SDW AIKO

English

简体中文
日本語

English

简体中文
日本語

Appearance

Getting Started

SQA Agent ​

SQA Agent is responsible for software quality assurance tasks, including file comparison check, deliverable inspection, change point check, Review Master review and other quality inspections for multiple divisions and business lines.

3.1.3 (Latest) ​

Release Date: 2026-10-23 Version Range: 3.1.2 (2026-09-04) → 3.1.3 @ 0b43565aCommits: 117 commits | Contributors: Jiang Long, Yuan Yihan, Li Guowang, Huang Yang, Quan Wei

Version Overview ​

3.1.3 is a milestone release that takes the Checklist knowledge base from zero to one and completes end-to-end validation. Using real business data from Division 5 as the baseline, this release connects the three-tier capability chain of checklist_kb knowledge base (v0.1 → v0.2.5) → checklist-generator generation engine (v2 general-purpose engine + check inspection mode) → review-master-builder inspection execution. It also provides the kb-curator / kb-eval / checklist-eval / check-eval evaluation skills, a persistent evaluation system, and functional scorecards, ultimately delivering them as an SQA Expert Agent release package (expert_agent). Meanwhile, the AERA review engine continues to advance with refactoring of document context and Domain profiles, as well as WebUI process lifecycle refactoring.

⚠️ Version Number Change: The Expert Agent release package version is aligned with the product roadmap, jumping from 1.1.1 to 3.1.3. All historical 1.x packages are superseded by 3.1.3.

🚀 Major Features ​

1. Checklist Knowledge Base (checklist_kb v0.1 → v0.2.5) ​

  • Technical solution finalized + feasibility validated (H1–H5), delivering the v0.1 knowledge base skeleton: directories + glossary + 13 rule cards + 336 drafts + zero-dependency tools + browser view.
  • Division 5 data consolidation: 336 drafts → 96 formal rule cards (absorbing 331); business-domain and toolchain tags added to rule cards, enabling in-vehicle software domain dimensions.
  • Generalized integration across Divisions 1–5: artifact glossary entries regrouped by family, cross-batch consolidation into cards, and completion of rule sets and documentation; a Division 3 checklist template was added.
  • Business materials reorganized: directories from all five divisions consolidated under data/<code>/; 47 registry paths rewritten and fully verified by md5; build_index embeds all 47 sources from the W5/VID/W5V2 batches.
  • SDW release capability: SKILL.md documentation + make_sdw_package.py packaging script, with release packages upgraded from v0.1.1 to v0.2.0.
  • Successive field-test fixes and releases: 0.2.1 (R-CASE-0002 consts restoration + validate params shape guardrail) → 0.2.2 (four SDW field-test comparison fixes + two human-only semantic cards) → 0.2.3 (Cm135 field-test fixes: Project/Phase machine regex + escape-column concept rebinding) → 0.2.4 (localization concept batch: eight concepts for cover-page approval fields/CS confirmation tables/load calculation tables + eight card remappings) → 0.2.5 (generic card family assignment fixes + source records, improving design_cs recall from 4 to 6 entries).
  • Improved usability: single-source Excel/browser data dictionary + Chinese columns + legend; source-text fidelity gaps fixed and a Reference comparison layer added.

2. checklist-generator Generation Engine and Check Inspection Mode ​

  • Added a self-contained checklist-generator skill (knowledge base + generation capability) and upgraded the v2 general-purpose engine: concept mapping / prototype execution / two-round confirmation process + CT reconciliation closure.
  • Check inspection mode (solution v1.4): directly produces a pending-confirmation version from deliverables plus a confirmation view; solution v1.5 final delivery contract: all machine-contract fields + confirm stage + template-semantic backfill.
  • B3 pattern inference implemented (gates 0–3 / new candidate files / suppression mode); card parameter fallback uses structured params on cards when thresholds are not met.
  • Generator engine fix batch: judgment confidence guardrails + confirmation persistence (remembered but not silent) + embedded images in the confirmation view.
  • Inspection confirmation view with five levels + better missing than wrong gatekeeper + unrecognized type gatekeeper (feedback P-02/03/04 mechanisms implemented).
  • Coverage transparency + guidance for uncovered areas: coverage subcommand / three-option gate guidance / upfront notice; checklist generation integrated into the existing review-master-builder skill (Division 5 RMB used for the inspection tool flow).

3. Evaluation System: kb-eval / checklist-eval / check-eval + Functional Scorecards ​

  • Added the kb-eval comprehensive knowledge-base evaluation skill: four-stage evaluation (static examination → generation equivalence → fixture trial run + OK/NG labeled set + mutation injection → regression gate); ct_spec maps new cards 1–20/22/23, and all baseline refresh gates pass.
  • Added the checklist-eval generation capability evaluation skill + dual-version test reports (HTML for presentation + MD for archiving); supports no-judgment mode (①-1 coverage still evaluated, ②-3 truthfully marked not applicable), with the case3 design-document-family scenario included in the package.
  • Persistent evaluation system implemented: check-eval + eval-all + shared comparator, with regression baselines restored for checklist-eval; first check-eval run (case1 inspection tool builder full-flow execution + inspection report and NG/OK dataset), with layered attribution rules (skipped/neutral cases receive no attribution; responsibility is assigned only for genuine misjudgments).
  • Functional scorecards (implementation of Chapter 6 of the master framework): manifest scoring criteria + aggregator, with scores of 95/62/100 for the three functions.
  • Evaluation credibility strengthened: code-version anchors + uncollected-dimension labels + two-tier treatment for small samples.

4. SQA Expert Agent Release Package (expert_agent → 3.1.3) ​

  • Added an SQA Expert Agent packaging pipeline: template + three skills + KB subset, with the 1.0.0 archive included in the repository.
  • Iteration path: 1.0.2 (removed the kbs list from frontmatter and wrote the four-level KB source arbitration rules into the persona) → 1.1.0 (KB 0.2.4 + five-level system/gatekeeper/coverage + persona obligation to disclose coverage) → 1.1.1 (KB 0.2.5) → 3.1.3 (version aligned with the product roadmap).
  • Expert package text CRLF→LF fix: the SDW engine's frontmatter key-value parser is sensitive to a trailing \r, causing zero_config to be dropped and the package to be incorrectly judged as "skeleton not adapted."

5. ASPICE Alignment and Test Datasets ​

  • ASPICE v4.1 coverage analysis and completion plan (v1.1): current-state matrix + five completion contracts; process-mapping spot check (30-card triangulation), with the feasibility gate for Contract ① passed.
  • Test datasets: added tests/sdw_20261016 (SDW manual test package using real Division 5 data), reorganized into three raw files for case1/2/3 (deliverable + manually created checklist + manual inspection results); case4 abnormal-scenario robustness test—bad samples fed into three entry points, producing errors without false judgments.
  • Documentation system: master framework for building and evaluating the three Skills (v1.0 → v1.2) + evaluation-metric descriptions for the three skills + consolidated LLM evaluator templates (methodology for six general evaluator templates); added root-level AGENT.md development principles + three process red lines for generated skills; added inspection business swimlane diagram and architecture design v2 (asset-factory positioning).

