New BrowserRenderer class uses Playwright to render JavaScript-heavy
documentation sites (React, Vue SPAs) that return empty HTML shells
with requests.get(). Activated via --browser flag on web scraping.
- browser_renderer.py: Playwright wrapper with lazy browser launch,
auto-install Chromium on first use, context manager support
- doc_scraper.py: browser_mode config, _render_with_browser() helper,
integrated into scrape_page() and scrape_page_async()
- SPA detection warnings now suggest --browser flag
- Optional dep: pip install "skill-seekers[browser]"
- 14 real e2e tests (actual Chromium, no mocks)
- UML updated: Scrapers class diagram (BrowserRenderer + dependency),
Parsers (DoctorParser), Utilities (Doctor), Components, and new
Browser Rendering sequence diagram (#20)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
8 diagnostic checks: Python version (3.10+), package install, git,
14 core deps, 10 optional deps, API keys, MCP server, output dir.
Each check reports pass/warn/fail with --verbose for extra detail.
Exit code 0 if no critical failures, 1 otherwise.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New bundled workflow `prompt-injection-check` scans scraped content for
prompt injection patterns (role assumption, instruction overrides,
delimiter injection, hidden instructions, encoded payloads) using AI.
Flags suspicious content without removing it — preserves documentation
accuracy while warning about adversarial content. Added as first stage
in both `default` and `security-focus` workflows so it runs automatically
with --enhance-level >= 1.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PR #326 added _validate_name() to create/update/delete but missed
get_workflow_tool, which would raise an unhandled ValueError instead of
returning a user-friendly error. Added try/except handling and 6 tests
covering all 4 tool functions with malicious names.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two call sites (_run_c3_analysis in unified_scraper.py and _analyze_c3x in
unified_codebase_analyzer.py) still passed the old enhance_with_ai and ai_mode
kwargs which were replaced by enhance_level. This caused a TypeError when
running C3.x codebase analysis.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PyGithub's get_languages() returns raw API JSON which in some environments
includes non-integer metadata keys (e.g., "url"), causing a TypeError in
sum(). Now filters to integer values only before calculating percentages.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The scraper previously reported len(visited_urls) as "Scraped N pages"
even when save_page() silently skipped pages with empty content (<50
chars). For JavaScript SPA sites this meant "Scraped 190 pages" followed
by "No scraped data found!" with no explanation.
Changes:
- Added pages_saved/pages_skipped counters to DocToSkillConverter
- save_page() now increments pages_skipped on skip, pages_saved on save
- New _log_scrape_completion() reports "(N saved, M skipped)" breakdown
- SPA detection warns when all/most pages have empty content
- build_skill() error now explains empty content cause when pages skipped
- Updated both sync and async scrape completion paths
- 14 new tests across 4 test classes (counting, messages, SPA, build)
Fixes#320
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase 1 - OpenCode Integration:
- Add OpenCodeAdaptor with directory-based packaging and dual-format YAML frontmatter
- Kebab-case name validation matching OpenCode's regex spec
Phase 2 - OpenAI-Compatible LLM Platforms:
- Extract OpenAICompatibleAdaptor base class from MiniMax (shared format/package/upload/enhance)
- Refactor MiniMax to ~20 lines of constants inheriting from base
- Add 6 new LLM adaptors: Kimi, DeepSeek, Qwen, OpenRouter, Together AI, Fireworks AI
- All use OpenAI-compatible API with platform-specific constants
Phase 3 - CLI Agent Expansion:
- Add 7 new install-agent paths: roo, cline, aider, bolt, kilo, continue, kimi-code
- Total agents: 11 -> 18
Phase 4 - Advanced Features:
- OpenCode skill splitter (auto-split large docs into focused sub-skills with router)
- Bi-directional skill format converter (import/export between OpenCode and any platform)
- GitHub Actions template for automated skill updates
Totals: 12 --target platforms, 18 --agent paths, 2915 tests passing
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Python 3.14's urlparse() raises ValueError on URLs with unencoded
brackets that look like malformed IPv6 (e.g. http://[fdaa:x:x:x::x
from docs.openclaw.ai llms-full.txt). sanitize_url() called urlparse()
BEFORE encoding brackets, so it crashed before it could fix them.
Fix: catch ValueError from urlparse, encode ALL brackets, then retry.
This is safe because if urlparse rejected the brackets, they are NOT
valid IPv6 host literals and should be encoded anyway.
Also fixed Discord e2e tests to skip gracefully on network issues.
Fixes#284
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous fix (a82cf69) only addressed anchor fragment stripping but
left the fundamental problem: _convert_to_md_urls() blindly appended
/index.html.md to ALL non-.md URLs from llms.txt. This only works for
Docusaurus sites — for sites like Discord docs it generates mass 404s.
