Team roles: cloud driver + verified local executors, quant-ladder catalog, and a coding bench gate - #133
Merged
Merged
Conversation
…ding bench as the promotion gate /team lets the cloud driver (e.g. pipenetwork glm-5.2) delegate execution to local models with zero endpoint typing: pick a name (or just `local`) and hi does backend detection, quiet download, hi-local serve, health wait, and role wiring, narrating phases with an in-place progress line. Bare `/team delegate|explore` opens a picker over the catalog, annotated per machine. Catalog entries now carry pipenetwork's whole MLX quant ladders (Laguna S 2.1 at 2/3/4/6/8bit, Nemotron Nano 4B at 4/8bit, GLM-5.2 REAP50, Nemotron Ultra 550B): auto picks the family by preference order and the highest quality quant that fits; `name@quant` forces a rung; explicit oversized picks fall to the smallest published quant. CUDA machines skip MLX-only families instead of failing at setup. `hi team-bench` benchmarks local models on verifiable coding tasks (codegen/bugfix/edit compiled + asserted by rustc, strict-JSON tool-call fidelity) and is the catalog's promotion gate: auto only serves models that passed it on real hardware. Live verification on a 64GB M-series Mini: coder-32b 4/4 (~7 tok/s), nemotron-4b@8bit 2/4 (~36 tok/s, over-reasons on codegen), nemotron-30b 1/4 (~17 tok/s) — and it caught laguna-s generating ~0 tok/s through the MoE expert-streaming path and qwen3.6-35b unable to load (nvfp4 unsupported by hi-mlx). Unverified/broken entries are explicit-pick only, honestly labeled. Hardening that fell out of live testing: the model cache is shared at ~/.hi/models (cwd-relative downloads duplicated 50GB checkpoints per project; legacy dirs keep working), model_present rejects partial downloads (aria2 control files, missing index shards), hi-local's default output ceiling rises 2048 → 8192 (it truncated real work mid-function), and delegate/explore routes ride SubagentRoute with cloud keys never leaked to local endpoints. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This was referenced Jul 27, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
/team— cloud driver, local execution. Routedelegate(write-capable) andexplore(recon) subagents to local models with zero endpoint typing:/team delegate localauto-sizes to the machine; bare/team delegateopens a picker over the supported catalog annotated with what fits and what's downloaded. Provisioning (backend detection → quiet download → hi-local serve → health) runs in the background with phase narration and an in-place progress line; the role wires itself when ready. Explicit endpoints still work for power users, and cloud API keys are never sent to local endpoints.Quant-ladder catalog. Entries carry pipenetwork's full MLX quant ladders (Laguna S 2.1 at 2/3/4/6/8bit, Nemotron Nano 4B, GLM-5.2 REAP50, Nemotron Ultra 550B — every repo verified against the HF listings, no guessed names). Auto picks the family by preference order and the highest-quality quant that fits;
name@quantforces a rung; CUDA machines skip MLX-only families instead of failing at setup.hi team-bench— the promotion gate. Benchmarks local models on machine-verified coding tasks: codegen and bugfix compiled and asserted by rustc, a precise mechanical edit that must survive textual invariants and compile, and strict-JSON tool-call fidelity. One server at a time, honest pass/fail plus tok/s, per-task winners, JSON reports under~/.hi/bench/.Measured on a 64GB M-series Mini
The bench is why the auto ladder is trustworthy: it caught laguna-s serving at ~0 tok/s (loads and health-checks green, but the 118B MoE rides expert streaming that's pathological beyond wired memory) and qwen3.6-35b failing to load entirely (nvfp4 quant unsupported by hi-mlx). Those entries — and bench-flunking nemotron-30b — are explicit-pick only with honest labels until the gaps close. Auto lands exclusively on bench-verified models.
Hardening from live testing
~/.hi/modelsroot (cwd-relative caching duplicated 50GB checkpoints per project); pre-existing local dirs keep working.model_presentnow rejects partial downloads: aria2 control-file remnants and safetensors-index shards that never started./team <role>with no model prints the sized catalog instead of "unknown role".Tests
1,642 tests pass across hi-agent / hi-tools / hi / hi-tui / hi-local-core, including new coverage for quant-ladder resolution,
@quantparsing, CUDA family skipping, partial-download rejection, the shared-cache fallback, bench validators, and the picker flow.🤖 Generated with Claude Code