<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[DEIXEI]]></title><description><![CDATA[DEIXEI]]></description><link>https://deixei.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:52:05 GMT</lastBuildDate><atom:link href="https://deixei.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Your coding agent is privileged automation: a workstation baseline for Claude Code]]></title><description><![CDATA[Treat the coding agent as privileged automation: managed policy, locked merges, command hooks, sandbox limits and telemetry for Claude Code.]]></description><link>https://deixei.hashnode.dev/your-coding-agent-is-privileged-automation-a-workstation-baseline-for-claude-code</link><guid isPermaLink="true">https://deixei.hashnode.dev/your-coding-agent-is-privileged-automation-a-workstation-baseline-for-claude-code</guid><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><category><![CDATA[Devops]]></category><category><![CDATA[Developer Tools]]></category><dc:creator><![CDATA[deixei]]></dc:creator><pubDate>Tue, 17 Nov 2026 20:08:52 GMT</pubDate><content:encoded><![CDATA[<p>A coding agent reads the repository, runs shell commands with the developer's privileges, edits code and calls external tools. On the workstation, the least production-like place it runs, it is often left ungoverned. But distance from production does not remove the path to production.</p>
<p>This is a condensed version of a baseline written for a regulated firm running Claude Code on managed Windows with WSL2, Linux workstations and hosted CI. It is vendor-specific on purpose.</p>
<h2>The settings file is not the boundary</h2>
<p>Managed settings are a client-side control, not a security boundary. A local administrator can edit the managed source, and managed settings bind only Claude Code. Every high-risk context needs an OS or network control behind the policy file. Also: only managed settings and hooks enforce. CLAUDE.md, rules, skills and subagents steer the model, so a control self-assessment should never claim a control on the strength of guidance.</p>
<h2>Five layers</h2>
<ul>
<li>Identity and tenancy: SSO, a pinned login method and organisation, workload credentials for pipelines, no raw API keys on endpoints.</li>
<li>- Managed policy: one version-controlled document, delivered by device management and locked.</li>
<li>- Runtime isolation: the OS sandbox, plus dev containers for high-risk repositories.</li>
<li>- Network and data: egress allow-lists, tenant restrictions, secret placement, short local retention.</li>
<li>- Monitoring and evidence: OpenTelemetry export, hook events to the SIEM, and host process telemetry as the independent record.</li>
</ul>
<h2>Lock the merge</h2>
<p>Permission rules merge across scopes. A deny anywhere wins, but allows from user and project files combine with managed allows unless you lock them. The allowManaged*Only flags close that separately for permission rules, hooks and MCP servers: locking the rules does not lock the MCP allow-list. Check it by sampling /permissions on real workstations and reading the source column.</p>
<h2>Denies are weaker than they look</h2>
<p>Evaluation runs deny, then ask, then allow. But rule matching is fragile: a deny on curl does not catch /usr/bin/curl or an sh -c wrapper. Use whole-command denies, never allow-list environment runners such as npx or docker exec, and put real parsing in a PreToolUse command hook. Denies don't cover arbitrary subprocesses or network access; the sandbox and the egress allow-list carry that load, and the agent should not be able to ask its way out of the sandbox.</p>
<h2>Command hooks enforce, HTTP hooks carry evidence</h2>
<p>HTTP hooks fail open by default, so enforcement belongs in command hooks set to block on failure. Hooks and MCP servers run outside the sandbox, so review hook scripts as privileged code. Sending prompt content to the SIEM is a works-council and DPIA decision, not a default.</p>
<h2>Know where the sandbox stops</h2>
<p>It covers shell commands and their children only; file tools, MCP and hooks run outside it. Native Windows and WSL1 have no sandbox, so make it fail rather than fall back silently. Default read access covers much of the machine, and a Docker socket is effectively host access.</p>
<h2>Before you rely on it</h2>
<p>This baseline is a design proposal to pilot, not a certified configuration. Setting names, minimum versions and defaults change quickly, so confirm each one on the build you pin.</p>
<p>The full baseline, with the threat model, identity routes, the seven-hook bundle and the regulated overlay: <a href="https://www.deixei.com/ai-delivery-factory/build/coding-agent-workstation/">https://www.deixei.com/ai-delivery-factory/build/coding-agent-workstation/</a></p>
