agent stuff

Author: tksuoran

.agents/commands/mind-sets/audit.md

@@ -348,7 +348,7 @@ Assign a score from 1 (critical) to 5 (excellent) for each area:
 
 ## Workflow
 
-1. **Gather context** -- Use AskUserQuestion to collect missing project information
+1. **Gather context** -- ask the user to provide missing project information
 2. **Read documentation** -- Analyze all provided files and configurations
 3. **Load foundations** -- Read `project-foundations-{language}.md` for the project's primary language
 4. **Execute audit** -- Follow the 9-section structure systematically

.agents/commands/x-audit.md

@@ -32,13 +32,13 @@ say so explicitly and restart from STEP 0.
 
 ## Your task
 
-- Read `.claude/commands/mind-sets/audit.md` (the audit methodology)
-- Read `.claude/commands/mind-sets/project-foundations-cpp.md` (the 16 C++ foundations the audit checks against -- it `$include`s `project-foundations.md`, read that too if referenced)
+- Read `.agents/commands/mind-sets/audit.md` (the audit methodology)
+- Read `.agents/commands/mind-sets/project-foundations-cpp.md` (the 16 C++ foundations the audit checks against -- it `$include`s `project-foundations.md`, read that too if referenced)
 - If a Memory Bank exists in this repo, read it first for project context; if not, skip -- this command does not require one
-- **Optional project settings:** if `.claude/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@lint` / `@code-nav` values instead of the generic CMake/CTest commands
+- **Optional project settings:** if `.agents/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@lint` / `@code-nav` values instead of the generic CMake/CTest commands
 - **Evidence Discipline (no false negatives):** follow the "Evidence Discipline" section of audit.md -- use the language server / code intelligence (LSAI, clangd) not a single grep; scope to OWN code (exclude vendored / `_deps` / `.cpm_cache` / generated trees -- they inflate counts); NEVER emit a `Missing`/`Absent` verdict without running and quoting the targeted negative check (tests: `ctest -N`, `gtest_discover_tests` / `catch_discover_tests`, `enable_testing` + `*/test*/` targets); report a have/total ratio (e.g. "8/36 libs have tests") as `Partial`, never `Missing`
 - **Code-level diagnostics (use LSAI to find problems IN the code, not just the build):** run `mcp__lsai__lsai_diagnostics` (minSeverity warning) and scope the result to OWN code (exclude vendored / SDK / generated / known-rotten trees). Discard the `-Wc++98-compat` / `-Wpre-c++NN-compat` pedantry the analyzer emits under its near-`-Weverything` profile, then report the SUBSTANTIVE warnings -- sign/signedness conversion, field/local shadowing, float `==`/`!=`, unsafe-buffer-usage / pointer arithmetic, unhandled `switch` enums, `const`-dropping casts, unused includes -- as a "Code-Level Diagnostics" finding with `file:line`. Tie any defect the real build hides (warnings-as-errors not on that compiler) back to Foundation #1.
-- Use AskUserQuestion to gather project context if not provided (what the project is, primary build system, target platforms, what worries the user)
+- Ask the user for project context if not provided (what the project is, primary build system, target platforms, what worries the user)
 - Confirm activation with "# Architecture Audit active (C++)"
 
 ## Scope reminder (C++ specifics)

.agents/commands/x-debug.md

@@ -32,9 +32,9 @@ say so explicitly and restart from STEP 0.
 - Skip the BASE_RULE and I will kick your honest ass straight back to STEP 0. Stop, reset your head, and do it right -- anything you produce otherwise is invalid.
 
 ## Prerequisites
-- Read `.claude/commands/mind-sets/debug-protocol.md` (Protocol D -- the methodology, single source of truth for the rules)
+- Read `.agents/commands/mind-sets/debug-protocol.md` (Protocol D -- the methodology, single source of truth for the rules)
 - If a Memory Bank exists, read it for context; otherwise skip (not required)
-- **Optional project settings:** if `.claude/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@asan` / `@code-nav` / `@logs` / `@platform` values instead of the generic commands below
+- **Optional project settings:** if `.agents/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@asan` / `@code-nav` / `@logs` / `@platform` values instead of the generic commands below
 
 ## Purpose
 

.agents/commands/x-develop.md

@@ -32,12 +32,12 @@ say so explicitly and restart from STEP 0.
 - Skip the BASE_RULE and I will kick your honest ass straight back to STEP 0. Stop, reset your head, and do it right -- anything you produce otherwise is invalid.
 
