Evidence: tool-list POSITION affects tool attention (Qwen3.8-27B) — descriptions + tail placement fix "tool not in kit" failures

[simongonzalezdc]

We measured a serving-side tool-attention failure class on Qwen3.8-27B (llama-server, --jinja, tools rendered by the chat template) that looks like model misbehavior but is payload-structure dependent. Sharing the evidence since it may affect how templates render large tool lists.

Phenomenon: with an 8-tool payload, the model verbally refused early-listed tools — claiming read_file/bash/write_file (positions 1-3) "aren't wired up" while successfully using tools at positions 4-8. 0/6 via the full payload; 15/15 via a 1-tool bare payload on the same server.

Controlled matrix (24 runs + 27 curl controls): NOT tool count (8 tools with read_file moved LAST: 3/3; first-3-only: 3/3), NOT context/style, NOT effort, NOT fresh-session. Winner: early, description-less tools lose salience. Validated fix (6/6 raw-API): add a one-line description to each bare tool spec. Alternative validated 3/3: reorder the problem tool to the end of the list.

Follow-up at 13 tools (descriptions everywhere): a prescribed-by-name mid-list pair (read_symbol/code_graph at positions 6-7) fired 0/5 — model claimed they were absent while registered. Moving that pair to the TAIL of the tools array: 3/3 clean, zero off-script calls. Same server, same protocol.

Hypothesis: later tools in the rendered template output receive more attention; description-less tools early in a long list are the worst case. Templates rendering tools as one compact JSON blob may amplify position effects versus per-tool natural-language rendering.

Numbers + methodology: https://github.com/KyaniteLabs/qwen38-27b-strix-halo (VALIDITY-MAP + WAVE2 ledgers). Hardware: Strix Halo (gfx1151/HIP), Q4_K_XL, current llama-server.