AI Code Assistant 使用报告AI Code Assistant Usage Dashboard

2026-05-14 03:34:34 | 1568 sessions | 89 projects | claudecode V3.1

富分析报告 / Rich Usage Report

中文优先的 claudecode 使用诊断:把成本、项目、工具治理与 vibe coding 效率串成可打印的行动报告。

总 TokensTotal Tokens
6.50B
1568 个会话 / 1239 个 subagent1568 sessions / 1239 subagents
估算成本Estimated Cost
$18,028.81
账单成本 $6,046.47billed $6,046.47
缓存命中率Cache Hit Rate
33.9%
缓存读取 2.19Bcache read 2.19B
Subagent 占比Subagent Share
23.9%
1.55B tokens1.55B tokens
CB:OutputCB:Output
194.9:1
评估:acceptableassessment: acceptable
附件事件Attachment Events
5575
Web search 0 / fetch 0web search 0 / fetch 0

1. 执行摘要1. Executive Diagnosis

本报告将当前 claudecode 使用从“花了多少 token”提升为“哪些工作方式最值得优化”。核心信号包括总消耗、估算与账单成本差距、缓存复用、subagent 占比、附件事件,以及项目集中度。This report reframes claudecode usage from raw token spend into work-pattern optimization. Core signals include total usage, estimated versus billed cost, cache reuse, subagent share, attachment events, and project concentration.

总 TokensTotal Tokens
6.50B
1568 个会话 / 1239 个 subagent1568 sessions / 1239 subagents
估算成本Estimated Cost
$18,028.81
账单成本 $6,046.47billed $6,046.47
缓存命中率Cache Hit Rate
33.9%
缓存读取 2.19Bcache read 2.19B
Subagent 占比Subagent Share
23.9%
1.55B tokens1.55B tokens
CB:OutputCB:Output
194.9:1
评估:acceptableassessment: acceptable
附件事件Attachment Events
5575
Web search 0 / fetch 0web search 0 / fetch 0

2. 效率机会2. Efficiency Opportunities

Vibe coding 效率的关键不是单纯降低 token,而是降低无效上下文构建、重复熟悉、压缩损失和等待中断。重点关注 CB:Output、forget curve、hourly productivity 与 compaction waste。The goal is not simply fewer tokens; it is less wasteful context building, less repeated re-familiarization, lower compaction loss, and fewer interruptions. Focus on CB:Output, forget curve, hourly productivity, and compaction waste.

CB:Output
194.9:1
assessment acceptable
Refam Tokens
1.55B
67 events
Subagent ROI
moderate
1.55B tokens
Avg Efficiency
5.2/10
median 5.1

3. 项目使用策略3. Project Portfolio Strategy

项目层面应优先治理 Pareto 头部项目:这些项目决定大部分成本,也最适合沉淀固定上下文、命令、测试入口和任务拆分规范。生命周期与连续性可以帮助判断项目是持续投入、短期爆发还是反复重启。At project level, prioritize the Pareto head: these projects drive most cost and benefit most from stable context, command lists, test entry points, and task-splitting rules. Lifecycle and continuity show whether projects are sustained, bursty, or repeatedly restarted.

Top Projects / 重点项目

ProjectSessionsTokensCostLast Used
101code2021.31B$5,200.79
token101-v2.1365545.01M$3,982.22
ameureka-media-agent1062.01B$2,059.18
ameureka-opc10797.16M$1,089.77
waza-x-rebuild43131.34M$820.42
_requierement22104.08M$781.28
open-waza58600.81M$536.24
box-with3188.67M$433.90
guochunlin39401.14M$414.69
mempalace1157.35M$339.53
deepseek-huawei-v462139.97M$334.95
waza-x2163.29M$306.99

4. 工具与治理4. Tools & Governance

工具治理关注权限模式、分支、native/internal 来源、API 错误与数据质量。目标是减少被权限、错误分支、网络/API 重试和不稳定环境打断的次数。Tool governance covers permission mode, branch, native/internal source, API errors, and data quality. The goal is fewer interruptions from permissions, wrong branches, API retries, and unstable environments.

