Tutorials, tech notes and build logs — how Wakii works under the hood and how to get the most out of an agentic IDE.
tech notes2026-10-25
unlazy: completion discipline for agents — runnable gates before workunlazy is not the image lazy-loading library the brief guessed — the GitHub API confirms it is an anti-laziness skill for AI agents: write acceptance gates before the work, re-verify everything that passed, and refuse your own evidence when it does not reproduce.7 min readread→tech notes2026-10-25Cumora: a rising AI repo, read through its own anti-patterns docCumora — team chat where AI agents are first-class teammates — is three weeks old with 3.5k stars, but the best read is docs/COORDINATION.md: a refreshingly honest anti-patterns log. Inside: HELD envelopes, hold-token overrides, and five rules that keep the prompt short.8 min readread→tech notes2026-10-24OpenBot: CopilotKit's sample agent — every action through one gateCopilotKit shipped OpenBot — a self-hosted AI coworker template: each bot gets its own computer, and every action passes a single gateway that denies first and records always. A close read of the gateway, the take-the-wheel handoff, and governed generative UI.8 min readread→tech notes2026-10-24ASu-skills: nine job-hunting skills and the skills-as-content waveASu-skills packages a job-hunting workflow into nine AI skills and distributes them to five harnesses from a single source, with a registry reconciled by CI. Inside: the skills-as-content phenomenon and the anti-drift mechanism Wakii itself needs.7 min readread→tech notes2026-10-23DSH Desktop: a plugin-first desktop for an agent harnessDSH Desktop wraps DeepSeek Harness in a thin Electron host where the desktop shell itself is an unprivileged plugin — pinned upstream, stable/beta channels split by a fail-closed protocol, and pull requests drafted by an agent.7 min readread→tech notes2026-10-23Tabby: self-hosted code assistant — inference on your own machineRun a code assistant on your own infrastructure with Tabby: one binary, two commands, models running locally — and how to read a license when the GitHub API says NOASSERTION.7 min readread→tech notes2026-10-22Continue: an AI assistant inside your existing IDE — the config-hub lessonHow Continue embedded an AI assistant into VS Code and JetBrains through a single config file — plus the operational lesson from a 35-thousand-star project going read-only.7 min readread→tech notes2026-10-22Lapce: a Rust editor built for performanceLapce's architecture from real code: a rope buffer with revisions, a UI split into its own process over RPC, plugins running in a WASI sandbox — how an editor stays fast as features grow.6 min readread→tech notes2026-10-21Helix: the modal editor that needs no configurationSelection before action, LSP and tree-sitter built into the core, a single config.toml — three Helix design decisions that keep an editor lean in the age of configuration sprawl.6 min readread→tech notes2026-10-21Yazi: a terminal file manager written in RustInside Yazi's architecture — a 31-crate workspace, an async scheduler with priorities, image-protocol negotiation with a fallback chain, and Lua plugins with a package manager.6 min readread→tech notes2026-10-20bat: cat with syntax highlighting and gitHow bat works — syntect highlighting, in-process git markers via gitoxide, plain-text fallback when piped — and the lesson of output that adapts to its consumer.6 min readread→tech notes2026-10-20Starship: the fast, cross-shell promptHow Starship works — one Rust binary across ten shells, 100+ signal modules, TOML config that degrades safely — and the prompt as workspace-state signal in a multi-worktree workflow.6 min readread→tech notes2026-10-19ripgrep: the fast search that feeds agent contextHow ripgrep earns its speed — gitignore-aware defaults, work-stealing parallel traversal, benchmarks that publish their own cliffs — and what it means for an agent's context loop.6 min readread→tech notes2026-10-19code-server: VS Code on a remote serverTaking apart code-server — VS Code vendored as a submodule, HTTP and WebSocket on separate channels, a heartbeat file for liveness — and the lesson of running the editor next to the worktree.7 min readread→tech notes2026-10-18ast-grep: find and fix code by ASTast-grep searches and rewrites code by syntax structure: patterns look like code, complex rules live in YAML files, and sg scan runs in CI as a machine gate. This post reads the repo's actual code to see how the mechanism works.8 min readread→tech notes2026-10-18DeepCode: multi-agent research coding from HKUDSDeepCode turns papers into runnable code with a seven-agent pipeline and a hard plan gate; this post reads repo health from GitHub API numbers instead of star counts.7 min readread→tech notes2026-10-17CAMEL: large-scale collaborative agent researchCAMEL turns a two-agent conversation into a data-generation pipeline with machine filters and standardized benchmarks — synthetic data is the product, and quality is measured by gates, not vibes.8 min readread→tech notes2026-10-17Qwen-Agent: bound to the Qwen stack, coupling as a featureQwen-Agent skips multi-model support and binds to one family: a tool-call template that matches each model's format, a self-declared tool catalog that hard-fails on duplicates — and it runs as the real backend of Qwen Chat. A study in when coupling is a feature.6 min readread→tech notes2026-10-16OpenAI Agents SDK: minimal primitives, tracing as a featureReading the OpenAI Agents SDK code: one Agent dataclass, handoffs as schema-carrying tools, guardrails running in parallel, and tracing shipped as its own product feature — compared with how Wakii places machine gates before human ones.8 min readread→tech notes2026-10-16PydanticAI: agents with tight typesPydanticAI ports the FastAPI typing playbook to agents: tool args and structured output are pydantic contracts, errors caught at the boundary instead of exploding far away at runtime. This post dissects the validation-first mechanism, dependency injection, and delegation-style multi-agent.8 min readread→tech notes2026-10-15Exo: an AI cluster from the devices you already ownExo discovers devices via multicast, models the cluster as a graph, and splits models by each machine's real memory. Inside the discovery, placement, and leader election code of a p2p cluster.8 min readread→tech notes2026-10-15Jan: local-first AI assistant — the desktop app ships its own engineJan embeds a llama.cpp engine inside a desktop app and reaches every engine — local or cloud — through a single interface. The packaging lesson: privacy is an architectural decision, not a settings toggle.8 min readread→tech notes2026-10-14crewAI: role-based agent teams — role/goal/backstory as datacrewAI splits each agent's role into three declarative fields and ships two runtimes: Crew for autonomy, Flow for sequential control. Reading the real code to see why roles-as-data lowers the cost of understanding.7 min readread→tech notes2026-10-14Nanobot: when less code is an architecture decisionA deep dive into HKUDS/nanobot: a personal agent driven by a single config.json, a measurable few-thousand-line core, Markdown personas — lessons in keeping a surface small on purpose.7 min readread→tech notes2026-10-13MetaGPT: a software company simulated by agents, run on SOPsMetaGPT replaces free-form agent chat with encoded SOPs: every role hands off through structured artifacts — PRD, API design, tasks, tests — whose schema is real code. Inside the 70,263-star framework that has slowed down lies a lesson about artifact contracts for multi-agent processes.7 min readread→tech notes2026-10-13AutoGen: Microsoft's multi-agent framework, disassembledAutoGen — 60,866 stars — rewrote its entire core at v0.4: from conversation-driven orchestration to an event-driven actor runtime. This post dissects that mechanism, the boundaries of the rewrite, and the repo's two-layer licensing.8 min readread→tech notes2026-10-12Neovim: an editor extended by Lua and its communityNeovim split from Vim and turned the editor into a platform: a Lua API in the core, a built-in LSP client, and Vim compatibility kept as a contract. Three patterns worth taking back to the project you maintain.7 min readread→tech notes2026-10-12vLLM: high-throughput LLM serving where memory is the bottleneckPagedAttention manages KV-cache the way an OS manages pages. Inside vLLM: why LLM serving throughput is a memory problem, the rhythm of 8 releases in 90 days, and the road from a SOSP 2023 paper to a 91k-star serving engine.8 min readread→tech notes2026-10-11mcp-chrome: when the agent drives your own Chromemcp-chrome turns your real Chrome into an MCP server — the agent reuses your tabs, cookies and logins instead of a separate browser. A dissect of the extension + native host architecture, with the development cadence as of 2026-09-08.8 min readread→tech notes2026-10-11MCP registry: the central directory of MCP serversInside modelcontextprotocol/registry: server.json declares its own schema version, reverse-DNS server names require ownership proof, and the metaregistry model points at packages instead of hosting code.7 min readread→tech notes2026-10-10MCP for Beginners: Microsoft's