<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Alexandre Mahe — Notes and Thoughts</title><description>Notes and essays by Alexandre Mahe.</description><link>https://alexandre.mahe.io/</link><language>en</language><item><title>An agent&apos;s context: client knowledge, memory and permissions</title><link>https://alexandre.mahe.io/blog/an-agents-context/</link><guid isPermaLink="true">https://alexandre.mahe.io/blog/an-agents-context/</guid><description>The agent explores the code on its own. Refining its context means bringing it what it can&apos;t find: knowledge of the client, a memory of its sessions and permissions around them, calibrated to confidentiality and compliance.</description><pubDate>Fri, 20 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The agent &lt;a href=&quot;https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;explores the codebase on its own&lt;/a&gt;: it picks the commands, knows where to look, reads the git history, runs the tests and, integrated into the editor, uses the same tools as the developer, up to driving a browser or taking screenshots with &lt;a href=&quot;https://github.com/microsoft/playwright-mcp&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Playwright MCP&lt;/a&gt;. That part of the context is the job of the agent and the harness. What&apos;s left to refine is what it won&apos;t find in any file.&lt;/p&gt;
&lt;p&gt;Prototyping a page, writing interface copy, generating a document in the company&apos;s colours: each of these tasks requires knowing its brand guidelines. Without them, the agent produces something correct, but anonymous. Knowing the client&apos;s interests, positioning and priorities also yields better-scoped tasks, with milestones that make sense to the client, not just to the code.&lt;/p&gt;
&lt;p&gt;Training an open model on the company&apos;s data could give better results, but it&apos;s a project in its own right, rarely within a freelancer&apos;s reach. A freelancer stands out elsewhere: anyone can have a well-tooled agent; deep knowledge of the trade and the client, far fewer can.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;This knowledge makes the difference, but what you give the agent, and how it accesses it, depend on confidentiality, compliance and permissions.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-ontology-behind-authorised-access&quot;&gt;The ontology, behind authorised access&lt;/h2&gt;
&lt;p&gt;This body of knowledge (the client&apos;s business, brand, structure, governance, interests, positioning, vision) has a name: the ontology. It&apos;s often the most useful context you can give an agent, and the most sensitive.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/an-agents-context/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;I keep the ontology separate, often in a private repository, and the agent accesses it through an &lt;a href=&quot;https://modelcontextprotocol.io/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;MCP&lt;/a&gt; server, protected by &lt;a href=&quot;https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;the protocol&apos;s authorisation&lt;/a&gt; when it&apos;s remote, and by system permissions when it&apos;s local. The agent gets only the part of the ontology its task justifies, instead of the whole corpus pasted into its prompt, and the client&apos;s knowledge stays under the client&apos;s control.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://arxiv.org/abs/2005.11401&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;RAG&lt;/a&gt; becomes useful in two cases: when the corpus is too large to browse, like years of documentation or meeting notes, and when every access an agent makes to a piece of data must be authorised and traced. Retrieval then goes through a service that checks permissions before answering.&lt;/p&gt;
&lt;h2 id=&quot;calibrate-to-sensitivity&quot;&gt;Calibrate to sensitivity&lt;/h2&gt;
&lt;p&gt;The brand&apos;s colours, typography and tone are often public: a short file in the repository is enough, and the agent reads it on every task. The client&apos;s interests, priorities and roadmap are confidential: they go through the authorised MCP, and only for the tasks that need them. Its users&apos; personal data has no place in the ontology: if a task needs it, it&apos;s anonymised before it leaves the machine, or processed by a local model, as in &lt;a href=&quot;../client-data-and-llms/&quot;&gt;the post on verification&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;So the same knowledge takes three different paths. The choice doesn&apos;t depend on what would be most convenient for the agent, but on what the client has authorised.&lt;/p&gt;
&lt;p&gt;What enters the context can also carry instructions. A document from the corpus, a response from an MCP server or an old session can contain a prompt injection; a source only enters the context if it&apos;s trusted, or after filtering by the proxy.&lt;/p&gt;
&lt;h2 id=&quot;a-memory-for-introspection&quot;&gt;A memory for introspection&lt;/h2&gt;
&lt;p&gt;Going back over a previous session helps the agent recover a forgotten task or a security warning that was never followed up. That memory has to exist first, and not everything should be lumped together.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The session archive.&lt;/strong&gt; Every session leaves a trace: the prompt, the response, the tokens consumed, the timestamp. I store it with the client&apos;s project, in git or in a git-ignored folder when it must not leave the machine: it contains the client&apos;s data, and leaves with the rest at the end of the engagement. When these files pile up, a &lt;a href=&quot;https://www.postgresql.org/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;PostgreSQL&lt;/a&gt; database takes over: the same content, but queryable, and protected by the database&apos;s access rights.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Context shared between agents.&lt;/strong&gt; An agent and its subagents need a common base, which can live in that same database. Placed at the start of requests, it benefits from &lt;a href=&quot;https://www.anthropic.com/news/prompt-caching&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;prompt caching&lt;/a&gt;, whose lifetime is measured in minutes.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Task tracking.&lt;/strong&gt; Checkboxes in tickets and statuses that move: the agent knows where it stands, so does the developer, and the next session picks up without re-explaining.&lt;/p&gt;