Compactions
351
139 sessions
Avg Max Fill
22.2%
context saturation
Tokens Lost
66.80M
estimated compaction loss
Spawn Parents
117
1239 children
API Errors
864
retry attempts 864
Attachment Types
22
159 sessions

5. 下一步行动清单5. Action Plan

下一次 session 立即执行Do this in the next session

  • 下次开始大型任务前,先让 claudecode 明确目标、验收标准和不做事项,减少重复熟悉上下文。Before the next large task, state the goal, acceptance checks, and non-goals to reduce re-familiarization.
  • 把高 token 项目拆成“探索、实现、验证”三个阶段,探索阶段优先用 subagent,主会话保留决策与执行。Split high-token projects into exploration, implementation, and validation; use subagents for exploration and keep decisions in the main session.
  • 当 CB:Output 偏高时,先要求产出短计划和文件定位,再允许大范围读取,避免把上下文预算消耗在无效扫描上。When CB:Output is high, ask for a short plan and file targets before broad reads to avoid spending context on low-yield scanning.
  • 每周复盘 Top 项目、Top 会话、压缩浪费和审批模式,把高成本但低产出的工作方式改成模板化流程。Review top projects, top sessions, compaction waste, and approval modes weekly, then turn high-cost/low-output patterns into reusable workflows.

每周复盘Weekly review

  • 检查 Project Efficiency 表中的高成本项目,判断是否需要项目级 README、任务模板或常用命令清单。Check high-cost projects in Project Efficiency and decide whether they need a project README, task template, or command checklist.
  • 关注 Forget Curve:反复重新熟悉的项目应补充稳定上下文文件,而不是每次重新让模型扫描。Watch the Forget Curve: projects with repeated re-familiarization should get stable context files instead of repeated scans.
  • 审查 Approval Mode 和 Git Branch 分布,减少在错误分支或过严权限模式下反复中断的工作。Review approval mode and git branch distribution to reduce interruptions from wrong branches or overly strict modes.
  • 对 subagent ROI 做抽样复盘:保留高价值探索,避免把可直接完成的小任务委派出去。Sample subagent ROI: keep high-value exploration and avoid delegating small tasks that can be completed directly.

Executive Overview

Top-line usage, cost, cache, subagents, and daily trend.

Total Tokens
6.50B
1568 sessions / 1239 subagents
Est. Cost
$18,028.81
billed $6,046.47
Cache Hit Rate
33.9%
cache read 2.19B
Projects
89
models 14
Subagent Share
23.9%
1.55B tokens
Attachments
5575
web search 0 / fetch 0
Platform
AI Code Assistant (Native + Internal)
Period
2026-02-27
Total Sessions
1.6K
Total Tokens
6.50B
Total Est Cost
$18,028.81
Billed Cost
$6,046.47
Top Project
101code (28.8%)
Top Model
opus (70.4%)
Native Vs Internal
936/632
Cache Hit Rate
33.9%
Output Input Ratio
0.0051
Anomaly Count
13
Subagent Sessions
1.2K
Attachment Events
5.6K
Json Dimensions
40
Dashboard Pages
7
Forecast Method
linear_regression
53 data points
Next 7d
1.58B
last 7d 523.17M
Next 30d
7.97B
slope/day 3.45M
Weekly Growth
-75.2%
prev 7d 2.11B

Project Token Distribution (Treemap)

Token Flow (Sankey)

Daily Token Trend

Cost Scenarios

ScenarioCostCacheAssumption
conservative$18,028.81observedUse observed cache_read and list pricing
moderate$9,915.8560% effective cacheAssume subscription and higher prompt-cache utilization reduce input cost
high_cache$6,310.0885% effective cacheAssume long sessions with high reusable context cache

Time Patterns & Risk

When usage happens, where spikes occur, and what the forecast says.

