Practical blogs

Sharing frontier AI technologies to help you become better at using AI.

AI workflow Build in public AI docs macOS utility
Claude Code data-residency cost Claude Code 2.1.239: reconcile the US-only inference premium before trusting the budget cap Verify policy, effective inference geography, local cost estimates, budget-cap behavior, and provider usage with a cache-aware four-surface reconciliation. By Ahab DeepSeek Harness extension contracts DeepSeek Harness extensions: choose Tool, MCP, or LLM Adapter without breaking replay Choose the right extension seam, build a typed Tool, bridge an MCP server, implement an LLM Adapter, and verify schema, streaming, replay, errors, and release compatibility with twelve canaries. By Ahab DeepSeek Harness delegation reliability DeepSeek Harness subagents, jobs, and workflows: delegate without losing ownership Compare six providers, separate Subagents from Jobs and Workflows, and verify authority, concurrency, cancellation, resume, and terminal states with six canaries. By Ahab DeepSeek Harness session reliability DeepSeek Harness sessions: resume, compact, and recover without losing replay evidence Separate raw events, durable storage, model-visible surface, compaction checkpoints, and spill artifacts; then verify crash recovery, fork, replay, and backup with twelve canaries. By Ahab DeepSeek Harness tool security DeepSeek Harness permissions and sandbox: trace every tool call before granting access Trace pre-execute policy, one-shot approval, monotonic guards, sandbox confinement, filesystem freshness, and immutable results with nine cross-platform fail-closed canaries. By Ahab DeepSeek Harness architecture DeepSeek Harness “Everything Is a Plugin”: choose the right Cordis extension seam Map plugins, services, events, effects, profiles, and bundles; choose an extension seam with a decision tree and seven lifecycle/HMR canaries. By Ahab DeepSeek Harness safety DeepSeek Harness Developer Preview: install it with a safe rollback boundary Pin rc.7, isolate DSH_HOME and the workspace, verify permissions and credential boundaries, and keep a reversible exit before trying real code. By Ahab Claude prompt asset migration Claude Workbench is now Playground: export saved prompts before September 1 Export legacy Workbench JSON before September 1, package prompts with parameters and evals, remove three retired API callers, and replay a golden set through the Messages API. By Ahab AI agent trust-boundary upgrade Claude Code 2.1.234 security upgrade: test paths, marketplace hosts, and redaction before rollout Verify NT-path rejection, effective marketplace hosts, MCP and approval redaction, and background-subagent denies with five payload-free canaries before restoring autonomy. By Ahab AI-native Git hosting rollout Cursor Origin vs GitHub: mirror repositories without creating two sources of truth Keep GitHub authoritative, prove code and PR round trips, preserve CI and secrets, test access revocation, and detach only as an explicit migration. By Ahab Gemini image API migration Imagen 4 API shutdown: migrate to Gemini 3.1 Flash Image without a blind model-ID swap Replace three Imagen 4 IDs scheduled for shutdown and the old predict contract, replay a paired image fixture, compare accepted-image cost, canary one lane, and drain queued jobs. By Ahab AI activity memory privacy ChatGPT Computer History: use an allowlist before Codex learns from your activity Map capture, server summarization, local plaintext memories, and later reuse; start with include-only sources, verify pause and deletion, test prompt injection, and keep rollback. By Ahab AI review evidence and growth control Qwen Code 0.21.12: require review evidence and stop autofix loops before they bloat a PR Require an executed witness for Critical findings, split source and test growth budgets, verify Critical-only behavior, and hand non-converging loops to a maintainer. By Ahab Claude Code Linux memory control Claude Code 2.1.233 on Linux: cap Bash tool memory before a build stalls the host Size CLAUDE_CODE_TOOL_MEMORY_LIMIT from a healthy build, verify cgroup membership and memory.max, run positive and disposable OOM canaries, and keep rollback. By Ahab ChatGPT Linux desktop preview ChatGPT for Linux public preview: verify the browser boundary before rollout Use the official Linux download, verify supported distributions, test the built-in browser and Chrome separately, prove native-app control stays unavailable, and keep rollback. By Ahab Grok long-running agent migration Grok 4.6 vs 4.5: migrate