@codex review

@codex review

@codex review

@codex review

@codex review

@codex review

@codex review

Exact head 09a854d4a477ac2f94686f13cf7371b81f131b66 contains the latest review repairs.

Confirmed and fixed:

• 796c45d92 preflight empty anthropic images: native Anthropic image blocks with data: "" set the vision requirement but were omitted from preflight. A model switch could evict the resident model before translation dropped the empty image. _anthropic_local_image_payloads() now passes every string to the existing decoder, so invalid empty data returns HTTP 400 before any load. The route regression covers native base64, data URLs, and empty native base64 with zero load calls and the resident model unchanged.
• 26df333f4 restrict conditional audio model discovery: the Transformers audio registry included speech classification, speech seq2seq, and waveform model types that the chat backend cannot serve. Local discovery now admits only the supported conditional audio paths, CSM and Whisper. Focused regressions prove Speech2Text and MusicGen checkpoints stay out of /v1/models and cannot trigger a destructive auto-switch.
• 09a854d4a deduplicate resident non-gguf catalog rows: a manually loaded LM Studio checkpoint could appear twice as a path-derived basename and a scanned namespaced id, both marked loaded. The catalog now drops the exact resident non-GGUF scan row when the orchestrator entry is already present. Regressions preserve context metadata and prove a distinct unloaded path with the same basename remains listed.

Investigated and rejected:

• BOM-prefixed config.json discovery was not changed. Every supported Transformers line reads local configs as plain UTF-8, and pinned mlx-lm does the same. Accepting BOM only in discovery would advertise the checkpoint, evict the resident model, then fail the real load.

Verification at this head:

• 26 focused resolver, catalog, and Anthropic route cases passed.
• All 20 test_openai_catalog.py cases passed.
• Ruff passed on all four touched Python files.
• Python 3.9 parsed all four touched Python files.
• Fresh independent post-push resolver, routing, and CI reviews are running. Final upstream and staging CI will be tracked after convergence.

Why this diff is large:

• The first eight feature commits already contained 1,796 additions and 149 deletions across 18 files. The model discovery and auto-switch feature crosses local resolution, API routing, catalog output, request preflight, settings, and locale labels.
• The current diff is 2,745 additions and 209 deletions across 26 files. Repair and convergence added 949 insertions and 60 deletions.
• Most of that growth is regression coverage. test_openai_auto_switch.py grew by another 613 insertions and 5 deletions, and test_openai_catalog.py grew by 66 insertions. The route implementation grew by 155 insertions and 31 deletions.
• The remaining added files came from real merge and CI failures: the frontend runtime-key cycle, the cross-browser dismissal probe, source-encoding checks, and the official tool-activity browser smoke.
• The latest three confirmed fixes added 102 insertions and 13 deletions across four files. Ninety-five of those insertions are focused tests.

There are no model files, generated bundles, new dependencies, or unrelated product features in the diff. The size is mainly the original broad feature plus defensive regression coverage. Extra agent review passes are stopped; only existing convergence threads and CI remain.

@codex review

Convergence fix a00f3417b handles the final OpenAI empty-image edge at exact head a00f3417b. An empty data:image/...;base64, part now reaches non-GGUF preflight, returns HTTP 400, makes zero load calls, and preserves the resident model. Five focused image-preflight cases, Ruff, and Python 3.9 parsing passed.

@codex review

Convergence fix bc7f2b1e2 rejects unsupported remote image URLs before a non-GGUF model switch. The regression covers remote and empty data URLs, returns HTTP 400, makes zero load calls, and preserves the resident model. Six focused image-preflight cases, Ruff, and Python 3.9 parsing passed.

@codex review

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: bc7f2b1e2a

@codex review

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

Reviewed commit: bc7f2b1e2a

Final repair and convergence summary

Final head: 19d172b5c6acef84cc9ed707b33ac242b8494f6e

Confirmed defects fixed

• Local MLX checkpoints were absent from /v1/models and could not be selected by an OpenAI-compatible request. The resolver now indexes complete local non-GGUF checkpoints, rejects incomplete, adapter-only, and Diffusers trees, and routes each resolved target to the correct backend.
• Model identity could drift across catalog rows, request matching, response bodies, the API monitor, unloads, failed loads, and aliases with the same basename. The fixes use exact paths and advertised ids, restore the previous alias after a failed load, publish completed switches before late cancellation, and deduplicate only the exact resident non-GGUF row.
• Audio discovery admitted checkpoint families that the chat backend cannot serve. Discovery now limits conditional audio models to supported CSM and Whisper paths while keeping unsupported Speech2Text and MusicGen checkpoints out of the switchable catalog.
• Media validation sometimes happened after a switch had already evicted the resident model. OpenAI, Anthropic, and GGUF paths now reject malformed, empty, mixed, unsupported, or remote image and audio payloads before any model load. Valid prior-turn images remain allowed for audio switches.
• Local revision scans and source encoding checks had CI compatibility gaps. Local checkpoints now skip inappropriate GGUF revision scans, the memory contract declares its encoding, and CPU replay metadata matches the launcher.
• Frontend CI exposed a runtime-key import cycle and gaps in the browser smoke wiring. The storage key moved to an acyclic module, tool activity runs in official CI, and the dismissal probe now requires menu closure plus focus restoration before Space can reopen the menu.
• Latest main conflicted in studio-frontend-ci.yml and probe_dismiss_guard.py. The merge keeps main's current --json Chromium tool-activity invocation and its cleaner focus facts while preserving the stricter pre-Space closure assertion.

Why the diff is large

The original feature already crossed local model resolution, API routing, catalog output, request preflight, MLX loading, settings copy, and twelve locale files. Repair rounds added focused regression coverage for every destructive switch edge case and the CI-only frontend failures. The final GitHub diff is 2,691 additions and 192 deletions across 19 files; most growth is tests. There are no model artifacts, generated bundles, new dependencies, or unrelated product features. The local conflict merge lists 41 files because it incorporates current main; those upstream files disappear from the PR diff against that same base.

Final verification

• 372 focused backend catalog, routing, auto-switch, and API monitor regressions passed on the final head.
• 8 focused workflow and Playwright-suite wiring tests passed.
• Ruff, Python 3.9 syntax parsing, YAML parsing, ancestry checks, and git diff --check passed.
• Focused fork evidence: https://github.com/mahiatlinux/unsloth/actions/runs/33144588261
• Review state: approved, mergeable, and 0 unresolved review threads.

All reviewed and fixed, LGTM. CI results