跪拜 Guibai
← All articles
Android

SmartPerfetto 1.19.0 Pivots to Pi Agent Core and Releases Perfetto-Skills for Agent-Local Trace Analysis

By Gracker ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Unifying on Pi Agent Core across Web, CLI, and API removes the friction of coordinating multiple model runtimes, while the deleted claim-verification flow signals a shift toward treating AI output as forensic reporting rather than proof. The standalone Perfetto-Skills path makes the analysis methodology available inside existing agent environments without adopting the full product.

Summary

SmartPerfetto 1.19.0 makes Pi Agent Core the default analysis engine across Web, CLI, and API, replacing a multi-runtime setup and removing the model declaration and claim-verification process that previously tried to prove every conclusion. The release also upgrades to Perfetto v59.1, adding evidence for process lifecycle, JobScheduler, Bitmap live sets, Doze states, CPU PMU counters, and camera/rendering frame beats while keeping clear boundaries on what each observation can prove.

Development around TraceDemo produced concrete fixes: targeted regression tests for P1–P5 and F1–F6 issues improved tool invocation, SQL guidance, and error reporting, but the release notes are explicit that these were not full-batch reruns. Final verification claims 405 corpus expectations, 2013 SQL contract executions, and 6/6 scenario trace regressions passing, with honest caveats on what those numbers represent.

Separately, Perfetto-Skills ships as an independent Agent Skill for developers who prefer to stay in their existing Codex, Claude Code, or OpenCode environment. It packages analysis workflows, SQL, and evidence rules without requiring the SmartPerfetto backend, and is pinned to the 1.19.0 commit with its own test suite.

Takeaways
— Default analysis engine switched to Pi Agent Core in 1.18.0; 1.19.0 continues from that migration with additional fixes.
— The model declaration and claim-verification flow was removed in 1.18.0, replaced by deterministic forensic gathering via Skills/SQL.
— Perfetto upgraded to v59.1, adding process lifecycle and JobScheduler evidence.
— New analysis capabilities include Bitmap live set shapes, Doze light/deep idle labels, CPU PMU counter deltas, and camera frame-beat measurements.
— TraceDemo testing recorded 88 synthetic samples as complete in fixround88, 9/9 targeted deliveries for P1–P5, and 29/1 success/failure in unseen30 with 3 isolated replays.
— Final release verification: 405 corpus expectations, 2013 SQL contract executions, and 6/6 scenario trace regressions passed.
— Perfetto-Skills is a standalone agent skill for Codex, Claude Code, or OpenCode, pinned to commit 070053917 with 514 tests and 1 skipped.
— Install Perfetto-Skills with `npx [email protected] add Gracker/Perfetto-Skills --skill perfetto-performance-analysis -a codex -a claude-code -a opencode -y`.
— The release provides native packages for Windows x64, macOS arm64, and Linux x64; npm CLI requires Node.js 24.
— Old Providers follow migration rules; PMU, heap, or presentation data not recorded in past traces cannot be recovered by upgrading.
Conclusions

Dropping claim-verification is an honest acknowledgment that AI-generated narratives cannot guarantee causal proof; structuring output as forensic evidence with explicit limitations is more valuable than forcing every statement into a verifiable claim.

The release notes are unusually transparent about test limitations, explicitly stating that targeted regressions are not full-batch reruns and that completing a run does not equal diagnostic accuracy.

Splitting Perfetto-Skills from the full SmartPerfetto product creates a lower-friction adoption path for developers already invested in other agent environments, while keeping the methodology and SQL knowledge portable.

The migration to Pi Agent Core centralizes the model loop while leaving policy, tools, and identity in the product, which simplifies maintenance but also means all channels share the same failure modes and cost boundaries.

Concepts & terms
Pi Agent Core
The default model request and tool-loop engine in SmartPerfetto 1.19.0, replacing multiple model runtimes. It handles the iterative process of invoking tools and feeding results back to the model, while the product retains control over policies, tools, SQL, identity, and reporting.
Perfetto-Skills
A standalone Agent Skill package from Gracker that provides trace analysis workflows, SQL, and evidence rules for use inside existing agent environments like Codex, Claude Code, or OpenCode, without requiring the full SmartPerfetto backend or UI.
Claim-verification flow
A former SmartPerfetto process where the model declared conclusions and evidence relationships in a specific format, and the product parsed, repaired, and verified those claims. It was removed in 1.18.0 in favor of deterministic forensic gathering and explicit reporting of limitations.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