&lt;p&gt;The ontology, for its part, stays separate, with its restricted access: it&apos;s not working memory, it&apos;s reference knowledge.&lt;/p&gt;
&lt;h2 id=&quot;an-ontology-in-three-files&quot;&gt;An ontology in three files&lt;/h2&gt;
&lt;p&gt;Without an MCP server or RAG, a few short files are enough to start. The public folder lives in the repository; the confidential one, elsewhere, behind authorised access:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-text&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span&gt;client/&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;  brand.md        colours, typography, tone&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;  business.md     what it does, what it stands for, what it refuses&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;  priorities.md   confidential: roadmap, interests, to serve over MCP&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;The confidential file goes through the &lt;a href=&quot;https://github.com/modelcontextprotocol/servers/tree/main/src/filesystem&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;reference MCP server&lt;/a&gt;, restricted to its folder, with its write tools denied in the agent&apos;s permissions:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-bash&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;claude&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; mcp&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; add&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; ontology&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; npx&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; -y&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; @modelcontextprotocol/server-filesystem&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; ~/clients/acme/ontology&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-json&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;deny&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: [&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;mcp__ontology__write_file&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;mcp__ontology__edit_file&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#6E011A, #FFB1AF);--shiki-light:#6E011A;--shiki-light-font-style:italic;--shiki-dark:#FFB1AF;--shiki-dark-font-style:italic&quot;&gt;...&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;And a ticket the agent keeps up to date itself, so the next session picks up without re-explaining:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-md&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;-&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt; [&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;x&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;] Parse the CSV export&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;-&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt; [ ] Add tests for invalid dates&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;-&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt; [ ] Validate the format with the client&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;These files stay short: the ETH Zurich study cited in &lt;a href=&quot;../a-minimal-harness/&quot;&gt;the post on the harness&lt;/a&gt; showed that &lt;a href=&quot;https://arxiv.org/abs/2602.11988&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;automatically generated context files&lt;/a&gt; slightly degrade results while increasing cost. The rest gets added when the project calls for it, starting with what the client has authorised.&lt;/p&gt;</content:encoded><category>agents</category><category>context-engineering</category><category>mcp</category><category>rag</category><category>memory</category></item><item><title>A minimal harness: reuse what exists, keep the agent in check</title><link>https://alexandre.mahe.io/blog/a-minimal-harness/</link><guid isPermaLink="true">https://alexandre.mahe.io/blog/a-minimal-harness/</guid><description>Mapping all development know-how into skills costs more than it returns. I reuse what exists, write only the high-value skills, and the harness is mainly there to keep the agent in check: permissions, tests, oversight.</description><pubDate>Thu, 19 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;When an agent gets stuck, the easy fix is to add an instruction: a git skill, a Scrum skill, a skill per API. A few months later, the developer experience has turned into a museum of instructions to maintain, half of them out of date.&lt;/p&gt;
&lt;p&gt;Those instructions aren&apos;t free. An ETH Zurich study shows that LLM-generated &lt;a href=&quot;https://arxiv.org/abs/2602.11988&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;context files like AGENTS.md&lt;/a&gt; slightly lower agents&apos; success rate and raise inference cost by more than 20%; written by developers, they bring a modest gain, at a higher cost. A frontier model already knows git and the common conventions. For a freelancer, productivity comes less from the number of instructions than from verified tools they know and can defend in front of a client.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;I reuse what exists, and only write the skills that bring real value.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;using-a-skill-well&quot;&gt;Using a skill well&lt;/h2&gt;
&lt;p&gt;A context file helps when it brings what the model doesn&apos;t know, and hurts when it repeats what it already knows. A &lt;a href=&quot;https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;skill&lt;/a&gt; follows the same rule, with one extra mechanism: the agent first sees only its name and description, and loads the rest only when needed. So the description decides whether the skill gets used; the main file stays short, and the details go into files it references. When I write one, the description says when to use it rather than what it is, and a deterministic step becomes a script rather than a procedure the model would have to follow to the letter.&lt;/p&gt;
&lt;p&gt;For an API released after training, a skill isn&apos;t the best answer: &lt;a href=&quot;https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Vercel measured&lt;/a&gt; that a compact index in the context did better than a skill the agent forgot to load. Documentation increasingly ships through &lt;a href=&quot;https://llmstxt.org/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;llms.txt&lt;/a&gt; or &lt;a href=&quot;https://modelcontextprotocol.io/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;MCP&lt;/a&gt; servers such as &lt;a href=&quot;https://github.com/upstash/context7&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Context7&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id=&quot;team-skills-personal-skills&quot;&gt;Team skills, personal skills&lt;/h2&gt;
