Future SmartPerfetto updates will stay in this post, with new entries first. It combines the earlier update articles into a record of version ranges, dates, and important changes.
Coverage currently ends at v1.15.1 (2026-10-06). This is the last version covered here, rather than a claim about the latest release. Downloads are on GitHub Releases.
v1.12.0 → v1.15.1 | 2026-09-17 to 2026-10-06
This period changed source and knowledge selection, answer delivery, analysis metrics, and installation diagnostics.
- Select source and documents per turn: controls beside the composer select codebases and knowledge bases independently. On-demand source lookup can find declarations and trace sites at different reading depths. Reading budgets and permission to send content to a Provider remain separate.
- Document folders as knowledge bases: registration, rights confirmation, indexing, and provider-send consent are separate steps. v1.15 removed the bundled Android Internals Wiki Pack and
smp knowledge-pack; existing users need to register document directories they have permission to use. - Answer delivery before verification finishes: supported runtimes show temporary streaming drafts, then the answer and its verification status. Drafts are not saved as history. All five runtimes use shared finalization; reports and snapshots involving private source or knowledge are creator-only.
- Analysis corrections: scrolling considers presentation cadence alongside frame duration. Click response separates physical input, app delivery, and candidate frame associations. Startup, CPU, and click details use UPID to isolate process instances. Heap analysis adds class growth, reachable instances, and dominated size.
- CPU cluster load fixes: numerator and denominator use a consistent clock basis, correcting suspend-spanning intervals and CPU identities across machines. Unresolved data produces unavailable reasons and null metrics.
- Retired entry points: old SQL, template-analysis, and Wiki APIs and model-calling Skill steps were retired. The development-mode
/uploadsstatic route, which bypassed authentication and ownership checks, was removed. Check the versioned migration notes before upgrading integrations. - v1.15.1 installation and test fixes (10-06):
smp doctor --format jsonadds a working native SQLite probe. npm installation checks exercise a separate consumer project and real SQL. Affected tests use their own temporary databases and directories. - v1.12.1 (09-21): disk preflight checks, startup wakeup-chain and sleep attribution, hypothesis re-solving, and prompt budgets previously described as post-release main-branch work shipped in this version.
The historical DeepSeek long-report, OpenCode startup-body, and OpenAI scrolling claim checks still had failures. These features do not establish real-provider acceptance across every runtime.
Details: v1.15.1 Changelog.
v1.7.0 → v1.12.0 | 2026-08-21 to 2026-09-17
This period covers v1.8.0, v1.8.1, v1.8.2, v1.8.3, v1.8.4, v1.9.0, v1.10.0, v1.11.0, and v1.12.0.
| Version | Main changes |
|---|---|
| v1.8.x | Upload or replace either trace in an arbitrary two-trace workspace, swap baseline and comparison, and restore layout; reduce reliance on prebuilt source indexes; sync Perfetto v58.2 |
| v1.9.0 | Explicit Auto / Fast / Full modes with intent retained in follow-ups; source provenance in reports |
| v1.10.0 | Durable investigation checkpoints retain hypotheses, evidence, and pending work for later turns; authorized source folders remain usable on demand without an active index |
| v1.11.0 | Bind answer claims to retained tool evidence and run bounded semantic review; retain verification results in Web, CLI, HTML reports, and snapshots; remove Claude SDK’s implicit fallback to local Claude Code login |
| v1.12.0 | More specific claim-format, transport, and retry diagnostics; evidence-triggered scrolling backpressure investigation; system-analysis E2E scenario definitions and smp probe strategy diagnostics |
Checkpoints support later turns; they do not automatically resume background work. Fast can still call a model. Verification states such as partial and not_checked retain their meaning and do not count as complete acceptance.
Details: v1.12.0 Changelog.
v1.1.1 → v1.7.0 | 2026-07-17 to 2026-08-21
The focus was cross-platform delivery, large traces, source analysis, and controlled feedback workflows.
| Date (2026) | Version | Main changes |
|---|---|---|
| 07-18 | v1.2.0 | Signed Android Internals Knowledge Pack bundled with releases; this mechanism was later removed in v1.15 |
| 07-18 to 07-28 | v1.2.x | Chinese analysis UI, Android Skill safeguards, optional Qoder runtime, update notices, and stronger cross-platform release processes |
| 07-28 | v1.3.0 | Isolated Trace Processor RPC sessions for large-trace dual windows; native folder picker |
| 08-03 | v1.4.0 | Controlled Self-Evolution proposals, evaluation, and manual application; Agent-assisted feedback drafts; opt-in enterprise OIDC and workspace isolation |
| 08-10 | v1.5.0 | Code-Aware analysis, source-aware reports, Mermaid hardening, and Windows DPAPI fixes; date follows the Changelog |
| 08-14 | v1.6.0 | Explicit Pi Provider configuration, Windows data-directory policy, dual-trace identity, and final-report quality checks |
| 08-21 | v1.7.0 | Qoder BYOK routing, DeepSeek / Pi finalization fixes, and license, publication-consent, and privacy-review records for six real-trace fixtures |
The period also improved macOS signing and notarization, native Windows checks, archive checksums, and update notices. Update checks only notify. Self-Evolution application and rollback require explicit enablement and manual action; the feedback entry only produces drafts for review.