official MCP curriculumRead Microsoft's 13-module curriculum as an artifact: what it teaches, in which order, how each lesson is designed, and what that says about the maturity of the MCP ecosystem.8 min readread→tech notes2026-10-10grok-build: when a model house writes its own harnessReading grok-build as a data point of the model-house harness pattern: a one-way mirror of an internal monorepo, a crate that reads Claude Code and Codex sessions, and ACP as the port to outside editors.7 min readread→tech notes2026-10-09Crush: a coding agent with Charm's terminal DNAA close read of Crush's interaction design — side-by-side diffs, structured question dialogs, shared workspaces — and what a terminal-native coding agent teaches about pairing UX.8 min readread→tech notes2026-10-09codebase-memory-mcp: long-term memory for agents that read codeA pure-C MCP server compiles your codebase into a SQLite knowledge graph; a git-aware watcher re-indexes only what changed so the memory never goes stale. We read the actual code: storage, watcher, team-shared artifact.8 min readread→tech notes2026-10-08Aider: git-native AI pair programming in your terminalAider turns each AI edit into a diffable, revertable git commit, compresses the codebase into a ~1,000-token repo map, and forces models to emit code in a fixed format — three mechanisms worth studying in any agent harness.8 min readread→tech notes2026-10-08Goose: the agent that automates dev work on your machinegoose runs directly on your machine — desktop app, CLI, API — turns repetitive dev chores into schedulable YAML recipes and extends through MCP extensions. A deep dive into the architecture and what an agentic IDE can learn.7 min readread→tech notes2026-10-07Headroom: a context compression layer for coding agentsHeadroom compresses the tool output, logs, and conversation history your agent reads before they reach the model — same answers, a fraction of the tokens. Inside: architecture, two safety invariants, and a public self-audit.7 min readread→tech notes2026-10-07OpenHands: from research project to self-running agent platformHow OpenHands separates the agent runtime from its control center: a sandboxed agent-server behind GUI, API and CLI, architecture guarded by a CI test, and six releases in twenty days.7 min readread→tech notes2026-10-06Cline: the coding agent that lives inside VS CodeA deep dive into Cline's architecture — 67,661 stars, Apache-2.0: per-run checkpoints as private git refs, transactional restore, a Plan/Act boundary enforced at the tool layer, and one agent core powering CLI, desktop, and SDK.8 min readread→tech notes2026-10-06fzf: the fuzzy finder that shaped the modern command lineInside fzf, the small utility approaching 13 years of age: a Smith-Waterman variant, a matcher parallelized per CPU core, a three-goroutine architecture — and the packaging lesson of shipping one binary.8 min readread→tech notes2026-10-05The official MCP servers: the protocol's living specificationReading the seven reference servers of modelcontextprotocol/servers as a specification you can run: what the protocol chose to standardize, how each server's architecture contains risk, and how deprecation is actually operated.7 min readread→tech notes2026-10-05Zed: a high-performance editor written in RustZed makes performance an architectural decision: a hand-written GPU UI framework, custom buffer data structures, and collaboration built into the buffer core. This repo read shows each layer chosen for speed.7 min readread→tech notes2026-10-04Gemini CLI: Google's agent terminalDeep-dive Gemini CLI — 106,867 stars, Apache-2.0: a big-vendor harness that ships its entire code in the open while standing behind Google's own model, read from the policy engine to the repo's own maintenance bots.7 min readread→tech notes2026-10-04Awesome MCP servers: a map of the MCP ecosystemReading awesome-mcp-servers — 94,614 stars, 3,862 entries — as a health index of the MCP ecosystem: which categories are crowded, how fast new entries appear, and how the list describes itself with icons and auto-computed scores.7 min readread→tech notes2026-10-03Claude Code: Anthropic's agent coding tool, seen from the outsideA 144k-star repo public on GitHub that ships no tool source and attaches no standard license — yet a 387-entry changelog and 13 official plugins still reveal the architecture from the outside.8 min readread→tech notes2026-10-03llama.cpp: efficient LLM inference from CPU to the edgeInside the original C/C++ inference engine of local LLM: quantization from 1-bit to 8-bit, 17 backends from x86 CPUs to phones, and a strict AI contribution policy worth studying.8 min readread→tech notes2026-10-02LangChain: from chains library to agent engineering platformLangChain at 145,927 stars, MIT-licensed: version 1.x moved the old chains into langchain-classic and kept the agent runtime — create_agent, middleware, LangGraph, LangSmith. An architecture lesson in how a framework cuts itself down.7 min readread→tech notes2026-10-02Ollama: local LLMs in one command — then coding agents tooBehind ollama run sits a real VRAM-fitting scheduler; and since 2026, ollama launch claude wires Claude Code straight to your local model.7 min readread→tech notes2026-10-01OpenCode: a vendor-independent coding agent for your terminalA deep dive into OpenCode's architecture — 205,815 stars, MIT: one core that runs against any model provider, per-agent permission rulesets, and measurable compaction thresholds, read straight from the code on GitHub.7 min readread→tech notes2026-10-01DeepSeek Harness: the agent architecture inside an open modelDissect the everything-is-a-plugin architecture of DeepSeek Harness — 55 Cordis packages, twelve releases in 25 days — and the patterns Wakii is taking from it.8 min readread→build log2026-09-30Building Wakii in the open — log #4The fourth log of building Wakii in the open: twenty longform posts written with their own editorial framework — a matrix locked in advance, a word-band lint, a claims registry — closed by PR #3, then an integration pass after the merge.7 min readread→build log2026-09-30Building Wakii in the open — log #3The third log of building Wakii in the open: story FI-349 redesigned the entire blog — category taxonomy, a two-locale article surface, hero tiles, hreflang — and closed as a single PR. With a sourced before/after comparison.7 min readread→tech notes2026-09-29CI gates: machine-checked quality before the human gateBefore any human gate looks at anything, the machine gate chain has already run: this post reads two real chains — this site's build chain and the product's release pipeline — plus the rare spot where a workflow states what it will not do: deploy only runs when its condition is met, otherwise it no-ops.8 min readread→tech notes2026-09-29Release roundup 1.4.x: what actually shippedAn inventory of what the Wakii 1.4.x release line actually shipped: two desktop releases and one Android pre-release, read straight from the public release notes — every feature tied to a verbatim notes line or a commit, including the gap that is stated openly.7 min readread→tech notes2026-09-28The w. monogram: one icon, two reposOne letter w. drawn as SVG geometry lives in two repos with two build pipelines: the site consumes vectors for favicon and OG image, the desktop app compiles icns/png from Icon Composer — this post reads both ends to show why one master suffices.7 min readread→tech notes2026-09-28Upstream sync: living with your parent projectThree remotes, a mirror branch and a dev branch, two divergence-counting commands — this post dissects how Wakii lives alongside its upstream stablyai/orca, using real git output, a timeline diagram, and a table of the conflict layers you actually hit.7 min readread→tech notes2026-09-27The auto-update feed, from release to your machineThe full update lifecycle in real code: version-named release assets, a 24-hour main-process check, feed pinning to a concrete tag, and safe quit-and-install.8 min readread→tech notes2026-09-27Why fork under MIT: open source on purposeWakii forks Orca publicly under MIT: this post reads the LICENSE file, the site footer credit, and the README to show a public fork is a commitment — users can build from source, and every credit line has an address.7 min readread→tech notes2026-09-26Wakii site: bento layout and design tokensInside the design system of wakii.xyz: one token file as the contract, a 12-column bento grid in plain CSS, and the min-width: 0 lesson when a cell has to host a fixed 680px board.7 min readread→tech notes2026-09-26Wakii site: motion and a two-locale i18nThe two layers that decide how wakii.xyz feels: a data-attribute motion util that respects prefers-reduced-motion, and a two-layer i18n architecture — routing prefixes plus two typed string packs.7 min readread→tech notes2026-09-25Computer-use architecture: from intent to UI actionThis post dissects Wakii's computer-use pipeline in the public repo: an agent decides an action, something validates its parameters, an OS-specific native module executes it, and the outcome is evidenced before it travels back to the agent.7 min readread→tech notes2026-09-25Relay cloud architecture: signals between machinesThis