Forecast Method
linear_regression
53 data points
Next 7d
1.58B
last 7d 523.17M
Next 30d
7.97B
slope/day 3.45M
Weekly Growth
-75.2%
prev 7d 2.11B

Weekday x Hour Heatmap

Monthly Trend

Anomaly Detection (Scatter)

High Cost Sessions

SessionProjectCostTokensSigma
8485b7d0-b663-4503-835c-32ef3bbf7b4c101code$1,527.42188.10M21.9
030347fa-cba2-4b68-9a6f-9a473aaa6dd5101code$1,032.11106.23M14.8
f8ec3229-3959-4a56-a5c8-9f5391ca104d101code$768.49110.24M10.9
agent-acompact-b8419e1a2423f8d6token101-v2.1$655.9579.54M9.3
ffa9ba93-9f12-47cf-9dcf-7aaa7f74e749101code$633.7278.61M9.0
0e33b65f-52e8-4013-8b08-8d7cc536191aameureka-media-agent$418.65324.06M5.9
abb41946-ad03-4154-9030-85006bf130e0ameureka-media-agent$341.45436.82M4.7
c0dd786a-4e1f-481b-a6fe-417bc934109awaza-x$305.68162.00M4.2
583a8ae8-22ed-461c-ac91-ffba1d0151c1ameureka-media-agent$276.69274.60M3.8
e6c13430-3875-4385-8fc2-6b4dbcd79aaaguochunlin$276.36275.07M3.8
57a407d4-7980-47f0-880c-2dacad840b2fmempalace$268.1235.29M3.7
e356f207-7833-4f6d-8d58-7389c9209903box-with$264.65128.05M3.6
agent-a8bf9facabdff3942token101-v2.1$204.4215.17M2.8

Projects, Sessions & Depth

Project concentration, lifecycle, task mix, session depth, and continuity.

Pareto Analysis by Project

Task Classification

Project Lifecycle (Gantt)

Session Depth Distribution

Project Efficiency

ProjectSessionsTokensTok/hCostModel
101code2021.31B3.11M$5,200.79GLM-5.1-Tencent
token101-v2.1365545.01M21.09M$3,982.22Opus (4.6)
ameureka-media-agent1062.01B4.04M$2,059.18GLM-5.1-Tencent
ameureka-opc10797.16M7.93M$1,089.77Opus (MaaS)
waza-x-rebuild43131.34M29.47M$820.42Opus (4.6)
_requierement22104.08M27.82M$781.28Opus (4.6)
open-waza58600.81M6.02M$536.24GLM-5.1-Tencent
box-with3188.67M34.66M$433.90Opus (4.6)
guochunlin39401.14M9.90M$414.69GLM-5.1-Tencent
mempalace1157.35M6.17M$339.53Haiku (4.5)
deepseek-huawei-v462139.97M22.35M$334.95Haiku (4.5)
waza-x2163.29M10.20M$306.99Opus (4.6)
chatgpt-business4820.76M7.71M$251.51Opus (4.6)
v2.04327.31M23.50M$218.87Haiku (4.5)
craft-agent-zero-trust-v3.2.0-main17130.83M5.81M$130.16GLM-5.1-Tencent
youtube-media-Waza16170.74M4.57M$121.93Kimi-K2.5
docs1217.73M39.51M$119.43Opus (4.6)
~ (home)21114.06M6.26M$117.57GLM-5.1-Tencent
frontend37.43M31.60M$113.27Opus (4.6)
auto-register-main2012.98M9.45M$94.49glm-5

