[offbyonebit]

This adds post-0.9.1 core features, one commit per area, so it can be reviewed or cherry-picked a piece at a time. This is 1.1 material; it should not land during the 1.0 soak.

What's in it

Plugin platform (db5c858)

• New arc_llama.plugin_api (v1.0) gives plugins a public surface: require_admin, gpu_lease(), models(), gpus(), base_url, and subscribe() for model_loaded / model_stopped / model_load_failed.
• Plugins can declare requires_api. Incompatible plugins show as incompatible instead of failing at runtime.
• Plugin pages: same-origin /plugins/... links on the dashboard card.
• arc-llama plugin new NAME scaffolds a working package with a test. arc-llama plugin list shows status.
• Plugin health appears in doctor and the support bundle.

Model types (29dcf48)

• Projectors (mmproj) pair automatically from the model's folder, using a family-name match. A generic mmproj-F16 pairs only when the folder has a single model family. Pairing works on scan, add, and HF download.
• Models with a projector launch with --mmproj, and the projector counts toward the VRAM fit. Image requests to text-only models get a 400 before anything loads.
• Split GGUFs register as one model, with size and fit calculated across all shards. HF downloads fetch every shard, and a missing shard fails with shard_missing.
• /v1/rerank serves models that have reranking = true.
• /v1/models metadata now includes capabilities and ctx.

Runtime updates (79c5422)

• arc-llama runtime update installs the new build beside the current one. It checks the build's --help against the flags your recipes use, then runs a canary: it launches your smallest model on a scratch port and asks for a reply. It only switches if both pass.
• arc-llama runtime rollback restores the previous binary and backend.

Serving (4b56801)

• API keys are stored hashed in <state_dir>/api-keys.json and managed with arc-llama keys or /admin/api-keys. Once a key exists, remote callers of /v1 and /api need it. Loopback callers never do.
• serve --lan asks for confirmation and creates a key if none exists.
• On a remote browser, the chat page asks for the key once.
• Behavior change: llama-server backends now always bind to 127.0.0.1. Before this, --host 0.0.0.0 exposed every raw backend port to the network with no auth.

Chat (d6d6014)

• Stop keeps the partial answer.
• Regenerate and Edit on the latest turn.
• The context estimate updates while you type.
• Per-model presets (system prompt and temperature).
• Image attachments for vision models.

Tuning and hardware (a366fe1, 3ca1c4c)

• benchmark --depths measures decode speed at real prefill depths (Benchmark generation speed at realistic context depths #80).
• The tune fingerprint now includes the kernel, compute runtime, Level Zero, IGC, Mesa, and oneAPI versions, so a driver update triggers a retune (Invalidate tuned recipes after driver or oneAPI changes #85).
• speculative --auto benchmarks every fitting draft plus n-gram and keeps the fastest.
• Experimental tensor split across GPUs with tensor_split + split_gpus.
• Hourly performance history is persisted, served at /admin/metrics/history, and shown as a sparkline on the dashboard.

Model library (10c3f3c)

• Dashboard search of HF GGUF repos, with a quant list and fits / tight / too-big badges.
• Background download with automatic registration.
• Disk usage panel with delete. Deletes only touch files inside models_dir that no other model shares.

Partly covers #81, #83, #84, #80, #85, and #2 (the code path only; it hasn't been run on 2+ cards).

Testing

• ruff check src tests registry-repo/scripts and mypy --check-untyped-defs src are clean.
• pytest tests/: 1243 passed, 1 skipped (same skip as main).
• Browser suite: 17/17. Seven new tests were added, and I confirmed the chat tests fail against the old chat.js.
• Bare wheel installed with no extras on Python 3.13. All new CLI --help commands run.
• The runtime canary is tested end to end against a stand-in llama-server script (POSIX only; skipped on Windows).

Not tested on real hardware: vision inference, rerank, tensor split, a canary against a real llama.cpp build, and LAN access from a phone. These need an Arc box before merging.

Known caveats

• Keys protect /v1 and /api only. Plugin routes still handle their own auth, as before. A reverse proxy on the same host connects from loopback and skips the key check (documented in docs/security.md).
• Fit badges in the library use the generic KV-per-token constant, because the GGUF header isn't downloaded yet. Registration recomputes context from the real metadata.
• The first three commits on this branch carry the wrong author and AI trailers. They need a history rewrite plus force-push before merge, if wanted.

---

Generated by Claude Code