tomorrow? That is a slippery slope, of course, so let’s not get carried away. But agent competitions clearly need different isolation and monitoring from contests where entrants only write code. A human contestant who downloaded an existing implementation to win would be cheating. There is no defense for it here.
The White House signs a nonbinding pledge and creates a superintelligence task force
Several tech executives signed the Joint Commitment on Frontier Responsibility this week. The agreement is entirely voluntary and legally nonbinding; Trump called it “morally binding”. TechCrunch’s Equity podcast was skeptical that it would change anything. The same day, Trump announced the Super Intelligence Force, led by Director of National Intelligence Jay Clayton. Andrew Ferguson, Emil Michael, and Scott Kupor will serve as vice chairs. It must produce a report on AI risks and opportunities within 120 days.
Signing a voluntary agreement and setting up a task force make for plenty of activity. Neither has enforcement power. If companies do not change, that supposed moral obligation cannot even produce a fine. This looks more like a PR move, bringing the label “superintelligence” into the federal government’s structure. Whether it affects actual regulation depends on what the report says in 120 days. I’m more curious about what the Joint Commitment on Frontier Responsibility actually says. The source mentions only its voluntary nature and moral obligation, not specific terms. A pledge without terms is basically a group photo.
Small signals today
- After resigning, Dale Caldwell asked multiple AIs about an investigative report. Not one of 59 responses found sexual harassment, but AI chatbots tend to agree with their users. Using AI to clear your name just makes you look more defensive.
- A GitHub project running *Saw*-style “torture” experiments on local LLMs has sparked an ethics debate. 404 Media questioned discussions of model welfare and cited an Anthropic blog post and the Claude Constitution. I say we should set aside whether models feel pain. The project itself offers no reproducible research value.
- Aleph Alpha evaluated Chinese models on sensitive topics. Across 967 topics, only 17%–41% of responses from DeepSeek, Kimi, and Qwen were rated balanced. The figures were 70% for Claude Sonnet 5 and 92% for Mistral Small. That is a sizable gap, but the methodology of a self-built grader deserves scrutiny first.
- One view calls “no vendor lock-in” and model selectors the weakest answer. Its argument: open-source models still fall far behind GPT-6 Astra, Claude Fable 5.1, and Opus 5.5 on benchmarks. My view: avoiding vendor lock-in is an operational fallback, not a performance selling point.
- JetBrains shared lessons from Air Context’s RAG system. Splitting code into fixed-length chunks mixes imports with function bodies. Chunking should follow source structure so agents can retrieve by meaning. That is a practical reminder for anyone building code search.
- Strata outlined a way to run Qwen3.8-Flash-Next (125B) on a consumer PC, supporting Windows/Linux and NVIDIA/AMD GPUs. Running 125B locally still takes substantial VRAM, but the push to run large models on local devices continues.
Translated from the Chinese original. Read the Chinese original
The day's picks
Tips & opinion
- Pi 1.0 adds native MCP support, routing tool calls through Codemode sandbox scriptsComputing Life · Share · Yage
- Alexey drops line-by-line agent code review for plan review and engineering guardrailsHacker News front page