Changes:
- _convert_to_md_urls() now strips anchors and deduplicates only,
preserving original URLs as-is instead of appending /index.html.md
- New _has_md_extension() helper uses urlparse().path.endswith(".md")
instead of error-prone ".md" in url substring matching
- Fixed ".md" in url checks at 4 locations (lines 465, 554, 716, 775)
- Removed 24 lines of dead commented-out code
- Added real-world e2e test against docs.discord.com (no mocks)
- Updated unit tests for new behavior (32 tests)
Fixes#277
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The unified scraper's _scrape_documentation() was creating temp configs
in flat/legacy format (no "sources" key), causing doc_scraper's
ConfigValidator to reject them. Wrap the temp config in unified format
with a "sources" array. Also remove dead code branches and fix a
pre-existing test that didn't clear GITHUB_TOKEN from env.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add MiniMax AI as LLM platform adaptor
Original implementation by octo-patch in PR #318.
This commit includes comprehensive improvements and documentation.
Code Improvements:
- Fix API key validation to properly check JWT format (eyJ prefix)
- Add specific exception handling for timeout and connection errors
- Remove unused variable in upload method
Dependencies:
- Add MiniMax to [all-llms] extra group in pyproject.toml
Tests:
- Remove duplicate setUp method in integration test class
- Add 4 new test methods:
* test_package_excludes_backup_files
* test_upload_success_mocked (with OpenAI mocking)
* test_upload_network_error
* test_upload_connection_error
* test_validate_api_key_jwt_format
- Update test_validate_api_key_valid to use JWT format keys
- Fix test assertions for error message matching
Documentation:
- Create comprehensive MINIMAX_INTEGRATION.md guide (380+ lines)
- Update MULTI_LLM_SUPPORT.md with MiniMax platform entry
- Update 01-installation.md extras table
- Update INTEGRATIONS.md AI platforms table
- Update AGENTS.md adaptor import pattern example
- Fix README.md platform count from 4 to 5
All tests pass (33 passed, 3 skipped)
Lint checks pass
Co-authored-by: octo-patch <octo-patch@users.noreply.github.com>
* fix: improve MiniMax adaptor — typed exceptions, key validation, tests, docs
- Remove invalid "minimax" self-reference from all-llms dependency group
- Use typed OpenAI exceptions (APITimeoutError, APIConnectionError)
instead of string-matching on generic Exception
- Replace incorrect JWT assumption in validate_api_key with length check
- Use DEFAULT_API_ENDPOINT constant instead of hardcoded URLs (3 sites)
- Add Path() cast for output_path before .is_dir() call
- Add sys.modules mock to test_enhance_missing_library
- Add mocked test_enhance_success with backup/content verification
- Update test assertions for new exception types and key validation
- Add MiniMax to __init__.py docstrings (module, get_adaptor, list_platforms)
- Add MiniMax sections to MULTI_LLM_SUPPORT.md (install, format, API key,
workflow example, export-to-all)
Follows up on PR #318 by @octo-patch (feat: add MiniMax AI as LLM platform adaptor).
Co-Authored-By: Octopus <octo-patch@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: octo-patch <octo-patch@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace → (U+2192) with -> in argparse help strings. Windows cmd uses
cp1252 encoding which cannot render unicode arrows, causing --help to
crash with UnicodeEncodeError.
Fix CI failures: version hardcoded in _version.py fallbacks and
test assertions (test_package_structure, test_cli_paths) still
referenced 3.2.0 after the version bump.
PyGithub's PaginatedList slicing (issues[:max_issues]) may fail with
'list index out of range' on some PyGithub versions or when repos
have no issues. Replace with itertools.islice() which works reliably
with any iterable, including PaginatedList.
Bug reported by @dream0438-cmd in PR #269.
Closes#269
## Summary
Add `skill-seekers sync-config` subcommand that crawls a docs site's navigation,
diffs discovered URLs against a config's start_urls, and optionally writes the
updated list back with --apply.
- BFS link discovery with configurable depth (default 2), max-pages, rate-limit
- Respects url_patterns.include/exclude from config
- Supports optional nav_seed_urls config field
- Handles both unified (sources array) and legacy flat config formats
- MCP tool sync_config included
- 57 tests (39 unit + 18 E2E with local HTTP server)
- Fixed CI: renamed summary job to "Tests" to match branch protection rule
Closes#306
The original fix (741daf1) only patched LlmsTxtParser._clean_url(),
which covers URLs extracted directly from llms.txt content. But URLs
discovered from .md files during BFS crawl (_extract_markdown_content)
and from HTML pages (extract_content) bypass _clean_url() entirely.
When those pages contain links with square brackets (e.g.