6. Continued AERA Review Engine Iteration ​

  • Document context and Domain profile refactoring, with reports from historical rounds cleaned up.
  • WebUI process lifecycle refactoring and SQLite commit semantics fixes.

✨ Platform and Engineering Improvements ​

  • The review-master-builder inspection skill was moved into skills/ and integrated with checklist generation; the shared runtime envs/ was briefly removed from the repository and then restored to Git tracking (user decision: deployment is by clone, and the runtime travels with the skill).
  • kb-curator skill: general capability for building/updating/extending knowledge bases, supporting complete construction from scratch (init_kb + metamodel seed + full pipeline with --kb-root), with the provenance tool trace.py added.
  • Evaluation/validation reports were refined over multiple rounds: overview + per-module details (CT/six families/OK-NG/mutations/B3/gates each in its own chapter), quick metric definitions (numerator/denominator + interpretation), full totals before module-level breakdowns, and sample-threshold criteria explained.
  • .gitignore completed for checklist-eval run snapshots, Excel lock files, and .gitlab_token (MR credentials excluded from the repository).

🐛 Key Fixes ​

  • Fixed KB generic card family assignments (source-ledger family = applicable family) + source records, improving design_cs recall from 4 to 6 entries.
  • Normalized A-CONTROL-STATE control-label matching—the root cause of 1–5 false positives (card-key shape drift).
  • Layered check-eval attribution rules—skipped/neutral cases receive no attribution; responsibility is assigned only for genuine misjudgments.
  • Expert package CRLF→LF: a trailing \r in frontmatter caused the SDW engine to drop zero_config and incorrectly report "skeleton not adapted."
  • Fixed checklist_kb source-text fidelity gaps + added a Reference comparison layer.
  • Fixed image-placeholder consensus derivation issues found during W5V2 six-family end-to-end validation.

🧹 Cleanup and Refactoring ​

  • Cleaned repository intermediate artifacts: replaced 745 absorbed drafts with a manifest, removed one-off VID analysis files, and removed Division 5 build pyc files.
  • De-customized checklist-generator skill documentation—moved evaluation figures/datasets out and deleted old runtime artifacts from the skill.
  • Archived and uploaded the handover AGENT source code; removed the code directory (using Temp instead) and sqa-assistant; cleaned up unused files.

Upgrade Notes ​

  1. The Expert Agent release package version is now aligned with the product roadmap: 1.1.1 → 3.1.3. Use the 3.1.3 package consistently; historical 1.x packages are no longer maintained.
  2. Checklist knowledge-base business materials are consolidated under data/<code>/; 47 registry paths have been rewritten and fully verified by md5, so references to old paths must be updated.
  3. Evaluation/validation reports now include code-version anchors and labels for uncollected dimensions. Interpret small-sample dimensions using the two-tier criteria, and use the anchored version in each report for cross-version comparisons.
  4. Evaluation skills output both HTML and MD: HTML for presentations, MD for archiving and adjustment.

3.1.2 ​

Release Date: 2026-09-04 Version Range: 3.1.1 (2026-08-13) → 3.1.2 @ 6ad5bdc7Commits: 14 commits | Contributors: Li Guowang, Huang Yang, chuangyum, Yuan Yihan, Quan Wei

Version Overview ​

3.1.2 is an in-depth iteration of AERA (AI Self-Evolving Design Review Engine) V3.9, centered on upgrading the review core: comprehensive upgrades to the review engine, rule system, and report generation; delivery of the Excel refactoring version snapshot (AERA v3.9), dedicated historical-candidate verification, and a design-baseline CS switch; refactoring of context construction and the fact graph (score-based truncation selection + cross-sheet references); and completion of the SKILL.md rewrite and version 1.5 engineering refinements.

🚀 Major Features ​

1. Comprehensive Review Engine Upgrade ​

  • Comprehensively upgraded the review engine, rule system, and report generation, rewrote the SKILL.md documentation, added review rules, and adjusted numbering.
  • Implemented dedicated historical-candidate verification and a design-baseline CS switch, improving preview fidelity and evidence-chain reliability.
  • Context construction and fact graph refactoring: introduced score-based truncation selection and cross-sheet references, and added architecture documentation and audit reports.

2. Excel Refactoring Version Snapshot (AERA v3.9) ​

  • Uploaded the Excel refactoring version snapshot—AERA v3.9 (including README instructions), with the pre-refactoring baseline version backed up and archived.

3. Skill Engineering and Performance Optimization ​

  • Two rounds of updates for version 1.5.
  • Standardized the scope of skill modifications (reference skill aligned with the current process); merged code performance improvements: concise HTML, module decoupling, function refactoring, and HTML expansion.

Upgrade Notes ​

  1. This release mainly upgrades the internals of the AERA review engine; after upgrading, a regression validation pass on existing review jobs is recommended.

3.1.1 ​

Release Date: 2026-08-13 Version Range: 3.1.0 (2026-07-17) → 3.1.1 @ 1d7f9cd7Commits: 142 commits | Contributors: Li Guowang, Quan Wei, chuangyum, Huang Yang, Yuan Yihan

Version Overview ​

3.1.1 is a dense iteration of Auto-Evolving Review Agent (AERA, AI Self-Evolving Design Review Engine). Compared to the Alter Ego Review Agent seven-stage pipeline in the 3.0.x era, this version completes a continuous evolution from V3.5 → V3.6 → V3.7 → V3.8 → V3.9 → "Eleventh Iteration": the brand and architecture are unified as AERA / Auto-Evolving-Review-Agent-V3, a new Vision Understanding (VLM) pipeline and Excel sheet-by-sheet understanding are added, vector recall is migrated to WSL Milvus with hardened degradation probes, the report side completes a redesign + evidence chain root fix, and product capabilities such as multi-person collaborative annotation, organization-level rule governance, and Learning Center are added.

⚠️ Breaking Change: The legacy Alter_Ego_Review_Agent_V3 has been fully migrated and renamed to Auto-Evolving-Review-Agent-V3; the term "host Agent" is unified as "AI review engine"; LLM gateway configuration is consolidated into a single source provider.json, and configuration-class environment variables no longer take effect.

🚀 Major Features ​

1. AERA Branding and Architecture Unification ​

  • AI Self-Evolving Design Review Engine (AERA) V3.5 → V3.9 → "Eleventh Iteration" continuous fusion: brand renaming, vector model switch, organization-tree rule filtering, product template soft deletion.
  • Legacy Alter_Ego_Review_Agent_V3 migrated to Auto-Evolving-Review-Agent-V3, legacy directory removed and dual-tree synchronized.
  • Unified term "AI review engine" replaces "host Agent", deprecated files cleaned up and product introduction page added.
  • Report structure restructured, new #dev hidden layer added, dual-tree synchronized and aligned with the latest version.