&lt;p&gt;Team skills describe what you can&apos;t guess by reading the code: how the team works with it, its conventions, its review steps. They live in the repository and serve everyone.&lt;/p&gt;
&lt;p&gt;Personal skills carry my shortcuts, my methods, my way of breaking down a task, as long as they stay compatible with the project. They follow me from one project to the next and end up defining my own augmented developer agent.&lt;/p&gt;
&lt;p&gt;It&apos;s also a matter of intellectual property. What comes from the company (its workflows, its conventions, its domain) belongs to it; what comes from my know-how belongs to me. Skills that distil years of method should help me, not feed the client&apos;s agents after the engagement: my personal skills stay in my own repository.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/a-minimal-harness/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;h2 id=&quot;influencing-the-first-inference-as-little-as-possible&quot;&gt;Influencing the first inference as little as possible&lt;/h2&gt;
&lt;p&gt;&lt;a href=&quot;https://www.mariozechner.at/posts/2025-11-30-pi-coding-agent/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Pi&lt;/a&gt; takes this logic all the way: &lt;a href=&quot;https://www.ai.engineer/talks/RjfbvDXpFls-pi-coding-agent&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;a short system prompt and four tools&lt;/a&gt;, almost nothing before the user&apos;s prompt.&lt;/p&gt;
&lt;p&gt;The Claude subscription &lt;a href=&quot;https://www.theregister.com/2026/02/20/anthropic_clarifies_ban_third_party_claude_access&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;only covers Claude Code and Claude.ai&lt;/a&gt;, not Pi. Nothing stops you from designing the harness with Pi, then porting it to Claude Code as skills, hooks and &lt;a href=&quot;https://code.claude.com/docs/en/subagents&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;subagents&lt;/a&gt;, which keep the main session clean.&lt;/p&gt;
&lt;h2 id=&quot;security-first-experience-second&quot;&gt;Security first, experience second&lt;/h2&gt;
&lt;p&gt;Context engineering decides what reaches the model at each step; the harness decides how the agent works, including the mechanisms that build that context.&lt;/p&gt;
&lt;p&gt;My two priorities, in that order. Security first: strict permissions, tests and CI that check the result without adding anything to the prompt, and tools reviewed before they are reused. A third-party skill can contain scripts, an MCP server is code running with the agent&apos;s permissions, and either one can slip instructions into its context: I treat them as dependencies. Developer experience second: a repository that is pleasant to work in is easier to maintain, for the human and the agent alike. Between the two, prompts stay minimal.&lt;/p&gt;
&lt;p&gt;A rule that must always apply has no place in a prompt. Claude Code &lt;a href=&quot;https://code.claude.com/docs/en/hooks&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;hooks&lt;/a&gt; enforce it deterministically: block a command, format an edited file, run the tests before handing back control. Other agents have their equivalents, under other names.&lt;/p&gt;
&lt;p&gt;The harness also covers moving from a local session to a remote one, talking to an agent running in a loop, and cost oversight, which goes through the proxy described in &lt;a href=&quot;../client-data-and-llms/&quot;&gt;the post on client data&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/a-minimal-harness/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;h2 id=&quot;a-comfortable-agent-in-three-blocks&quot;&gt;A comfortable agent in three blocks&lt;/h2&gt;
&lt;p&gt;Working preferences, in a few lines of the context file. Mine:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-text&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span&gt;Break the work into small tasks and parallelise what can be parallelised.&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;Run the tests at the end of each task, not after every change.&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;Keep a progress report up to date in PROGRESS.md.&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;Ask a question rather than guess at a vague requirement.&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;Permissions, in &lt;code&gt;.claude/settings.json&lt;/code&gt;, versioned with the repository; the &lt;a href=&quot;https://code.claude.com/docs/en/permissions&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;documentation&lt;/a&gt; details the syntax. Pi has none by default; &lt;a href=&quot;https://pi.dev/packages/pi-permission-modes&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;extensions&lt;/a&gt; add them.&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-json&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;allow&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: [&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(npm run test *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(git diff *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(git commit *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#6E011A, #FFB1AF);--shiki-light:#6E011A;--shiki-light-font-style:italic;--shiki-dark:#FFB1AF;--shiki-dark-font-style:italic&quot;&gt;...&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;],&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;ask&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;:   [&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(git push *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(npm install *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;],&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;deny&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;:  [&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Read(./.env)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Read(./secrets/**)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;Bash(curl *)&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#6E011A, #FFB1AF);--shiki-light:#6E011A;--shiki-light-font-style:italic;--shiki-dark:#FFB1AF;--shiki-dark-font-style:italic&quot;&gt;...