Skip to content

diaznakh/researchcomb

v0.5.23MIT

Search indexes, verify claims against sources, and produce evidence-backed research documents.

GitHub

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project uses three-part version numbers of the form major.minor.patch as stored in plugin.json and skills/researchcomb/VERSION.

Entries below are taken from git history and the GitHub Releases for tags v0.5.21, v0.5.22, and v0.5.23. Dates are the annotated-tag dates.

Unreleased

Added

  • Opt-in on-disk cache for safe_fetch.py (--cache / RESEARCHCOMB_CACHE=1), stored under $XDG_CACHE_HOME/researchcomb (or ~/.cache/researchcomb; %LOCALAPPDATA%\researchcomb on Windows), keyed by the final validated URL after redirects, with a 24-hour TTL. RESEARCHCOMB_CACHE_DIR overrides the location. Cache hits copy into a fresh output file. --no-cache / RESEARCHCOMB_NO_CACHE=1 forces a new fetch. The cache is not read or written when --allow-private is on. Error responses, HTTP 429, non-2xx, and Cache-Control: no-store are not stored.

  • CI installs pinned jsonschema==4.26.0 so the evidence-schema cross-check runs on GitHub Actions.

  • Keep a Changelog covering tags v0.5.21, v0.5.22, and v0.5.23 from git history and the GitHub Releases for those tags.

  • Live GitHub Actions status badge for .github/workflows/test.yml in the README header.

  • CONTRIBUTING.md, .github/ISSUE_TEMPLATE/ issue forms, and .github/pull_request_template.md.

  • Draft 2020-12 JSON Schema for evidence.json at skills/researchcomb/references/evidence.schema.json, linked from evidence-record.md, with a standard-library test that validates every fixture evidence.json. Closed objects (source, claim, evidence items, search entries, citation trails, trail leads, and search rounds) set additionalProperties to false. Schema tests cross-check with jsonschema when that library is importable and include invalid evidence fixtures.

Changed

  • Rename focused workflows to the comb metaphor and fold similar commands: add comb-part, comb-snag, comb-sweep, and comb-weave; merge comb-blueprint and comb-rerun into comb-trace; keep comb-align, comb-blueprint, comb-critique, comb-manuscript, comb-rerun, and comb-survey as deprecated aliases. Result is eleven workflows plus six aliases (#8).
  • Pin README install instructions to ResearchComb v0.5.23 at commit c43d1dcaef4d930c463fe51ab31ae60a1999fc73 (#4).
  • Reframe comb-paper for academic integrity: remind users to check institutional policy, refuse producing a finished coursework essay for submission, and append an ## AI Assistance Disclosure section (#5, #6).
  • Rewrite focused-workflow descriptions in ResearchComb's own voice (#7).
  • Update the README local-checker count to 32 after adding a disclosure word-count test (#6).

0.5.23 - 2026-10-11

GitHub Release: ResearchComb v0.5.23.

Merged pull requests listed on that release:

  • #1 Fix word counting, HTTP headers, stale docs, test coverage, and add 3 owner-approved features
  • #2 Release v0.5.23: Fix proxy support, accept HTTP 2xx, improve guidance
  • #3 Security hardening: DNS rebinding fix, API etiquette, prompt injection guidance, CI