2. Vision Understanding (VLM) Pipeline ​

  • New Vision Understanding (VLM) pipeline, supports PPT / Word / PDF media understanding and generates image / chart findings.
  • Excel two-stage media recognition chain: whole-sheet preview cache, formula explicitization, VLM sheet understanding enhancement and test suite hardening.
  • Visual evidence chain connected to HTML reports, fixes crop marker height calculation and slicing planning infinite loop (out-of-bounds merged cells).
  • Vision understanding timeout and concurrency configurable in WebUI, batches split by payload volume, contract failures truthfully recorded as degradation.

3. Sheet (Sheet-by-Sheet) Understanding and Concurrency Speedup ​

  • New timeout budget, concurrent overview and progress callbacks, in-flight progress reporting and front-end progress display refactored.
  • tile_planner merged-area and visual object row-window clipping optimization.
  • Bounded concurrency at three points (vision / overview / embedding) + rendering fallback receipt; independent timeout added to core analysis.

4. WSL Vector Recall and Historical Issue Engine ​

  • Vector retrieval migrated to WSL Milvus: configurable WSL probe timeout added (default 12s → 30s), long path support, pymilvus/milvus-lite dependencies auto-installed.
  • Root fix for vector backend silent degradation: real retrieval probe, failure logging, prohibiting small indexes from overwriting shared libraries.
  • Vector model switched to text-embedding-3-large, new check templates and hyperparameter presets added.
  • Historical issues: configurable historical depth verification added, historical candidate pipeline configuration and dimension validation, multi-document parallel design pre-review and historical lexical degradation recall; corrected historical TOP to contain only verified gaps and added screenshot PDF fallback.
  • New classification vocabulary and classification observability, issue classification refactored to SQLite.

5. Report Redesign and Evidence Chain Root Fix ​

  • Three report redesign items landed + two evidence chain defects root-fixed.
  • Fixed four evidence chain breaks (broken images / dead links / fake tables / machine sheet id), restoring machine sheet id to real worksheet names before display.
  • Summary page metric distortion fixed (green KPI ambiguity, rule pass rate exceeding 100%).
  • Reports display actual judgment basis and retrieval coverage, worksheet role determination + historical evidence package framework page down-weighted.
  • WebUI adds asset routes within the report directory, images no longer break.

6. Collaboration, Governance and Learning Center ​

  • Implements multi-person collaborative annotation, issue AI analysis, governance model configuration and library optimization.
  • Supports organization-level rule scanning, fixes historical matching classification and report UI issues.
  • "Eleventh Iteration": Learning Center, design fact graph refresh, performance logging and parallel understanding.
  • WebUI adds runtime configuration persistence, developer environment settings management, job switcher, progress bar and run log expansion.

✨ Platform and Engineering Improvements ​

  • Gateway single source: gateway configuration changed to provider.json, configuration-class environment variables removed; unified configuration priority and re-issued stage start events.
  • New setup-env one-click environment setup command; startup script robustness improvements, Web UI uses custom dialogs instead of native prompt/confirm.
  • Excel streaming parsing and batch review optimization, fixes OOM and improves prompts; parallel processing, Chinese cleansing and masking enhanced to improve report quality.
  • Enhanced observability, WSL recovery, task_id reverse lookup and front-end refactoring; job polling connection fault tolerance, server-side request queue expanded.
  • Other divisions: Division 5 CT / MD / skills multiple checklist optimizations and script updates; Division 1 CS, check-designviewpoint 1.3/1.4 etc. synchronized; skill supports directory form, route merging skill solution and first Japanese version.

🐛 Key Fixes ​

  • WSL vector backend silent degradation / fail-fast false positives root-fixed; reranker score filtering and wsl_probe_timeout validation hardened.
  • Enrichment retry, rerank fault tolerance and front-end log convergence; COM thread initialization fixed.
  • Issue DB import field name fixed, portable version runtime robustness hardened; report manual status dropdown aligned with backend four-state definition.
  • Fixed 16MB limit BUG, _historical_aux_projection docstring syntax error.
  • 06_Items_for_Improvement ten consistency issues fixed + dual thresholds.

🧹 Cleanup and Refactoring ​

  • Migrated and removed legacy Alter_Ego_Review_Agent_V3, organized AERA project documentation and archived historical architecture (v3.1 / v3.2).
  • Overall removed generalized_miss_detection generic missed-detection supplementary round; recall unit excludes mechanical diff tables and windowizes paragraph-level sources.
  • Deleted qc_report.py quality sampling script, cleaned up expired reports and implementation documents, organized raw issue analysis files.

Upgrade Notes ​

  1. The legacy Alter_Ego_Review_Agent_V3 call path is no longer valid; please use Auto-Evolving-Review-Agent-V3 instead.
  2. Vector recall depends on WSL Milvus; please run setup-env first and confirm the embedding model is text-embedding-3-large; WSL probe timeout defaults to 30s.
  3. LLM gateway configuration must be unified into provider.json; legacy configuration-class environment variables no longer take effect.

3.1.0 ​

Release Date: 2026-07-17 Version Range: v3.0.6 (2026-07-07) → 3.1.0 @ 6a3ede3eCommits: 51 commits | Contributors: chuangyum, Quan Wei, Li Guowang, Yuan Yihan, Huang Yang

Version Overview ​

3.1.0 is a transitional version in which Alter Ego Review Agent is refactored from the "seven-stage pipeline" to the "V3 capability-layer". The core change is decomposing the fixed seven stages of 3.0.6 into composable capability-layer units (knowledge pack → evidence gate → question loop → viewpoint review → final decision), landing the AI Alter Ego refactoring V3; on the recall side the rerank judge ablation and switch is completed, on the report side it moves to a LangChain document-style left-right layout, and Review Master Builder is pushed toward config-driven + new UI.

⚠️ Change Notice: The alter-ego-v3 "review on first run" package is added along with historical architecture backups; the external prototype source Common Design Review SKILL (about 50MB, not required at runtime) has been removed from the repository.

🚀 Major Features ​

1. AI Alter Ego Refactoring V3 (capability-layer Units) ​

  • AI Alter Ego refactoring V3 landed: added the alter-ego-v3 "review on first run" package, added the Alter Ego Review Agent project directory and historical architectures.
  • Introduced the capability-layer unit layer: added viewpoint review and final report capability units, refactored the design review evidence judgment logic.
  • Introduced the knowledge-pack layer and integrated it into the review workflow.
  • Integrated the project evidence collection agent, refactored the question loop and final decision flow.
  • Added question_loop and unit_05_evidence_gate module specification documents; release readiness gate and offline deterministic product contract validation.

