docs: rewrite AGENTS.md for the four-model, model-library era - #10
Conversation
The file still described a Gemma-only, run-and-report checkout. Now it covers: all four pinned families with per-family install docs and disk ranges; the ChatTemplate targets in the layout; the model library (root scanning, manifest-based detection, toolbar picker) versus explicit --model paths for CLI/server; the --verify integrity modes with strict full-sha256 as default; env-gated real-model test suites; the benchmark baselines pointer; and the current app shell (sidebar, model picker, streaming markdown transcript, glass composer, realtime inspector metrics). The model-run safety rules, server binding rules, and reporting requirements are carried over unchanged.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7367da3fa5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| roots — the `Mference.libraryRoot` default if set, the package checkout's | ||
| `scratch/`, and `~/Library/Application Support/Mference` — and auto-adopts |
There was a problem hiding this comment.
The app never reads a Mference.libraryRoot default (rg -n 'libraryRoot' only finds this new AGENTS entry); AppModelLocation.storageKey is Mference.modelDirectory, and AppModelLocation.resolve(...) chooses one explicit model directory/package scratch/Application Support path rather than scanning a library root. Agents following this instruction will set or debug a preference the app ignores, which makes model-location automation unreliable.
Useful? React with 👍 / 👎.
| installed models; its toolbar picker switches between families and offers | ||
| downloads for missing ones. The CLI and server take an explicit `--model` |
There was a problem hiding this comment.
The current Mac UI does not have a toolbar picker that switches families or lists downloads for missing models: the header renders a static ModelStatusBadge, and model selection is via AppModelInstallDescriptor.selected at launch or the “Choose Existing Model…” path. Documenting this as existing behavior will misdirect agents working on model-selection or install flows, and the same wording should be corrected wherever it appears in this file.
Useful? React with 👍 / 👎.
|
Re the two Codex P2 comments ( |
1 participant
What changed
AGENTS.md still described a Gemma-only checkout whose purpose was "running and reporting existing behavior". Rewritten to match the current repo:
Sources/ChatTemplate/(Core/UI/Mac); example commands use the--modelinstaller flag and note the ~15 GB → ~148 GB download range with pointers to the DSV4/Inkling pre-install docs.defaults write Mference model …/MFERENCE_MODELselection, and the--verifyintegrity modes (strictfull-sha256default,trusted-receiptopt-in, from fix: Inkling FP16 expert-row clip that rendered as "!!!!"; speed up prefill #7).scratch/gemma4.gturboto "the model the run needs"; notes env-gated real-model suites (e.g.MFERENCE_INKLING_GTURBO); adds the benchmarks baseline pointer.Note on ordering
The
Sources/ChatTemplate/layout line, the library/picker description, and the app-controls section describe #6 — merge #6 before (or with) this PR so AGENTS.md doesn't get ahead of the tree.Verification
ruby Scripts/check_markdown_links.rb— 26 files, all local links and anchors resolve.🤖 Generated with Claude Code