issue
MANDATORY: Follow the complete shared lifecycle contract in ${CLAUDE_PLUGIN_ROOT}/skills/shared/run-lifecycle.md with declared skill issue.
Issue Skill
MANDATORY: READ ENTIRE FILE before composing user-facing prose: ${CLAUDE_PLUGIN_ROOT}/skills/shared/readability-style.md.
Create one or more GitHub issues in the current repository with LLM-based semantic duplicate detection. Two modes:
- Single mode (no
--input-file): a free-form description is the issue body. - Batch mode (
--input-file FILE): parse a multi-item markdown file (OOS format from/implement, or a generic### <title>+ body fallback) and create N issues in one pass.
Both modes run the same 2-phase dedup pipeline against a newest-first snapshot of at most 100 issues: open issues plus closed issues inside the default 90-day window. --no-dedup skips Steps 4–5 entirely. Phase 1 triages by title; Phase 2 reads full bodies + comments for shortlisted candidates and filters. Dedup fails open: any helper failure (network, rate limit, gh auth) produces a warning on stderr and falls through to create-all.
Internal umbrella mode. --dependency-only is reserved for /umbrella. It suppresses duplicate candidates and duplicate verdicts while retaining external and intra-batch dependency analysis. Unlike --no-dedup, every required snapshot, agent analysis, and validation must complete; unavailable, malformed, or incomplete analysis aborts before creation and before the caller-provided sentinel is written. --exclude-issue N excludes the umbrella from candidate generation, validation, and emitted dependency edges. Existing --no-dedup callers retain their skip-both behavior.
Default-on dependency analysis (issue #546): unless --no-dedup is set, every /issue invocation analyzes the new item(s) against the open issues admitted by the same 100-issue snapshot and detects pairs where (a) running them in parallel would risk merge conflicts, or (b) one clearly requires the other to land first. For each detected pair, /issue applies a hard GitHub-native blocker dependency via the Issue Dependencies REST API on the dependent ("client") issue. In batch mode, dependency analysis also covers intra-batch edges. --no-dedup skips both dedup and dependency analysis (Steps 4–5); all other invocations run the analysis unconditionally. Dependency-write failures use a hard-fail-with-retries contract (3 tries with 10s/30s pre-retry sleeps; on exhaustion, best-effort close the just-created orphan, increment ISSUES_FAILED, continue to the next item; process exits non-zero iff ISSUES_FAILED>0 at end). See ## Dependency Analysis below for the full contract.