2. Recall rerank Judge Ablation and Switch ​

  • rerank judge switched by default to gpt-5.6-sol (dual-metric ablation champion, R12 verified on all five documents + user cross-check passed), with the ablation panorama index SPEC added.
  • v2 history-matcher integrates book-level rerank + rich profiling, with concurrency acceleration.
  • Archived recall ablation reports + reproduction tools into the repository.

3. Reports and Review Master Builder ​

  • Reports changed to a LangChain document-style left-right layout, report module updated.
  • Review Master Builder: config.json based + new UI + new HTML integration, default check mode preferred, intermediate viewpoint confirmation step optimized; fixed issues such as rich text deletion, images not extracted, and async generators not closed.

✨ Platform and Engineering Improvements ​

  • Enhanced Excel input security protection and removed the numpy dependency.
  • Enhanced the installation flow and configuration interface, updated the default embedding model.
  • Unified file path prefixes in SPEC.md and added wildcards; updated architecture diagrams while keeping old-version backups.
  • Division synchronization: CT review automation tool, checklist batch update, CS design update, common viewpoint deliverable update.

🧹 Cleanup and Refactoring ​

  • Removed the external prototype source Common Design Review SKILL (about 50MB, not required at runtime) and redundant files.
  • AI Alter Ego review refactored to the V3 version, old architecture archived as backup.

v3.0.6 ​

Release Date: 2026-07-07 Version Range: v3.0.5 (2026-06-18) → 3.0.6 @ 0950b40fCommits: 121 commits | Contributors: Li Guowang, Quan Wei, Huang Yang, Yuan Yihan, chuangyum

Version Overview ​

3.0.6 is a milestone-level refactoring. Compared to the POC/experimental phase of 3.0.x, this version officially lands the complete Alter Ego Review Agent (AI Alter Ego Design Review) system: from "universal lossless parsing" to "historical issue recall", then to "seven-stage review pipeline" and "final decision report", forming an end-to-end deliverable review capability. It also consolidates the Common Design Review SKILL universal skill package, project centralization, and deployment optimization.

⚠️ Breaking Change: POC directories (poc / poc_kb) and the old bundled_units and old SKILL directories have been removed and refactored into the bundled_units seven-stage architecture; old call paths are no longer available.

🚀 Major Features ​

1. Alter Ego Review Agent (AI Alter Ego Design Review) ​

  • New project structure and core module initialization, with CLAUDE.md project rules document and per-stage READMEs.
  • Seven-stage artifact pipeline: extends seven-stage artifact production, with real-machine integration (WSL vector delegation, contract and visibility completion).
  • v2 core refactoring (3.0.6 wrap-up): refactors core review capability, introduces best-fact selection and semantic explanation; adds evidence gate closure plan, issue loop closure action and answer sufficiency output.
  • Adds design pre-review prompt files.

2. Final Decision and Report ​

  • Adds final decision build script and unit tests, rewrites HTML report rendering and conclusion selection logic, adds historical match item construction.
  • Refactors image optimization logic, simplifies format processing (including EMF conversion bug fix).

3. Historical Issue Recall Engine (03_history_matcher) ​

  • Recall engine generalization: scale-resistant, multi-template generalization, removes template specialization, fixes truncation order.
  • Knowledge base expansion: issue_master from 397 → 1086 entries (689 internal review entries + LLM gap-filling).
  • 02 summary injection recall + single container fault tolerance / offline fallback + domain normalization + supporting evaluation tools.
  • Embedding model correction: default model corrected to text-embedding-3-large (empirical), with container-level degradation alerts added.
  • bge-reranker local re-ranking evaluation: tested negative benefit, temporarily suspended (evaluation document archived in section 9.5).
  • Milvus vector store changed to build artifact, removed Git tracking and added to .gitignore.

4. Universal Lossless Parser (Universal Parser / GDM) ​

  • GDM lossless parsing extended to Word / PDF / TXT / MD, covering images, shapes, text boxes, with pluggable OCR / VLM.
  • Parser refactored into a single file, removed review_agent dependency.
  • Large file chunking (normalized text layer, coarse-to-fine + capping, universal for all formats).
  • Excel / PPT image hierarchy and pointing recognition; fixes get_slide_texts() issue that could only extract text boxes / placeholders / AutoShapes.

5. Common Design Review SKILL (Common Design Review Skill Package) ​

  • Introduces common design review SKILL module, rewritten according to standard skill spec.
  • Template commonization: extraction / script / model three types of templates unified; IO changed from single ExcelIO to multi-source IO_list, supporting simultaneous inspection of PPT / Excel and related materials.
  • Integrates initial version of five-division solution.

✨ Platform and Engineering Improvements ​

  • Project centralization: first version of project center implementation (GUID + FileID), field organization and centralization solution design.
  • Deployment optimization: in-project .tools priority (compatible with old projects), dependencies extracted as whl package installation; template copy 5s → 0.4s.
  • Runtime directory naming refactoring, automatically adapts to Windows path length limits.
  • Timeout optimization.
  • Radar chart visualization v1, HTML template field organization.

🐛 Key Fixes ​

  • Seven-stage real-machine integration fixes: WSL vector delegation / EMF conversion bug / contract and visibility completion.
  • Embedding default model correction + container-level degradation alerts.
  • get_slide_texts() extraction range fix.
  • POC recall chain de-template specialization, truncation order fix, Word image recognition symmetry.

🧹 Cleanup and Refactoring ​

  • Removes POC (poc / poc_kb, both merged), old SKILL directories, old bundled_units.
  • Archives Q&A-style and reference framework code to deprecated directory.
  • Cleans up unused CLI acceptance test files, Office edit lock files (~$).

Upgrade Notes ​

  1. Old POC and old SKILL call paths are invalid, please use bundled_units seven-stage entry.
  2. Recall module needs to confirm embedding model is text-embedding-3-large.
  3. Deployment please use in-project .tools and whl package installation process.

