When BuildCompactWinGrap returns {} (WPF overlays, Eye window), PrintMultiWindowTargets skips AutoRegisterBug for verif...
a11y eval action is REMOVED (hard-blocked). Returns false immediately with [REMOVED] banner + migration hint. MCP path...
When hwnd: is in the grap, CheckMissingScope is skipped. When --force is passed, ReadinessContext.Force=true bypasses I...
Key insights from Korean YouTube interview: Oh My Claude Code (Heo Ye-chan) + Oh My Open Agent (Kim Yeon-gyu) on AI har...
Twice-daily (11:00, 18:00) AI news digest -- see ai-news-to-suggests for the full pipeline including Gemini filter and...
wkappbot agent 'task' - tier uses Claude-standard canonical flags as universal interface. All AI types (haiku/sonnet/op...
5-step checklist before every Agent() call. Prevents claude-md-sync / delegation / pace / brief guard failures that was...
Anthropic Computer Use lets Claude drive a real desktop via screenshot + mouse/keyboard. Two paths: (1) API beta tool c...
Find and interpret WKAppBot logs across command-specific files, runtime history, errors.jsonl, skill invocation logs, a...
Automatically capture user prompts, tool calls, outcome scores, and durable patterns, then route reusable knowledge to...
End-to-end CDP ask pipeline health check via wkask.ps1. Run after hot-swap, Chrome accumulation, or before triad sessio...
Reduce ask-gpt calls when triaging wkappbot suggest backlogs. Use when you have a page of suggestions and want one GPT...
T2 domain tier. Step-by-step: batch GPT ranking, concurrent per-item queries, timeout handling, pipeline health check,...
T3 reference tier. Canonical rules 2026-05-20, duplicate step archive, calibration history.
Create a multi-agent analysis team (orchestrator + sub-agents) using only Claude chat. Anthropic official format: YAML...
CDP quick-ref: port isolation rule, command table, typical flow, error codes. For step-by-step: cdp-command-guide-howto...
T2 domain tier. Full command reference: navigate, html, capture, eval-js, tabs, close, url, title. cdp-mon anomaly repo...
T3 reference tier. Post-reboot recovery, IME daemon watch, tab accumulation bug, JS alert detection, verified tests.
Pattern for automating Chrome pages that show alert/confirm/beforeunload dialogs. Uses Page.enable + event loop to catc...
Runtime.evaluate with focus-risky expressions (.focus()/.click()/dispatchEvent etc) now triggers pre-minimize dance bef...
Keep CLAUDE.md Pending in sync with actual work. CLAUDE.md is canonical -- if not there it does not exist.
Newbie orientation and mandatory rules guide. Explains two-layer CLAUDE.md system global vs project repo root. Lists ke...
AI-only pre-save gate that rejects text files containing characters not representable in CP949. Use before Codex, Claud...
PreToolUse hook + Codex wrapper logs all AI tool calls with timestamp and CWD slug to D:/GitHub/ai-audit.log.
PreToolUse hook intercepts taskkill and Stop-Process. Processes outside CWD require user approval; inside CWD allowed s...
T2 domain tier. Examples, body lines, template, scope, best practices, full harness:skill usage.
T3 ref. isEditOp history, lock-clear hygiene, global CLAUDE.md integration, Catch-22 fix.
Design rationale + migration guide for running Claude Code via terminal (AppBot-mediated) instead of VS Code extension....
80-90% of tokens are spent on file exploration, not writing code. Techniques: (1) skill-search-first prevents redundant...
AGENTS.md rules that guide Codex to use Claude-named cmd wrappers instead of native write_file/read_file tools. Require...
How Codex should create, edit, verify, install, and re-read AppBot skills.
Which tools Codex actually calls, ranked by frequency. Key input for harness coverage planning.
Codex shell_command tool flows through wkharness but bypasses all guards -- gap analysis and one-line fix.