&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;A &lt;code&gt;Read&lt;/code&gt; rule does not, however, block a &lt;code&gt;cat .env&lt;/code&gt; run through the shell: the real protection is still having no secrets in the container, as in &lt;a href=&quot;../running-untrusted-repos/&quot;&gt;the first post&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Finally, a reputable skill collection rather than a home-grown library. &lt;a href=&quot;https://github.com/obra/superpowers&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Superpowers&lt;/a&gt;, by Jesse Vincent, covers design, TDD, debugging and review; &lt;a href=&quot;https://github.com/anthropics/skills&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Anthropic&apos;s official repository&lt;/a&gt; is the reference for writing your own. Review before installing, like any dependency:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-bash&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;/plugin&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; install&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; superpowers@claude-plugins-official&lt;/span&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;   # Claude Code&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;pi&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; install&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; git:github.com/obra/superpowers&lt;/span&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;            # Pi&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;Next post: &lt;a href=&quot;../an-agents-context/&quot;&gt;context, meaning what the agent knows about the client, and its memory&lt;/a&gt;.&lt;/p&gt;</content:encoded><category>agents</category><category>harness</category><category>skills</category><category>claude-code</category><category>pi</category></item><item><title>One model per role, with just enough intelligence</title><link>https://alexandre.mahe.io/blog/minimum-sufficient-intelligence/</link><guid isPermaLink="true">https://alexandre.mahe.io/blog/minimum-sufficient-intelligence/</guid><description>Spec-driven development with less human in the loop: each step becomes a role, served by the minimum intelligence that does the job. The developer becomes product owner and settles early who validates what.</description><pubDate>Wed, 18 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Competition between freelancers and client expectations push for more output. A freelancer doesn&apos;t have to bend to AI: they can make it converge on their own practices and know-how. To reduce the human&apos;s share of the loop, the first step is to copy the stages of continuous development: specify, break down, code, test, review, decide, maintain. What changes is who holds each stage.&lt;/p&gt;
&lt;p&gt;If agents are the developers, the natural framework is spec-driven development. You discuss the specs at length, add the context of the domain, the partnerships and the methods, then derive the user stories, tasks and tests from them. The developer no longer types every line, but spots vague requirements and missed cases: they become a product owner with an engineering background. This orchestrator stance keeps what makes the job enjoyable: understanding the problem, deciding, building.&lt;/p&gt;
&lt;p&gt;Addy Osmani sums up the split: agents can own &lt;a href=&quot;https://www.oreilly.com/radar/own-the-outer-loop/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;the inner loop&lt;/a&gt;, where you investigate, code, test and report, but humans keep the outer loop, the one of decisions and accountability. A single expensive model for everything costs too much; a single cheap model misses what takes thinking.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Each role gets the cheapest model that does the job correctly.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;each-step-becomes-a-role&quot;&gt;Each step becomes a role&lt;/h2&gt;
&lt;p&gt;The loop works through the tasks one by one, runs some in parallel, and waits when it has to. Claude Code already chains a good part of this work within a single session; separating the roles is the harness&apos;s job, the topic of the next post.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/minimum-sufficient-intelligence/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;h2 id=&quot;three-tiers-of-models-and-a-decision-model&quot;&gt;Three tiers of models, and a decision model&lt;/h2&gt;
&lt;p&gt;I think in three tiers, valid with any provider, or mixing them: frontier models, second-tier models, and lighter models. At Anthropic, today these are Fable, Opus and Sonnet; the names will change, the logic won&apos;t.&lt;/p&gt;
&lt;p&gt;Roles that take thinking, such as writing the specs or reviewing the code, benefit from a frontier model, with a high rather than maximum reasoning effort. The code can go to a cheaper second-tier model. For small, well-defined tasks, a lighter model is enough.&lt;/p&gt;
&lt;p&gt;Decisions don&apos;t need text. Choosing what comes next, judging a risk, accepting or rejecting: a decision model such as &lt;a href=&quot;https://typesafe.ai/blog/introducing-system-one-models-and-jev&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Jev&lt;/a&gt;, from TypeSafe, returns a typed choice with a confidence score directly, without going through an LLM that generates JSON. A general-purpose AI coordinates the agents, under human supervision.&lt;/p&gt;
&lt;p&gt;Compliance sometimes narrows the choice: if the client requires an attested service, as in &lt;a href=&quot;../client-data-and-llms/&quot;&gt;the post on verification&lt;/a&gt;, what&apos;s left is mostly open models.&lt;/p&gt;
&lt;h2 id=&quot;a-long-loop-that-runs-without-the-developer&quot;&gt;A long loop that runs without the developer&lt;/h2&gt;
&lt;p&gt;An agent loop takes time. The developer checks in on it without waiting for a slow model to finish its answer, and the developer experience has to follow that pace.&lt;/p&gt;
&lt;p&gt;With my subscription, I use &lt;a href=&quot;https://code.claude.com/docs/en/claude-code-on-the-web&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Claude Code cloud sessions&lt;/a&gt;, and other providers offer the equivalent. Claude clones the repository into its own VM, takes ownership of the code, contributes and commits. On a client&apos;s code, that VM at the provider needs the same client consent as any LLM. I follow it and steer it from &lt;a href=&quot;https://code.claude.com/docs/en/mobile&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;the mobile app&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;These agents need the relevant knowledge and the session history, with nothing local: the computer must be able to shut down. Rather than handing them my own credentials while I sleep, a secrets manager issues them, per task, short-lived access scoped to the data they need.&lt;/p&gt;