Top Sessions

SessionProjectTokensDateModelSubagent
8485b7d0-b663-4503-835c-32ef3bbf7b4c101code188.10M2026-04-12 10:48Opus (4.6)False
030347fa-cba2-4b68-9a6f-9a473aaa6dd5101code106.23M2026-04-17 02:42Opus (4.6)False
f8ec3229-3959-4a56-a5c8-9f5391ca104d101code110.24M2026-04-17 08:18Opus (4.6)False
agent-acompact-b8419e1a2423f8d6token101-v2.179.54M2026-04-17 08:18Opus (4.6)True
ffa9ba93-9f12-47cf-9dcf-7aaa7f74e749101code78.61M2026-04-12 13:17Opus (4.6)False
0e33b65f-52e8-4013-8b08-8d7cc536191aameureka-media-agent324.06M2026-05-03 16:05Opus (4.6)False
abb41946-ad03-4154-9030-85006bf130e0ameureka-media-agent436.82M2026-05-01 17:59Sonnet (4.6)False
c0dd786a-4e1f-481b-a6fe-417bc934109awaza-x162.00M2026-05-01 02:48Opus (4.6)False
583a8ae8-22ed-461c-ac91-ffba1d0151c1ameureka-media-agent274.60M2026-04-25 06:16GLM-5.1-TencentFalse
e6c13430-3875-4385-8fc2-6b4dbcd79aaaguochunlin275.07M2026-05-01 01:56GLM-5.1-TencentFalse
57a407d4-7980-47f0-880c-2dacad840b2fmempalace35.29M2026-04-16 18:31Opus (4.6)False
e356f207-7833-4f6d-8d58-7389c9209903box-with128.05M2026-04-21 16:41Opus (4.6)False
agent-a8bf9facabdff3942token101-v2.115.17M2026-04-13 11:33Opus (4.6)True
637a7f72-d506-4db7-ae53-290383ad5ecdopen-waza279.38M2026-04-27 16:36GLM-5V-TurboFalse
96c4cd83-a8b9-44aa-9fa8-5bee67daf344box-with56.48M2026-04-21 14:01Opus (4.6)False
db5eaf10-f55f-49da-a772-5733374554d1ameureka-media-agent58.04M2026-05-10 04:18Opus (4.6)False
agent-af9d94e7cf6b29cb3ameureka-opc8.74M2026-03-04 01:23Opus (MaaS)True
6c376494-f60a-4bfb-89a2-74c87b2bc8ffameureka-media-agent134.35M2026-04-14 05:54GLM-5.1-TencentFalse
85def55a-4266-4c0d-b889-833623a7b6cbopen-waza88.74M2026-05-05 11:31Opus (4.6)False
agent-a5125331c7f28eb13waza-x-rebuild17.31M2026-05-01 17:14Opus (4.6)True

Continuity by Project

ProjectSessionsAvg Gap hMedian hMin hMax h
token101-v2.13654.60.00.0690.7
101code2028.40.10.0581.8
ameureka-opc1071.60.00.068.4
ameureka-media-agent1066.10.10.096.4
deepseek-huawei-v4621.30.00.036.3
open-waza5810.80.10.0149.3
cc-audit-workspace550.00.00.00.1
chatgpt-business480.40.00.013.2
waza-x-rebuild430.40.00.06.0
v2.0434.20.10.0134.8
guochunlin3912.20.00.0227.5
_requierement220.10.00.00.2
remotion2216.80.10.0107.9
~ (home)2189.90.20.0914.4
auto-register-main200.30.10.01.3
craft-agent-zero-trust-v3.2.0-main170.90.10.05.8
youtube-media-Waza169.90.00.064.8
008-博客优化150.10.00.01.1
2026-04-28-short-happyhorse-1-0147.10.00.084.5
sany-endpoint-security1431.80.00.0356.0

Session Depth Statistics

MinP25MedianP75P90P95MaxMean
4220.5K754.3K2.10M5.02M14.03M436.82M5.29M

Productivity & Vibe Coding

Context-build efficiency, memory half-life, and session productivity.

CB:Output
194.9:1
assessment acceptable
Refam Tokens
1.55B
67 events
Subagent ROI
moderate
1.55B tokens
Avg Efficiency
5.2/10
median 5.1

Context Build vs Output Ratio

Project Forget Curve

Session Efficiency Map

Efficiency by Duration

Hourly Productivity Score

Vibe Coding Optimization Recommendations

INFO — Diagnostic
Reduce context re-build waste
→ Update memory/CLAUDE.md at end of each session; use /compact before long gaps; start sessions with 'recall context' instead of re-reading files
~$2,324/mo (assuming 10% cache)
INFO — Diagnostic
Split long sessions into focused 1-3h blocks
→ Use /compact proactively; end sessions at natural breakpoints; start fresh sessions with memory context
Est. 10-20% of long-session tokens
INFO — Diagnostic
Reduce compaction waste
→ Use more focused sessions; /compact manually before auto-compact kicks in; keep essential context in memory files
Est. 10-15% of input tokens in affected sessions
INFO — Diagnostic
Investigate API error patterns
→ Check network stability; review if specific models/sessions have higher error rates
Minimal direct cost, but improves experience

Context Saturation & Agents

Compaction pressure, subagent spawning, token waste, attachments, and API errors.