 ## Prerequisites
-- Read `.claude/commands/mind-sets/cpp-senior.md` (the C++ senior mindset -- the rules you code by)
-- Read `.claude/commands/mind-sets/common-coding-style.md` (cross-language rules: immutability, KISS/DRY/YAGNI, file limits)
-- Read `.claude/commands/mind-sets/project-foundations-cpp.md` (the foundations the code must satisfy)
-- If the methodology is TDDAB: read `.claude/commands/mind-sets/cpp-tddab-overlay.md`
+- Read `.agents/commands/mind-sets/cpp-senior.md` (the C++ senior mindset -- the rules you code by)
+- Read `.agents/commands/mind-sets/common-coding-style.md` (cross-language rules: immutability, KISS/DRY/YAGNI, file limits)
+- Read `.agents/commands/mind-sets/project-foundations-cpp.md` (the foundations the code must satisfy)
+- If the methodology is TDDAB: read `.agents/commands/mind-sets/cpp-tddab-overlay.md`
 - If a Memory Bank / plan file exists, read it for context; otherwise work from the plan the user points you to
-- **Optional project settings:** if `.claude/commands/cpp-project.md` exists, read it and use its `@build` / `@configure` / `@test` / `@lint` / `@asan` values in the per-block VERIFY loop instead of the generic CMake/CTest commands
+- **Optional project settings:** if `.agents/commands/cpp-project.md` exists, read it and use its `@build` / `@configure` / `@test` / `@lint` / `@asan` values in the per-block VERIFY loop instead of the generic CMake/CTest commands
 
 ## Purpose
 
@@ -84,4 +84,4 @@ non-trivial.
 - **Follow existing patterns and naming** in the target codebase
 - **Fix the cause, not the symptom** -- a defensive check that hides the root is rejected
 - **One block at a time** -- never batch blocks past a red build
-- **Ask when unsure** -- use AskUserQuestion rather than guessing the intent
+- **Ask when unsure** -- ask the user rather than guessing the intent

.agents/commands/x-review-plan.md

@@ -32,10 +32,10 @@ say so explicitly and restart from STEP 0.
 - Skip the BASE_RULE and I will kick your honest ass straight back to STEP 0. Stop, reset your head, and do it right -- anything you produce otherwise is invalid.
 
 ## Prerequisites
-- Read `.claude/commands/mind-sets/tddab-planner.md` + `.claude/commands/mind-sets/cpp-tddab-overlay.md` (the rules the plan must satisfy)
-- Read `.claude/commands/mind-sets/project-foundations-cpp.md` (project conformity)
+- Read `.agents/commands/mind-sets/tddab-planner.md` + `.agents/commands/mind-sets/cpp-tddab-overlay.md` (the rules the plan must satisfy)
+- Read `.agents/commands/mind-sets/project-foundations-cpp.md` (project conformity)
 - If a Memory Bank exists, read it for context; otherwise skip
-- **Optional project settings:** if `.claude/commands/cpp-project.md` exists, read it and use its `@code-nav` / `@build` / `@test` values where relevant
+- **Optional project settings:** if `.agents/commands/cpp-project.md` exists, read it and use its `@code-nav` / `@build` / `@test` values where relevant
 
 ## Steps
 

.agents/commands/x-tddab.md

@@ -32,11 +32,11 @@ say so explicitly and restart from STEP 0.
 - Skip the BASE_RULE and I will kick your honest ass straight back to STEP 0. Stop, reset your head, and do it right -- anything you produce otherwise is invalid.
 