&lt;p&gt;From my machine, I steer these agents through a main agent: costs, permissions, shared credentials. Giving each agent its own account, a &lt;a href=&quot;https://eips.ethereum.org/EIPS/eip-4337&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;programmable wallet&lt;/a&gt; with spending caps or a &lt;a href=&quot;https://www.w3.org/TR/did-core/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;decentralised identity&lt;/a&gt;, allows fine-grained control over its permissions and budget.&lt;/p&gt;
&lt;h2 id=&quot;who-decides-what&quot;&gt;Who decides what&lt;/h2&gt;
&lt;p&gt;The level of validation depends on the severity of a potential problem and the impact of the change.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/minimum-sufficient-intelligence/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;These classes are defined early: you identify where a check is needed, set its level with the client, then enforce it in code, with permissions, credentials and merge rights. A rule that exists only in a document relies on everyone&apos;s goodwill.&lt;/p&gt;
&lt;p&gt;Osmani proposes an &quot;accountability contract&quot; per codebase: the checks agreed for a change, the evidence the decision rests on, and the person who answers for it. That is what the client must be able to review. The classes also address what he calls the orchestration tax: human attention doesn&apos;t scale with the agents. The more precise they are, the less the developer reviews what doesn&apos;t need it, and the less cognitive debt they accumulate.&lt;/p&gt;
&lt;h2 id=&quot;classes-of-changes-in-the-repository&quot;&gt;Classes of changes in the repository&lt;/h2&gt;
&lt;p&gt;Without a fleet of agents, I start by listing the project&apos;s types of change and, for each, who validates: the agent alone, me, or the client. A &lt;a href=&quot;https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/about-code-owners&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;CODEOWNERS&lt;/a&gt; file records them in the repository:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-text&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span&gt;# .github/CODEOWNERS&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;*                @me&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;/infra/          @client/ops&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span&gt;/src/payments/   @client/tech-leads&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;A protected branch that requires code owner review enforces it. It&apos;s a basic solution: it only covers code review, file by file. A real project usually needs a more complete permission system, on the data itself: who, human or agent, can read or modify what, and with what validation, as in &lt;a href=&quot;../an-agents-context/&quot;&gt;the post on context&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Next post: &lt;a href=&quot;../a-minimal-harness/&quot;&gt;the harness, and why it&apos;s better not to overdo it&lt;/a&gt;.&lt;/p&gt;</content:encoded><category>agents</category><category>llm</category><category>models</category><category>governance</category><category>specs</category></item><item><title>Client data and LLMs: protect, verify, choose where inference runs</title><link>https://alexandre.mahe.io/blog/client-data-and-llms/</link><guid isPermaLink="true">https://alexandre.mahe.io/blog/client-data-and-llms/</guid><description>Data protection starts as soon as you write the system that processes it. You can stay compliant with a capable model if what the LLM does is verifiable; a proxy centralises costs, anonymisation and keys.</description><pubDate>Tue, 17 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;A client gives me access to their repository, and the temptation is to start Claude Code within the minute. Often, they provide their own infrastructure. If not, their constraints may call for a more elaborate inference stack.&lt;/p&gt;
&lt;p&gt;During development, the agent sees what the developer sees: the database dump pulled to reproduce a bug, the staging &lt;code&gt;.env&lt;/code&gt;, the logs copied from production. Nothing is in production, yet everything already goes into the model&apos;s context, and the contract almost never mentions it. The client expects speed and compliance: one must not cost the other.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Working on private data with a capable model while staying compliant is possible, if what the LLM does is verifiable, by a verifier that is itself verifiable.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;two-kinds-of-compliance-the-clients-and-mine&quot;&gt;Two kinds of compliance: the client&apos;s and mine&lt;/h2&gt;
&lt;p&gt;If the code or data sent to the LLM contain personal data, the client is the controller, the freelancer is their processor, and the LLM provider becomes a sub-processor: &lt;a href=&quot;https://www.cnil.fr/fr/reglement-europeen-protection-donnees/chapitre4#Article28&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Article 28 of the GDPR&lt;/a&gt; then requires the client&apos;s prior written authorisation. Without personal data, the provider&apos;s terms of use still add to the confidentiality commitments. And nothing may spill over from one project to another, side projects included: that is the job of the isolation described in &lt;a href=&quot;../running-untrusted-repos/&quot;&gt;the previous post&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Day to day, I use Claude, for its subscription. It is a consumer plan: &lt;a href=&quot;https://docs.claude.com/en/docs/claude-code/data-usage&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;training on sessions&lt;/a&gt; must be turned off, otherwise they may be used for it and retained for five years. If the client requires a data processing agreement (DPA), you need a commercial plan or the API. I only move to a European AI platform if they require their data to stay in the EU.&lt;/p&gt;
&lt;h2 id=&quot;a-proxy-between-the-agent-and-the-llm&quot;&gt;A proxy between the agent and the LLM&lt;/h2&gt;
