返回 Skill 列表
extension
分类: 开发与工程无需 API Key

perf-optimizer

针对OneKey移动应用(iOS)的系统性能优化和回归调试。使用场景:(1) 修复性能回退 - 当指标如tokensStartMs、tokensSpanMs或functionCallCount出现回退,需要将其恢复到正常水平时;(2) 提升基准性能 - 当需要优化冷启动时间或减少函数调用开销时;(3) 用户请求对应用启动或首页刷新流程进行性能优化/改进/调试。

person作者: jakexiaohubgithub

Performance Optimizer

Systematic workflow for diagnosing and fixing performance issues in the OneKey mobile app using the perf-ci infrastructure and performance-server tooling.

Overview

This skill provides a structured iterative approach to:

  • Establish performance baselines from existing sessions
  • Run controlled perf measurements (3 runs, median aggregation)
  • Analyze session data to identify bottlenecks
  • Make targeted code changes
  • Verify improvements against thresholds
  • Document all changes and results

Key Metrics:

  • tokensStartMs: Time when Home tokens refresh starts (lower is better)
  • tokensSpanMs: Duration of Home tokens refresh (lower is better)
  • functionCallCount: Total function calls during session (lower is better)

Success Criteria:

  • SUCCESS: Time metrics improve by ≥10%
  • 🌟 MINOR_IMPROVEMENT: Time unchanged but function calls reduce by ≥20% (safe, small-scope changes)
  • NO_IMPROVEMENT: Neither threshold met → revert changes

Workflow

Phase 1: Setup and Baseline

Step 1.1: Select Baseline Session

Ask user to choose a baseline session or help them select one:

# List recent sessions with key metrics
cat ~/perf-sessions/sessions.overview.jsonl | \
  jq -r '[.sessionId, .createdAt, .marks["Home:refresh:done:tokens"]] | @tsv' | \
  tail -20

User can specify:

  • A known good session (for regression fixes)
  • A recent session (for improvement work)
  • Let you choose a representative session

Step 1.2: Analyze Baseline

Extract baseline metrics from the session:

# Get detailed analysis
node development/performance-server/cli/derive-session.js <baseline-sessionId> \
  --pretty \
  --output /tmp/perf-baseline-derived.json

Read baseline metrics from ~/perf-sessions/<sessionId>/mark.log:

# Extract tokensStartMs (timestamp of Home:refresh:start:tokens)
grep "Home:refresh:start:tokens" ~/perf-sessions/<sessionId>/mark.log | jq '.timestamp'

# Extract tokensSpanMs (done - start)
grep "Home:refresh:done:tokens" ~/perf-sessions/<sessionId>/mark.log | jq '.timestamp'

# Count function calls
wc -l < ~/perf-sessions/<sessionId>/function_call.log

Create baseline metrics JSON for comparison:

echo '{"tokensStartMs": <start>, "tokensSpanMs": <span>, "functionCallCount": <count>}' > /tmp/baseline-metrics.json

Step 1.3: Initialize Documentation

Create session document at development/output/perf-optimization-<timestamp>.md using the template from references/template.md. Fill in:

  • Current date/time
  • Baseline session ID
  • Baseline metrics
  • Current branch name
  • Target (regression fix or improvement)

Phase 2: Iterative Optimization Loop

Maximum iterations: 10

For each iteration (run in a sub-agent):

Step 2.1: Run Performance Tests

The perf script automatically runs 3 times and aggregates results:

node development/perf-ci/run-ios-perf-detox-release.js

Output location: development/perf-ci/output/<jobId>/

  • report.json - Contains aggregated results in agg field
  • detox/runs.json - Contains individual run sessionIds

Extract current metrics from report.json:

# Read aggregated metrics directly
cat development/perf-ci/output/<jobId>/report.json | jq '{
  tokensStartMs: .agg.tokensStartMs,
  tokensSpanMs: .agg.tokensSpanMs,
  functionCallCount: .agg.functionCallCount
}' > /tmp/current-metrics.json

Step 2.2: Analyze Current Performance

For deeper analysis, run derive-session on individual sessions:

# Get sessionIds from the run
SESSIONS=$(cat development/perf-ci/output/<jobId>/detox/runs.json | jq -r '.runs[].sessionId')

# Analyze each session
for sid in $SESSIONS; do
  node development/performance-server/cli/derive-session.js $sid \
    --pretty \
    --output /tmp/perf-derived-$sid.json
done

