菜单
This skill should be used when the user asks to "analyze performance", "read flame instrumentation", "inspect flame dumps", "find frame spikes", or use the Adriatic instrument profile.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Velfi/adriatic --skill perf-anal命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Write, revise, and develop story material for Adriatic as an embedded narrative designer, including scenes, dialogue, character voices, letters, quest beats, story arcs, barks, lore, and implementation-ready conversation text. Use when the user invokes autowriter, asks to write or improve Adriatic fiction, requests narrative ideas or continuity checks, or needs story content fitted to the project's dialogue and campaign systems.
Build and run Adriatic's deterministic command-line capture and cinematic export tools to produce PNG screenshots, plant reference sheets, animated GIFs, MP4 videos, or vehicle turntable frames. Use when the user asks to capture, render, animate, or inspect a specific Adriatic scene, lab, mode, named target, camera angle, resolution, contact sheet, GIF, video, cinematic, or turntable from the built app. Do not use for a screenshot of the already-running game; use hotshot for that live workflow.
Launch Adriatic under LLDB, keep the debugger session alive while a human interactively reproduces a crash, then capture the stopped process state, diagnose the failure, and optionally implement and verify a fix. Use when the user says "watch me crash", asks the agent to observe or debug a live crash reproduction, or wants to reproduce an Adriatic crash while the agent waits in a debugger.
基于 SOC 职业分类
正在显示 SKILL.md
| name | perf-anal |
| description | This skill should be used when the user asks to "analyze performance", "read flame instrumentation", "inspect flame dumps", "find frame spikes", or use the Adriatic instrument profile. |
| version | 0.1.0 |
Use the instrument profile to collect Odin function instrumentation, then use the local flame analyzer to summarize frame cost, identify spikes, and inspect one frame's scopes. Keep the output in catermujo's source-of-truth format so the same analysis workflow works across both repositories.
Build and launch the instrumented application:
make instrument
Exercise the scene or reproduce the slow operation. Close the application through
its normal quit path so flame_graph_destroy runs and exports the session. Avoid
SIGINT or force-kill when the export matters; process termination skips Odin
defer cleanup.
Automatic export covers the complete retained history, currently 180 completed frames, without requiring a range selection.
The instrument build stages runtime assets in build/instrument. The application
changes its working directory to the executable directory, so the default dump
base resolves there:
build/instrument/flame.graph
build/instrument/flame.scopes.ndjson
build/instrument/flame.frames.ndjson
build/instrument/flame.folded
flame.graph is the base name. The analyzer accepts the base name or either
NDJSON sidecar and resolves the sibling scopes sidecar automatically.
Keep all instrumentation consumers on these sidecars:
flame.scopes.ndjson starts with one history_header record, followed by one
frame_scopes record per frame. Each scope record contains other_ms,
scope_count, and aggregated colored scopes.flame.frames.ndjson starts with the same header, followed by one frame
record per frame. Each frame contains raw nested slots with source locations,
tick bounds, relative start time, and duration.flame.folded contains semicolon-separated folded stacks with performance-counter
tick weights for standard flamegraph viewers.Use freq_hz from the header when converting start_ticks, end_ticks, or
folded weights. Adriatic's current clock uses nanosecond ticks and emits
freq_hz: 1000000000; do not assume a machine-specific frequency.
Run the copied analyzer directly from the repository root:
python3 tools/flame_dump.py summary build/instrument/flame.graph --top 12
The summary reports frame count, average frame/FPS, CPU/wait/GPU averages, the worst frame, top aggregated scopes, and worst frames. Feed the scopes sidecar directly when path resolution needs to be explicit:
python3 tools/flame_dump.py summary \
build/instrument/flame.scopes.ndjson \
--top 16
Analyze spikes with the default p95 threshold:
python3 tools/flame_dump.py spikes build/instrument/flame.graph --top 20
Use an explicit percentile or a fixed frame-time threshold. Select one mode per invocation:
python3 tools/flame_dump.py spikes build/instrument/flame.graph \
--percentile 99 --top 20
python3 tools/flame_dump.py spikes build/instrument/flame.graph \
--ms 20 --top 20
Inspect a particular frame after obtaining its ID from summary or spikes:
python3 tools/flame_dump.py frame build/instrument/flame.graph \
--frame-id 1842 --top 24
Accept the original odd-style single-dash spellings when reproducing an old
command, for example -path, -top, -p, -ms, and -frame_id.
Start with summary to establish whether the problem is broad frame cost or a
small number of spikes. Compare avg frame with worst frames; a large gap
indicates bursty work. Compare cpu, wait, gpu, and other before changing
code. Use top scopes for repeated cost and frame for one bad frame's nested
scope breakdown.
Treat other_ms as unbucketed frame time. It is not an additional measured
function; it is the remainder after the exported scope buckets. Treat GPU time
as unavailable unless gpu_valid is true.
Use flame.frames.ndjson when source-location nesting is needed. Use
flame.scopes.ndjson for fast aggregate reports. Use flame.folded when opening
the data in a standard folded-stack viewer.
Check the following in order:
make instrument completed and that the instrument binary is
build/instrument/adriatic.build/instrument and check for flame.scopes.ndjson.tools/flame_dump.py if the shell is not at
the repository root.Do not analyze flame.graph.json as the primary artifact. It is a legacy debug
export retained for the in-app developer button; the automatic instrument export
is the catermujo-compatible NDJSON plus folded set.
tools/flame_dump.py - Local analyzer implementing summary, spikes,
and frame commands against the source-of-truth sidecars.