Automation for triaging issues in Shiny team repositories.

The workflow in .github/workflows/team-issue-triage.yml runs weekly and on manual dispatch. Its allowlist, label taxonomy, safety rubric, and helper scripts live in .github/triage/.

See .github/triage/README.md for configuration, required secrets, state branch behavior, and validation commands.

The triage-verse CLI (Python, managed with uv) mirrors issues, PRs, and comments from every repo in config/repos.yaml into a local SQLite database. GitHub stays the source of truth; the mirror is derived data and can always be rebuilt.

uv sync                                      # one-time setup
uv run triage-verse sync --full                # initial backfill (resumable)
uv run triage-verse sync                       # incremental refresh (seconds-minutes)
uv run triage-verse verify-counts              # reconcile against GitHub search
uv run triage-verse analytics export           # burndown series -> .data/analytics.json
uv run triage-verse snapshot publish --dated   # upload to mirror-latest + dated tag
uv run triage-verse snapshot bootstrap         # fresh machine: pull mirror-latest

Cursors live in the mirror's repos table; --full ignores them. The backfill is resumable: re-running sync --full re-upserts idempotently, and interrupted incremental syncs simply continue from the last cursor.

config/repos.yaml ships the pilot trio active (reactlog, shinytest2, py-shinylive); uncomment the rest of the shinyverse when ready to run the full fleet.

Turns the mirror into triage proposals using local embeddings and a language model backend. Model and embedder config live in config/models.yaml.

uv run triage-verse embed                         # compute/update embeddings (local, free)
uv run triage-verse analyze --wait                # classify + dedup -> .data/proposals/
uv run triage-verse analyze-status                # in-flight batches + today's spend

Backends: The default backend: claude_cli uses claude -p on Claude Code auth (no API key needed, requires Claude CLI installed and enterprise subscription). To switch to the Anthropic Batch API, set backend: anthropic_batch and provide ANTHROPIC_API_KEY in the environment (see issue #18).

analyze is a resumable state machine: re-running it collects in-flight batches rather than resubmitting, so an interrupted run (or the future scheduled job) simply continues. Spend is metered to the mirror's spend table and capped by max_usd_per_day in config/models.yaml. Under backend: claude_cli, each claude -p call executes and bills synchronously; config/models.yaml's batch.workers controls how many run concurrently (default 1, this repo starts at 2). max_usd_per_day is checked before every new dispatch, bounding a tripped budget's overshoot (and a crash's loss) to at most workers items instead of the whole stage; use --limit to additionally bound a single run's spend (each call costs roughly $0.01-0.02).

Design: docs/superpowers/specs/2026-06-12-shinyverse-issue-triage-design.md. Open followups: docs/superpowers/plans/2026-06-12-triage-verse-followups.md.