v3.0.5 ​

Release Date: 2026-06-18 | Base Version: v3.0.4

1. Overview ​

ItemContent
Release Date2026-06-18
Versionv3.0.5
Base Versionv3.0.4

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Division 2 - SQA_Agent_Base Code Code Framework1. Added SQA_Agent_Base/Code/ complete SQA Agent code framework (4ca6c9e, 166d913):   agent/: llm_client.py, judge.py, planner.py, prompts.py, orchestrator.py, tool_schemas.py  checklist/loader.py, parsers/(excel_parser/ppt_parser/markitdown_fallback)  pipeline.py, runtime.py, config.py, cli.py, models.py, validators.py  report/: excel_report.py, markdown_report.py; skill/loader.py  tools/: registry.py, excel_tools.py, gitlab_tools.py; indexing/memory_index.py
2. Added technical solution and Codex execution manual, FileCompare-Checker-SKILL.md
3. Added test case data (TestCase/Design/Verification) and sample checklist
Division 2 - SQA_Agent_Base Code_cc mainframe Common Protocol1. Added Code_cc/mainframe/ common document parsing protocol framework (5b362d3):   gdm.py, parse.py, query.py, fieldmap.py, excel_profile.py  com_render.py, shape_resolve.py, llm.py, gitlab_rest.py  checklists/filecompare.py, run_filecompare.py
2. Added ARCHITECTURE.md, spec.md, technical_spec.md
Division 2 - Alter_Ego_Review_Agent GDM Parsing and Embedding Matching1. Added GDM lossless document parsing module (b4a152e): parsers/gdm/(gdm.py, parse.py, query.py, fieldmap.py, excel_profile.py)
2. Historical issue relevance matching upgraded to semantic embedding + LLM re-ranking + dimension-gated layering (0b74d01):   Added embedding_client.py, issue_embedding_store.py, issue_matcher.py, llm_client.py  Added scripts/embed_issues.py, requirements.txt
3. Added http_retry.py retry mechanism (687030c)
4. Added design document template and review issue summary xlsx (7198484)
5. Added historical issue relevance matching implementation SPEC.md (dacd639)
Division 1 - CVMS Configuration Change Management System1. Added CVMS/ configuration change management system (3ecb294):   Frontend: index.html, css/style.css, js/(app/recommend/seed/store/ui/views.js)  Backend: server/(server/auth/db/model/export/notify/xlsx.js)  public/app/, public/login.html, data/kotei.db
Division 1 - ct-checklist-b Binary Toolset1. Added ct-checklist-b/ toolset and configuration (0558283):   SKILL.md, check-config.json  Multiple .exe tools: excel_blank_checker, excel_filter, excel_to_markdown,   extract_shape/sheet_snapshots, filecompare_rule_check,   jira_get/search series, ppt_to_pdf, run_ct_checklist series, etc.
Division 5 - MT v2 ModuleTest review_master1. Added MT/v2/ ModuleTest review_master complete framework (a3a654b, b57ac6c):   checklist/Test Specification_KOTEI_PF_Delivery CheckList_ModuleTest.xlsx  review_master/app/: check_templates/, checks/(test_spec series check functions),   common/, config.py, infrastructure/, main.py, reporting/, routers/, services/, ui/
2. Added check items: file_name_format, document_no_extract, confidential_mark_image_extract,   asil_extract, design_document_name_matches, name_matches, no_and_object_id_match,   option_checkbox, project_phase_extract, wi_environment_ecu_dash

2.2 Changes ​

ModuleDetails
Division 5 - review-master-builder skill refactoring (d8f5540)1. SKILL.md/AGENT.md large-scale update, template check function refactoring
2. Added app/config.py, input_selector.py, output_writer.py
3. Added UI module: ui/init.py, review_window.py
4. refs/ directory renamed to references/
5. Added create_review_tool.cmd script
6. Updated check_templates/, checks/, common/, infrastructure/, reporting/, routers/, services/, etc.
Division 2 - SQA_Agent_Base directory reorganization and documentation cleanup1. Cleaned up old Workflow/Project Plan/Human-machine collaboration/prompts directories (34c6c7b)
2. Added 01-Requirements/spec.md requirement specification document (7614b47, 17142ca)
3. Added meeting minutes materials 20260615 (4e47cd8, 0e22ce7)
4. Meeting review materials reorganization: scattered files replaced with .7z archive + final issue list (68179c5)
5. Updated SQA full-link AINative integrated flowchart (7357200, 0b12c6a)
Division 5 - MD checklist and deliverable batch update (2d140b4)1. ModuleDesign checklist update
2. Added/updated multiple COMRouter ModuleUnitDesign.xlsx deliverables (12 files)

2.3 Removed ​

ModuleDetails
Deleted Items1. Removed Division 2/Requirement-Checker-b/ redundant b version (4cd3b09): .gitignore, README.md, SKILL.md, bin/exe
2. Cleaned up SQA_Agent_Base old directories: 02-Workflow/, 03-Project Plan/, 04-Human-machine collaboration/ all old files (34c6c7b)
3. Cleaned up SQA_Agent_Base/01-Requirements/ old SPEC.md/prompts/reference materials (0e22ce7)
4. Cleaned up scattered meeting review material files, replaced with .7z archive (68179c5)
5. Cleaned up TestCase temporary check result files (166d913)

2.4 Bug Fixes ​

ModuleDetails
Issue Fixes1. Fixed Alter_Ego_Review_Agent POC embedding_client/llm_client/http_retry related code (687030c)
2. Fixed Alter_Ego_Review_Agent historical issue matching implementation SPEC patch (687030c)
3. Fixed MT v2 review_master check item function registration (b57ac6c)

v3.0.4 ​

Release Date: 2026-06-12 | Base Version: v3.0.3

1. Overview ​

ItemContent
Release Date2026-06-12
Versionv3.0.4
Base Versionv3.0.3

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Division 2 - SQA_Agent_Base Full Link Solution1. Added Division 2/SQA_Agent_Base/ full-link SQA manual process solution:   01-Requirements: SPEC.md, prompt.md, V-shaped development process.xlsx, automotive software comparison reference materials  02-Workflow: SQA-full-link-Workflow.drawio, manual Workflow table version/description, prompts  03-Project Plan: SQA-full-link-manual-process-organization-project-plan.md  04-Human-machine collaboration: SQA_Agent_human-machine-collaboration-proposal.md 2. Added SQA_Agent_Base meeting review materials (authoritative 37cefa8):   June 15 Monday conference presentation files (html/speech), difficulties and solutions, detailed development spec, V-shaped flow chart/three-party panorama 3. Added SQA-full-link_AINative-integrated-flowchart.drawio (0b12c6a)
Division 2 - Alter_Ego_Review_Agent POC1. Added Alter_Ego_Review_Agent/POC/ AI Alter Ego review Agent POC core process (51fd7e3):   SKILL.md, agent.py, base.py, plugin.py  core/: config.py, errors.py, types.py, workflow.py  parsers/: changelist_parser.py, document_parser.py  reporting/: artifact_writer.py, markdown_renderer.py, result_builder.py  services/: conclusion_engine.py, precheck_engine.py, requirement_context.py, knowledge_base.py  scripts/: build_sqlite_knowledge_base.py  配套测试: test_artifact_writer/changelist_parser/config/document_parser/precheck_conclusion/reporting/types/workflow 2. Added 03_SQLite_POC/ SQLite knowledge base construction: alter_ego_review_knowledge_poc.db, schema.sql, preview.xlsx (a1687c9)
Division 5 - SealCheck Signature Check Tool1. Added Division 5/MD/SealCheck/ signature check tool (245defd):   checkseal_app/: gui.py, main.py, seal_checker.py, config.py, excel_common.py, change_history_template.py  dist/CheckSeal.exe executable, config.json configuration  signature.xlsm sample file
Division 5 - MD review_master new check items1. Added multiple ModuleDesign check functions (b77e788, 90a9b41):   check_2_1_to_2_5 series: file list version matching, function graph name space check, function list consistency with related graph, red version corresponding sheet check  check_2_7 series: definition table description non-empty, element count corresponds to 2_8 sheet, template sheet deletion  check_2_8: title matches sheet number/variable name  check_2_9: definition table details non-empty  check_3x: function information matches 2_1~2_5 function list  check_variable_element series: numbering and counting, template and redundancy  check_sheet_top_module_info_matches_cover 2. Added multiple cover_sheet_image extraction artifacts (reports/extracted_artifacts/)

