v0.50.277 — model-picker shared-reference fix (supersedes #1511)

[nesquena-hermes]

Release v0.50.277 — model-picker shared-reference fix (supersedes PR #1511)

Single self-built fix replacing contributor PR #1511 by @lost9999. Their PR's diagnosis pointed at the wrong layer; I dug into the actual root cause from first principles.

Bug shape

When multiple "auto-detected" providers (Ollama / HuggingFace / custom OpenAI-compatible endpoints / Google Gemini CLI / Xiaomi / etc.) all fell through to the unconfigured-provider branch in api/config.py:get_models_grouped(), every group ended up sharing the SAME auto_detected_models list reference AND the SAME dicts inside.

When _deduplicate_model_ids() then mutated those dicts to add @provider_id: prefixes and provider-name parentheticals, the changes were applied to every group that referenced the same dict.

Visible symptom (vishnu via Discord, relayed to PR #1511):

Deepseek V4 Flash (Xiaomi) (Ollama) (HuggingFace) (Google-Gemini-Cli)

Hidden symptom (worse, never reported as a bug because users couldn't tell): the id field also collapsed to @xiaomi:deepseek-v4-flash (whichever provider_id won the alphabetical-first race) on every group. Selecting the model under any group silently routed the request to the WRONG provider.

Why PR #1511's fix was insufficient

The contributor's PR removed the label-suffix logic in _deduplicate_model_ids(). That would have made labels look clean (Deepseek V4 Flash instead of accumulated parentheticals) — but the IDs would still have been corrupted. The user would see clean labels and silently get their model requests routed to the wrong provider.

The actual fix

api/config.py:2078 — wrap auto_detected_models in copy.deepcopy() when assigning to a group:

                else:
                    if auto_detected_models:
+                       # Per-group deep copy so subsequent mutation by
+                       # _deduplicate_model_ids() (which prefixes ids with
+                       # @provider_id:) does not bleed into other groups
+                       # that also fall through to this branch.
                        groups.append(
                            {
                                "provider": provider_name,
                                "provider_id": pid,
-                               "models": auto_detected_models,
+                               "models": copy.deepcopy(auto_detected_models),
                            }
                        )

The existing _deduplicate_model_ids() logic is unchanged and correct. The bug was in the assignment site (each group sharing the same list/dicts), not the dedup function itself.

The single-parenthetical disambiguation in labels is retained because the composer chip at static/index.html:441 shows the model label without the optgroup header context — Deepseek V4 Flash (Ollama) is more useful there than ambiguous Deepseek V4 Flash.

First-principles investigation (verified empirically)

I read the dedup function and noticed it only ever appends ONE provider name per location it visits — it cannot accumulate four on its own. So either (a) the model dicts are shared by reference, or (b) the function is re-called, or (c) the report is exaggerated.

I wrote a Python repro and tested all three. Scenario (a) — shared references — reproduces vishnu's exact symptom. I then traced the 5 group-build paths in get_models_grouped():

Path	Line	Status
OpenRouter	2027-2030	new dicts via list-comp ✓
ollama-cloud	2038-2041	new dicts via list-comp ✓
_PROVIDER_MODELS / cfg	2055	explicit copy.deepcopy() ✓
Named custom (custom:slug)	2019	each _slug has its own list ✓
Else (auto-detected fall-through)	2078	shared auto_detected_models ✗

Only the else-branch was broken. Single-line fix.

Tests (4 new in tests/test_issue1511_dedup_shared_reference.py)

1. test_groups_have_independent_model_lists — structural invariant: no two groups share a list or dict by identity.
2. test_unconfigured_providers_no_shared_dedup_bleed — end-to-end against the corrected code: 4 colliding providers each get their OWN @provider_id: prefix and exactly ONE parenthetical. Negative: every label has at most ONE (.
3. test_shared_reference_pre_fix_demonstrates_corruption — direct evidence of the broken state when references ARE shared. Documents reasoning.
4. test_get_models_grouped_unconfigured_providers_get_independent_dicts — production-path regression guard added in-release per Opus SHOULD-FIX. Inspects the live source via inspect.getsource() for the literal copy.deepcopy(auto_detected_models) call AND runs an end-to-end smoke of the fixed assignment loop. A future refactor that removes the deepcopy will fail this test immediately.

Full suite: 3929 passed (was 3925 → +4 new). Zero regressions.

Reviews

• Pre-release Opus advisor: SHIP AS-IS with one SHOULD-FIX (test coverage gap). The SHOULD-FIX was absorbed in-release as test fix(api): resolve model provider from config to prevent misrouting #4 above (the production-path regression guard), per the reviewer-flagged-fix-in-release-not-followup policy (<20 LOC, clearly defensive, narrow scope). Verified all 5 group-build paths and confirmed only the else-branch was the shared-ref site.
• Self-built: per independent-review policy, self-built PRs need either nesquena APPROVED OR Opus advisor pass. Opus pass + production-path regression test is the gate here.

Behavior change worth noting

For users running multiple unconfigured auto-detected providers (e.g. Ollama + HuggingFace + custom OpenAI-compat endpoints), this release fixes silent model-routing. Pre-fix, selecting a duplicate model under any group routed the request to the alphabetical-first provider regardless of which group the user clicked. Post-fix, requests route to the correct provider. This is a fix, not a regression — but if a user was unknowingly relying on the broken routing, the model they see selected is the model they get.

Diff

api/config.py                                  |   11 ++
tests/test_issue1511_dedup_shared_reference.py |  217 ++++++++++++++++++++
CHANGELOG.md                                   |   12 ++
ROADMAP.md                                     |    2 +-
TESTING.md                                     |    4 +-
5 files changed, 245 insertions(+), 1 deletion(-)

Branches & cleanup

• Closes fix: remove redundant provider label concatenation in _deduplicate_model_ids #1511 (with credit + explanation)
• Tag: v0.50.277
• Stage branch: stage-277
• Worktree: /tmp/wt-1228-fix (cleaned up post-merge)