Review — end-to-end ✅ (clean approve, regression test pushed)

What this ships

#1188 — _findModelInDropdown() step-3 fuzzy match was over-broad: stripping the trailing version segment from target (e.g. gpt-5.5 → base gpt.5) and matching against any option that startsWith(base) || includes(base). For gpt.5.5, the live Nous model @nous:openai/gpt-5.4-mini (norm: gpt.5.4.mini) starts with gpt.5 → false match. The chip showed "GPT-5.4 Mini (via Nous)" for a session storing gpt-5.5.

7-line fix in static/ui.js:103-110: use the FULL normalized target as the prefix when base.length > 4 and base !== target. Only fall back to the shorter base when it's a bare root (length ≤ 4) where stripping was effectively a no-op.

Traced against upstream hermes-agent

Pure WebUI display logic. Zero agent coupling — this is the chip-label resolver in the model picker.

End-to-end trace

The fix's decision matrix:

The includes half of the original disjunction is gone. That was actually the only path that would match e.g. user-typed mini to gpt-5.4-mini — but that's a degenerate case that user likely never hits, and dropping it tightens the match. ✅

Bug-confirmed harness — 9/9 PR, 7/9 master

Built a Node behavioural harness running the live _findModelInDropdown against fake <select> options:

PASS [gpt-5.5 should NOT match gpt-5.4-mini (issue #1188)] → null
PASS [gpt-5.5 finds @nous:openai/gpt-5.5 (exact post-norm prefix)]
PASS [gpt finds gpt-5.4-mini (bare root prefix)]
PASS [gpt-5 finds gpt-5.4-mini (base=gpt is bare root)]
PASS [claude-opus-4.7 should NOT match claude-opus-4.6]
PASS [claude finds claude-opus-4.6 (bare root)]
PASS [claude-opus finds claude-opus-4.6 (base===target since no version)]
PASS [exact match short-circuits]
PASS [unrelated target returns null]

On master, two fail: gpt-5.5 → @nous:openai/gpt-5.4-mini and claude-opus-4.7 → claude-opus-4.6. Exactly the issue's described over-match shape.

What I pushed — 6126552

The PR didn't add a regression test. I added tests/test_issue1188_fuzzy_match.py — 9 tests running the live function via Node so the real regex/normalization rules are exercised (no Python mirror to drift). 2 tests lock the over-match cases as None; 7 lock the preserved-good cases (bare-root prefix, exact short-circuit, unrelated). CI re-ran green on 6126552.

Edge-case trace

Tests

• My new test_issue1188_fuzzy_match.py: 9/9 pass.
• Local full suite: 2637 passed (+ my 9 new tests = 2646), only the unrelated pre-existing macOS test_sprint3 failure that PR #1186 fixes.
• CI on PR after my push: ✅ test (3.11), ✅ test (3.12), ✅ test (3.13).

Other audit — confirmed correct

• JS syntax: node --check passes on ui.js.
• No agent coupling: pure UI display code.
• useBase threshold of 4 chars: covers the realistic bare-root names (gpt, claude, gemini, llama, qwen — wait gemini is 6, llama is 5 — those wouldn't trigger useBase). Let me re-check: target='gemini' → base='gemini' (no version) → base===target → useBase=true regardless of length 6. ✅ Same for llama. The length<=4 check is a separate path for "target with version stripped to bare 1-3-letter root".

Minor observations (non-blocking)

• The threshold of 4 chars is arbitrary; bare 5-letter roots (llama) hit base===target (no version to strip) so they go through the useBase branch anyway. The threshold is mostly defensive against pathological short inputs.
• Step 3's existence at all is a fuzzy-match fallback — exact match (step 1) and provider-prefix match (step 2) handle the common cases. If users find this still over-matches in some other shape, dropping step 3 entirely would be the next move.

Recommendation

Approved. Tight 7-line fix with clear decision logic. The useBase = base.length<=4 || base===target predicate correctly distinguishes "stripping changed nothing meaningful" (bare roots) from "stripping is now over-loosening the match" (versioned IDs). Behavioural harness directly confirms 2 master failures match the bug shape and all 9 cases pass on PR. Pushed regression test runs the live function via Node so the rules can't silently drift. CI green; no agent coupling. Parked at approval — ready for the release agent's merge/tag pipeline.