CLAUDE.md CronCreate declarations + harness auto-detection replaces custom scheduler. Recording execution in CLAUDE.md...
Operator reference for diagnostic and introspection verbs: whisper (real-time speech plus offline pipeline), validate (...
ListenAsync spawns 32 concurrent AcceptLoop workers each owning a NamedPipeServerStream. Burst of parallel --sudo Conne...
contextWarnedPcts is pre-populated from runtime/context_warned.json on Eye startup. SaveContextWarnedEntry writes each...
When eye-evict fires, processes with null GetProcessCommandLine output are admin-elevated; skip window check before evi...
Commit message rules for this repo. English only -- Korean wastes 2-3x tokens in git history.
T1 core. Sonnet architects, Haiku prototypes in safe sandbox, wkcodex applies to production. Never let Haiku touch orig...
T2 domain tier. Brief templates for Haiku prototype + wkcodex apply. When to prototype vs direct apply. Anti-patterns.
T3 ref. Real session examples (successes + failures), failure mode analysis, evolution notes.
When you need to verify a new guard, skill, or CLI actually works, delegate the mechanical execution to Haiku. Haiku do...
After two same-kind failures on a mechanical CLI call (suggest/skill add/edit/mail/git commit with arg or validation er...
Build a compound-interest knowledge base from any raw sources (web clips, YouTube transcripts, docs). No RAG, no embedd...
Manage LinkedIn profile via CDP WebSocket. Profile at linkedin.com/in/***. Sections lazy-load on scroll. Use comtypes U...
Execution checklist for porting a multi-tab assistant interface into AppBot, focused on adapters, event schema, isolate...
Design a VS Code extension style multi-tab assistant interface as an integration reference for AppBot, focusing on prov...
Denying native Edit/Write is a CODEX-only mechanism (codex-tool-wrappers); never inject Edit/Write into the Claude .cla...
PreToolUse hook: blocks Bash/PowerShell until CLAUDE.md CronCreate entries are registered in session JSONL. Global hook...
T1 core tier. Session-start decision rules: when to run mandatory loads, gg trigger, bare-keyword routing. Execution de...
When Opus (or any model) encounters a wkappbot bug, UX flaw, retry-prone pattern, unhelpful error, or improvement idea,...
QMD: BM25 + vector semantic + optional LLM reranking local search engine. Pre-indexes C# source, skill JSONs, and docs....
auto_dismiss_hts_popups.py steals keyboard/mouse focus every 1-5s via WM_SETFOCUS (lines 454,534,677). Focusless replac...
Complete proven recipe to RESOLVE stale BUG-AUTO ASK CDP FOCUS-STEAL timeout noise suggests (especially MERGEx2 survivo...
Shared append helper for runtime logs and JSONL files. If the target file was created on a previous day, rotate it into...
Tracking suggest resolution for codex skill review
Temporary skill to clear stale session lock from wkappbot-sdk 03:25 session
Scan recent session logs for repeated successful patterns, score them for durability, and promote only stable knowledge...
Resolve and suppress same-version skill sync warnings by recording chosen ids and verifying a clean follow-up sync.
Inspect wkappbot skill sync split-brain warnings, extract the conflicting ids, and decide the minimal resolve list befo...
Auto-detect .cs source changes, map to required skills, verify harness:skill registration in CLAUDE.md, alert on missin...
Suggest = AI-logged bug/improvement record. Triage with merge, resolve with evidence script. SUBMIT RULE: use wkappbot-...
5-step preflight to pass all resolve guards first try. Covers commit keyword, class size, evidence filename, hidden-win...
T3 reference tier. KEYWORD GUARD FAIL amend pattern, RIQUA items, verified test history.
Resolve CDP bug suggests with evidence script validation
T2 HowTo: suggester confirm flow, delegation patterns, skill-news pre-check, stale-resolve recipe for suggest-triage-pl...