long-running agents by accepted-task economics Run six paired long tasks, measure context and cache growth, calculate accepted-task economics, verify partner availability separately, and keep 4.5 rollback. By Ahab Gemini agent model migration Gemini 3.7 Flash vs 3.6: migrate by accepted-task cost, not launch claims Run six paired agent tasks, normalize thinking and service tiers, calculate accepted-task cost, verify Copilot separately, and keep 3.6 rollback ready. By Ahab DeepSeek agent rollout DeepSeek V4 Pro GA: verify Responses API, thinking effort, and peak pricing before rollout Pin the 0813 snapshot, test Codex and tool state, map low/high/max effort, convert UTC price windows, and compare Pro with Flash by accepted-task cost. By Ahab Portable AI agent plugins Qwen Code 0.21.11: install Agent Plugins v1 without assuming every capability works Separate native package loading from runtime compatibility, test Skills and MCP beside ignored components, and gate path, credential, state, and rollback boundaries. By Ahab AI gateway stream reliability Claude Code 2.1.229: keep Vertex and Bedrock gateway streams alive during long thinking Map the gateway path, correlate one session, prove SSE liveness across the old idle boundary, classify failures, and roll out without hiding total-timeout or auth problems. By Ahab AI browser memory debugging Chrome DevTools MCP 1.7: debug memory leaks without reading raw heap snapshots Capture three heap snapshots, isolate retained growth by context and object, follow the retaining path, replay the fix, and close every snapshot. By Ahab Local AI coding privacy GitHub Copilot for JetBrains: use Ollama without confusing local inference with local memory Prove Ollama serves the selected model locally, audit Copilot Memory separately, test fail-closed behavior, and keep both controls reversible. By Ahab AI action sandbox recovery Claude Code 2.1.227: restore Bash in Claude Code Action without dropping the sandbox Verify the fixed CLI actually restores Bash under allowed_non_write_users, keep subprocess isolation on, and gate rollout with six canaries. By Ahab AI agent LAN access security Qwen Code 0.21.9 Local Control: pair a phone without exposing your agent Use a trusted-LAN threat model, verify the QR token and active-session handoff, run six canaries, and prove clean revocation before phone access. By Ahab AI API latency economics GPT-5.6 Fast mode long context: price the 272K threshold before rollout Compare Standard and Fast prices above 272K, verify the returned service tier, ramp traffic safely, and gate adoption on cost per accepted result. By Ahab AI approval automation security Codex CLI 0.147 --approve-for-me: verify Auto-review before unattended runs Keep workspace-write boundaries, test denials and fake-secret handling, verify resume and fork parity, and retain one-command rollback with six canaries. By Ahab Managed agent cost control Claude Managed Agents session budgets: prove the hard stop before unattended runs Size a list-cost cap, measure multiagent overshoot, verify budget_reached events, and gate every resume decision with six canaries. By Ahab AI workspace trust security Qwen Code 0.21.8: verify nested workspace trust before loading project env Test most-specific trust precedence, safe-mode parity, and both parent/child env paths before restoring unattended qwen serve workflows. By Ahab AI agent infrastructure security Claude Code 2.1.224 self-hosted environments: verify the trust boundary before rollout Map what stays local, isolate each runner and credential, deny egress, verify session identity, and test failure recovery before production. By Ahab AI model rollout verification Kimi K3 in GitHub Copilot: verify rollout status before enabling it Reconcile the paused rollout, live docs, plan, policy, picker, pricing, and five bounded canaries before changing a team workflow. By Ahab AI agent permission security Claude Code 2.1.223: verify the permission-bypass fixes before restoring autonomy Upgrade every launcher, keep bypass mode disabled, and run harmless Unicode, command-parser, agent-policy, and workflow-sandbox canaries. By Ahab AI spend observability Cloudflare AI Gateway User Insights: attribute spend before chasing anomalies Establish identity coverage, investigate session-cost anomalies, add per-user budgets, and test every 429 and billing boundary. By Ahab AI platform infrastructure OpenAI Terraform provider: import API resources without destructive drift Inventory ownership, import existing resources, require a no-op plan, separate credentials, and