<p>This cross-post was condensed from the original with AI assistance.</p>
]]></content:encoded></item><item><title><![CDATA[The Agent-Speed SDLC: letting coding agents move fast without losing the audit trail]]></title><description><![CDATA[Make every SDLC step a commit, govern agent context, distrust same-source tests, and apply eight controls so coding agents stay auditable.]]></description><link>https://deixei.hashnode.dev/the-agent-speed-sdlc-letting-coding-agents-move-fast-without-losing-the-audit-trail</link><guid isPermaLink="true">https://deixei.hashnode.dev/the-agent-speed-sdlc-letting-coding-agents-move-fast-without-losing-the-audit-trail</guid><category><![CDATA[AI]]></category><category><![CDATA[Devops]]></category><category><![CDATA[software architecture]]></category><category><![CDATA[code review]]></category><dc:creator><![CDATA[deixei]]></dc:creator><pubDate>Sun, 15 Nov 2026 10:53:48 GMT</pubDate><content:encoded><![CDATA[<p>Coding agents don't create new kinds of delivery failure. They push four old ones past what our controls were sized for: volume (more change than reviewers can absorb), plausibility (output looks finished before it is proven correct), drift (generated work converges on model defaults instead of your architecture) and attribution (authorship and accountability blur).</p>
<p>The answer isn't to slow agents down. It is to keep intent, review, identity and evidence connected to every change.</p>
<h2>Make every step a commit</h2>
<p>The chain is intent → spec → plan → diff and tests → PR and review findings → release and evidence → monitoring → new intent. Every arrow is a commit, so the chain itself becomes the audit trail.</p>
<ul>
<li>intent.md: the problem, outcome, users and constraints in the originator's words. The product owner accepts it; risk tier and data class are assigned.</li>
<li>- spec.md: testable requirements and design.</li>
<li>- plan.md: files, order, tests, assumptions, risks and what is out of scope. The accountable human approves it before any code is written.</li>
<li>- Tests, PR, review and release records, each naming the exact upstream version it was built from.</li>
</ul>
<p>Two boundary rules matter most: the agent that authors a change cannot approve it, and it may act only up to the production gate.</p>
<h2>Govern the context the agent reads</h2>
<p>Three artefacts answer the questions an agent would otherwise guess at: a layered standing instruction file (how we build here), the specification (what to build) and the recipe (the one approved assembly, release and evidence pattern). Architecture decisions are walls, enforced by hooks, permissions or CI. Guidance is direction, given as advisory skills. Put a rule in the wrong layer and it either blocks good work or gets ignored.</p>
<h2>Don't trust tests that share the change's source</h2>
<p>Tests an agent writes from the same spec measure faithfulness to the spec, not correctness of the product. Rank tests by trust: written or approved by someone who knows the product today; owned by a downstream consumer; drawn from an independent source such as incident cases; and, lowest, generated with the change. Protect the oracle: test changes are reviewed separately, and any change to a single decisive value (a tolerance, threshold or limit) goes to a human, however small. In one composite example, a four-line tolerance change backed by 14 green same-source tests was auto-approved and widened a cash-break band about a hundredfold for weeks.</p>
<h2>Eight controls, two per failure mode</h2>
<ul>
<li>Volume: automated review gates before a human looks; an AI-era PR template.</li>
<li>- Plausibility: two human reviewers on critical-function paths; mandatory dynamic security testing on auth, payment and customer-data changes.</li>
<li>- Drift: architecture rules as code at PR time; a pinned library allow-list per repository.</li>
<li>- Attribution: agents commit under their own registered identity with a named human of record; deployment uses federated identity with no stored secrets.</li>
</ul>
<h2>The test of readiness</h2>
<p>A pilot is ready to expand when an independent reviewer can trace intent → plan → diff → tests → evaluation → approvals → release, and name the agent and the human at every boundary.</p>
<p>The full page, with roles, failure modes, acceptance checks and recovery, is part of the AI Delivery Factory reference: <a href="https://www.deixei.com/ai-delivery-factory/delivery/agent-speed-sdlc/">https://www.deixei.com/ai-delivery-factory/delivery/agent-speed-sdlc/</a></p>
<p>This cross-post was condensed from the original with AI assistance.</p>
]]></content:encoded></item></channel></rss>