Operator playbook for reviewing wkappbot suggestions: classify duplicates, prioritize by impact, keep evidence requirem...
T3 Reference: guard chain workarounds, known bugs, proven patterns for suggest-triage-playbook
How-to guide: merge, resolve, check, evidence script rules, guardrails. See suggest-workflow for core principles.
History, verified tests, lessons learned. Core rules in suggest-workflow; how-to in suggest-workflow-howto.
How to write evidence scripts for suggest resolve. The script must demonstrate regression check passing before 2nd-part...
Integrate browser-control MCP patterns into WKAppBot's Eye MCP architecture to extend CDP-based automation with standar...
Extend WKAppBot's MCP server to expose desktop control actions (click/type/screenshot) as MCP tools callable by externa...
Borrow self-healing CDP reconnect patterns for WKAppBot CdpClient to handle Chrome restarts without losing prompt-pump...
Audit OculOS's MCP tool surface against WKAppBot's a11y command set and promote missing desktop-control primitives to t...
Extract CDP session self-healing patterns into WKAppBot's CdpClient to replace current manual reconnect logic. Source:...
Compare UIA traversal and fallback strategy against WKAppBot's 5-tier element search to identify gaps or borrowable pat...
Operational cheatsheet for wkappbot a11y in Claude and Codex workflows. Covers principles (inspect first, resolve grap...
Reference and history: deprecated commands, verified tests, USAGE-REMOVED patterns, audit notes.
How-to reference for a11y sub-actions: discovery, window-state, input, read, wait, common flags.
wkappbot ask delegates questions to GPT/Gemini/Claude. triad = all three in parallel. agent = autonomous sub-agent loop...
Smoke test script for ask optimization: dispatch latency, CDP port isolation, a11y eval block, cdp field in grap. Run:...
ARCHIVED: moved to cheap-commit-todo-gate
PreToolUse gate that blocks git commit from Haiku/Sonnet sessions unless CLAUDE.md Pending was edited in the last 30 tu...
WKAppBot Chrome profile auth sync: master->project copy on launch, project->master SQLite merge on exit. Use when sessi...
Claude Code PreToolUse hook: block PowerShell Stop-Process targeting wkappbot processes; companion audit-log traces all...
When Claude hits context limit, claude-proxy intercepts the error and hands off to wkappbot agent gemini --new-session...
Project CLAUDE.md applies mandatory agent delegation for all wkappbot commands. Haiku for read/survey, Codex for execut...
Read .wkappbot/hq/logs/errors.jsonl, aggregate by command pattern, file suggest for patterns >=5 hits. Claude main duty...
wkappbot file tools for filesystem access. edit auto-creates .bak/ backup. glob MUST use **/ prefix. After edit, U+FFFD...
T2 usage recipes for wkappbot file tools: read/write/edit, glob, grep, batch edits.
T3 gotchas for wkappbot file tools: grep starvation fallback, encoding safety, backup/undo.
GRAP = Grab Accessible Pattern -- JSON5-extended universal UI element address. {JSON5-window}#AbsTagPath: identifies an...
When ctx% exceeds 8MB, prepare handoff. Compress to essentials -- precision beats length. The next Claude must hit the...
Pattern for making Claude skills evolve automatically. CLAUDE.md harness:learn directives trigger on tool completions,...
Documents the mandatory format for harness:skill rules in CLAUDE.md. Pattern must use 3 required skills on single lines...
Documents the mandatory format for harness:skill rules in CLAUDE.md. Pattern must use 3 required skills on single lines...
wkappbot logcat tails .log files with regex filtering. OR patterns via semicolon. Glob segments auto-expanded. --past N...
T2 how-to for md-heal: the step-by-step CLAUDE.md heal procedures (scan repo MD, drift-check vs README and AGENTS, gg v...
T3 reference for md-heal: the file role map (CLAUDE.md vs AGENTS.md vs README.md), token-optimization rule, and the tod...