post dissects Wakii's relay cloud architecture through the public code itself: why the phone and the desktop never connect directly, how the splice state machine pairs two sessions, and how admission budgets and close codes keep one cell from failing silently.8 min readread→tech notes2026-09-24Wakii's Electron process modelHow Wakii splits main, preload and renderer, seen in real code: where the sandbox flags live, what the 79 bridge contracts expose, and why renderer code has no Node.js.8 min readread→tutorial2026-09-24Diagnosing Wakii failures: three layers, six situationsA diagnosis order from orca status through the terminal to logs, then the six failures you meet most: agents not starting, worktree creation failing, a missing CLI, dead SSH terminals, browser no tab, and GitHub rate limits.6 min readread→tutorial2026-09-23Wiring Linear and GitHub: issue to worktree, worktree to PRTwo connections in Settings → Integrations: create a worktree from an issue with a prefilled name and Linear's suggested branch name, per-team opt-in status sync, and an agent that reads Linear through orca linear.8 min readread→tutorial2026-09-23Running sessions side by side: split panes and per-worktree portsWatch several agents at once with nested split panes, keep a dev port per worktree, drive sessions from the CLI, and clean up when the work is done.6 min readread→tutorial2026-09-22Cloud relay: agents on a server you own, UI anywherePair two machines through a Remote Orca Server or orca serve — the server keeps the runtime, sessions, and agents; clients are just the UI. Plus pointing the CLI at a remote runtime and the safety rules.7 min readread→tutorial2026-09-22The worktree workflow: one task, one worktree, day to dayCreate, work, review, ship, and clean up a worktree in Wakii — picking the right start-from, sharing node_modules and .env across worktrees, the CLI table, and the way back to plain git when you need it.7 min readread→tutorial2026-09-21Your first custom skill: SKILL.md, try it, share itCreate a skill of your own — one directory, one SKILL.md file — verify it with orca skills installed, edit it without a build step, and share it behind a revocable unlisted link.7 min readread→tutorial2026-09-21SSH worktrees: keep the runtime local, run the work remoteAdd an SSH target, create the worktree on the remote machine, and keep the local feel: file-event sync, status chips, sessions that survive disconnects, port forwarding, and fixing a remote missing its toolchain.7 min readread→tutorial2026-09-20Your first story, end to end: what to click, what to typeA hands-on first run, step by step: state your idea in the ⚡ Workflow tab, watch impact and the plan land on Linear, follow SFs across the 🌳 Story bracket, answer gates B0–B5, and receive a single PR.8 min readread→tutorial2026-09-20Pair your phone with the Wakii desktop over QRInstall the Android app from the APK on GitHub Releases, scan the QR your desktop shows to pair, and put four things to work right away: the tiered story view, gate resolution with free-text + confirm, gate open/closed notifications, and guard codes.6 min readread→tech notes2026-09-19Notifications and shortcuts: gates without watching the screenAn agent stops at a gate awaiting your approval — you do not have to babysit the screen. This post dissects the notification loop, gate-open → deep-link → resolve → gate-closed, with real sources, plus the ⌘⇧⌫ workspace shortcut on desktop.7 min readread→tutorial2026-09-19Installing and updating Wakii from zeroFrom an empty machine to a running app: check Node 24, pnpm and git, grab the version-named prebuilt download, understand what the first launch installs by itself — and always know exactly which release you are running.7 min readread→tech notes2026-09-18Native computer use: agents that reach the desktopWakii lets agents reach the real desktop: the src/main/computer/ module picks a provider per platform, validates parameters before execution and stamps explicit verification on each action. Plus the CI path that shipped it and the removal of the silent left-click fallback.7 min readread→tech notes2026-09-18Per-workspace env: one environment per worktreeMultiple agents on one machine sharing an environment means fighting over env vars, dependencies, and credentials. Wakii ties the environment to the workspace instead: recipes declared in orca.yaml, on-demand disposable runtimes created fresh per workspace, base and auth snapshots reused.8 min readread→tech notes2026-09-17AI diff annotation: reviews that carry their reasonsPin review notes to the diff line and ship them straight back to the running agent. This post dissects what shipped in v1.4.199 — the note composer, the card pinned into the diff, GitHub comments re-attached on the PR page — straight from the code.8 min readread→tech notes2026-09-17Android emulator: a device inside the appChecking a mobile flow no longer means holding a device and typing adb commands outside the agent loop: the orca-emulator-android skill lets the agent boot AVDs, tap, swipe, type, install apps and read logcat from inside the working session.7 min readread→tech notes2026-09-16Design Mode: point at the UI, hand the agent the contextFlip the Design Mode toggle in Wakii's browser, click the broken UI element, and its DOM, computed CSS, screenshot — even the source file/line when a source map exists — land in the agent chat as one attachment.8 min readread→tech notes2026-09-16SSH worktrees: remote code, isolated environmentsImport SSH targets straight from ~/.ssh/config and keep isolation for repos on internal servers or remote dev boxes — and when a clone fails, the error says where it ran.10 min readread→tech notes2026-09-15Terminal splits: many sessions, one windowRunning several agent sessions with a single terminal means waiting your turn. This post dissects Wakii's terminal splits from the code side: split as a runtime operation, a layout that survives restarts, and an activation order worth benchmarking.6 min readread→tech notes2026-09-15Skill /prompt-master: a prompt is a productprompt-master turns a rough idea into one production-ready prompt: extract the intent first, identify the target tool, lock the output format up front — a single prompt that works on the first paste.8 min readread→tech notes2026-09-14Skill /figma-orientation: read Figma like an agentA router skill that reads your Figma intent before the agent calls a tool: the failure class it blocks, the decision tree it walks, and the cases where a project doesn't need it.8 min readread→tech notes2026-09-14Skill /graph-engineering: turning complex systems into graphsGraph engineering has two halves: knowledge graphs answer what an agent remembers, task graphs answer how agents coordinate — the 9-stage pipeline, the diamond pattern, and the kit's nine-agent team read as a running task graph.8 min readread→tech notes2026-09-13Skill /mock-prototype: prototype before a line of production codeThree self-contained HTML directions posted as links that open right in the browser; you pick one before any refinement — the mock never enters the codebase, what crosses over is a direction doc.8 min readread→tech notes2026-09-13Skill /web-design-guidelines: the web rulebook at handReview UI code against 105 concrete rules — focus, aria, forms, animation: the failures a screenshot can't show. Rules vendored from an MIT upstream, runs offline, output is a file:line list.8 min readread→tech notes2026-09-12Skill /design-taste-frontend: audit the built UI before it shipsAn audit-first skill that only loads after the UI exists: read the brief to infer intent first, scan for named slop signals, then hold the page at a pre-flight gate before it ships.9 min readread→tech notes2026-09-12Skill /image-to-code: from screenshot to componentRead a reference image like an art director: one large image per section, hierarchy – spacing – tokens extracted, then code checked against the image — plus the boundary of when not to use this skill.7 min readread→tech notes2026-09-11Skill /gpt-taste: breaking AI's statistical defaultsWhy AI-built UIs all look alike: six-line headings, repeated left/right layouts, cheap meta labels. Skill /gpt-taste breaks those statistical defaults with scripted randomization and enforceable hard rules.8 min readread→tech notes2026-09-11Skill /frontend-design: aesthetics as decisions, not preferencesThe /frontend-design skill turns taste into a chain of justified decisions: ground the design in the subject's own world, pick typography and palette for this brief, and allow exactly one aesthetic risk — as long as you can defend it.8 min readread→tech notes2026-09-10Skill /orca-superpowers-workflow: the bridge between workflow and appOne command wraps the full six-phase pipeline from idea to Linear Done — and at each transition point, the skill fires a bridge into Linear, worktrees, the task DAG and gates. This post reads the skill source to unpack the mechanism.8 min readread→tech notes2026-09-10The /story-workflow skill: an epic, many SFs, one destinationType /story-workflow and a large feature runs as a story: one Linear epic, many sub-features in parallel, one destination branch. This post reads the skill's actual source — the entry