Added

  • Opt-in sandbox/proxy fetch mode (--allow-private / RESEARCHCOMB_ALLOW_PRIVATE=1) that honours HTTP_PROXY, HTTPS_PROXY, and NO_PROXY and allows non-global resolved addresses; record "fetch_mode": "allow_private" on the source when used.
  • tests/local_proxy.py helper for CONNECT-proxy testing.
  • Search checker acceptance of API hosts for Crossref (api.crossref.org), OpenAlex (api.openalex.org), Semantic Scholar (api.semanticscholar.org), Europe PMC (www.ebi.ac.uk/europepmc), and PubMed (eutils.ncbi.nlm.nih.gov).
  • PASS line that reports searched, inaccessible, and unavailable counts separately.
  • comb-check steps to identify document version (preprint, accepted manuscript, version of record) and compare a DOI's reference list with Crossref metadata.
  • Optional evidence-record fields documented in the source table: checked_date, source_type, commit, version, and reference_issues.
  • comb-paper case in tests/check_results.py.
  • GitHub Actions workflow .github/workflows/test.yml (push and pull request to main, Python 3.11, python3 -m unittest discover -s tests -p 'test_*.py', contents: read, actions pinned by commit SHA).
  • SECURITY.md with private vulnerability reporting, supported versions, scope, and known limitations.
  • Prompt-injection guidance: treat fetched pages, PDFs, repositories, and search results as untrusted data, not instructions.
  • API etiquette notes in the search-source guide (Crossref mailto, OpenAlex API key after the polite-pool change, NCBI rate limits, Semantic Scholar limits, backoff on HTTP 429).

Changed

  • Manuscript word counting strips table pipe characters and recognizes additional bibliography heading aliases, including Works Cited.
  • safe_fetch.py sends User-Agent: ResearchComb.
  • Documentation skill and test counts updated to thirteen workflow skills.
  • Root .gitignore patterns anchored so tracked fixtures such as tests/fixtures/workflows/evidence.json are not ignored.
  • Version set to 0.5.23 in plugin.json and skills/researchcomb/VERSION.

Fixed

  • safe_fetch.py proxy support, including HTTPS targets through an HTTP CONNECT proxy; HTTPS proxy URLs remain rejected.
  • HTTP 2xx responses accepted (not only 200).
  • Stub detection and temporary-file handling in safe_fetch.py.
  • Proxy environment variables no longer disable private-address checks outside opt-in mode.
  • Europe PMC direct URLs must use the /europepmc path prefix.
  • README and other stale documentation corrections from the v0.5.23 release commits.

Security

  • Connections pin to the validated address after DNS (including --allow-private non-proxy paths) to prevent DNS rebinding.
  • --allow-private continues to block loopback (127.0.0.0/8, ::1), link-local / cloud metadata (169.254.0.0/16, fe80::/10), and unspecified addresses (0.0.0.0/8, ::). The unspecified-address block covers the full 0.0.0.0/8 range, not only 0.0.0.0.

0.5.22 - 2026-10-09

GitHub Release: ResearchComb v0.5.22 (compare link only: v0.5.21...v0.5.22).

Changed

  • comb-critique requires checking each central claim against its main-text location, accessible supplementary material, and the cited source; classify findings as demonstrated error, unsupported claim, or concern needing a check; add ML-specific inspection items; and end with an inspection inventory of supplementary material, data, code, and cited sources.
  • Version set to 0.5.22 in plugin.json and skills/researchcomb/VERSION.
  • README install pin updated to v0.5.22.

0.5.21 - 2026-10-09

GitHub Release: ResearchComb v0.5.21 (compare link only: v0.5.20...v0.5.21).

Added

  • paper-search mode in check_completion.py for open-ended comb-paper discovery.
  • Evidence-record citation_trails, trail_leads, and search_rounds for paper search: trail kind is references, cited_by, or related; trail outcome is screened or inaccessible (with a reason when inaccessible); each trail lead has a title and status included (with source_id), excluded (with a reason), or inaccessible; each search round has a nonnegative new_relevant count, and the final count is zero only after screening encountered relevant leads.
  • Completion-gate tests that require all eleven scholarly source attempts and screened citation trails.

Changed

  • Open-ended comb-paper discovery attempts every scholarly source in the search guide (with a separate site: fallback when direct search fails), screens backward references and forward or related-paper trails, queues encountered relevant leads, and continues until a further round yields no new relevant leads. A five-paper checker PASS is only a floor. Inaccessible trails or leads fail the paper completion gate.
  • Version set to 0.5.21 in plugin.json and skills/researchcomb/VERSION.
  • README install pin updated to v0.5.21.