Release v0.51.295 — stage-3739/3742 (model-pick revert fix #3739 + session-status revert #3742)

[nesquena-hermes]

Release stage for v0.51.295 — two small independent fixes batched (no file overlap).

#3739 → fixes #3737 — model picker silently reverts explicit cross-family pick

v0.51.290 regression (#3448 profile-aware resolution). @someaka. Honors an explicit pick across BOTH resolver branches + a toast on genuine mismatch.

Deep-review notes (this stage):

• Revert-guard caught the original PR's stale base (v0.51.267) — a naive merge would have reverted ~14K lines incl. Redesign live-to-final assistant replies #3401. Self-rebased onto fresh master.
• The author's 1st fix only patched the legacy bare-prefix branch and MISSED the profile-provider branch (fix(#3405): respect profile provider in session model resolution #3448) that is the actual bug(models): picked model reverts to previous/default model on send (likely #3448 profile-aware resolution, v0.51.290) #3737 trigger; author then pushed commits covering both branches.
• Codex caught that the client explicit_model_pick flag never engaged in the normal flow (boot.js onchange cleared the pending-pick marker before send() could read it). Fixed: marker is recorded in onchange, kept, and consumed once in send().
• Resolver live-verified: explicit=True honors gpt-5.5 under an anthropic profile; explicit=False still repairs to default.

#3742 → reverts #3570 — manual per-session status labels

Maintainer-authored, pure deletion (+4/−118). localStorage-only (no cross-device persistence, read as a bug) + crowded the root context menu.

Gate (all green)

• Full suite: 8090 passed, 0 failed
• Codex regression (diff vs master): SAFE TO SHIP
• Opus correctness: SAFE TO SHIP
• ESLint runtime / scope-undef / browser-smoke: CLEAN
• revert-guard: origin/master is an ancestor (no stale-base revert risk)

Tests added: test_issue3737_explicit_pick_client_wiring.py (4), author's resolver tests in test_provider_mismatch.py (2), refresh-persistence test realigned.

Known non-blocking follow-ups: mismatch toast string not yet i18n-wrapped; explicit-pick is single-use per send (net improvement, server-persist is a future enhancement). Closes #3737. Reverts #3570.

[greptile-apps]

Greptile Summary

This release batches two independent fixes: #3739 corrects the v0.51.290 regression where an explicitly chosen cross-family model was silently reverted on send() (root cause: onchange cleared the pending-pick marker before send() could read it; server lacked the explicit_model_pick flag in both resolver branches), and #3742 reverts the localStorage-only per-session status labels from #3570.

• fix: honor explicit model pick, suppress silent revert on cross-family selection (#3737) #3739 — explicit-pick wiring: modelSelect.onchange no longer clears _pendingSessionModel; send() now reads it, gates explicit_model_pick on a model+provider match, clears it once, and passes the flag to /api/chat/start; _resolve_compatible_session_model_state honors the flag in both the profile-aware branch (bare cross-family names) and the legacy bare-prefix branch. A mismatch toast is shown when the server still normalizes despite the flag. The slash-qualified cross-provider path in the profile branch (line 2044) does not yet have the same guard, leaving that narrower case unaddressed.
• revert(sidebar): remove manual session status labels (#3570) #3742 — session-status revert: Removes JS state, cycle logic, context-menu entries, sidebar badge render, CSS, and all 12 locale string sets for the manual status feature; the deletion is clean and symmetric.

Confidence Score: 4/5

Safe to merge for the reported bug; one narrow cross-provider edge case remains open.

The core timing fix (marker surviving onchange → send) and the two server-side resolver guards are correct and well-tested. The one gap is the profile-aware branch at line 2044: slash-qualified cross-family models (e.g., openai/gpt-5.4-mini under an anthropic profile) still bypass the explicit_model_pick guard and get reverted to the profile default, so the silent-revert symptom from #3737 can still occur for that model format.

api/routes.py line 2044 — the slash-qualified cross-family profile branch has no explicit_model_pick guard; all other changed files are clean.