/api/[v1]/users), httpx raises 'Invalid IPv6 URL' on fetch.
Fix: add a shared sanitize_url() utility in cli/utils.py that
percent-encodes [ and ] in path/query components, and apply it at
every URL ingestion point:
- _enqueue_url(): main chokepoint — all discovered URLs pass through
- scrape_page(): safety net for start_urls that skip _enqueue_url
- scrape_page_async(): same for async mode
- dry-run sync/async paths: direct fetches that also bypass _enqueue_url
LlmsTxtParser._clean_url() now delegates bracket-encoding to the
shared sanitize_url() (DRY), keeping only its malformed-anchor
stripping logic.
Added 16 tests: sanitize_url unit tests, _clean_url bracket tests,
_enqueue_url sanitization tests, and integration test verifying
markdown content with bracket URLs is handled safely.
Fixes#284
Follow-up to PR #309 (perf: optimize with caching, pre-compiled regex,
O(1) lookups, and bisect line indexing). These fixes were committed to
the PR branch but missed the squash merge.
Review fixes (credit: PR #309 by copperlang2007):
1. Rename _pending_set -> _enqueued_urls to accurately reflect that the
set tracks all ever-enqueued URLs, not just currently pending ones
2. Extract duplicated _build_line_index()/_offset_to_line() into shared
build_line_index()/offset_to_line() in cli/utils.py (DRY)
3. Fix pre-existing bug: infer_categories() guard checked 'tutorial'
but wrote to 'tutorials' key, risking silent overwrites
4. Remove unnecessary _store_results() closure in scrape_page()
5. Simplify parser pre-import in codebase_scraper.py
Benchmark stabilization:
- test_benchmark_metadata_overhead was flaky on CI (106.7% overhead
observed, threshold 50%) because 5 iterations with mean averaging
can't reliably measure microsecond-level differences
- Fix: 20 iterations, warm-up run, median instead of mean, threshold
raised to 200% (guards catastrophic regression, not noise)
Ref: https://github.com/yusufkaraaslan/Skill_Seekers/pull/309
Update version across pyproject.toml, _version.py fallbacks,
CHANGELOG.md, and README badges for v3.2.0 release.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Skip OCR on WEBCAM/OTHER frames (eliminates ~64 junk results per video)
- Add _clean_ocr_line() to strip line numbers, IDE decorations, collapse markers
- Add _fix_intra_line_duplication() for multi-engine OCR overlap artifacts
- Add _is_likely_code() filter to prevent UI junk in reference code fences
- Add language detection to get_text_groups() via LanguageDetector
- Apply OCR cleaning in _assemble_structured_text() pipeline
- Add two-pass AI enhancement: Pass 1 cleans reference Code Timeline
using transcript context, Pass 2 generates SKILL.md from cleaned refs
- Update video-tutorial.yaml prompts for pre-cleaned references
- Add 17 new tests (197 total video tests), 2540 tests passing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- UP024: Replace `(OSError, IOError)` with `OSError` in video_setup.py
- E402: Use existing `os` import instead of `import os as _os` in video_visual.py
- SIM103: Inline condition in `_detect_gpu()` return statement
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Auto-detects NVIDIA (CUDA), AMD (ROCm), or CPU-only GPU and installs the
correct PyTorch variant + easyocr + all visual extraction dependencies.
Removes easyocr from video-full pip extras to avoid pulling ~2GB of wrong
CUDA packages on non-NVIDIA systems.
New files:
- video_setup.py (835 lines): GPU detection, PyTorch install, ROCm config,
venv checks, system dep validation, module selection, verification
- test_video_setup.py (60 tests): Full coverage of detection, install, verify
Updated docs: CHANGELOG, AGENTS.md, CLAUDE.md, README.md, CLI_REFERENCE,
FAQ, TROUBLESHOOTING, installation guide, video dependency plan
All 2523 tests passing (15 skipped).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix extract_visual_data returning 2-tuple instead of 3 (ValueError crash)
- Move pytesseract from core deps to [video-full] optional group
- Add 30-min timeout + user feedback to video enhancement subprocess
- Add scrape_video_impl to MCP server fallback import block
- Detect auto-generated YouTube captions via is_generated property
- Forward --vision-ocr and --video-playlist through create command
- Fix filename collision for non-ASCII video titles (fallback to video_id)
- Make _vision_used a proper dataclass field on FrameSubSection
- Expose 6 visual params in MCP scrape_video tool
- Add install instructions on missing video deps in unified scraper
- Update MCP docstring tool counts (25→33, 7 categories)
- Add video and word commands to main.py docstring
- Document video-full exclusion from [all] deps in pyproject.toml
- Update parser registry test count (22→23 for video parser)
All 2437 tests passing, 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sync with latest development changes including ruff formatting,
bug fixes, and pinecone adaptor additions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
DocToSkillConverter has self.skill_dir (string), not self.output_dir.