2.2 Changes ​

ModuleDetails
Division 5 - review-master-builder state machine normalization1. State machine specification, structured SKILL.md (added validation, constraints) 2. JSON extraction: fixed quote truncation when PowerShell passes inline Python code, changed double quotes to single quotes 3. Created permanent script to template when checklist_analysis.md updates to avoid creating temporary wheels every time 4. Mode 2 interaction display optimization, shows "you can..." to guide users 5. Added scripts: update_check_status.py, verify_batch.py, verify_extraction.py 6. create_review_tool moved from .py to .ps1 7. Added reference documents: ps51_inline_python_quoting.md, ps_file_array_params.md, shell_dollar_escaping.md
Division 2 - Alter_Ego_Review_Agent solution and process improvement1. Improved overall solution SPEC.md, human-machine collaboration process (drawio/md), interactive design review Workflow 2. Updated AI Alter Ego review Agent development plan, project analysis documentation 3. Updated memory.md memory document, data preparation requirements, full process description for database interaction 4. Fixed POC skill relative path handling and parsing and workflow testing (aa75542) 5. Refactored Code directory into Q&A-style subdirectories (40a9c43) 6. Reorganized SQLite POC product directory: 04_SQLite_POC -> 03_SQLite_POC (35ca81c), cleaned old directory (15ced91)
Division 1 - check-deliverables second edition1. check-deliverables-b updated binary and runtime configuration 2. Added script splitting: check_engine.py, content_extractor.py, ctx_extract_shape_snapshots.py, do_export_slides.py, read_selective.py, report_generator.py, run_check_wrapper.py 3. Added corresponding -win-x64.exe compiled file 4. Added README.md and Python source code to check-deliverables main directory 5. Added usage.md, packaging.md, report-format.md

2.3 Removed ​

ModuleDetails
Deleted Items1. Removed entire Division 2/Alter_Ego_Review_Agent/03_Template-and-Import/ directory: DB_FIELD_MAPPING.md, DB_IMPORT_FLOW.md, import template 2. Removed Division 2/Alter_Ego_Review_Agent/05_PostgreSQL-Delivery/ directory 3. Removed 03_SQLite_POC old running database and reports (15ced91) 4. Removed SQA manual full-link Workflow old files (replaced by new version) 5. Removed Division 5 create_review_tool.py (migrated to .ps1) 6. Removed Division 1 check-deliverables/scripts/parse_directory.cjs

2.4 Bug Fixes ​

ModuleDetails
Issue Fixes1. Fixed POC skill relative path handling and parsing and workflow testing (aa75542) 2. Fixed review-master-builder PowerShell inline Python quote truncation issue (2b179bd) 3. Fixed Alter_Ego_Review_Agent POC reporting/config/workflow related tests (a1687c9)

v3.0.3 ​

Release Date: 2026-06-05 | Base Version: v3.0.2

1. Overview ​

ItemContent
Release Date2026-06-05
Versionv3.0.3
Base Versionv3.0.2

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Division 2 - Requirement-Checker New Addition1. Added Division 2/Requirement-Checker/ requirement management table inspection Agent: Agent/Plugin framework: agent.py, plugin.py, agent.yaml, manifest.yaml inspection core library: lib/runner.py, lib/spec_contract.py, lib/reporting.py, lib/semantic_matching.py batch processing and tool gateway: scripts/requirement_checker_batch.py, scripts/export_user_report.py, scripts/tools_gateway.py supporting documents and tests: AGENTS.md, SKILL.md, requirement-management-table-inspection-spec.md, tests/test_requirement_checker.py 2. Added Division 2/Requirement-Checker-b/ independent executable version: bin/requirement-checker-b-win-x64.exe (413eccf5), can run without Python environment
Division 2 - Alter_Ego_Review_Agent initialization1. Added Division 2/Alter_Ego_Review_Agent/Code/, imported AI Alter Ego review Agent prototype (ba46607, a421efb): orchestration and workflow: core/orchestrator.py, core/workflow.py data access: db/sqlite_client.py, db/postgres_client.py, db/repositories.py skill modules: skills/parse_changelist.py, skills/retrieve_knowledge.py, skills/generate_questions.py, skills/score_review.py, skills/compose_report.py supporting tests: tests/test_design_review_workflow.py, tests/test_parse_changelist.py, tests/test_runtime_999d.py, tests/test_skill_entrypoint.py
Division 1 - 605 delivery snapshot and check-deliverables extension1. Added Division 1/605/ directory, supplemented check-deliverables and check-deliverables-b packaging and script delivery snapshot (3c26938) 2. Added/updated check-deliverables-b binary and runtime configuration, continuously iterates check-config.json and usage documentation
Division 5 - Review Master assets and template enhancement1. Added Division 5/MD/review_master/ example project, sample results and report materials 2. review-master-builder template additions: scripts/gen_check_items_html.py scripts/preview_checklist.ps1 assets/viewpoint_confirmation.html, assets/viewpoint_confirmation_template.html tracing directory base artifacts: tracing/check_generation_log.md, tracing/check_items.html, tracing/checklist_analysis.md
Version Comparison1. Version comparison: v3.0.2 -> v3.0.3 (2026-06-05) 2. Release interval: 2026-06-01 ~ 2026-06-05 3. Number of commits: 34 (v3.0.2..v3.0.3)

2.2 Changes ​

ModuleDetails
Requirement-Checker inspection logic continuous enhancement1. Supports merged display of same exception type locations in reports, result structure clearer 2. Responsible field check changed from single column to full column check within group 3. Added reverse consistency check: no need for design reason outputs warning when has value but design status not "no design required" 4. lib/reporting.py, lib/runner.py, lib/spec_contract.py and documentation/tests updated synchronously
check-deliverables structure reorganization1. check-deliverables reclaimed from old Agent plugin structure to lighter Skill + script organization 2. Kept and strengthened parse_directory.cjs and parse_directory-win-x64.exe paths 3. report-format.md returned to main directory, configuration file adjusted synchronously
review-master-builder output link optimization / Alter Ego skill entry improvement1. HTML generation stability optimization, template script and description updated synchronously 2. Trace files moved from project root to tracing/, facilitates centralized management of running artifacts 3. Startup link scripts bootstrap.ps1, run_review.ps1, start_review.bat updated 4. Skill entry and plugin skeleton completed, removed MCP shell dependency 5. README.md and test_skill_entrypoint.py updated, entry behavior clearer