Details: v1.7.0 Changelog.
v1.0.28 → v1.1.1 | 2026-06-04 to 2026-07-17
This period added dual windows, deterministic evidence answers, and private analysis context while unifying shared behavior across four Agent runtimes. The start date follows the original article’s reporting cutoff; v1.0.28 was released on 06-02.
| Date (2026) | Version | Main changes |
|---|---|---|
| 06-05 | v1.0.29 | Stronger memory-analysis evidence and corrected thread-blocking reason semantics |
| 06-06 | v1.0.30 | Perfetto v56.0 sync |
| 06-06 | v1.0.31 | Real DeepSeek path for Agent E2E verification |
| 06-07 | v1.0.32 | Runtime engine reorganization, cancellation propagation, third-party runtime state continuation, and misdiagnosis patterns |
| 06-07 | v1.0.33 | Exclude runtime data from portable packages |
| 06-17 | v1.0.34 | DeepSeek analysis fixes and turn-budget adjustments |
| 06-19 | v1.0.35 | Always-loaded strategy Core with on-demand Detail, smaller tool descriptions, and summaries with paginated references for large SQL results |
| 06-22 | v1.0.36 | Case knowledge, Self-Evolution V1, and memory-layer fixes |
| 06-22 | v1.0.37 | macOS portable signing-resource fixes |
| 07-04 | v1.0.38 | Quick Mode evidence answers and receipts, Provider preset updates, and Perfetto v57.1 |
| 07-11 | v1.0.39 | Dual-trace workspace and eight AI workflows under RFC-0025 |
| 07-17 | v1.1.0 | Shared run specification and private-context authorization; Camera / Heap / GPU scenarios, Android 17 rendering knowledge, trace corpus, and public Skill export; Perfetto v57.2 |
| 07-17 | v1.1.1 | Fix Docker OpenCode runtime links and arm64 build paths |
Quick Mode uses zero model turns only for supported intents with sufficient evidence. The workspace then paired the current trace with a historical reference; arbitrary two-trace input arrived in v1.8.0. Camera, Heap, and GPU analysis still require the relevant captured data.
Details: v1.1.1 Changelog.
v1.0.7 → v1.0.28 | 2026-05-17 to 2026-06-04
v1.0.28 was released on 2026-06-02; the original update covered work through 06-04.
- Smart mode inventories detected scenes before deeper analysis. A scene preview does not establish a root cause.
- Selection questions and CLI carry timeline ranges or event context into analysis.
smp capture android, terminal analysis, comparisons, and session artifacts reuse the backend core. - More scenario evidence rules cover Power, ANR, Input, Display, IO, Network, and capture completeness. Failed SQL diagnostics are kept separate from successful query evidence.
- Four runtimes—Claude, OpenAI, Pi Agent Core, and OpenCode—share tools and public output rules, with further fixes for report completeness, provenance, and session recovery.
Details: v1.0.28 Changelog.
First update after open source | 2026-04-29 to 2026-05-17
The original title referred to v1.0.7 (released 2026-05-14). Its body described the 05-17 source snapshot, including later main-branch work, so the full list is not a promise about the v1.0.7 package.
- Analysis snapshots, similar-case retrieval, and cross-window result comparison make completed results reusable.
- Claude / OpenAI runtimes, Provider Manager, and runtime diagnostics become central parts of the workflow.
- SQL stdlib checks, evidence IDs, provenance, and sentence-level references help trace conclusions back to tables.
- Rendering-pipeline teaching links explanations to observed events and identifies missing signals.
- Docker, three-platform portable packages, source, and CLI distribution improve, alongside fixes for large uploads, startup compatibility, session isolation, and frontend prebuild assets.
Historical snapshot: 2467b95b.
Project links
Repository · Downloads · Issues · Introduction
This is a summary of historical changes, without rerunning device, model, or release acceptance. Use the selected version’s documentation for installation and migration. Source and background knowledge do not replace evidence from the current trace, and missing data is not zero.