test each removal behavior. By Ahab AI connector security Cursor Google Workspace plugins: connect Gmail, Drive, and Calendar safely Inventory the live MCP tools, isolate untrusted content, stage writes, and keep final sends and edits behind human review. By Ahab MCP replay safety Qwen Code 0.21.5: stop duplicate MCP writes after disconnects Force a post-commit disconnect, verify trust and annotations, and reconcile unknown outcomes before any MCP write is retried. By Ahab AI agent credential security Claude Code 2.1.221: mask credential files before sandboxed commands read them Use structured extraction, narrow injection hosts, disposable sentinels, and eight fail-closed gates before authenticated sandbox rollout. By Ahab Multi-agent prompt isolation Qwen Code 0.21.4: test forked subagent prompt isolation Use three disposable sentinels, bounded history inspection, and eight rollout gates before trusting same-turn parallel forks. By Ahab Codex model migration Codex GPT-5.4 retirement: migrate signed-in workflows to Terra or Luna Inventory workspace defaults, saved settings, managed policy, custom agents, and scheduled tasks without rewriting unaffected API routes. By Ahab AI coding agent security Devin Local 3.6.27: test the symlink write boundary before auto-approving edits Verify the four protected native tools with a disposable sentinel lab, then keep shell writes and undocumented paths behind separate controls. By Ahab AI code review verification Qwen Code 0.21.3: verify PR test plans before trusting /review Validate Test Plan claims, replay failures against the merge base, and prove the test harness before accepting an AI review verdict. By Ahab Embodied AI migration Gemini Robotics ER 2 vs Streaming: migrate before ER 1.6 shuts down Choose the standard or Streaming endpoint, calculate the 2x token-price change, and pass eight physical-action safety tests before August 31. By Ahab AI video API rollout Grok Imagine Video 1.5: verify reference images and 1080p before rollout Resolve the release-to-doc mismatch, calculate per-clip cost, and pass an eight-case API canary before shipping. By Ahab AI coding plan decision Cursor Start vs Pro in India: test the ₹649 plan before treating it as cheap Pro Compare the confirmed plan boundaries, then use a 12-task, seven-day accepted-work test instead of guessing from an undisclosed quota. By Ahab Open-model deployment Kimi K3 local deployment: check hardware before downloading 1.56 TB Use the weight-file total, 64+ accelerator guidance, license, engine, context, and seven deployment gates to decide whether self-hosting is realistic. By Ahab AI review finalization Qwen Code 0.21.1: stop polling CI inside AI review agents Split AI code judgment from CI waiting with a commit-bound handoff, deterministic finalizer, closed-green gate, and eight acceptance tests. By Ahab Mobile AI code review Cursor for iPad: review cloud-agent PRs without merging blind Bind evidence to the current commit, review by risk, keep feedback bounded, and pass eight gates before merging a Cursor agent PR from iPad. By Ahab AI coding plugin rollout Codex CLI 0.146 Agent Plugins: migrate manifests and publish safely Choose the smallest plugin shape, separate portable and Codex-only manifests, test a local marketplace, gate workspace publishing, and pass eight rollout checks. By Ahab AI speech-to-text workflow GPT Transcribe vs GPT Live Transcribe: choose the right speech-to-text workflow Route completed files and live audio correctly, test context hints, reconcile streaming events, and keep specialized fallbacks with eight production gates. By Ahab MCP implementation migration Cloudflare Agents 0.20: migrate to MCP SDK v2 without breaking legacy clients Choose a stateless or temporary legacy lane, pin the exact MCP SDK, verify edge and OAuth boundaries, and drain old sessions with eight canaries. By Ahab AI agent governance GitHub Copilot managed settings: govern the app and cloud agent safely Separate app access from runtime controls, verify key coverage across clients, and roll out plugin and approval guardrails with six canaries. By Ahab AI agent verification Qwen Code 0.21 Goal evidence: verify completion before trusting it Classify evidence provenance, bind it to one Goal revision, cap the packet, verify independently, and keep human authority over terminal decisions. By Ahab AI coding workflow GitHub Copilot for Linear: a safe issue-to-draft-PR checklist Turn one bounded Linear issue into a reviewable draft PR with explicit branch, test, permission, cost, and rollback gates. By Ahab