&lt;p&gt;Routing requests through a proxy gives a single point of control. There you track and cap costs per project, anonymise the prompt on one server rather than on every workstation, and filter prompt injections before inference, which reduces the risk without removing it. The provider&apos;s keys stay there, stored per project; the agent only gets a proxy key, scoped and revocable. The proxy assumes API access: on the subscription, Claude Code talks directly to the provider.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/client-data-and-llms/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;h2 id=&quot;a-verifier-that-is-itself-verifiable&quot;&gt;A verifier that is itself verifiable&lt;/h2&gt;
&lt;p&gt;That leaves proving what the proxy did. A log is still the word of whoever keeps it: a record only counts as evidence if the third party behind it is verifiable.&lt;/p&gt;
&lt;p&gt;That is the role of a TEE, an isolated hardware region that even the host cannot read, whose remote attestation proves which code is running. The proxy goes there, as frozen code, logging the prompt and the response, then signing the lot. &lt;a href=&quot;https://docs.phala.network/confidential-ai-inference/confidential-ai-api&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Phala Network&lt;/a&gt; offers this as a service: a signing key is born inside the TEE, and every response is signed.&lt;/p&gt;
&lt;p&gt;The guarantee is strong, not absolute: in 2025, the &lt;a href=&quot;https://thehackernews.com/2025/10/new-teefail-side-channel-attack.html&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;TEE.Fail&lt;/a&gt; attack extracted attestation keys, NVIDIA GPUs included, through physical access to the server. A TEE does not spare you from choosing a host you trust.&lt;/p&gt;
&lt;h2 id=&quot;four-ways-to-run-inference&quot;&gt;Four ways to run inference&lt;/h2&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/client-data-and-llms/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A local model, offline and memoryless.&lt;/strong&gt; To summarise a CV or sort HR documents, &lt;a href=&quot;https://ollama.com&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Ollama&lt;/a&gt; on a machine cut off from the network is enough. Less brilliant, but nothing leaves.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Anonymised data, sent to an attested service.&lt;/strong&gt; The proxy replaces names, addresses and numbers with placeholders, then restores the real values in the response. The limit: TEE services mostly offer open models, not the most advanced ones.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A model hosted, by the freelancer or by the client, in a verifiable VM.&lt;/strong&gt; The model runs in a &lt;a href=&quot;https://docs.phala.network/confidential-ai-inference/host-llm-in-tee&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;confidential VM&lt;/a&gt;, on GPUs that &lt;a href=&quot;https://arxiv.org/html/2409.03992v2&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;support a TEE&lt;/a&gt;, from a reproducible image. Inference becomes verifiable, and the host has no access to the client&apos;s data, short of a hardware attack.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A proprietary model, such as Claude.&lt;/strong&gt; Inference is not private: the provider sees what you send it. It is what I use daily, without a proxy, so without proof: I keep it for code without personal data, with the client&apos;s agreement. When the stakes justify it, every call can go through the API and the TEE proxy, which logs it, signs it and publishes a salted hash of it, timestamped by a third party such as &lt;a href=&quot;https://opentimestamps.org/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;OpenTimestamps&lt;/a&gt; and stored on &lt;a href=&quot;https://docs.ipfs.tech/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;IPFS&lt;/a&gt;, for example. If a response is aberrant, the company can prove what it asked and what it received, without exposing the user&apos;s data.&lt;/p&gt;
&lt;p&gt;Homomorphic encryption, which computes on encrypted data without ever seeing it in the clear, will go further, but for now it remains &lt;a href=&quot;https://arxiv.org/abs/2601.18511&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;far too slow&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id=&quot;what-i-set-up-for-a-client&quot;&gt;What I set up for a client&lt;/h2&gt;
&lt;p&gt;During development: synthetic or anonymised data, secrets out of the agent&apos;s reach, access limited to the task, and training turned off on a consumer subscription. In the contract, one sentence stating which LLM touches the code: &quot;I use Claude Code on my clients&apos; codebases&quot;.&lt;/p&gt;
&lt;p&gt;When the client goes through the API, an open-source proxy such as &lt;a href=&quot;https://docs.litellm.ai/docs/simple_proxy/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;LiteLLM&lt;/a&gt; covers anonymisation with &lt;a href=&quot;https://docs.litellm.ai/docs/proxy/guardrails/pii_masking_v2&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Presidio&lt;/a&gt; in a few lines:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-yaml&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;guardrails&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;:&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;  - &lt;/span&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;guardrail_name&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;presidio-pii&quot;&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;    litellm_params&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;:&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;      guardrail&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;presidio&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;      mode&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;pre_call&quot;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;Each project then gets its own virtual key, with a capped budget; &lt;a href=&quot;https://github.com/diegosouzapw/OmniRoute&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;OmniRoute&lt;/a&gt; offers the equivalent.&lt;/p&gt;
&lt;p&gt;When everything to do with the client lives in their container, an end-of-engagement destruction or return clause takes minutes to honour: archive, hand back, delete. What was sent to an LLM provider, however, follows that provider&apos;s own retention period.&lt;/p&gt;