Important Files Changed

Filename	Overview
api/routes.py	Adds explicit_model_pick parameter to _resolve_compatible_session_model_state and wires it through _handle_chat_start; both the profile-aware branch (cross-family bare name) and the legacy bare-prefix branch now honor the flag — but slash-qualified cross-family models in the profile branch remain unguarded.
static/boot.js	Removes the premature _clearPendingSessionModel call from modelSelect.onchange; the marker now survives until send() consumes it, fixing the timing bug that was the root cause of #3737.
static/messages.js	send() now reads and consumes the pending explicit-pick marker, sends explicit_model_pick to the server, and shows a mismatch toast on revert; marker is cleared before the API call (single-use on failure) and the toast string is not yet i18n-wrapped.
static/sessions.js	Clean deletion of the manual-status localStorage state, cycle logic, context-menu entries, and sidebar badge render from #3570 — no logic concerns.
static/i18n.js	Removes the four session_status_* locale keys from all 12 locales; deletion is symmetric and complete across all locale blocks.
tests/test_issue3737_explicit_pick_client_wiring.py	New static source-structure tests guard the onchange-not-clear and send-consumes-marker invariants; assertions are precise and lock the exact fix from regressing silently.
tests/test_provider_mismatch.py	Two new resolver tests verify that explicit_model_pick=True preserves a cross-family bare model and explicit_model_pick=False still normalizes; clean and targeted.
tests/test_model_selection_refresh_persistence.py	Updated the existing assertion to now assert _clearPendingSessionModel is absent from onchange; correctly reflects the intent of the fix.
static/style.css	Removes the five session-manual-status CSS rules; clean deletion, no concerns.
CHANGELOG.md	Adds a Removed entry for the session-status revert; description is accurate and thorough.

Sequence Diagram

sequenceDiagram
    participant U as User
    participant DS as modelSelect.onchange (boot.js)
    participant SS as /api/session/update
    participant SN as send() (messages.js)
    participant CS as /api/chat/start
    participant R as _resolve_compatible_session_model_state

    U->>DS: picks cross-family model (e.g. gpt-5.5)
    DS->>DS: _rememberPendingSessionModel(sid, model, provider)
    DS->>SS: POST /api/session/update
    Note over DS: marker NOT cleared here (fix for #3737)
    U->>SN: submits message
    SN->>SN: _readPendingSessionModel(sid)
    SN->>SN: "_explicitPick = (pending matches current model)"
    SN->>SN: _clearPendingSessionModel(sid) [consumed]
    SN->>CS: "POST {model, explicit_model_pick: true}"
    CS->>R: "_resolve_compatible_session_model_state(..., explicit_model_pick=True)"
    alt profile-aware branch — cross-family bare name
        R-->>CS: (model, profile_provider, False) honored
    else legacy bare-prefix branch
        R-->>CS: (model, requested_provider, False) honored
    end
    CS-->>SN: "{effective_model, ...}"
    alt "effective_model != sentModel"
        SN->>U: showToast(model changed — profile provider mismatch)
    end

Loading

Comments Outside Diff (1)

1. api/routes.py, line 2044-2046 (link)

   Slash-qualified cross-family model not guarded by explicit_model_pick

   The explicit_model_pick guard at line 2023 only fires when the model name starts with a bare prefix (gpt, claude, gemini) and model_family is detected. For a slash-qualified model like openai/gpt-5.4-mini picked explicitly under an anthropic profile, model_family resolves to "" (the bare-prefix loop doesn't fire since the string starts with "openai/"), so execution falls through to this block and the model is still reverted to the profile default — the same silent revert bug bug(models): picked model reverts to previous/default model on send (likely #3448 profile-aware resolution, v0.51.290) #3737 fixed for bare names. Slash-qualified cross-provider sessions would remain broken after this PR.

_{Reviews (1): Last reviewed commit: "test(#3737): realign refresh-persistence..." | Re-trigger Greptile}

[greptile-apps]

The mismatch toast message is a raw English string with string concatenation and is not wrapped with the t() i18n helper. The PR notes this as a known follow-up, but worth tracking as a concrete anchor. The t('model_mismatch_toast') key would also need to be added to i18n.js for all locales.

Suggested change

	showToast('Model '+_sentModel+' changed to '+startData.effective_model+' — profile provider mismatch', 5000);
	showToast((t('model_mismatch_toast')||'Model {0} changed to {1} — profile provider mismatch').replace('{0}',_sentModel).replace('{1}',startData.effective_model), 5000);

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Marker consumed before API call — explicit pick is single-use on failure

_clearPendingSessionModel is called before await api('/api/chat/start', …), so if the network request fails or the server returns an error, the pending-pick marker is already gone. A retry send() would compute _explicitPick = false and the server would silently revert the cross-family model back to the profile default. The PR notes this as an intentional "single-use per send" trade-off with a future server-persist path, but it could confuse users who retry after a transient failure.