 ## Prerequisites
-- Read `.claude/commands/mind-sets/tddab-planner.md` (the base TDDAB methodology -- single source of truth for plan structure)
-- Read `.claude/commands/mind-sets/cpp-tddab-overlay.md` (C++ specifics: Catch2/CTest, expected/optional, sanitizers in VERIFY)
-- Read `.claude/commands/mind-sets/project-foundations-cpp.md` (the foundations the plan must respect)
+- Read `.agents/commands/mind-sets/tddab-planner.md` (the base TDDAB methodology -- single source of truth for plan structure)
+- Read `.agents/commands/mind-sets/cpp-tddab-overlay.md` (C++ specifics: Catch2/CTest, expected/optional, sanitizers in VERIFY)
+- Read `.agents/commands/mind-sets/project-foundations-cpp.md` (the foundations the plan must respect)
 - If a Memory Bank exists, read it for context; otherwise skip
-- **Optional project settings:** if `.claude/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@asan` values in the VERIFY steps of the plan
+- **Optional project settings:** if `.agents/commands/cpp-project.md` exists, read it and use its `@build` / `@test` / `@asan` values in the VERIFY steps of the plan
 
 ## Purpose
 
@@ -47,7 +47,7 @@ context -- no dependency on a later block.
 
 ## Steps
 
-1. **Clarify the feature** -- use AskUserQuestion for anything ambiguous (scope,
+1. **Clarify the feature** -- ask the user about anything ambiguous (scope,
    target API, where it plugs into the existing code).
 2. **Decompose bottom-up** -- list the smallest provable units. For C++ that is
    usually: value types/DTOs -> domain functions -> the service/composition.

.agents/skills/erhe-cpp-debugging/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: erhe-cpp-debugging
-description: Debug an erhe C++ crash, abort, VERIFY/ERHE_FATAL, thrown exception, hang or deadlock at the source level with a real callstack and live variable values, instead of guessing from logs/log.txt and a stripped backtrace. On macOS drive lldb from the Bash tool (preferred: the embedded Python SB API emitting JSON); on Windows drive the live Visual Studio via the visualstudio MCP server (breakpoints, launch-under-debugger, step, callstack, locals). Use this for "editor.exe exited non-zero / [crash] unhandled exception / a minidump appeared / a VERIFY or FATAL fired / a C++ exception 0xe06d7363 / it deadlocked or hung" -- the minidump and log are a starting hypothesis, the live callstack + locals are the diagnosis. For GPU "renders wrong" bugs use erhe-renderdoc-gpu-debug instead; this skill is for CPU-side control flow and state.
+description: Debug an erhe C++ crash, abort, VERIFY/ERHE_FATAL, thrown exception, hang or deadlock at the source level with a real callstack and live variable values, instead of guessing from logs/log.txt and a stripped backtrace. On macOS drive lldb from a shell (preferred: the embedded Python SB API emitting JSON); on Windows drive the live Visual Studio via the visualstudio MCP server (breakpoints, launch-under-debugger, step, callstack, locals). Use this for "editor.exe exited non-zero / [crash] unhandled exception / a minidump appeared / a VERIFY or FATAL fired / a C++ exception 0xe06d7363 / it deadlocked or hung" -- the minidump and log are a starting hypothesis, the live callstack + locals are the diagnosis. For GPU "renders wrong" bugs use erhe-renderdoc-gpu-debug instead; this skill is for CPU-side control flow and state.
 ---
 
 # erhe C++ debugging (lldb on macOS / Linux, Visual Studio MCP on Windows)
@@ -11,20 +11,20 @@ variable values at the fault site. Do NOT try to reason a crash out from
 `logs/editor_crash_*.dmp` minidump -- they are a hypothesis, not the diagnosis.
 
 This is the authoritative macOS/Linux lldb run-book. The Windows VS-MCP prose
-(full tool list, typical flow, gotchas) lives in CLAUDE.md under "Debugging on
+(full tool list, typical flow, gotchas) lives in AGENTS.md under "Debugging on
 Windows / MSVC (Visual Studio MCP server)".
 