rubric, the bracket file format, context packs, and the three-layer watchdog.8 min readread→tech notes2026-09-09Skill /brainstorm: from raw idea to a verified spec/brainstorm interrogates an idea with questions before it lets any code be written: one question at a time, a six-section spec committed to the repo, two or three approaches with trade-offs — inside the mechanism, with a real example.8 min readread→tech notes2026-09-09The /writing-plans-linear skill: plans that live on LinearTurn a spec into a plan written for someone with zero context: bite-sized tasks, a silent Write mode and a Publish mode that puts the plan on Linear for whoever comes next.8 min readread→tech notes2026-09-0850 projects shaping agentic coding — the September 2026 mapA map of 50 open-source projects around agentic coding: harnesses, multi-agent frameworks, MCP, editors, local inference, terminal — star counts pulled straight from the GitHub API on 2026-09-08.7 min readread→build log2026-09-07Building Wakii in the open — log #1The first log of building Wakii: FI-305 ships story view and gates to mobile, release 1.4.199 is out, and the blog you're reading is itself a running story.2 min readread→build log2026-09-07Building Wakii in the open — log #2The second log of building Wakii in the open: the blog grows from 10 to 30 posts through two stories sharing one parity gate, plus the numbers as of the writing date — releases, skills, story CLIs — every number sourced so you can rerun it.7 min readread→tutorial2026-09-07A tour of Wakii's public skillsWalk the /skills/ page and this site's own skills.ts: the public skills catalog in three groups, each card carrying its command, description, and how it works — with real grep transcripts so you can recount the numbers at reading time.9 min readread→tutorial2026-09-06Reviewing AI agents from your phone: Stories and decision gates in your pocketPair your phone with your desktop over a QR code, read story progress in the Stories tab, and resolve decision gates right from your phone — agents never wait for you to come back to the keyboard.3 min readread→tech notes2026-09-06Anatomy: the bilingual RSS feedThis post dissects the blog's rss.xml.js: one feed for both locales, no <language> tag, and every item carries a guid that is an absolute permalink — with a grep transcript from the built feed so you can verify it yourself.7 min readread→tech notes2026-09-05Anatomy: the OG article contract on this blogThis post dissects the blog's Open Graph contract straight from the repo's source: absolute og:image, og:type article, article:published_time — with a grep transcript from a real build so you can verify it yourself.6 min readread→tech notes2026-09-04The story workflow: from a one-line idea to a merged PRThe anatomy of the Wakii story workflow: brackets split into tiers, a DAG schedules parallel SF agents, decision gates keep humans in the loop, and everything converges on one PR.3 min readread→build log2026-09-04Wakii in production: hub-storehub-store, a warehouse-operations platform, is built end to end with Wakii's story workflow — and every process trace (brackets, specs, plans) is public on GitHub. This post walks the project's seven-bracket journey; each GitHub link carries a verbatim quote so you can diff the words against the real files.8 min readread→tech notes2026-09-03Decision gates: why Wakii's AI agents always stop to askInside decision gates: pending guards reject bad resolves with explicit error codes, notifications announce gate-open and gate-closed, and why supervised still beats autonomous.2 min readread→build log2026-09-03Case study: this blog is itself a storyThe blog you are reading was built with the same story workflow its posts describe — and the evidence sits in a public repo. This post opens the machinery up: a parity gate wired into the build, an OG contract pinned as a comment on the code, and a single PR closing the first story.7 min readread→tech notes2026-09-02Forking an IDE while keeping up with upstream: lessons from WakiiWakii forks Orca: main is the release line, fork-synced ff-only every day, while story branches run parallel to upstream — and the strategy when conflicts erupt.2 min readread→build log2026-09-02Two releases in one day — cadence falls out of the processOn September 5, Wakii's public release history recorded two releases and one Android pre-release in a single day. This post reads the real transcript to explain why shipping cadence falls out of the process, and what version-named assets change when you download the app.6 min readread→tech notes2026-09-01Convergence QA — the last tier proves it holdsThe last tier of a story adds no features — it proves that everything already built still holds. This post dissects a real convergence QA: assemble on frozen contracts, measure regression against a tag-grouped baseline, and require every diff to trace back to a source.8 min readread→tech notes2026-08-31Done means evidence, not a self-reportAn agent saying done is convincing — until it breaks on first use. This post redefines done: self-reports don't count; an independent verifier reruns the acceptance criteria against a written rubric before any verdict lands.7 min readread→tech notes2026-08-31One destination branch, one PR: where a story convergesA long story leaves dozens of commits scattered across branches — what do you review, and where? This post walks through the convergence discipline of the story workflow: every sub-feature merges into one destination branch, story/<epic>-<slug>, the story closes with a single PR, and the merge commit is a milestone you can trace backward.8 min readread→tutorial2026-08-29Linear as external memory for the agent teamA chat log is a timeline, not a state — once the context window runs out, the agent loses everything. This post walks through the mechanism that makes Linear the team's external memory: sub-issue bodies written in a standard field frame, state written to be reproduced, and an audit trail that marks instead of deletes.8 min readread→tech notes2026-08-28Real parallelism through worktree isolationRunning several agents on one repo while sharing a directory means fighting over files and polluting each other's history. This post dissects Wakii's isolation mechanism — one worktree per SF, one branch per worktree — with live git transcripts, and shows why the atomic commit is the unit of rollback.7 min readread→tech notes2026-08-27Brackets and tiers: a map for long-running projectsLong projects rarely die from a shortage of work — they die from losing the map. This post puts two real brackets side by side — 3 SFs and 28 SFs — to show the same skeleton at both scales: epic on top, tiers below, dependency edges between.7 min readread→tech notes2026-08-26Controlled rework — revert is a featureSpecs changing mid-story is normal; patching on top of patches is the real danger. This post covers controlled rework: revert to the last green commit, re-execute as one reviewable unit — through the real case of the landing switching from direction v1 Terminal Mono to v2 Bento Premium.6 min readread→tech notes2026-08-25Defensive by designWakii's workflow assumes the agent itself will make mistakes — so it makes mistakes cheap and visible: nothing deleted, unknowns flagged instead of guessed, new commands dry-tested first. This post covers those three habits and the merge conflict resolved by keeping both sides.6 min readread→tech notes2026-08-24The story learning loopThe next story should know what the first story paid to learn. This post walks Wakii's learning loop: the three-question post-task ritual, story memory that keeps its provenance, and how one note about a renamed CLI flag saved every later story from tripping on it.5 min readread→tech notes2026-08-23Watchdog: idle is not deadA silent agent is not necessarily stuck — it might be running a long build. The watchdog checks three layers (commits, terminal, Linear) before declaring a stall: healthy silence is left alone, real wedges resume from the last green commit. This post tells both real cases and the resume mechanism that loses no work.5 min readread→tech notes2026-08-22Gates, not trust — and Rule 0Every sub-feature passes five tool-enforced gates and the B0–B5 checks before it counts as done — and Rule 0 demands opening the browser and seeing the result. This post walks the mechanism: what each gate checks, which verdicts decide a story, and the real case where a flow check caught a stale preview server.6 min readread→tech notes2026-08-21Nine agents, separated powersA Wakii story run is staffed by nine agents — the analyst doesn't code, the writer doesn't approve its own work, the tester doesn't fix. This post dissects that separation of powers: the full role table, how information flows between roles, and a real case where a reviewer caught what the executor had signed off on.6 min readread→tutorial2026-08-20Your first agent team has no setup stepOpen Wakii for the first time and the superpowers kit installs itself into ~/.claude/ — skills, nine agents, and the story-* command-line tools — without touching your existing config. This post takes the zero-setup mechanism apart: what ships in the kit, why re-running never breaks anything, and how to verify every claim with real commands.5 min readread→
