Make the skills more agent agnostic (trying to make claude and codex agreeing on them)

Merged
YA
8d ago
+3659-5552

Make Goldener research skills usable across agents while retaining their research criteria, repository formats, and authorized delivery behavior. Instructions describe capabilities rather than specific interaction or connector APIs; each installable skill includes the resources it needs.

  • Add the shared sync contract and its six generated standalone copies. Restore explicit relevance and identity checks, bibliography/researcher formats, source trust rules, worktree handling, branch/commit/PR naming, issue completion rules, and recovery after partial delivery. Imported ideas retain the existing Question / Date / Author / Context format with escaped source text and provenance in the run record and PR.
  • Bundle the full paper-review workflow, template, and helper with short-paper-review, and add its target-format guide. skills/tools/bundle_resources.py regenerates these resources and derives marketplace descriptions from SKILL.md frontmatter, which is the source of truth.
  • Use explicit, overridable budgets: 30 author checks, 200 abstract checks, and 300 network requests by default; theme pool/seed caps are also overridable. The three-paper qualification for discovered researchers and two-level recursion remain. Unresolved namesake ambiguity can require a question, now disclosed in the researcher description.
  • For large unscoped lab lists, declare an eight-calendar-year window including the current year. Completing that window permits issue closure; the PR and issue outcome must identify the exact years and excluded older output. Explicit ranges and all-history requests override this default, and incomplete requested coverage keeps issues open.
  • Keep Goldener's established non-profit, open-source positioning. Outreach uses neutral sender wording unless the user supplies a representative role. PR researcher lookup fetches the recorded head/base commits into a scratch object store and extracts only regular root/topic research documents, rejecting unsafe paths and symlinks.
  • Fix Unicode title matching and validate explicit repository roots using the same markers as automatic discovery. Proposal reports include snapshot/retrieval provenance and explain any shortfall against the requested proposal count. Short summaries disclose AI generation by default, with the user override stated consistently in frontmatter and instructions.
  • Merge current main, preserving Thunderbird signature handling and draft Date/Message-ID headers. Mailbox config returns 0 for success, 2 for absent/partial configuration, and 1 for errors; it resolves the password once, includes the signature source, preserves missing-env-file paths, and suppresses credential/provider details.

Validation: 16 local regression tests passed; all 10 skills passed the skill-creator validator; generated-resource/catalog checks, relative references in 19 Markdown files, and git diff --check passed. Tests use local fixtures and mocked mailbox/vault access; no live outreach or publishing workflows were executed.