 ## Pick the platform path
 
-- **macOS / Linux -> lldb via the Bash tool.** `/usr/bin/lldb` ships with the
+- **macOS / Linux -> lldb via a shell.** `/usr/bin/lldb` ships with the
   Command Line Tools (no install). All three Debug builds have full symbols:
   `build_xcode_{metal,opengl,vulkan}/src/editor/Debug/editor`.
 - **Windows -> the `visualstudio` MCP server** (`mcp__visualstudio__*`). It drives a
   live VS instance: `debugger_add_breakpoint` -> `debugger_launch` -> poll
   `debugger_status` until it breaks -> `debugger_get_callstack` /
   `debugger_get_locals` / `debugger_evaluate` -> `debugger_stop`. Always
   `solution_info` first to confirm which solution is loaded; for headless-only
-  repros open `build_vs2026_vulkan_headless/erhe.slnx`. (CLAUDE.md has the full
+  repros open `build_vs2026_vulkan_headless/erhe.slnx`. (AGENTS.md has the full
   tool list and flow.) `lldb-mi`/GDB-MI do not exist here; on Windows use VS-MCP.
 
 ## macOS: which lldb interface (most efficient first)

.agents/skills/erhe-headless-verify/SKILL.md

@@ -6,7 +6,7 @@ description: Verify editor behavior end-to-end without a display -- build the he
 # erhe headless verification (headless Vulkan build + in-editor MCP)
 
 The condensed run-book for the loop used to verify editor changes headlessly.
-Full reference: CLAUDE.md "In-editor MCP server".
+Full reference: AGENTS.md "In-editor MCP server".
 
 ## Step 1 -- build
 
@@ -84,7 +84,7 @@ py -3 scripts/mcp_call.py get_scene_nodes b64:<base64-of-{"scene_name":"Default
   `transform_selection` it, then re-screenshot and diff by eye.
 - Missing capability? Add a new MCP tool to
   `src/editor/mcp/mcp_server.{hpp,cpp}` rather than working around it --
-  CLAUDE.md calls this out as first-class debugging infrastructure.
+  AGENTS.md calls this out as first-class debugging infrastructure.
 
 ## Step 4 -- clean up (always)
 

.agents/skills/erhe-quest-launch/SKILL.md

@@ -57,7 +57,7 @@ Capture files land in `logs/quest_<YYYYMMDD>_<HHMMSS>.log` (gitignored) and are
 The script enforces these invariants -- understand them before changing it:
 
 - **One extra process, ever.** The single backgrounded `adb logcat` is the only process owned by this workflow. Re-running `quest_logcat.sh` kills the previous streamer before starting a new one; `stop` kills without starting a replacement. The shared `adb` server daemon is not "ours" -- it persists across sessions regardless.
-- **No leaks across Claude Code sessions.** The PID is tracked in `logs/.logcat.pid` (a real file in the repo, gitignored), so even a fresh shell finds and stops the prior capture.
+- **No leaks across AI client sessions.** The PID is tracked in `logs/.logcat.pid` (a real file in the repo, gitignored), so even a fresh shell finds and stops the prior capture.
 - **No PID-recycle hazard.** Before killing, the script confirms the recorded PID still points at an `adb` process via `ps -p PID`; if the PID was recycled to something unrelated, it leaves the new owner alone and just removes the stale PID file. Git Bash's `ps` reports the executable path only (no argv), so it matches on `adb` -- fine because every other adb call in this workflow is short-lived.
 
 ## Filtering a capture to erhe-only lines

.agents/skills/erhe-renderdoc-gpu-debug/SKILL.md

@@ -225,5 +225,5 @@ rect) settle it without any CPU-readback ambiguity.
   pre-tonemap) -- read the **stats**, do not judge brightness by the PNG. The ID
   textures are RGBA8/UNORM, so their PNGs are literal.
 - `save_texture` path: forward slashes, parent dir must exist, `overwrite:true`.
-- The proxy leaves `qrenderdoc` running across Claude Code sessions;
+- The proxy leaves `qrenderdoc` running across AI client sessions;
   `renderdoc_shutdown` only kills an instance the proxy itself launched.