Continues #2100 under the margin the envelope trim (#2123) created:
58 999 to 59 079 against the 60 000 ceiling, spending under half the
room and leaving ~920.
Picked by evidence, not traffic alone: gnubok_search_tools was sent a
nonexistent `offset` in prod today, so its examples show the actual
levers (query, detail, limit) and the description of the mistake; the
create/complete document-upload pair's examples ARE the two-step flow,
same upload_id and file_name on both sides;
list_uncategorized_transactions is the highest-traffic read (15 122
calls/30d) and gets the pagination shape; link_document_to_voucher gets
the minimal linking call.
All seven pass the input-examples validation suite (same unknown-key
guard the server runs, plus required/type/enum/pattern and the
placeholder-id check), and the pinned tool list is updated.
Claude-Session: https://claude.ai/code/session_01L3P2hr19PhQuCoTSGoegcY
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
#2066 asked for input_examples on the top ~20 tools by call volume. The
binding constraint turned out to be budget, not writing: after #2089
reclaimed 3 763 tokens, its own policy required ratcheting the tools/list
ceiling down with it, so the real headroom was 317 tokens. Ten examples
across five tools cost 199, leaving ~118. The ceiling is not raised.
Tools picked from 30 days of mcp.tool_called crossed with the
combinations the descriptions already warn about and callers still get
wrong: account_override without an explicit vat_treatment (books gross,
no moms line), representation without deltagare and syfte, confirmed on
a high-risk approval, a balanced voucher where the moms leg is its own
line, and get_kpi_report, where one caller sent `metric` 604 times over
seven days to a tool whose only parameter is period_id.
Examples are also surfaced in the unknown-parameter error. That costs
nothing in tools/list, because it only ships on the response to a call
that already failed, and it reaches the caller that most needs it: a key
list told DueCue's agent which parameter was wrong but not what a
correct call looks like, and the same rejected call repeated for a week.
Every example is validated against its own schema by the same
findUnknownArgKeys guard the server runs, plus required/type/enum/pattern
checks. An example our own boundary would reject is worse than none: it
teaches the exact mistake the guard then punishes. That test caught three
invented enum values in this change's own first draft.
Claude-Session: https://claude.ai/code/session_01L3P2hr19PhQuCoTSGoegcY
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