unlazy: completion discipline for agents — runnable gates before workunlazy is not the image lazy-loading library the brief guessed — the GitHub API confirms it is an anti-laziness skill for AI agents: write acceptance gates before the work, re-verify everything that passed, and refuse your own evidence when it does not reproduce.7 min readread→tech notes2026-10-25Cumora: a rising AI repo, read through its own anti-patterns docCumora — team chat where AI agents are first-class teammates — is three weeks old with 3.5k stars, but the best read is docs/COORDINATION.md: a refreshingly honest anti-patterns log. Inside: HELD envelopes, hold-token overrides, and five rules that keep the prompt short.8 min readread→tech notes2026-10-24OpenBot: CopilotKit's sample agent — every action through one gateCopilotKit shipped OpenBot — a self-hosted AI coworker template: each bot gets its own computer, and every action passes a single gateway that denies first and records always. A close read of the gateway, the take-the-wheel handoff, and governed generative UI.8 min readread→tech notes2026-10-24ASu-skills: nine job-hunting skills and the skills-as-content waveASu-skills packages a job-hunting workflow into nine AI skills and distributes them to five harnesses from a single source, with a registry reconciled by CI. Inside: the skills-as-content phenomenon and the anti-drift mechanism Wakii itself needs.7 min readread→tech notes2026-10-23DSH Desktop: a plugin-first desktop for an agent harnessDSH Desktop wraps DeepSeek Harness in a thin Electron host where the desktop shell itself is an unprivileged plugin — pinned upstream, stable/beta channels split by a fail-closed protocol, and pull requests drafted by an agent.7 min readread→tech notes2026-10-23Tabby: self-hosted code assistant — inference on your own machineRun a code assistant on your own infrastructure with Tabby: one binary, two commands, models running locally — and how to read a license when the GitHub API says NOASSERTION.7 min readread→tech notes2026-10-22Continue: an AI assistant inside your existing IDE — the config-hub lessonHow Continue embedded an AI assistant into VS Code and JetBrains through a single config file — plus the operational lesson from a 35-thousand-star project going read-only.7 min readread→tech notes2026-10-22Lapce: a Rust editor built for performanceLapce's architecture from real code: a rope buffer with revisions, a UI split into its own process over RPC, plugins running in a WASI sandbox — how an editor stays fast as features grow.6 min readread→tech notes2026-10-21Helix: the modal editor that needs no configurationSelection before action, LSP and tree-sitter built into the core, a single config.toml — three Helix design decisions that keep an editor lean in the age of configuration sprawl.6 min readread→tech notes2026-10-21Yazi: a terminal file manager written in RustInside Yazi's architecture — a 31-crate workspace, an async scheduler with priorities, image-protocol negotiation with a fallback chain, and Lua plugins with a package manager.6 min readread→tech notes2026-10-20bat: cat with syntax highlighting and gitHow bat works — syntect highlighting, in-process git markers via gitoxide, plain-text fallback when piped — and the lesson of output that adapts to its consumer.6 min readread→tech notes2026-10-20Starship: the fast, cross-shell promptHow Starship works — one Rust binary across ten shells, 100+ signal modules, TOML config that degrades safely — and the prompt as workspace-state signal in a multi-worktree workflow.6 min readread→tech notes2026-10-19ripgrep: the fast search that feeds agent contextHow ripgrep earns its speed — gitignore-aware defaults, work-stealing parallel traversal, benchmarks that publish their own cliffs — and what it means for an agent's context loop.6 min readread→tech notes2026-10-19code-server: VS Code on a remote serverTaking apart code-server — VS Code vendored as a submodule, HTTP and WebSocket on separate channels, a heartbeat file for liveness — and the lesson of running the editor next to the worktree.7 min readread→tech notes2026-10-18ast-grep: find and fix code by ASTast-grep searches and rewrites code by syntax structure: patterns look like code, complex rules live in YAML files, and sg scan runs in CI as a machine gate. This post reads the repo's actual code to see how the mechanism works.8 min readread→tech notes2026-10-18DeepCode: multi-agent research coding from HKUDSDeepCode turns papers into runnable code with a seven-agent pipeline and a hard plan gate; this post reads repo health from GitHub API numbers instead of star counts.7 min readread→tech notes2026-10-17CAMEL: large-scale collaborative agent researchCAMEL turns a two-agent conversation into a data-generation pipeline with machine filters and standardized benchmarks — synthetic data is the product, and quality is measured by gates, not vibes.8 min readread→tech notes2026-10-17Qwen-Agent: bound to the Qwen stack, coupling as a featureQwen-Agent skips multi-model support and binds to one family: a tool-call template that matches each model's format, a self-declared tool catalog that hard-fails on duplicates — and it runs as the real backend of Qwen Chat. A study in when coupling is a feature.6 min readread→tech notes2026-10-16OpenAI Agents SDK: minimal primitives, tracing as a featureReading the OpenAI Agents SDK code: one Agent dataclass, handoffs as schema-carrying tools, guardrails running in parallel, and tracing shipped as its own product feature — compared with how Wakii places machine gates before human ones.8 min readread→tech notes2026-10-16PydanticAI: agents with tight typesPydanticAI ports the FastAPI typing playbook to agents: tool args and structured output are pydantic contracts, errors caught at the boundary instead of exploding far away at runtime. This post dissects the validation-first mechanism, dependency injection, and delegation-style multi-agent.8 min readread→tech notes2026-10-15Exo: an AI cluster from the devices you already ownExo discovers devices via multicast, models the cluster as a graph, and splits models by each machine's real memory. Inside the discovery, placement, and leader election code of a p2p cluster.8 min readread→tech notes2026-10-15Jan: local-first AI assistant — the desktop app ships its own engineJan embeds a llama.cpp engine inside a desktop app and reaches every engine — local or cloud — through a single interface. The packaging lesson: privacy is an architectural decision, not a settings toggle.8 min readread→tech notes2026-10-14crewAI: role-based agent teams — role/goal/backstory as datacrewAI splits each agent's role into three declarative fields and ships two runtimes: Crew for autonomy, Flow for sequential control. Reading the real code to see why roles-as-data lowers the cost of understanding.7 min readread→tech notes2026-10-14Nanobot: when less code is an architecture decisionA deep dive into HKUDS/nanobot: a personal agent driven by a single config.json, a measurable few-thousand-line core, Markdown personas — lessons in keeping a surface small on purpose.7 min readread→tech notes2026-10-13MetaGPT: a software company simulated by agents, run on SOPsMetaGPT replaces free-form agent chat with encoded SOPs: every role hands off through structured artifacts — PRD, API design, tasks, tests — whose schema is real code. Inside the 70,263-star framework that has slowed down lies a lesson about artifact contracts for multi-agent processes.7 min readread→tech notes2026-10-13AutoGen: Microsoft's multi-agent framework, disassembledAutoGen — 60,866 stars — rewrote its entire core at v0.4: from conversation-driven orchestration to an event-driven actor runtime. This post dissects that mechanism, the boundaries of the rewrite, and the repo's two-layer licensing.8 min readread→tech notes2026-10-12Neovim: an editor extended by Lua and its communityNeovim split from Vim and turned the editor into a platform: a Lua API in the core, a built-in LSP client, and Vim compatibility kept as a contract. Three patterns worth taking back to the project you maintain.7 min readread→tech notes2026-10-12vLLM: high-throughput LLM serving where memory is the bottleneckPagedAttention manages KV-cache the way an OS manages pages. Inside vLLM: why LLM serving throughput is a memory problem, the rhythm of 8 releases in 90 days, and the road from a SOSP 2023 paper to a 91k-star serving engine.8 min readread→tech notes2026-10-11mcp-chrome: when the agent drives your own Chromemcp-chrome turns your real Chrome into an MCP server — the agent reuses your tabs, cookies and logins instead of a separate browser. A dissect of the extension + native host architecture, with the development cadence as of 2026-09-08.8 min readread→tech notes2026-10-11MCP registry: the central directory of MCP serversInside modelcontextprotocol/registry: server.json declares its own schema version, reverse-DNS server names require ownership proof, and the metaregistry model points at packages instead of hosting code.7 min readread→tech notes2026-10-10MCP for Beginners: Microsoft's official MCP curriculumRead Microsoft's 13-module curriculum as an