feat: integrate Perfetto SQL stdlib modules (P0+P1) - #12
Closed
mufans wants to merge 98 commits into
Closed
Conversation
refactor: merge post-refactoring fixes and improvements into master
refactor: merge P0 fixes and architecture improvements
refactor: P0 quality fixes, collector enhancements, and tooling improvements
…accuracy - Extract actual method names from anonymous inner class names (e.g. Outer$1) and stack traces, storing the enclosing method as context_method - Tag XML layout inflate calls as [XML布局] with call count and total time - Reorder report sections to put attribution data first, preventing truncation of core input when total content exceeds token budget - Add completeness check rule to report prompt to avoid missing entries - Switch CLI entry point to graph module and add hatchling build config Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
refactor: 工具增强与匿名内部类归因支持
Covers plugin system, URL API, postMessage, trace_processor HTTP mode, proposed local Web Server bridge architecture, and integration plan with existing LangGraph pipeline. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Document the perfetto-plugin/build.sh usage, prerequisites, build steps, and the --skip-clone flag for incremental builds. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…fetto UI Add SI Bridge plugin for Perfetto UI enabling interactive frame-level analysis: users select a time range in Perfetto UI and get AI-powered analysis with source attribution via WebSocket bridge. New components: - perfetto-plugin/: TypeScript plugin + build.sh for self-hosted Perfetto UI - frame_analyzer agent: queries slice data, runs attribution, LLM analysis - bridge_server: serves Perfetto UI + WebSocket bridge to CLI agent - /open, /close, /frame CLI commands - HDC command-line research doc for HarmonyOS support Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
feat: Perfetto UI 交互帧分析 + 工具链增强 + 文档更新
Replace report example with actual 2026-04-15 analysis output, update quick start examples, improve table alignment, clean up Todo checkbox syntax, and remove unreferenced screenshot. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: update README and add Perfetto UI interactive analysis docs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…te analysis Add a fast path in attributor that skips LLM for straightforward java type searches using direct Glob→Grep→Read. This reduces latency and cost for cases without anonymous inner classes or unknown methods. Add thread state analysis to perfetto collector that shows Running/S/D distribution per SI$ slow slice, helping distinguish slow code from blocked threads. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Feat/call stack attribution
# Conflicts: # src/smartinspector/agents/attributor.py
…r class handling - Fix SI$block# slices (dur≈0) being squeezed out by LIMIT 50 SQL query: split into two queries prioritizing SI$ slices over system slices - Fix anonymous inner class method resolution: always set context_method from _extract_method_from_anonymous, only override method_name with stack - Add lightweight LLM analysis (_analyze_snippets) for fast-path results to replace raw source code with analysis text in source_snippet - Fix _build_frame_hints: sort SI$ slices by effective dur DESC, aggregate repeated calls with cumulative duration and call count - Fix dur extraction from SI$block# tag suffix (#NNms) in frame_analyzer - Filter out negligible dur slices (< 0.01ms) from attribution - Show "(匿名内部类, 定义在 method 内)" label for inner class results - Add debug logging for Step 3 LLM input/output Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous query required thread_state entries to be entirely within the slice window (ts >= start AND ts + dur <= end), which missed almost all entries since thread_state records commonly straddle slice boundaries (e.g. a long Running state starting before the slice). Changed to overlap-based calculation: SUM(MIN(end, slice_end) - MAX(start, slice_start)) with proper handling for dur < 0 (still-running entries). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two bugs in collect_thread_state state normalization: 1. State distribution used overwrite instead of accumulation: when multiple raw states mapped to the same normalized name (e.g. R and R+ both → Running), the second would overwrite the first instead of summing. This caused incorrect percentages when both R and R+ appeared. 2. S+ state was not mapped to Sleeping: S+ (interruptible sleep, preemptible) is common in Android apps but was left as-is instead of being normalized to "Sleeping". Added comprehensive tests for normalization, overlap calculation, and filter conditions. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…s, and perfetto auto-degradation P1-4: Add dependency reference search in Attributor Agent — after finding target files, extract import statements for project-internal classes and R.layout XML references, read related files, and include as dependency_context for richer AI diagnosis. P1-5: Add package_list cold-start fallback — new _resolve_target_process() method tries process table first, falls back to package_list table for UID lookup when process table is empty during cold start. Annotates view_slices with target process info and includes resolved process in metadata. P1-6: Add SELinux cat-pipe bypass — when stdin-pipe perfetto config fails, fallback to pushing config file and using cat pipe (cat config | perfetto -c -). P1-7: Add perfetto collection auto-degradation — when config mode fails entirely, degrade to simpler cmdline mode with just atrace categories. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add design document for thread state blocking analysis and redesign.
