Mining

Mine for transferable decisions, not code volume.

Choose what is worth reading

• Candidate discovery and a one-line registry entry are read-only triage. Survey or deep-read only a repository the user named or approved in learnings/SOURCES.md.
• Start with a survey of the tree, active code paths, documentation, examples, tests, and license. Spend deep-reading effort only where the survey finds a decision that could change a Robium skill.
• Prefer official/vendor sources. Use comparative reads when independent projects answer the same question; use a re-crawl when a previously mined source changed.
• Read WORKFLOW.md for the survey, deep-read, comparative, and re-crawl mechanics.

Keep only reusable evidence

• Keep patterns, orderings, failure discriminators, integration shapes, and configuration decisions that transfer beyond the source project.
• Drop project names, incidental ports, arbitrary tuning, style, and other local choices unless they explain a reusable boundary.
• Every observation cites a pinned repository commit, path, line range, and a verbatim source excerpt that the citation verifier can find.
• Read EVIDENCE.md for source-authority thresholds, conflicts, licensing, placement, and new-skill proposals.

Preserve the learning boundary

• Output gitignored local observations and source-registry updates, never direct skill edits. Absorption belongs to learning-loop and remains separately reviewed.
• Compare findings with existing skills and observations. Record disagreement with both provenances; do not silently replace Robium field evidence with an upstream idiom.
• Point to source code by default. Vendor code only when it is short, genuinely useful as a maintained example, and its license and attribution permit it.
• Do not crawl issue trackers or commit history as part of the current mining workflow.

Done

• The registry records the pinned commit, crawl date, status, and the skill or reference files that ultimately retained the finding. Local observation IDs are not repository history.
• Citation and observation checks pass for every retained finding.
• Each finding names the owning skill or explains why no current skill owns it.
• Temporary clones and survey reports remain outside the committed plugin.
• End with a short handoff: report how many local observations were saved and how many are ready versus tentative, then say that the next step is to run learning-loop to distill ready findings into existing skills or route a justified new-skill proposal through skill-author. Ask whether the user wants that absorption run in a background subagent; do not start it without approval.