Claude Code security Claude Code 2.1.219: lock sandbox network egress with a strict allowlist Inventory required hosts, prove the deny path, keep MCP separate, and run a reversible canary before trusting a Claude Code sandbox network policy. By Ahab AI agent dependency migration Cloudflare Agents SDK + AI SDK v7: a safe package-pair migration checklist Match peer majors, test one real agent path, run a canary, and keep a lockfile rollback before moving a Cloudflare Agents app to AI SDK v7. By Ahab AI coding model rollout Claude Opus 5 in GitHub Copilot: a safe model rollout checklist Verify access, select bounded repository tasks, compare evidence, cap the trial, and keep a rollback path before making Claude Opus 5 part of a Copilot workflow. By Ahab AI issue automation GitHub Issue Agent Automations: confidence and approval rollout checklist Start GitHub Issue AI triage with suggestions, a narrow action matrix, confidence thresholds, rationale review, and an eight-issue canary. By Ahab MCP protocol migration MCP 2026-07-28: migrate to stateless requests with conformance tests Replace protocol sessions with self-contained requests, preserve explicit application state, run official conformance tests, and stage a seven-case reversible rollout. By Ahab AI agent orchestration Grok Build Workflows: fan out parallel agents with verification Choose the right task shape, design independent verification, cap fan-out, lock permissions, and run a six-case canary before saving a reusable workflow. By Ahab AI cost control OpenAI API hard spend limits: cap cost without a surprise outage Set organization and project caps with early alerts, quota-aware 429 handling, safe fallbacks, and a six-case failure drill. By Ahab AI model routing Cursor Router: choose Cost, Balance, or Intelligence with an eval gate Choose an Auto mode with a 30-task eval, accepted-change economics, visible model attribution, six canary tests, and fixed-model rollback. By Ahab Gemini API migration Gemini 3.6 Flash vs 3.5 Flash-Lite: migrate and route by task Choose each model by workload, remove deprecated API fields, calculate accepted-task cost, and stage a six-case reversible rollout. By Ahab Claude Code operations Claude Code 2.1.217: cap subagent concurrency, depth, and spend Control subagent breadth, nesting, and print-mode budgets with conservative workload profiles, five safe rollout tests, evidence, and rollback gates. By Ahab AI model rollout Kimi K3 API: test context, cost, agent loops, and rollout Adopt Kimi K3 with a model-routing table, cost worksheet, correct history and tool-loop contracts, multimodal limits, and a seven-case rollout gate. By Ahab Computer-use agents Qwen CUA Driver 0.7.3: test coordinates and MCP payload filtering Validate opt-in coordinate mapping and reversible MCP text filtering with a five-case rollout gate covering pixel, zoom, binary, signing, and rollback boundaries. By Ahab Claude Code security Claude Code 2.1.216: test sandbox and worktree boundaries Upgrade safely with a five-case regression matrix for filesystem exceptions, network egress, worktree redirects, symlink writes, and resumed-agent restrictions. By Ahab AI model rollout Grok 4.5 for coding agents: evaluate cost, tools, and rollout Test Grok 4.5 on real repository tasks with explicit reasoning, a seven-case eval, the 200k context price boundary, and staged promotion gates. By Ahab Claude Code workflow Claude Code verification: run /verify and /code-review explicitly Claude Code 2.1.215 stopped invoking two review skills automatically. Rebuild a reliable finish gate with explicit commands, deterministic CI, and recorded evidence. By Ahab Agent memory migration Claude agent memory: migrate before July 22 Update memory-store headers, path prefixes, depth, ordering, and saved cursors before the list API behavior changes. By Ahab AI code review GitHub Copilot code review: custom instructions, firewall, and runners Configure head-branch instructions, a dedicated review environment, firewall allowlists, and runner boundaries without treating AI comments as merge approval. By Ahab Supply chain security Dependabot cooldown: configure the new 3-day default Choose a Dependabot cooldown policy, configure narrow exceptions, avoid invalid zero-day settings, and keep a dependency PR merge gate. By Ahab AI security workflow GitHub Copilot security review vs code scanning autofix Choose when to run Copilot /security-review, PR code review, AI security detections, CodeQL, and agentic autofix without losing human merge evidence. By Ahab AI crawler policy Cloudflare