Added project rules, overview, structure, conventions, logging rules, architecture, commands, configuration, known issues, and build notes to CLAUDE.md.
…ead_state for blocking details Replace sched-based thread state inference with __intrinsic_thread_state table queries that expose blocked_function, io_wait, and waker_utid fields. This enables actionable blocking analysis instead of generic Running/Sleeping labels. Falls back to legacy thread_state table when __intrinsic_thread_state is unavailable. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both --src and --source-dir are accepted (--src is the primary name). Updated CLAUDE.md docs to use --src. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
DeepSeek thinking mode adds `reasoning_content` to AIMessage responses. On subsequent API calls, if the message history contains these but gets trimmed, the API rejects with 'reasoning_content must be passed back'. Fix: strip reasoning_content from every LLM response before appending to message history, and also from trimmed messages during window trimming. Also relax fast-path: allow anonymous inner classes ($ in class_name) to use deterministic search. extract_class() already strips the $ suffix for glob search, and context_method handles method-level grep. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When getResourceEntryName() fails on the app Context (e.g. for android.R layouts like simple_list_item_1), try Resources.getSystem() as fallback before falling back to hex ID. This produces readable tag names like 'SI$inflate#simple_list_item_1#RecyclerView' instead of 'SI$inflate#0x109010a#null'. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
For SI$block slices with pure anonymous classes (e.g. MainActivity$1), extract_method returns 'unknown' and _extract_method_from_anonymous finds no enclosing method. This blocks fast path and causes LLM path to produce vague 'unknown' attribution. Fix: when method_name is 'unknown' for an anonymous class block, walk the stack trace to find the caller method from the same class (e.g. 'loadAndDisplayItems' from 'MainActivity.loadAndDisplayItems'). Set it as context_method and use 'run' as method_name, enabling fast-path deterministic search to locate the source code. Stack trace used: DataRepository.buildPayload → DataRepository.loadItemsJson → MainActivity.loadAndDisplayItems → MainActivity$1.run Now correctly attributes to MainActivity.loadAndDisplayItems. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous commit's edit replaced extract_class with _extract_caller_from_stack, leaving extract_class's body as dead code after a return statement. Restore the def line. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…attribution, and report section ordering - Fix metadata query: column 'key' -> 'name' (Perfetto metadata table schema) - Fix heap_graph_field error: add pre-check for heap_graph_object table existence - Fix thread_state 0 entries: _resolve_main_utid now tries 3 strategies (name='main', name=package_name, lowest tid in target process) - Remove structured output probe in attributor to prevent DeepSeek thinking mode activation and reasoning_content errors - Add CPU hotspot extraction to attribution pipeline for background thread attribution - Reorder report sections to match CLAUDE.md priority (thread state before frame timeline) - Add [CPU热点] tag in report formatter Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
feat: metric QA, startup analysis, attribution improvements, and collector bug fixes
- Create si_tag.py with SITag dataclass and parse_si_tag() function
that does single-pass parsing of all SI$ tag types (block, RV,
inflate, view, handler, db, net, img, touch, default)
- Rewrite extract_class(), extract_method(), extract_fqn() as thin
wrappers around parse_si_tag(), eliminating ~280 lines of duplicated
parsing logic from attribution.py
- Move _split_fqn_method, _extract_method_from_anonymous, system class
constants to si_tag.py as single source of truth
- Add SITag.is_system and SITag.is_system_method properties for
declarative system class checking
- Rewrite is_system_class(), is_system_method(), classify_search_type()
to use parse_si_tag() internally
- Add 90 new unit tests:
- test_si_tag.py: 63 tests covering all SI$ tag types, edge cases,
backward-compatible wrappers, and integration
- test_deterministic.py: 27 tests for severity classification,
call-chain distribution, RV hotspots, CPU hotspots, thread state,
IO slices, memory analysis, frame budget detection, compute_hints
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…own) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…startup.*) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… slices.time_in_state) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…or tree (android.memory.heap_graph.*) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ger) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
mufans
force-pushed
the
feat/stdlib-modules
branch
from
May 12, 2026 02:13
658b5f0 to
5fbf427
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
集成13个Perfetto SQL Standard Library模块，大幅增强SmartInspector的trace分析能力。
P0 模块（核心能力）
P1 模块（场景增强）
Stats
INCLUDE PERFETTO MODULEwith stdlib tables/viewsTest Plan