2.3 Removed ​

ModuleDetails
Deleted Items1. Removed old Agent/Plugin related files under Division 1/check-deliverables/: agent.py, base.py, plugin.py, manifest.yaml context_manager.py, tool_manager.py tools/ full tool module prompt/SYSTEM.md, check-deliverables.spec 2. Removed Division 1/check-deliverables-b/check-config.template.json (configuration simplified) 3. Cleaned temporary lock files (like ~$*.xlsx) and some intermediate files

2.4 Bug Fixes ​

ModuleDetails
Issue Fixes1. Fixed missing image issue in Review Master report (0fef198) 2. Fixed Requirement-Checker responsible field judgment and exception prompt logic deviation (cd79f93) 3. Fixed and enhanced Review Master HTML generation stability (c8f769a) 4. Continuously fixed check-deliverables-b skill and packaging configuration availability (1df40e2, 413eccf5)

v3.0.2 ​

Release Date: 2026-05-29 | Base Version: v3.0.1

1. Overview ​

ItemContent
Release Date2026-05-29
Versionv3.0.2
Base Versionv3.0.1

2. Release Notes ​

2.1 New Features ​

ModuleDetails
Division 1 — check-deliverables-b1. Added Division 1/check-deliverables-b/: deliverable inspection B-version tool, as independent executable version of check-deliverables
Division 2 — ChangeList_Checker1. Added Division 2/ChangeList_Checker/: change point list (new variant table) 10-point inspection Skill
Division 5 — Review Master Matrix1. Added Division 5/: new division directory, hosts review master system for MT and UDS two business lines 2. Added skills/mt_review_master/: MT checklist item-by-item inspection implementation Skill 3. Added skills/uds_review_master/: UDS checklist item-by-item inspection implementation Skill 4. Added skills/review-master-builder/: generic Review Master project scaffolding generator 5. Added script/MT/review_master/: MT review master script version, contains 20+ specific check functions (file name format, confidentiality mark extraction, ASIL extraction, test outline matching, tool version check, etc.) 6. Added script/UDS/review_master_script/: UDS review master script version, contains 20+ UDS specific check functions (font color, BCI marking, floating evaluation shape, signal direction, RAM assignment, etc.) 7. Added MT/UDS checklist and deliverable data: MT/checklist/MT_checklist.xlsm, UDS/input/checklist/UDS_checklist.xlsm and multiple deliverable Excel samples
check-deliverables enhancement1. Added check-deliverables.spec: PyInstaller packaging specification file, supports packaging check-deliverables as Windows x64 independent executable 2. Added knowledge/HUD_20260522_検査報告.html: deliverable inspection report HTML template (666 lines), replaces original report-format.md 3. Added check-config.json / check-config.template.json updates: configuration-driven enhancement
filecompare_checker enhancement1. Added lib/orchestration/output_layout.py: unified output layout module, standardizes inspection result output format and directory structure

2.2 Changes ​

ModuleDetails
filecompare_checker orchestration layer comprehensive upgrade1. Memory-aware document gating (caca061): dynamically adjusts actual document concurrency based on system memory water level, supplements concurrency count, memory usage and skip reason during runtime 2. Unified _jobs and deliverable Log directory parsing: batch scripts and user report export use new directory structure 3. Enhanced evidence reuse between inspections: 3, 11, 12 checks reuse deterministic evidence, 6/7 check data linkage enhanced 4. claude_runtime.py: runtime logic adjusted (190+ lines changed) 5. concurrency.py: concurrency control logic refactored (118 lines changed) 6. executor.py: executor logic updated (61 lines changed) 7. inspection_logic.py: inspection logic significantly adjusted (218 lines changed) 8. paths.py: path parsing refactored (58 lines changed) 9. planner.py / reporting.py / user_reporting.py: planning and report modules updated synchronously 10. filecompare_batch.py: batch script significantly enhanced (445+ lines changed), supports new directory structure and memory-aware scheduling 11. export_user_report.py: user report export adapted to new layout
check-deliverables comprehensive refactoring1. agent.py / agent.md: Agent core logic and prompt rewritten (268+ lines changed) 2. context_manager.py: context manager updated 3. tool_manager.py: tool manager updated 4. SYSTEM.md: system prompt completely rewritten (335 lines changed) 5. knowledge.md: knowledge content significantly updated (220 lines changed) 6. All tool modules updated: config_tools.py, directory_tools.py, pptx_tools.py, report_tools.py, xlsx_tools.py, init.py all modified 7. parse_directory.cjs: directory parsing script refactored (245 lines changed) 8. manifest.yaml: version information updated
review-master-builder UTF-8 strategy upgrade1. SKILL.md and AGENT.md added UTF-8 First Policy (ed6b848): explicitly specifies UTF-8 as default encoding for all text interactions, standardizes PowerShell environment variable settings (PYTHONIOENCODING / Console.InputEncoding / Console.OutputEncoding), defines standard fallback path for terminal garbled (Python → UTF-8 temporary file → Read tool)

2.3 Removed ​

ModuleDetails
Deleted Items1. check-deliverables/SKILL.md: old Skill description file deleted (function inlined into agent.md / SYSTEM.md) 2. check-deliverables/knowledge/SoCAMET-C_DRIMOD-CSTD_VMID.drawio: old drawio diagram deleted (migrated to check-deliverables-b/knowledge/) 3. check-deliverables/report-format.md: old Markdown report format deleted (replaced by HTML template)

2.4 Bug Fixes ​

ModuleDetails
Issue Fixes1. Seal analysis did not fall back to cover screenshot when no seal screenshot (e8e4642): fixed inspection_runner seal analysis logic, correctly falls back to cover screenshot when seal screenshot does not exist 2. Temporary file and image processing issues (1bcdcae): fixed temporary file cleanup and image reference related issues 3. UTF-8 encoding garbled (ed6b848): fixed Chinese/Japanese output garbled issue on Windows terminal, unified encoding processing strategy through UTF-8 First Policy,优先 writes UTF-8 temporary file then reads, avoids relying on terminal rendering 4. Relative path reference issue (ed6b848): fixed incorrect relative path reference in project generated by review-master-builder, tests passed 5. bat startup script issue (ed6b848): fixed compatibility issues of batch startup scripts such as start_review.bat, tests passed