AI bot policies for indie sites Separate Search, Agent, and Training crawlers before changing Cloudflare AI bot settings, with a path-by-path policy matrix for indie sites. By Ahab Agent-ready web WebMCP agent-ready website security checklist Choose a safe first WebMCP tool, keep website actions bounded, and test the browser-agent contract with DevTools and Lighthouse before shipping. By Ahab AI application security CodeQL system prompt injection: a JavaScript fix checklist Find untrusted data flowing into system prompts and tool descriptions, choose the right fix, and close the alert with static and runtime evidence. By Ahab AI coding workflow GitHub Copilot in VS Code: control browser tools, parallel agents, and cost A practical control checklist for using Copilot repository overviews, browser tools, parallel sessions, model choices, and cost visibility without losing review evidence. By Ahab AI model migration GitHub Models retirement: migrate before July 30, 2026 A practical migration checklist for finding GitHub Models dependencies, choosing a replacement, adding a provider adapter, testing brownouts, and cutting over safely. By Ahab AI model workflow GPT-5.6 Sol vs Terra vs Luna: which model should indie developers use A practical comparison of GPT-5.6 Sol, Terra, and Luna across price, availability, reasoning, Codex and API use, and routing decisions for indie products. By Ahab AI search SEO llms.txt vs robots.txt for AI search: an indie site decision guide A practical decision guide for using llms.txt as an agent-readiness map while keeping robots.txt as the policy layer for AI search, training crawlers, and user-triggered fetches. By Ahab AI model workflow GPT-5.6 model preview: an indie developer migration checklist What OpenAI has officially confirmed about GPT-5.6 Sol, Terra, and Luna, plus a practical migration checklist for routing, fallbacks, evals, prompt caching, and agent permissions. By Ahab Developer tools security How to sandbox an AI coding agent before opening an untrusted repo A practical zero-trust checklist for scanning untrusted repos, isolating secrets, limiting network access, reviewing scripts and MCP config, and verifying agent changes before setup. By Ahab AI coding workflow Chrome DevTools MCP for AI coding agents A practical debugging workflow for using Chrome DevTools MCP with AI coding agents: browser evidence first, bounded sessions, Playwright handoff, and merge-ready verification. By Ahab Developer tools security Vet npm packages before installing them in AI-built apps A practical npm dependency intake workflow for checking AI-suggested packages, install scripts, provenance, OSV data, and lockfile risk before running npm install. By Ahab AI coding workflow AGENTS.md vs CLAUDE.md vs Copilot instructions A practical setup guide for deciding where AI coding-agent rules belong across Codex, Claude Code, Cursor, GitHub Copilot, scripts, hooks, and checks. By Ahab Build in public Why my Product Hunt launch failed: 10 votes, 1 download A candid indie-developer launch recap on low Product Hunt votes, weak conversion, no vote-buying, and what to fix after a failed launch. By Ahab Social preview SEO Slack OG image not showing: debugging checklist A practical workflow for fixing missing Slack preview images by checking raw HTML, image delivery, redirects, headers, and cache behavior. By Ahab AI search SEO AI search SEO checklist for indie sites A practical checklist for adapting indie product pages to Google AI Overviews and AI Mode without chasing GEO hacks. By Ahab macOS utility Locating the input position on macOS A real debugging note on Accessibility caret bounds, Electron edge cases, coordinate systems, and working with an AI agent without letting it guess. By Ahab macOS utility Building Paste Switch from idea to macOS app A build-in-public story about product scope, clipboard watching, shortcuts, replacement logic, and macOS permissions. By Ahab AI workflow Build a Record & Replay Mac app A technical product plan for browser + desktop recording, semantic replay, BYOK models, privacy, and distribution. By Ahab AI workflow Codex Record & Replay principles Understand how a demonstrated workflow becomes a reusable skill and how to build a browser + computer-use MVP. By Ahab Build in public Building MD+HTML Reader in public A first indie product launch, from local reader to payment, analytics, updates, and promotion. By Ahab AI docs Read AI-generated Markdown and HTML on Mac A practical workflow for reviewing plans, reports, Mermaid diagrams, and HTML artifacts. By Ahab