Repository navigation
Conversation
The utility-model wrapper (the /drain and /print seam) moves out of modules/halogen.nix into modules/utility-model.nix, so disabling the engine no longer removes it. lib/local-model.nix is the one place for the endpoint (http://strix:8731) and the served id (qwen3.8-flash-next); Gufo answers 404 to any other id and 400 to none. The think=false flag now becomes chat_template_kwargs.enable_thinking=false, the control Gufo honours; it ignores the top-level enable_thinking Halogen took. Refs #525. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tly off screen-ocr, handwriting intake, call-diarize cleanup and the academic OCR drivers name qwen3.8-flash-next and switch thinking off through chat_template_kwargs (screen-ocr and intake sent Halogen's top-level enable_thinking, which Gufo ignores, so they would have started thinking; intake also sent reasoning_effort medium and Halogen's drafter field). Intake's idle check reads Gufo's probes: /ready for the model, /v1/models for image input, /metrics (requests_processing, requests_deferred) for load, with new tests. call-diarize's cleanup label becomes local, so new evidence lands in cleanup/local/ and old calls keep cleanup/halogen/. Refs #525. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pi's local provider takes its endpoint and id from lib/local-model.nix, lists qwen3.8-flash-next plus the two operator-switched alternates (qwen3.8-27b, deepseek-v4-flash), and turns supportsReasoningEffort on, so pi's thinking level reaches the server (#424, carried by #525). The provider key stays halogen: substrate's runtimes, seats and the models catalogue name it, and substrate reads a missing provider as halogen. The models.toml seed, the monthly review's inference row and the flows member name qwen3.8-flash-next; seats only changes its label. Refs #525. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
drain, print and ai-memory wording (including the exact unavailable message the drain skill quotes, and its tests), the print README and the Sway OCR bind comment. Refs #525. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ches the module The flake asserts tie services.gufo's served model names and port to lib/local-model.nix, so the module and the consumers cannot drift apart (Gufo answers 404 to any other id). pi's qwen3.8-27b row uses the 131072 context the Gufo module declares for it. Refs #525. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # flake.nix
mecattaf
changed the title
mecattaf
marked this pull request as ready for review
October 21, 2026 19:19
This was referenced Oct 22, 2026
Closed
Owner
Author
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
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.
Halogen was still the resident GPU engine and callers named its old model ID. This cutover retires its module, container declarations and Strix model loans, enables Gufo on the same endpoint, and repoints callers to
qwen3.8-flash-next.Gufo runs one 102400-token session with thinking disabled, a fixed 1 GiB prefix cache and bounded queue/timeouts. GPU boot parameters are preserved. Fleet status and the local seat check Gufo readiness. The existing Substrate provider/seat names remain compatibility identifiers; the mutable catalog supplies Gufo's exact wire model and metadata.
restartIfChanged = falseprotects running puller jobs during the configuration switch.Validation: Gufo 0.20 loaded the existing NAS-borrowed GGUF, vision and MTP files in 25 seconds. Direct chat returned
GUFO_READY, the utility wrapper returnedUTILITY_READY, and installed Pi returnedPI_GUFO_READY. 41 utility tests and 14 fleet-status tests pass. Combined suite/cutover flake evaluation passes; focused Nix check builds validate units, model catalog, callers and topology.Activation is coordinated with suite PR579: stop the commissioning unit immediately before the declarative Gufo unit takes port8731. Halogen's exit137 evidence is preserved. Herdr, user manager, networking and active puller jobs are preserved; no reboot.
Part of #525 (fleet cutover #597).