v3.0.1 ​

Release Date: 2026-05-22 | Base Version: v3.0.0

1. Overview ​

ItemContent
Release Date2026-05-22
Versionv3.0.1
Base Versionv3.0.0

2. Release Notes ​

2.1 New Features ​

ModuleDetails
filecompare_checker / inspection system1. Inspection items expanded from 11 to 12, added "inspection 11 Filecompare manual matching result consistency check" and "inspection 12 requirement-code-FileCompare three-party consistency check (LLM judgment)" 2. Added evidence connection specification between inspections: fixed analysis order 1→8→9→10→11→12→2→3→4→5→6→7, display/report forced natural order 1→12 3. Added query language hard judgment rule: judges after removing path/sheet name/source document original text, explicitly passed via --query-language, prohibits model guessing 4. Report saving stage added "original text support for reasoning" required field, missing original text always treated as ⚠️ manual confirmation
filecompare_checker / runtime1. Integrated AIKO MCP plugin, supports AIKO_MODEL_SELECTOR / AIKO_MODELS_JSON / AIKO_LLM_MODEL / AIKO_LLM_API_KEY / AIKO_LLM_BASE_URL overlay 2. Added shared environment variable configuration system (SYSTEM_ENV_SETUP.md), separates LLM / OpenAI / Anthropic / GitLab / AIKO five connection layers 3. Added runtime memory persistence mechanism: forced writing to docs/<doc_id>/artifacts/memory/document_memory.json and inspection_<n>.memory.json, avoids rule loss due to long context compression 4. Added runtime heartbeat reporting (LLM_HEARTBEAT_INTERVAL_SECONDS) 5. Added automatic document concurrency adjustment based on memory
filecompare_checker / orchestration layerAdded lib/orchestration/ multi-Agent parallel orchestration framework, includes 18 modules such as adapters / anthropic_compat / canonical / claude_prompts / claude_runtime / concurrency / discovery / executor / inspection_logic / job_reporting / language / llm_runtime / planner / reporting / runtime_config / user_reporting
filecompare_checker / toolset1. Added GitLab evidence collector gitlab_evidence_collector.py, archives evidence by hash value and establishes hash ↔ requirement ID/Commit/code difference correspondence 2. Added Jira toolset: jira_get_custom_fields / jira_get_issue / jira_list_projects / jira_search 3. Added COM tools and Excel formula scanning: com_utils / excel_formula_scan / excel_row_extractor / runtime_stop 4. Added batch scripts: filecompare_batch / runtime_worker / export_user_report / verify_runtime_env
Project ArchitectureRestructured directory structure by division: Division 1 / Division 2 / Division 5, original filecompare_checker moved into Division 2/
Division 1 / check-deliverablesAdded deliverable (PPT/Excel/DRBFM) dynamic checklist check Skill, supports check-config.json configuration-driven, contains config_tools / directory_tools / pptx_tools / report_tools / xlsx_tools toolset
Division 1 / ct-checklistAdded unit test specification 8-item inspection Skill: cover sheet signature and date order, review/approval, title consistency, update history SHM ticket number, requirement_interface correspondence integrity, test case list formula error, result column and document path consistency
Division 1 / formula-r-checklistAdded formula R PPT V01–V33 complete 33-item inspection Skill, grouped by V01–V07 / V08–V09 / V10–V21 / V22 / V23–V31 / V32–V33
Division 1 / assistantAdded generic Agent runtime framework, includes complete modules: agent / base / context_manager / env_loader / runtime_paths / plugin / tool_manager

2.2 Changes ​

ModuleDetails
filecompare_checker / inspection 10Judgment method changed from "similarity scoring" to "LLM semantic judgment matching", canceled scoring system
filecompare_checker / inspection 12Report generator changed from rule matching to LLM judgment
filecompare_checker / language consistencyUpgraded from advisory rule to MANDATORY hard rule: framework text must use query language, data original retains source language; report file name follows query language (three sets of naming: Chinese/English/日本語)
filecompare_checker / runtime configurationruntime block moved from environment variables to agent.yaml (default_runtime / fallback_runtime / trace_cleanup_* / llm_heartbeat_interval_seconds / provider / openai.*), separates project defaults from shared values
filecompare_checker / parallel AgentChanged from built-in parallel logic to external generic parallel Agent package, decoupled and reimplemented internal parallel logic
filecompare_checker / GitLabUpdated GitLab server address and personal api_token, added GITLAB_SSL_VERIFY certificate verification switch

2.3 Removed ​

ModuleDetails
code_compareOriginal code comparison Skill top-level module completely deleted, capability merged into filecompare_checker GitLab evidence collection and inspection 11/12
filecompare_checker_DOriginal "version D" directory and commits_with_diff.txt and other attached files deleted
OtherSKILL.md.bak old backup, .idea/ IDE configuration directory removed from repository; requirements.runtime.txt dependencies tightened to only runtime requirements

2.4 Bug Fixes ​

ModuleDetails
Inspection 10Inspection 10 always failed to hit resulting in "⚠️ manual confirmation" — fixed after底层 replaced with LLM semantic judgment
Report LanguageMixed Chinese/Japanese in report — fixed by query language hard judgment + fixed label translation table + file name follows language
Context CompressionHard rules lost after long context compression — fixed by formal memory persistence to docs/<doc_id>/artifacts/memory/, prohibits using traces/ or tmp/ as only memory source
Inspection 12Similarity score BUG — fixed after report generator changed to LLM judgment
Inspection 3 / Inspection 4Seal detection cannot find seal / recognition error — fixed by seal context screenshot tool
Score CalculationFixed scoring logic error
Language / Report ModuleFixed language module judgment and report output format BUG, improved report output
GitLabFixed GitLab certificate verification failure issue (via GITLAB_SSL_VERIFY configuration)
Performance & PauseFixed too slow running and pause function BUG
Original Text FidelityFixed issue where inspection document original text was modified in report — inspection original must be word-for-word consistent with document source language
SKILL Semantic CompletenessInspection 11 originally depended on external description file caused SKILL own semantic incomplete, now external description inlined into SKILL.md
openpyxl CompatibilityAdded _openpyxl_patch.py patch solves specific openpyxl version compatibility issue

v3.0.0 ​

Release Date: May 15, 2026 | Base Version: -

1. Overview ​

ItemContent
Release DateMay 15, 2026
Versionv3.0.0
Base Version-

2. Release Notes ​

2.1 New Features ​

ModuleDetails
System Monitoring1. Completed task lifecycle management
2. Completed key operation logging
3. Completed key operation exception handling
Persistence1. Completed workspace persistence
2. Completed user session recovery functionality
Permission Management1. Completed user management functionality
2. Completed role authentication functionality
3. Completed permission management functionality
GUICompleted graphical interface for user management, role management, system settings, file management, persistence functionality, knowledge base, and Q&A assistant
Basic Function EngineCompleted MCP client integration

2.2 Changes ​

ModuleDetails
Knowledge Base1. Completed built-in knowledge base functionality
2. Completed knowledge base file management functionality
EB IntegrationCompleted EB automatic project opening, project-included script execution, and EB plugin invocation

2.3 Removed ​

ModuleDetails
NoneNone

2.4 Bug Fixes ​

ModuleDetails
API Gateway1. xxx
2. xxx

SDW AIKO Product Documentation

© 2026 SDW AIKO. All rights reserved.