Compactions
351
139 sessions
Avg Max Fill
22.2%
context saturation
Tokens Lost
66.80M
estimated compaction loss
Spawn Parents
117
1239 children
API Errors
864
retry attempts 864
Attachment Types
22
159 sessions

Subagent Spawn Graph (Sankey)

Context Saturation (Compaction)

Compaction vs Waste

Top Compaction Waste Sessions

SessionProjectCompactionsTokensMax Fill %Waste/Compaction
70fbd502-3ba2-4a0b-818c-6b3957b93778ameureka-media-agent119.41M17.0388.2K
972d0bbc-fe49-4770-a452-39736f29e482ameureka-media-agent123.47M16.7469.3K
17ec0c7c-f710-4e83-a169-a79372d785e2ameureka-media-agent138.83M16.9776.5K
40add9ed-ac1d-46aa-b4c6-8e377d837b65ameureka-media-agent3101.77M18.2678.4K
8f38fc82-1e7f-4bcd-aa20-94030e3b4eb2ameureka-media-agent268.35M16.7683.5K
6c376494-f60a-4bfb-89a2-74c87b2bc8ffameureka-media-agent6134.35M18.3447.8K
c32a8290-0144-4831-8923-55f88b6a45cdameureka-media-agent373.13M16.8487.6K
583a8ae8-22ed-461c-ac91-ffba1d0151c1ameureka-media-agent9274.60M17.0610.2K
d1acbc41-2c95-48a8-a1cc-0a117b683340ameureka-media-agent349.09M18.6327.3K
b10673fd-d535-4b6f-84bd-e987e12d4cd4ameureka-media-agent128.55M16.8570.9K
6247bbe9-018a-4f74-877d-6c0d54392329ameureka-media-agent466.16M18.6330.8K
e64a10f6-d3cc-4afb-95a4-33318102d2d8ameureka-media-agent142.95M17.3858.9K
4e6a29df-2f12-429a-b10b-1c2e880956bbameureka-media-agent391.61M16.8610.7K
3936696d-90e3-4e22-8f97-e8d153012caaopen-waza343.76M16.7291.7K
637a7f72-d506-4db7-ae53-290383ad5ecdopen-waza13279.38M19.3429.8K
a8322e5a-294e-45ce-b697-b86502a03d8copen-waza125.35M17.3507.1K
42be1e42-a743-4ae4-bad2-126ca4244cee101code121.77M17.2435.5K
bc7d6fc8-b8e5-45df-a20f-dc1841deda25101code136.55M16.9731.1K
1850eccd-c81c-4d87-91ea-eb401a7bbf04101code294.00M16.8940.0K
63bdc6a1-b7f7-45a3-a0b0-e939d55e122b101code133.57M20.7671.4K

Attachment Types

TypeCountPct
task_reminder256045.9
file104218.7
nested_memory5189.3
skill_listing3566.4
hook_success1873.4
queued_command1512.7
compact_file_reference1252.2
edited_text_file1122.0
mcp_instructions_delta1031.8
invoked_skills821.5
command_permissions671.2
date_change671.2
plan_mode_exit631.1
plan_mode531.0
plan_file_reference430.8
hook_additional_context200.4
hook_blocking_error130.2
ultrathink_effort40.1
plan_mode_reentry40.1
auto_mode20.0

Environment, Source & Governance

Source split, approval policies, branches, CLI versions, and data quality.