artifact: what it teaches, in which order, how each lesson is designed, and what that says about the maturity of the MCP ecosystem.8 min readread→tech notes2026-10-10grok-build: when a model house writes its own harnessReading grok-build as a data point of the model-house harness pattern: a one-way mirror of an internal monorepo, a crate that reads Claude Code and Codex sessions, and ACP as the port to outside editors.7 min readread→tech notes2026-10-09Crush: a coding agent with Charm's terminal DNAA close read of Crush's interaction design — side-by-side diffs, structured question dialogs, shared workspaces — and what a terminal-native coding agent teaches about pairing UX.8 min readread→tech notes2026-10-09codebase-memory-mcp: long-term memory for agents that read codeA pure-C MCP server compiles your codebase into a SQLite knowledge graph; a git-aware watcher re-indexes only what changed so the memory never goes stale. We read the actual code: storage, watcher, team-shared artifact.8 min readread→tech notes2026-10-08Aider: git-native AI pair programming in your terminalAider turns each AI edit into a diffable, revertable git commit, compresses the codebase into a ~1,000-token repo map, and forces models to emit code in a fixed format — three mechanisms worth studying in any agent harness.8 min readread→tech notes2026-10-08Goose: the agent that automates dev work on your machinegoose runs directly on your machine — desktop app, CLI, API — turns repetitive dev chores into schedulable YAML recipes and extends through MCP extensions. A deep dive into the architecture and what an agentic IDE can learn.7 min readread→tech notes2026-10-07Headroom: a context compression layer for coding agentsHeadroom compresses the tool output, logs, and conversation history your agent reads before they reach the model — same answers, a fraction of the tokens. Inside: architecture, two safety invariants, and a public self-audit.7 min readread→tech notes2026-10-07OpenHands: from research project to self-running agent platformHow OpenHands separates the agent runtime from its control center: a sandboxed agent-server behind GUI, API and CLI, architecture guarded by a CI test, and six releases in twenty days.7 min readread→tech notes2026-10-06Cline: the coding agent that lives inside VS CodeA deep dive into Cline's architecture — 67,661 stars, Apache-2.0: per-run checkpoints as private git refs, transactional restore, a Plan/Act boundary enforced at the tool layer, and one agent core powering CLI, desktop, and SDK.8 min readread→tech notes2026-10-06fzf: the fuzzy finder that shaped the modern command lineInside fzf, the small utility approaching 13 years of age: a Smith-Waterman variant, a matcher parallelized per CPU core, a three-goroutine architecture — and the packaging lesson of shipping one binary.8 min readread→tech notes2026-10-05The official MCP servers: the protocol's living specificationReading the seven reference servers of modelcontextprotocol/servers as a specification you can run: what the protocol chose to standardize, how each server's architecture contains risk, and how deprecation is actually operated.7 min readread→tech notes2026-10-05Zed: a high-performance editor written in RustZed makes performance an architectural decision: a hand-written GPU UI framework, custom buffer data structures, and collaboration built into the buffer core. This repo read shows each layer chosen for speed.7 min readread→tech notes2026-10-04Gemini CLI: Google's agent terminalDeep-dive Gemini CLI — 106,867 stars, Apache-2.0: a big-vendor harness that ships its entire code in the open while standing behind Google's own model, read from the policy engine to the repo's own maintenance bots.7 min readread→tech notes2026-10-04Awesome MCP servers: a map of the MCP ecosystemReading awesome-mcp-servers — 94,614 stars, 3,862 entries — as a health index of the MCP ecosystem: which categories are crowded, how fast new entries appear, and how the list describes itself with icons and auto-computed scores.7 min readread→tech notes2026-10-03Claude Code: Anthropic's agent coding tool, seen from the outsideA 144k-star repo public on GitHub that ships no tool source and attaches no standard license — yet a 387-entry changelog and 13 official plugins still reveal the architecture from the outside.8 min readread→tech notes2026-10-03llama.cpp: efficient LLM inference from CPU to the edgeInside the original C/C++ inference engine of local LLM: quantization from 1-bit to 8-bit, 17 backends from x86 CPUs to phones, and a strict AI contribution policy worth studying.8 min readread→tech notes2026-10-02LangChain: from chains library to agent engineering platformLangChain at 145,927 stars, MIT-licensed: version 1.x moved the old chains into langchain-classic and kept the agent runtime — create_agent, middleware, LangGraph, LangSmith. An architecture lesson in how a framework cuts itself down.7 min readread→tech notes2026-10-02Ollama: local LLMs in one command — then coding agents tooBehind ollama run sits a real VRAM-fitting scheduler; and since 2026, ollama launch claude wires Claude Code straight to your local model.7 min readread→tech notes2026-10-01OpenCode: a vendor-independent coding agent for your terminalA deep dive into OpenCode's architecture — 205,815 stars, MIT: one core that runs against any model provider, per-agent permission rulesets, and measurable compaction thresholds, read straight from the code on GitHub.7 min readread→tech notes2026-10-01DeepSeek Harness: the agent architecture inside an open modelDissect the everything-is-a-plugin architecture of DeepSeek Harness — 55 Cordis packages, twelve releases in 25 days — and the patterns Wakii is taking from it.8 min readread→build log2026-09-30Building Wakii in the open — log #4The fourth log of building Wakii in the open: twenty longform posts written with their own editorial framework — a matrix locked in advance, a word-band lint, a claims registry — closed by PR #3, then an integration pass after the merge.7 min readread→build log2026-09-30Building Wakii in the open — log #3The third log of building Wakii in the open: story FI-349 redesigned the entire blog — category taxonomy, a two-locale article surface, hero tiles, hreflang — and closed as a single PR. With a sourced before/after comparison.7 min readread→tech notes2026-09-29CI gates: machine-checked quality before the human gateBefore any human gate looks at anything, the machine gate chain has already run: this post reads two real chains — this site's build chain and the product's release pipeline — plus the rare spot where a workflow states what it will not do: deploy only runs when its condition is met, otherwise it no-ops.8 min readread→tech notes2026-09-29Release roundup 1.4.x: what actually shippedAn inventory of what the Wakii 1.4.x release line actually shipped: two desktop releases and one Android pre-release, read straight from the public release notes — every feature tied to a verbatim notes line or a commit, including the gap that is stated openly.7 min readread→tech notes2026-09-28The w. monogram: one icon, two reposOne letter w. drawn as SVG geometry lives in two repos with two build pipelines: the site consumes vectors for favicon and OG image, the desktop app compiles icns/png from Icon Composer — this post reads both ends to show why one master suffices.7 min readread→tech notes2026-09-28Upstream sync: living with your parent projectThree remotes, a mirror branch and a dev branch, two divergence-counting commands — this post dissects how Wakii lives alongside its upstream stablyai/orca, using real git output, a timeline diagram, and a table of the conflict layers you actually hit.7 min readread→tech notes2026-09-27The auto-update feed, from release to your machineThe full update lifecycle in real code: version-named release assets, a 24-hour main-process check, feed pinning to a concrete tag, and safe quit-and-install.8 min readread→tech notes2026-09-27Why fork under MIT: open source on purposeWakii forks Orca publicly under MIT: this post reads the LICENSE file, the site footer credit, and the README to show a public fork is a commitment — users can build from source, and every credit line has an address.7 min readread→tech notes2026-09-26Wakii site: bento layout and design tokensInside the design system of wakii.xyz: one token file as the contract, a 12-column bento grid in plain CSS, and the min-width: 0 lesson when a cell has to host a fixed 680px board.7 min readread→tech notes2026-09-26Wakii site: motion and a two-locale i18nThe two layers that decide how wakii.xyz feels: a data-attribute motion util that respects prefers-reduced-motion, and a two-layer i18n architecture — routing prefixes plus two typed string packs.7 min readread→tech notes2026-09-25Computer-use architecture: from intent to UI actionThis post dissects Wakii's computer-use pipeline in the public repo: an agent decides an action, something validates its parameters, an OS-specific native module executes it, and the outcome is evidenced before it travels back to the agent.7 min readread→tech notes2026-09-25Relay cloud architecture: signals between machinesThis post dissects Wakii's relay cloud architecture through the public code