slugline-mcp: Instructions file for Claude Code

slugline-mcp — resume notes

Read TASKS.md first — it's the authoritative build checklist (one commit per task, top to bottom). This file is context to avoid re-deriving decisions already made. As of this writing: slugline-mcp v0.1.1 is published on real PyPI (https://pypi.org/project/slugline-mcp/) and TestPyPI, and every task on TASKS.md is checked off, including Phase 9 (the 0.1.1 patch release -- see "Patch release: v0.1.1" below). All commits are pushed to origin/main. There is no active TASKS.md work queued right now, and the reference index gap that used to be called out here is now closed -- see "Reference index: now live" below.

Never reuse cached git/GitHub credentials for anything beyond git push/ git pull. If a task needs GitHub API access (dispatching workflows, checking run status, etc.) and gh CLI isn't installed/authenticated, stop and ask the user -- don't look for a workaround via git credential fill or similar. This was a real mistake made and corrected during this project (see memory: credential-scope feedback) -- don't repeat it, and this applies beyond just this repo.

Reference index: now live. The real prebuilt reference index (the full MovieSum build, ~286,000 scenes across ~2,200 movies) was built on a rented RunPod GPU (see remote-index/DEPLOY.md) and is published as a GitHub Release asset -- tag data-v1 (deliberately separate from the package's own vX.Y.Z tags, so the index can be rebuilt/republished independently of a PyPI release): https://github.com/NalluriTanavreddy/slugline-mcp/releases/tag/data-v1. bootstrap.py downloads and unpacks it (~1.8GB, one time, with a tqdm progress bar) on first run -- a fresh uvx slugline-mcp install now returns real retrieval results out of the box, no manual index build needed. The README and demo docs have been updated to say so; don't reintroduce the old "empty results until you build one" framing.

Working agreement (from the user)

• One git commit per task in TASKS.md, using that task's commit message verbatim. Check the box in TASKS.md and commit that checkbox update in the same commit as the task's implementation.
• Work through tasks strictly in order, top to bottom.
• Stop and ask before starting Phase 6 (CI/CD) and Phase 8 (Publish) — both involve secrets / an irreversible public PyPI release.
• If anything is ambiguous or needs a judgment call (versioning, model choice, hosting, etc.), stop and ask rather than guessing.
• git push after finishing each phase (established during Phase 3/4).
• The user may interject mid-task with new requirements pasted inline. When that happens: insert the new task(s) into TASKS.md at the position they specify, and if the insertion point is before where work already is, implement it immediately (out of order) before continuing; if it's ahead of current progress, just queue it and reach it in normal order.
• Every commit message now ends with a Co-Authored-By: Claude Sonnet 5 <[email protected]> trailer (instruction given at the start of the Phase 5 session, applies from that point forward — earlier commits don't have it, don't go back and rewrite history to add it).