Retrieve upstream implementation evidence before coding with XERJ
Build a focused reference corpus, retrieve implementation passages with file-and-line citations, and check license constraints before adapting unfamiliar code.
npx skills add agentskillexchange/skills --skill retrieve-upstream-implementation-evidence-before-coding-with-xerj
Use this workflow before implementing against an unfamiliar API, porting an algorithm, or continuing a coding task that has stalled through repeated guesses. The agent selects reference repositories relevant to the specific problem, indexes a focused corpus, and retrieves implementation evidence before editing the target project.
What this skill actually does
Inputs are the coding question, the selected repository URLs or a locally downloaded corpus manifest, and a running XERJ node. Keep reference corpora separate from the working repository. The result is a short implementation brief containing relevant passages, source paths and lines, recorded revisions and licenses, and an explanation of what can be adapted.
- Choose a small corpus by problem domain. Register references with
xerj corpus add reference-task <git-url>..., or reproduce recorded revisions withxerj corpus add --from ./corpus.json. A manifest must be a local file. Use dedicated managed clones: reproducing a manifest can discard edits inside existing corpus clones. - Run
xerj corpus index reference-taskagainst the running node. Check its completion andxerj corpus listbefore trusting coverage. Indexing exit 3 indicates completion with skipped or junk files; it differs from query exit 3. - Ask a specific implementation question with
xerj code reference-task "the contract or algorithm needed". The default BM25 mode does not require configuring vector retrieval. Read the returned definitions and cite their file-and-line locators. - Explain the reference’s assumptions and license constraints before adapting its approach. Preserve provenance in the implementation brief. Irrelevant or weak matches require normal source inspection or another corpus. Query exit 1 is a miss, exit 2 an operational refusal or error, and exit 3 means the corpus is not loaded on this node.
Invoke this when the agent needs implementation precedent and a reproducible reference set. Ordinary file reads remain appropriate for small edits or APIs already understood. This entry covers reference selection, indexing, retrieval and the evidence handoff; it does not cover general database administration, shared agent memory, Corpus Hub publishing, automated code copying, or external feedback submission.
The native CLI can be called by multiple agent runtimes. Shell-free MCP clients can query an already prepared corpus through xerj_code_search using the upstream registration instructions. Keep node storage outside indexed trees. The documented --insecure startup is only for a trusted single-user loopback environment: it disables TLS and authentication and grants reachable callers full access. Shared deployments require the documented authenticated configuration.
Metadata and invocation documentation were verified against upstream; no runtime installation or security review was performed. See the pinned reference-coding skill and reviewed release.