Focus on these sections in the derived output:

  • slowFunctions: Functions taking the most cumulative time
  • homeRefreshTokens: What's consuming time in the critical refresh window
  • jsblock: Main thread blocks causing delays
  • repeatedCalls: Thrashing patterns or excessive re-renders
  • keyMarks: Critical milestone timing

Identify top 1-3 bottlenecks that are:

  • Taking significant time
  • Potentially optimizable
  • Within the critical path (Home refresh flow)

Step 2.3: Determine Action

Compare current metrics to baseline:

# Quick comparison
cat /tmp/baseline-metrics.json
cat /tmp/current-metrics.json

# Calculate deltas manually or use script in skill directory

Decision tree:

If current metrics show improvement over baseline:

  • SUCCESS (≥10% time improvement) → STOP, document success
  • 🌟 MINOR_IMPROVEMENT (≥20% function call reduction, time stable) → Create branch, commit, return to main branch, continue

If no improvement yet:

  • Continue to Step 2.4 (make changes)

If iteration count reaches 10:

  • Document findings and stop

Step 2.4: Make Code Changes

Based on analysis, make ONE targeted change per iteration:

Change types:

  1. Optimization: Remove redundant work, cache results, reduce allocations
  2. Add perfMark: Add marks to understand unclear bottlenecks better
  3. Both: Add marks + optimize in same area

Guidelines:

  • One change at a time (unless analysis proves multiple changes must work together)
  • Small, focused changes
  • Safe changes only (never break functionality)
  • Document rationale clearly

Adding perfMarks:

Use the performance utilities in packages/shared/src/performance/:

import { perfMark } from '@onekeyhq/shared/src/performance/perfMark';

// Add mark at a specific point
perfMark('MyComponent:operation:start');
// ... operation ...
perfMark('MyComponent:operation:done');

Naming convention: <Component>:<action>:<phase> (e.g., Home:refresh:start:tokens)

If adding perfMarks for investigation:

  1. Add marks around suspected bottleneck
  2. Run one perf cycle with marks
  3. Analyze new data with marks visible
  4. Then make code optimization
  5. Verify with another perf cycle

Step 2.5: Document Iteration

Update the session document with:

  • Analysis: Job ID, session IDs, median metrics, key findings from derive-session
  • Code Changes: File, location, change type, description, rationale
  • Verification Results: New job ID, metrics, deltas vs previous/baseline, verdict, action taken

Phase 3: Finalization

Step 3.1: Handle MINOR_IMPROVEMENT Branch

If any iterations resulted in MINOR_IMPROVEMENT:

git checkout -b perf/minor-<description>
git add <changed-files>
git commit -m "perf: <description>

Reduces function call count by X% while maintaining time metrics.

Reason: <brief explanation>
"
git checkout <original-branch>
git restore .

Document the branch name in the session document.

Step 3.2: Complete Documentation

Fill in the Summary section:

  • Total iterations run
  • Final result (SUCCESS with % improvement, or still investigating)
  • List all effective changes
  • List all ineffective changes with reasons
  • List any branches created
  • Next steps if incomplete

Key Files and Paths

Perf Infrastructure:

  • development/perf-ci/run-ios-perf-detox-release.js - Main perf runner
  • development/perf-ci/output/<jobId>/ - Job output directory
  • development/performance-server/cli/derive-session.js - Session analyzer
  • ~/perf-sessions/ - Session data storage (default)
  • ~/perf-sessions/sessions.overview.jsonl - Session index

Thresholds:

  • development/perf-ci/thresholds/ios.release.json - Release mode thresholds

Performance Utilities:

  • packages/shared/src/performance/perfMark.ts - Performance marking utility

References

  • references/template.md: Session documentation template
  • references/perf_tool_guide.md: Detailed guide to derive-session and analysis tools

Important Notes

  1. Run each optimization loop in a sub-agent to avoid context bloat
  2. Never commit changes unless SUCCESS or MINOR_IMPROVEMENT
  3. Always document failed attempts - helps avoid repeating ineffective changes
  4. Trust the data - if metrics don't improve, revert even if change "should" help
  5. Be patient - each perf run takes significant time (build + 3 runs); rushing leads to mistakes
  6. Focus on the critical path - Home screen tokens refresh is the key metric
  7. Watch for trade-offs - some optimizations might reduce one metric but increase another