The --chunk-for-rag flag on scrape command crashed with AttributeError.
Changed to Path(converter.skill_dir).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add FALLBACK_MAIN_SELECTORS constant and _find_main_content() helper to
eliminate 3 duplicated fallback loops in doc_scraper.py
- Move link extraction before early return in extract_content() so links
are always discovered from the full page, not just main content
- Fix single-threaded dry-run to extract links from soup (full page)
instead of main element only — fixes reactflow.dev finding only 1 page
- Add link extraction to async dry-run path (was completely missing)
- Remove main_content from get_configuration() defaults so fallback logic
kicks in instead of a broad CSS comma selector matching body
- Smart create --config routing: peek at JSON to determine unified
(sources array → unified_scraper) vs simple (base_url → doc_scraper)
- Update docs/user-guide/02-scraping.md and docs/reference/CONFIG_FORMAT.md
to use unified config format (legacy format rejected since v2.11.0)
- Fix test_auto_fetch_enabled and test_mcp_validate_legacy_config
Closes#300
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Stage 1 quality improvements from the Arbitrary Limits & Dead Code audit:
Reference file truncation removed:
- codebase_scraper.py: remove code[:500] truncation at 5 locations — reference
files now contain complete code blocks for copy-paste usability
- unified_skill_builder.py: remove issues[:20], releases[:10], body[:500],
and code_snippet[:300] caps in reference files — full content preserved
Enhancement summarizer rewrite:
- enhance_skill_local.py: replace arbitrary [:5] code block cap with
character-budget approach using target_ratio * content_chars
- Fix intro boundary bug: track code block state so intro never ends
inside a code block, which was desynchronizing the parser
- Remove dead _target_lines variable (assigned but never used)
- Heading chunks now also respect the character budget
Hardcoded language fixes:
- unified_skill_builder.py: test examples use ex["language"] instead of
always "python" for syntax highlighting
- how_to_guide_builder.py: add language field to HowToGuide dataclass,
set from workflow at creation, used in AI enhancement prompt
Test fixes:
- test_enhance_skill_local.py: rename test to test_code_blocks_not_arbitrarily_capped,
fix assertion to count actual blocks (```count // 2), use target_ratio=0.9
Documentation:
- Add Stage 1 plan, implementation summary, review, and corrected docs
- Update CHANGELOG.md with all changes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix#299: rename --chunk-size/--chunk-overlap to --streaming-chunk-size/
--streaming-overlap in arguments/package.py to avoid collision with the
RAG --chunk-size flag from arguments/common.py
- Phase 1a: make package_skill.py import args via add_package_arguments()
instead of a 105-line inline duplicate argparse block; fixes the root
cause of _reconstruct_argv() passing unrecognised flag names
- Phase 1b: centralise setup_logging() into utils.py and remove 4
duplicate module-level logging.basicConfig() calls from doc_scraper.py,
github_scraper.py, codebase_scraper.py, and unified_scraper.py
- Fix test_package_structure.py / test_cli_paths.py version strings
(3.1.1 → 3.1.2)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix 1 — gemini.py: replace deprecated gemini-2.0-flash-exp (404 errors)
with gemini-2.5-flash (stable, GA, Google's recommended replacement).
Closes#290.
Fix 2 — enhance dispatcher: implement the documented auto-detection that
was missing from the code. skill-seekers enhance now correctly routes:
- ANTHROPIC_API_KEY set → Claude API mode (enhance_skill.py)
- GOOGLE_API_KEY set → Gemini API mode
- OPENAI_API_KEY set → OpenAI API mode
- No API keys → LOCAL mode (Claude Code Max, free)
Use --mode LOCAL to force local mode even when an API key is present.
9 new tests cover _detect_api_target() priority logic and main()
routing (API delegation, --mode LOCAL override, no-key fallback).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
All scrapers (scrape, github, analyze, pdf) now share a common argument
contract via add_all_standard_arguments() in arguments/common.py.
Universal flags (--dry-run, --verbose, --quiet, --name, --description,
workflow args) work consistently across all source types.
Previously, `create <url> --dry-run`, `create owner/repo --dry-run`,
and `create ./path --dry-run` would crash because sub-scrapers didn't
accept those flags. Also fixes main.py _handle_analyze_command() not
forwarding --dry-run, --preset, --quiet, --name, --description to
codebase_scraper.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The create command crashed with 'Namespace' object has no attribute
'max_pages' because it accessed args.max_pages directly instead of
using getattr() like all other source-specific attributes in the
same method.
Closes#293
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