itself: why the phone and the desktop never connect directly, how the splice state machine pairs two sessions, and how admission budgets and close codes keep one cell from failing silently.8 min readread→tech notes2026-09-24Wakii's Electron process modelHow Wakii splits main, preload and renderer, seen in real code: where the sandbox flags live, what the 79 bridge contracts expose, and why renderer code has no Node.js.8 min readread→tutorial2026-09-24Diagnosing Wakii failures: three layers, six situationsA diagnosis order from orca status through the terminal to logs, then the six failures you meet most: agents not starting, worktree creation failing, a missing CLI, dead SSH terminals, browser no tab, and GitHub rate limits.6 min readread→tutorial2026-09-23Wiring Linear and GitHub: issue to worktree, worktree to PRTwo connections in Settings → Integrations: create a worktree from an issue with a prefilled name and Linear's suggested branch name, per-team opt-in status sync, and an agent that reads Linear through orca linear.8 min readread→tutorial2026-09-23Running sessions side by side: split panes and per-worktree portsWatch several agents at once with nested split panes, keep a dev port per worktree, drive sessions from the CLI, and clean up when the work is done.6 min readread→tutorial2026-09-22Cloud relay: agents on a server you own, UI anywherePair two machines through a Remote Orca Server or orca serve — the server keeps the runtime, sessions, and agents; clients are just the UI. Plus pointing the CLI at a remote runtime and the safety rules.7 min readread→tutorial2026-09-22The worktree workflow: one task, one worktree, day to dayCreate, work, review, ship, and clean up a worktree in Wakii — picking the right start-from, sharing node_modules and .env across worktrees, the CLI table, and the way back to plain git when you need it.7 min readread→tutorial2026-09-21Your first custom skill: SKILL.md, try it, share itCreate a skill of your own — one directory, one SKILL.md file — verify it with orca skills installed, edit it without a build step, and share it behind a revocable unlisted link.7 min readread→tutorial2026-09-21SSH worktrees: keep the runtime local, run the work remoteAdd an SSH target, create the worktree on the remote machine, and keep the local feel: file-event sync, status chips, sessions that survive disconnects, port forwarding, and fixing a remote missing its toolchain.7 min readread→tutorial2026-09-20Your first story, end to end: what to click, what to typeA hands-on first run, step by step: state your idea in the ⚡ Workflow tab, watch impact and the plan land on Linear, follow SFs across the 🌳 Story bracket, answer gates B0–B5, and receive a single PR.8 min readread→tutorial2026-09-20Pair your phone with the Wakii desktop over QRInstall the Android app from the APK on GitHub Releases, scan the QR your desktop shows to pair, and put four things to work right away: the tiered story view, gate resolution with free-text + confirm, gate open/closed notifications, and guard codes.6 min readread→tech notes2026-09-19Notifications and shortcuts: gates without watching the screenAn agent stops at a gate awaiting your approval — you do not have to babysit the screen. This post dissects the notification loop, gate-open → deep-link → resolve → gate-closed, with real sources, plus the ⌘⇧⌫ workspace shortcut on desktop.7 min readread→tutorial2026-09-19Installing and updating Wakii from zeroFrom an empty machine to a running app: check Node 24, pnpm and git, grab the version-named prebuilt download, understand what the first launch installs by itself — and always know exactly which release you are running.7 min readread→tech notes2026-09-18Native computer use: agents that reach the desktopWakii lets agents reach the real desktop: the src/main/computer/ module picks a provider per platform, validates parameters before execution and stamps explicit verification on each action. Plus the CI path that shipped it and the removal of the silent left-click fallback.7 min readread→tech notes2026-09-18Per-workspace env: one environment per worktreeMultiple agents on one machine sharing an environment means fighting over env vars, dependencies, and credentials. Wakii ties the environment to the workspace instead: recipes declared in orca.yaml, on-demand disposable runtimes created fresh per workspace, base and auth snapshots reused.8 min readread→tech notes2026-09-17AI diff annotation: reviews that carry their reasonsPin review notes to the diff line and ship them straight back to the running agent. This post dissects what shipped in v1.4.199 — the note composer, the card pinned into the diff, GitHub comments re-attached on the PR page — straight from the code.8 min readread→tech notes2026-09-17Android emulator: a device inside the appChecking a mobile flow no longer means holding a device and typing adb commands outside the agent loop: the orca-emulator-android skill lets the agent boot AVDs, tap, swipe, type, install apps and read logcat from inside the working session.7 min readread→tech notes2026-09-16Design Mode: point at the UI, hand the agent the contextFlip the Design Mode toggle in Wakii's browser, click the broken UI element, and its DOM, computed CSS, screenshot — even the source file/line when a source map exists — land in the agent chat as one attachment.8 min readread→tech notes2026-09-16SSH worktrees: remote code, isolated environmentsImport SSH targets straight from ~/.ssh/config and keep isolation for repos on internal servers or remote dev boxes — and when a clone fails, the error says where it ran.10 min readread→tech notes2026-09-15Terminal splits: many sessions, one windowRunning several agent sessions with a single terminal means waiting your turn. This post dissects Wakii's terminal splits from the code side: split as a runtime operation, a layout that survives restarts, and an activation order worth benchmarking.6 min readread→tech notes2026-09-15Skill /prompt-master: a prompt is a productprompt-master turns a rough idea into one production-ready prompt: extract the intent first, identify the target tool, lock the output format up front — a single prompt that works on the first paste.8 min readread→tech notes2026-09-14Skill /figma-orientation: read Figma like an agentA router skill that reads your Figma intent before the agent calls a tool: the failure class it blocks, the decision tree it walks, and the cases where a project doesn't need it.8 min readread→tech notes2026-09-14Skill /graph-engineering: turning complex systems into graphsGraph engineering has two halves: knowledge graphs answer what an agent remembers, task graphs answer how agents coordinate — the 9-stage pipeline, the diamond pattern, and the kit's nine-agent team read as a running task graph.8 min readread→tech notes2026-09-13Skill /mock-prototype: prototype before a line of production codeThree self-contained HTML directions posted as links that open right in the browser; you pick one before any refinement — the mock never enters the codebase, what crosses over is a direction doc.8 min readread→tech notes2026-09-13Skill /web-design-guidelines: the web rulebook at handReview UI code against 105 concrete rules — focus, aria, forms, animation: the failures a screenshot can't show. Rules vendored from an MIT upstream, runs offline, output is a file:line list.8 min readread→tech notes2026-09-12Skill /design-taste-frontend: audit the built UI before it shipsAn audit-first skill that only loads after the UI exists: read the brief to infer intent first, scan for named slop signals, then hold the page at a pre-flight gate before it ships.9 min readread→tech notes2026-09-12Skill /image-to-code: from screenshot to componentRead a reference image like an art director: one large image per section, hierarchy – spacing – tokens extracted, then code checked against the image — plus the boundary of when not to use this skill.7 min readread→tech notes2026-09-11Skill /gpt-taste: breaking AI's statistical defaultsWhy AI-built UIs all look alike: six-line headings, repeated left/right layouts, cheap meta labels. Skill /gpt-taste breaks those statistical defaults with scripted randomization and enforceable hard rules.8 min readread→tech notes2026-09-11Skill /frontend-design: aesthetics as decisions, not preferencesThe /frontend-design skill turns taste into a chain of justified decisions: ground the design in the subject's own world, pick typography and palette for this brief, and allow exactly one aesthetic risk — as long as you can defend it.8 min readread→tech notes2026-09-10Skill /orca-superpowers-workflow: the bridge between workflow and appOne command wraps the full six-phase pipeline from idea to Linear Done — and at each transition point, the skill fires a bridge into Linear, worktrees, the task DAG and gates. This post reads the skill source to unpack the mechanism.8 min readread→tech notes2026-09-10The /story-workflow skill: an epic, many SFs, one destinationType /story-workflow and a large feature runs as a story: one Linear epic, many sub-features in parallel, one destination branch. This post reads the skill's actual source — the entry rubric, the bracket file format, context packs, and the three-layer