Agent Skills › DotHarness/dotcraft › trace-review

trace-review

GitHub

用于审查 DotCraft 会话轨迹证据，评估可靠性、延迟、工具行为、Token 效率及 Prompt 缓存问题。通过解析事件文件生成包含确切事件 ID 的发现报告。

src/DotCraft.TraceViewer/Skills/BuiltIn/trace-review/SKILL.md DotHarness/dotcraft

Trigger Scenarios

需要审查不可变的会话轨迹数据分析 AI 交互的性能与稳定性问题检查 Token 使用效率或缓存命中率

Install

npx skills add DotHarness/dotcraft --skill trace-review -g -y

More Options

Non-standard path

npx skills add https://github.com/DotHarness/dotcraft/tree/main/src/DotCraft.TraceViewer/Skills/BuiltIn/trace-review -g -y

Use without installing

npx skills use DotHarness/dotcraft@trace-review

指定 Agent (Claude Code)

npx skills add DotHarness/dotcraft --skill trace-review -a claude-code -g -y

安装 repo 全部 skill

npx skills add DotHarness/dotcraft --all -g -y

预览 repo 内 skill

npx skills add DotHarness/dotcraft --list

SKILL.md

Frontmatter

{
    "name": "trace-review",
    "tools": "ReadFile, FindFiles, GrepFiles, SubmitTraceReview",
    "description": "Review one immutable DotCraft Session Trace for reliability, latency, tool behavior, token efficiency, and prompt-cache issues using an Evidence Bundle."
}

DotCraft Trace Review

Review the immutable Trace evidence in the current workspace. Stay within the recorded Session and submit only conclusions supported by exact Event ids.

Evidence layout

Start with manifest.json to confirm the Session, revision, and Event count.
Search events/index.jsonl to locate Event types, timing, model calls, tools, and correlation ids.
Each index entry points to events/NNNNNN/event.json for the complete structured Event fields.
Large fields are stored beside event.json as bounded content, tool-argument, tool-result, metadata, or final-system-prompt files. The fieldFiles object lists their order and total character count.
Read every chunk required for a conclusion. Do not infer omitted content from one chunk.

Review workflow

Confirm the manifest and scan the complete index before forming conclusions.
Review Error, ProviderError, abnormal terminal events, retries, and incomplete Turns for Reliability.
Compare Turn duration, provider attempts, tool duration, and recorded inactive intervals for Latency. Do not invent duration for point Events.
Correlate tool start and completion by callId. Check failures, repeated arguments, unchanged retries, and unusual call density for Tool behavior.
Compare input, fresh, cached, cache-write, output, and reasoning tokens across calls for Token efficiency.
Inspect cache diagnostics, prompt hashes, tool-schema hashes, and changed fields for Prompt cache behavior.
Open the exact Event details and large fields needed to verify each candidate Finding.
Finish by calling SubmitTraceReview with a concise summary and the validated Findings.

Finding rules

Use Major for an explicit failure or a high-confidence problem with material impact.
Use Minor for a localized problem or operational risk with observable impact.
Use Suggestion for a supported optimization opportunity that did not cause a current failure.
Use Confirmed when the cited Events directly establish the claim. Use Inferred when the claim is the strongest explanation but not directly recorded.
Use only these dimensions: Reliability, Latency, Tool behavior, Token efficiency, and Prompt cache.
Every Finding must cite at least one exact Event id. A range must use chronological start and end Event ids from this revision. Omit endEventId for a single-Event reference.
Do not turn missing Trace data, incomplete pairing, or unsupported assumptions into Findings.
Do not judge whether the original user task was completed correctly.
Do not claim access to source files, operational logs, configuration files, or the target workspace.