2026.04.05.3
Changes
- Reddit comment insights — top comments with scores, authors, and excerpts for highest-scoring results
- X/Twitter search with full engagement data — cookie-based GraphQL search with likes, reposts, replies
- Two-phase search — Phase 1 extracts entities (subreddits, handles, authors), Phase 2 does targeted follow-ups
- HN channel now includes
num_commentsin metadata - Claims pipeline carries engagement scores and cross-platform signals through to synthesis
- Reranking boosts cross-source convergent items as stronger evidence
- Channel selection considers query type (how-to, comparison, opinion, etc.)
Fixes
- CI author check no longer fails on merge commits from
actions/checkout
2026.04.05.2
Changes
- Per-platform engagement scoring with composite scores (relevance + recency + engagement)
- Cross-source convergence detection with “also on” annotations
- Search query type classification for smarter source prioritization
2026.04.05.1
Changes
- HuggingFace channel — search models and datasets by downloads
- Papers with Code channel — trending AI papers via HuggingFace Daily Papers
- 34 search channels total (was 32)
- Cleaned up 18 orphan skills (1,828 lines removed)
2026.04.04.9
Changes
- CI hygiene workflow — checks author identity, codenames, personal paths, container pins
- Git history fully rewritten — removed all PII from 270 commits
Fixes
- Removed literal PII from .gitleaks.toml detection rules
- Cleaned up 10 stale remote branches
2026.04.04.8
Changes
- Lazy-load channel plugins on first query for faster startup
- Added isort and bugbear rules to ruff linter
- Pyright now checks the correct directories
Fixes
- Pinned semgrep container image to SHA digest
- Fixed shell injection risk in release.yml
- Added .dockerignore to prevent secret leakage into Docker builds
- HTML-escaped title in report assembly to prevent XSS
- Added
set -euo pipefailto run_search.sh
2026.04.04.7
Changes
- Engine-level health tracking — dependent channels suspend together on engine failure
- Transient failures auto-retry once before giving up
- Channel health data persists immediately after every success/failure
2026.04.04.6
Fixes
- Search failures correctly distinguished from empty results
2026.04.04.5
Fixes
- Chinese channels no longer get suspended by English-session zero-yield
2026.04.04.4
Changes
- Pre-release test suite with
./scripts/release-test.sh - 6 user scenarios (Quick/Standard/Deep x EN/ZH) with Docker support
2026.04.04.3
Fixes
- Claims compression enforced to prevent Block 4 hangs at 120+ results
- All 6 pipeline blocks now have time limits
2026.04.04.2
Changes
- Haiku-compressed claims before synthesis
- Targeted gap-fill queries from coverage analysis
- Research context persistence across sessions
2026.04.04.1
Changes
- Delivery format choice: Markdown, Rich HTML, or Slides
- Language auto-detection — Chinese topics get Chinese output
- 6-block orchestrated pipeline with real-time progress
- 9 to 279 tests
2026.4.3
Changes
- Plugin release — install via
claude plugin marketplace add 0xmariowu/autosearch - 32 search channels with convention-based loader
- Chinese channels via Baidu Kaifa Developer Search
- Two-stage citation lock
- Model routing: Haiku for batch, Sonnet for synthesis
- Benchmark: AutoSearch 92% vs Native Claude 72%
2026.4.1
Changes
- Rubric AVO — auto-generate rubrics, score delivery, evolve skills
- Pipeline baseline: 0.880 pass rate on initial benchmark