&lt;p&gt;Next post: &lt;a href=&quot;../minimum-sufficient-intelligence/&quot;&gt;choosing the right model for each role in the development loop&lt;/a&gt;.&lt;/p&gt;</content:encoded><category>security</category><category>privacy</category><category>compliance</category><category>freelance</category><category>llm</category></item><item><title>Running an unknown repo without exposing your machine</title><link>https://alexandre.mahe.io/blog/running-untrusted-repos/</link><guid isPermaLink="true">https://alexandre.mahe.io/blog/running-untrusted-repos/</guid><description>Fake technical tests are a documented attack. Two settings limit the damage; the rest takes an isolated environment, with no secrets, even disposable, whose configuration comes from you and not from the repo.</description><pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;A recruiter offers me a web3 contract. The rate is good, the company looks legitimate, and the next step never changes: clone this repo, run it, fix a bug.&lt;/p&gt;
&lt;p&gt;Since 2023, the North Korean &lt;a href=&quot;https://labs.cloudsecurityalliance.org/research/csa-research-note-dprk-contagious-interview-cross-ecosystem/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Contagious Interview&lt;/a&gt; campaign has followed exactly this script: fake recruiters, credible companies, and a coding exercise that ships with its own backdoor, whose mechanics &lt;a href=&quot;https://www.microsoft.com/en-us/security/blog/2026/03/11/contagious-interview-malware-delivered-through-fake-developer-job-interviews/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Microsoft&lt;/a&gt; lays out in detail. More than 1,700 malicious packages have been tied to it since January 2025, and crypto developers, sitting on keys worth real money, are the primary targets.&lt;/p&gt;
&lt;p&gt;The method is mundane, and that is why it works. An npm dependency runs code on install, a &lt;a href=&quot;https://thehackernews.com/2026/07/north-korean-hackers-publish-108.html&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;VS Code task launches when the folder is opened&lt;/a&gt;, a few encoded lines hide in otherwise clean code. Once it runs, it&apos;s over: trojan, ransomware, or a quiet sweep of &lt;code&gt;.env&lt;/code&gt; files, SSH keys and the crypto wallet. With the competition out there, the temptation to say yes quickly to every technical test is real, and that is precisely what the attack exploits.&lt;/p&gt;
&lt;p&gt;For a freelancer, isolation is not just a personal precaution: the client expects their data not to cross from one project to another and access to stay strictly to what is needed, down to the inference that generates their code. And with agents running commands on my behalf, the question comes up every session.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;No code I didn&apos;t write runs on my machine.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;a-devcontainer-whose-configuration-you-write&quot;&gt;A devcontainer whose configuration you write&lt;/h2&gt;
&lt;p&gt;VS Code&apos;s &lt;a href=&quot;https://code.visualstudio.com/docs/editing/workspaces/workspace-trust&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Restricted Mode&lt;/a&gt; and &lt;code&gt;--ignore-scripts&lt;/code&gt; close the shortest paths. But as soon as the tests run, the code executes anyway: it has to execute somewhere else. In a &lt;a href=&quot;https://containers.dev/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;devcontainer&lt;/a&gt;, the code runs inside and the machine stays outside.&lt;/p&gt;
&lt;p&gt;On one condition: for an unknown repo, the &lt;code&gt;devcontainer.json&lt;/code&gt; is written by the attacker, and its &lt;a href=&quot;https://containers.dev/implementors/json_reference/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;&lt;code&gt;initializeCommand&lt;/code&gt;&lt;/a&gt; runs on the host before the container even exists. The configuration has to come from you.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://alexandre.mahe.io/blog/running-untrusted-repos/&quot;&gt;Diagram: open the post to see it.&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;I run several agents in parallel, each in its own container. Locally, the code stays on my disk, which lets me work offline, but it only executes inside the container and only moves between agents through git. Other agents run continuously in the cloud, with Claude Code or &lt;a href=&quot;https://earendil.com/posts/pi-durable/&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;Pi Durable&lt;/a&gt;. From my workstation, I manage their costs, their models and their permissions.&lt;/p&gt;
&lt;h2 id=&quot;closed-network-no-secrets-nothing-automatic&quot;&gt;Closed network, no secrets, nothing automatic&lt;/h2&gt;
&lt;p&gt;The network is closed by default. w3cj&apos;s &lt;a href=&quot;https://github.com/w3cj/devcontainer-features/blob/main/src/firewall/README.md&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;&lt;code&gt;firewall&lt;/code&gt; feature&lt;/a&gt; blocks everything except a few allowed hosts and, in verbose mode, reports every refused connection. You quickly find out that a &quot;simple technical test&quot; is very keen to talk to an unknown server. A backdoor that does run can only reach those hosts: the risk drops, but does not disappear.&lt;/p&gt;
&lt;p&gt;Secrets are compartmentalised. For trusted code, the container gets a development API key and a GitHub token scoped to that one repository. Otherwise, nothing.&lt;/p&gt;
&lt;p&gt;Nothing starts on its own: not the repo&apos;s tasks, not npm scripts outside an &lt;a href=&quot;https://docs.npmjs.com/cli/v11/commands/npm-install-scripts&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;explicit allowlist&lt;/a&gt;. The container has access to neither the home directory nor the Docker socket, and the SSH agent, which devcontainers share by default, is cut off unless explicitly requested.&lt;/p&gt;
&lt;p&gt;The agent itself reads the repo&apos;s files, which may contain instructions aimed at it. With no secrets and a closed network, a prompt injection has little to steal, and nowhere to send it.&lt;/p&gt;
&lt;h2 id=&quot;a-home-made-tool-not-a-security-product&quot;&gt;A home-made tool, not a security product&lt;/h2&gt;
&lt;p&gt;With agents, new projects multiply. So I vibe-coded &lt;code&gt;bs&lt;/code&gt;, a small tool that clones a repo and opens it in a devcontainer, with the language tooling and the security already in place, without installing Rust or Node on my machine. Good specs cover at least 80% of my bootstrapping needs; the rest gets adjusted from inside the devcontainer itself.&lt;/p&gt;
&lt;p&gt;&lt;code&gt;bs&lt;/code&gt; secures nothing on its own: Docker, proven base images and a readable firewall do that. The tool only makes the safe path short enough that I am not tempted to skip it, and trusting a repo remains a human decision. If the repo already contains a devcontainer, &lt;code&gt;bs&lt;/code&gt; rewrites it before opening: git will show a modified file, but the generated configuration is what runs.&lt;/p&gt;