watchdog.8 min readread→tech notes2026-09-09Skill /brainstorm: from raw idea to a verified spec/brainstorm interrogates an idea with questions before it lets any code be written: one question at a time, a six-section spec committed to the repo, two or three approaches with trade-offs — inside the mechanism, with a real example.8 min readread→tech notes2026-09-09The /writing-plans-linear skill: plans that live on LinearTurn a spec into a plan written for someone with zero context: bite-sized tasks, a silent Write mode and a Publish mode that puts the plan on Linear for whoever comes next.8 min readread→tech notes2026-09-0850 projects shaping agentic coding — the September 2026 mapA map of 50 open-source projects around agentic coding: harnesses, multi-agent frameworks, MCP, editors, local inference, terminal — star counts pulled straight from the GitHub API on 2026-09-08.7 min readread→build log2026-09-07Building Wakii in the open — log #1The first log of building Wakii: FI-305 ships story view and gates to mobile, release 1.4.199 is out, and the blog you're reading is itself a running story.2 min readread→build log2026-09-07Building Wakii in the open — log #2The second log of building Wakii in the open: the blog grows from 10 to 30 posts through two stories sharing one parity gate, plus the numbers as of the writing date — releases, skills, story CLIs — every number sourced so you can rerun it.7 min readread→tutorial2026-09-07A tour of Wakii's public skillsWalk the /skills/ page and this site's own skills.ts: the public skills catalog in three groups, each card carrying its command, description, and how it works — with real grep transcripts so you can recount the numbers at reading time.9 min readread→tutorial2026-09-06Reviewing AI agents from your phone: Stories and decision gates in your pocketPair your phone with your desktop over a QR code, read story progress in the Stories tab, and resolve decision gates right from your phone — agents never wait for you to come back to the keyboard.3 min readread→tech notes2026-09-06Anatomy: the bilingual RSS feedThis post dissects the blog's rss.xml.js: one feed for both locales, no <language> tag, and every item carries a guid that is an absolute permalink — with a grep transcript from the built feed so you can verify it yourself.7 min readread→tech notes2026-09-05Anatomy: the OG article contract on this blogThis post dissects the blog's Open Graph contract straight from the repo's source: absolute og:image, og:type article, article:published_time — with a grep transcript from a real build so you can verify it yourself.6 min readread→tech notes2026-09-04The story workflow: from a one-line idea to a merged PRThe anatomy of the Wakii story workflow: brackets split into tiers, a DAG schedules parallel SF agents, decision gates keep humans in the loop, and everything converges on one PR.3 min readread→build log2026-09-04Wakii in production: hub-storehub-store, a warehouse-operations platform, is built end to end with Wakii's story workflow — and every process trace (brackets, specs, plans) is public on GitHub. This post walks the project's seven-bracket journey; each GitHub link carries a verbatim quote so you can diff the words against the real files.8 min readread→tech notes2026-09-03Decision gates: why Wakii's AI agents always stop to askInside decision gates: pending guards reject bad resolves with explicit error codes, notifications announce gate-open and gate-closed, and why supervised still beats autonomous.2 min readread→build log2026-09-03Case study: this blog is itself a storyThe blog you are reading was built with the same story workflow its posts describe — and the evidence sits in a public repo. This post opens the machinery up: a parity gate wired into the build, an OG contract pinned as a comment on the code, and a single PR closing the first story.7 min readread→tech notes2026-09-02Forking an IDE while keeping up with upstream: lessons from WakiiWakii forks Orca: main is the release line, fork-synced ff-only every day, while story branches run parallel to upstream — and the strategy when conflicts erupt.2 min readread→build log2026-09-02Two releases in one day — cadence falls out of the processOn September 5, Wakii's public release history recorded two releases and one Android pre-release in a single day. This post reads the real transcript to explain why shipping cadence falls out of the process, and what version-named assets change when you download the app.6 min readread→tech notes2026-09-01Convergence QA — the last tier proves it holdsThe last tier of a story adds no features — it proves that everything already built still holds. This post dissects a real convergence QA: assemble on frozen contracts, measure regression against a tag-grouped baseline, and require every diff to trace back to a source.8 min readread→tech notes2026-08-31Done means evidence, not a self-reportAn agent saying done is convincing — until it breaks on first use. This post redefines done: self-reports don't count; an independent verifier reruns the acceptance criteria against a written rubric before any verdict lands.7 min readread→tech notes2026-08-31One destination branch, one PR: where a story convergesA long story leaves dozens of commits scattered across branches — what do you review, and where? This post walks through the convergence discipline of the story workflow: every sub-feature merges into one destination branch, story/<epic>-<slug>, the story closes with a single PR, and the merge commit is a milestone you can trace backward.8 min readread→tutorial2026-08-29Linear as external memory for the agent teamA chat log is a timeline, not a state — once the context window runs out, the agent loses everything. This post walks through the mechanism that makes Linear the team's external memory: sub-issue bodies written in a standard field frame, state written to be reproduced, and an audit trail that marks instead of deletes.8 min readread→tech notes2026-08-28Real parallelism through worktree isolationRunning several agents on one repo while sharing a directory means fighting over files and polluting each other's history. This post dissects Wakii's isolation mechanism — one worktree per SF, one branch per worktree — with live git transcripts, and shows why the atomic commit is the unit of rollback.7 min readread→tech notes2026-08-27Brackets and tiers: a map for long-running projectsLong projects rarely die from a shortage of work — they die from losing the map. This post puts two real brackets side by side — 3 SFs and 28 SFs — to show the same skeleton at both scales: epic on top, tiers below, dependency edges between.7 min readread→tech notes2026-08-26Controlled rework — revert is a featureSpecs changing mid-story is normal; patching on top of patches is the real danger. This post covers controlled rework: revert to the last green commit, re-execute as one reviewable unit — through the real case of the landing switching from direction v1 Terminal Mono to v2 Bento Premium.6 min readread→tech notes2026-08-25Defensive by designWakii's workflow assumes the agent itself will make mistakes — so it makes mistakes cheap and visible: nothing deleted, unknowns flagged instead of guessed, new commands dry-tested first. This post covers those three habits and the merge conflict resolved by keeping both sides.6 min readread→tech notes2026-08-24The story learning loopThe next story should know what the first story paid to learn. This post walks Wakii's learning loop: the three-question post-task ritual, story memory that keeps its provenance, and how one note about a renamed CLI flag saved every later story from tripping on it.5 min readread→tech notes2026-08-23Watchdog: idle is not deadA silent agent is not necessarily stuck — it might be running a long build. The watchdog checks three layers (commits, terminal, Linear) before declaring a stall: healthy silence is left alone, real wedges resume from the last green commit. This post tells both real cases and the resume mechanism that loses no work.5 min readread→tech notes2026-08-22Gates, not trust — and Rule 0Every sub-feature passes five tool-enforced gates and the B0–B5 checks before it counts as done — and Rule 0 demands opening the browser and seeing the result. This post walks the mechanism: what each gate checks, which verdicts decide a story, and the real case where a flow check caught a stale preview server.6 min readread→tech notes2026-08-21Nine agents, separated powersA Wakii story run is staffed by nine agents — the analyst doesn't code, the writer doesn't approve its own work, the tester doesn't fix. This post dissects that separation of powers: the full role table, how information flows between roles, and a real case where a reviewer caught what the executor had signed off on.6 min readread→tutorial2026-08-20Your first agent team has no setup stepOpen Wakii for the first time and the superpowers kit installs itself into ~/.claude/ — skills, nine agents, and the story-* command-line tools — without touching your existing config. This post takes the zero-setup mechanism apart: what ships in the kit, why re-running never breaks anything, and how to verify every claim with real commands.5 min readread→