Actionable checklist to heal project CLAUDE.md: scan for stale commands, check drift against README/AGENTS, validate gg...
How to deliver prompts to a running Claude Code / Codex window. Use prompt send for self-prompting, scheduled nudges, a...
Guard that warns on Edit/Write if skill news or skill list was not run in the last 60 minutes this session.
wkappbot slack send/reply for Slack communication. Eye = Slack daemon built-in (Socket Mode), no separate listener. Alw...
Stall-guard exemption rules for wk* AI tools: all wk* commands are exempt while pace is under limit, with per-tool budg...
wkappbot cdp tools (formerly 'web'). Exempt URL-takers: open, fetch, search, tabs, status, close. Tab-bound verbs (navi...
Index for wk AI routing, budgets, and recovery. Read T1/T2/T3 as needed.
Budget rules for cheap-first auto selection and cooldown handling.
Recovery notes for wrappers, cache files, and smoke tests.
Routing map for wk AI wrappers and model aliases.
ask gpt --timeout Ns can hang as a zombie consuming 4-5 GB RAM. Detection and immediate kill procedure.
Hard-won lessons from implementing find tree: grap candidate scoring, Afx MFC class detection, glob char-class pitfall...
Two new windows command modes: (1) --focus shows focused child hwnd + proper grap expression per window. (2) no-filter...
wkappbot file edit supports regex substitution, auto .bak/ backup, undo, and encoding-safe transactions. Use this inste...
wkappbot file edit --old-file pattern matching silently fails when old-file has BOM or CRLF/LF mismatch with target. wk...
Design spec for wkedit.sh failure detection + arg correction + retry. Based on live test results 2026-05-18.
Design for a shared append-only bulletin file that all running background agents reference live, force-fed by a PostToo...
Guard cluster quick-ref for wkharness.ps1. One-liner per guard + escape hatch. Details in howto and ref skills.
wkharness-guards per-guard unblock steps, escape hatch procedure, guard cascade recovery
How to write wkask/wkcdp-style QA scripts that find bugs cheaply. Key principle: state-check first (fast), eval-js seco...
wk-tool scripts are QA-first tools: they detect bugs, validate state, and report anomalies cheaply so Claude never wast...
Universal relay pattern for all wk*.ps1 tools. .sh for bash/Claude, .cmd for CMD/user, .ps1 for direct PS. wkwrap.sh --...
When stall-guard fires on repeated wkzombie runs, delegate to Opus Agent (not Haiku). Haiku subagents are blocked by wk...
Periodic repo health check covering GitHub issues (close Daily reports, triage real bugs), unanswered issue comments, e...
구현 방법이 불확실하거나 버그 원인이 안 잡히면 삼두(GPT+Gemini+Claude)에 먼저 물어봐라. 삽질 30분보다 질문 1분이 낫다. Claude 혼자 추론하면 컨텍스트 낭비 -- 삼두에 위임하면 plann...
삼두(ask triad)에서 얻은 유용한 인사이트는 바로 스킬로 저장. 다음 세션도 같은 고민 반복 안 하게. 오답노트+성공기록 = 집단지성 누적.
How wkappbot skill search ranks results: GlobCoverageScore formula, which fields count, why steps are invisible, and ho...
새 세션 시작 시 첫 5분이 전체 세션 효율을 결정한다. 순서 지켜서 빠르게 파악 후 슬랙 보고, 그 다음 바로 작업 돌입.
타겟 요소 구조를 모르면 삽질 확정. 자동화 코드 짜기 전 wkappbot a11y inspect 로 UIA 트리 확인 필수. automation_id, class, 패턴 지원 여부 먼저 파악.
UIA Invoke/Value/Toggle/Select = 포커스 불필요. SendInput/Hotkey = EnsureFocus 필요. 포커스 탈취는 사용자 작업 방해 -- 항상 Focusless 먼저 시도.