Source Comparison

Permission Mode Impact

Git Branch Token Distribution

Approval Mode Impact

Source Comparison Detail

SourceSessionsTokensCostBilledTools
native9363.14B$4,845.15$2,280.6430946
internal6323.36B$13,183.66$3,765.8330118

Version History

VersionSessionsInputCostPeriod
2.1.71266188.38M$1,033.482026-03-07~
2.1.1122362.05B$4,167.232026-04-16~
2.1.119810.66M$10.632026-05-01~
2.1.6313198.60M$1,175.342026-02-28~
2.1.78125633.57M$4,941.302026-04-12~
2.1.119109619.23M$490.592026-04-24~
2.1.668666.08M$682.612026-03-04~
2.1.9253445.71M$2,978.462026-04-07~
2.1.10538339.55M$334.782026-04-14~
2.1.13935194.99M$319.182026-05-11~
2.1.11434256.24M$259.172026-04-18~
2.1.11734218.92M$171.152026-04-22~
2.1.12126274.19M$276.362026-05-01~
2.1.1162396.96M$85.322026-04-21~
2.1.10922280.62M$208.152026-04-15~
2.1.1322134.09M$34.002026-05-06~
2.1.1262176.13M$77.012026-05-01~
2.1.1042095.24M$102.432026-04-12~
2.1.701716.48M$154.722026-03-06~
2.1.1381736.30M$79.742026-05-09~
2.1.1401668.35M$70.352026-05-13~
2.1.6996.11M$11.642026-03-05~
2.1.1208273.70M$276.692026-04-25~
2.1.1317563.8K$0.562026-05-06~
2.1.894407.2K$0.792026-04-01~
2.1.110412.80M$12.862026-04-16~
2.1.623780.0K$3.602026-02-27~
2.1.133233.76M$33.592026-05-09~
2.1.11821.55M$1.572026-04-23~
2.1.128138.70M$35.512026-05-05~

Approval Mode Detail

ModeSessionsTokensAvg/SessionAvg MinTok/h
unknown14411.48B1.03M4.114.88M
bypassPermissions1195.02B42.19M578.54.38M
default82.86M357.3K95.1225.5K

Quality

MetricValue
data_completeness{"sessions_with_tokens_pct": 78.4, "sessions_with_duration_pct": 98.1, "sessions_with_project_pct": 100.0, "sessions_with_model_pct": 97.7}
subagent_coverage{"subagent_sessions": 1239, "subagent_pct": 79.0, "subagent_tokens_pct": 23.9, "parents_detected": 117}
confidence_notes["JSONL token usage is authoritative for observed local events but may undercount provider-side cache discounts.", "Subagent parent detection is based on directory structure and may be approximate for direct project-level subagent folders.", "Cost scenarios are analytical estimates, not invoices."]

Diagnostics & Recommendations

Prioritized findings and action items.

Diagnostic Insights

CRITICAL — Cost
Re-familiarization waste: 1549M tokens from sessions starting after 24h+ gaps. These sessions re-read context that was previously built.
CRITICAL — Cost
57 sessions exceeded 4 hours. Long sessions have diminishing returns — compaction events lose context and require re-reading.
CRITICAL — Efficiency
351 context compaction events across 139 sessions. Each compaction loses ~60% of prior context, requiring re-reading (42M tokens estimated lost).
MEDIUM — Efficiency
864 API errors encountered. These cause retries and wasted context window usage.

Action Items

PriorityTitleImpactEffortSavingAction
P1Reduce context re-build wastehighmedium~$2,324/mo (assuming 10% cache)Update memory/CLAUDE.md at end of each session; use /compact before long gaps; start sessions with 'recall context' instead of re-reading files
P2Split long sessions into focused 1-3h blocksmediummediumEst. 10-20% of long-session tokensUse /compact proactively; end sessions at natural breakpoints; start fresh sessions with memory context
P2Reduce compaction wastemediummediumEst. 10-15% of input tokens in affected sessionsUse more focused sessions; /compact manually before auto-compact kicks in; keep essential context in memory files
P3Investigate API error patternsmediummediumMinimal direct cost, but improves experienceCheck network stability; review if specific models/sessions have higher error rates