&lt;p&gt;It is an assembly of configurations, easier to specify than to write, that generates files I review. I wrote it in Rust without knowing the language, in a devcontainer generated by &lt;code&gt;bs&lt;/code&gt; itself.&lt;/p&gt;
&lt;p&gt;For a technical test, assuming such tests are even still necessary, my version fits in one command:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-bash&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;# --typescript : TypeScript tooling and the common security layer&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;# --remote     : temporary VM, destroyed afterwards. Nothing on the local disk, and&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;#                a container escape lands on an empty machine&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;# --omniroute  : key generated by OmniRoute, to try LLMs without authenticating&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;#                or exposing your own key to a prospect&apos;s repo&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#66707B, #BDC4CC);--shiki-light:#66707B;--shiki-dark:#BDC4CC&quot;&gt;# --untrusted  : no environment variables or secrets from the host machine&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;bs&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --typescript&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --remote&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --omniroute&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --untrusted&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; https://github.com/malevolent/test.git&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;A devcontainer changes nothing for colleagues who don&apos;t use one: the whole team can improve the shared configurations, and the same container runs identically in the cloud.&lt;/p&gt;
&lt;h2 id=&quot;a-starting-configuration&quot;&gt;A starting configuration&lt;/h2&gt;
&lt;p&gt;Without any tooling: don&apos;t trust a folder in VS Code before reading the code, and install dependencies without scripts.&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-bash&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;npm&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; install&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --ignore-scripts&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;In the &lt;code&gt;devcontainer.json&lt;/code&gt;, written by you, the firewall takes a few lines; the &lt;a href=&quot;https://github.com/w3cj/devcontainer-features/blob/main/src/firewall/README.md&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;feature&apos;s documentation&lt;/a&gt; lists the options:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-jsonc&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt;&quot;features&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: {&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;  &quot;ghcr.io/w3cj/devcontainer-features/firewall:latest&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: {&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;    &quot;githubDomains&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt;true&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;&quot;npmRegistry&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt;true&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;&quot;claudeCode&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt;true&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;, &lt;/span&gt;&lt;span style=&quot;color:light-dark(#024C1A, #72F088);--shiki-light:#024C1A;--shiki-dark:#72F088&quot;&gt;&quot;verbose&quot;&lt;/span&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;: &lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt;true&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;  }&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#0E1116, #F0F3F6);--shiki-light:#0E1116;--shiki-dark:#F0F3F6&quot;&gt;}&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;To open an unknown repo without using its configuration, the &lt;a href=&quot;https://github.com/devcontainers/cli&quot; target=&quot;_blank&quot; rel=&quot;noopener noreferrer&quot;&gt;official CLI&lt;/a&gt; accepts a file you wrote yourself, kept outside the repo:&lt;/p&gt;
&lt;pre class=&quot;shiki shiki-themes github-light-high-contrast github-dark-high-contrast&quot; style=&quot;background-color:light-dark(#ffffff, #0a0c10);--shiki-light-bg:#ffffff;--shiki-dark-bg:#0a0c10;color:light-dark(#0e1116, #f0f3f6);--shiki-light:#0e1116;--shiki-dark:#f0f3f6&quot; tabindex=&quot;0&quot;&gt;&lt;code class=&quot;language-bash&quot;&gt;&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;npx&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; @devcontainers/cli&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; up&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --workspace-folder&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; .&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --config&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; ~/devcontainers/untrusted.json&lt;/span&gt;&lt;/span&gt;
&lt;span class=&quot;line&quot;&gt;&lt;span style=&quot;color:light-dark(#702C00, #FFB757);--shiki-light:#702C00;--shiki-dark:#FFB757&quot;&gt;npx&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; @devcontainers/cli&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; exec&lt;/span&gt;&lt;span style=&quot;color:light-dark(#023B95, #91CBFF);--shiki-light:#023B95;--shiki-dark:#91CBFF&quot;&gt; --workspace-folder&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; .&lt;/span&gt;&lt;span style=&quot;color:light-dark(#032563, #ADDCFF);--shiki-light:#032563;--shiki-dark:#ADDCFF&quot;&gt; bash&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;
&lt;p&gt;On sensitive client code, no LLM without the client&apos;s prior consent, let alone a free one. For a tool like &lt;code&gt;bs&lt;/code&gt;, this post can serve as a starting point for a discussion with your own agent.&lt;/p&gt;
&lt;p&gt;Next post: &lt;a href=&quot;../client-data-and-llms/&quot;&gt;what happens to a client&apos;s data when an LLM reads their code, and how a freelancer stays compliant&lt;/a&gt;.&lt;/p&gt;</content:encoded><category>agents</category><category>security</category><category>devcontainers</category><category>supply-chain</category></item></channel></rss>