normalizeOrderUnknownFailure replaced the agent's message with a canonical
sentence, so the diagnostic #210 added never reached the database. Across the
61 PURCHASE_ORDER_PAYMENT_REPEATED failures on production between 2026-09-07
and 09-11, not one carried paymentBackAttempts or
consecutivePaymentSamplesAfterBack, which left three very different causes
indistinguishable: the Back never left the payment activity, the page rendered
slower than the ~1.1s window the agent allows, or the page was already the
order page and no marker matched. Those need different fixes, and there was no
way to tell which one to make.
The canonical sentence stays authoritative and leads the message — agent text
is free-form and the wording per error code has to stay stable — but the agent
evidence is now appended behind it, flattened to one line (control characters
and U+FEFF removed, whitespace collapsed) and truncated to the column's 1000
runes with the canonical part kept whole. The sibling "failed" branch already
stored agent text verbatim, so this adds no new trust assumption.
That branch also only wrote the failure onto the task, never the attempt, so
every attempt it produced carried a NULL error_code and was invisible to
attempt-level statistics — the spec-panel failures on production are exactly
that. It now writes both.
This restores observability only; it does not reduce how often
PAYMENT_REPEATED happens. Tuning the agent's sampling window is deliberately
left out until the next production failure shows real values.
Wiki updated first: Android-Agent-API-Contract@ba259178.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
Three defects that only a real MySQL run could show.
The table had no version-local migration. Adding the model to
migrations.Migrate covers databases built from scratch, but every existing
one already has the older versions recorded in sys_migration, so that list
never runs again there. cmd/migrate's post-migration table check caught it:
"迁移后仍缺少表：syb_product_filter". The new version file creates the table
and seeds the same eight rules with FirstOrCreate on the normalized key, so
running both paths against one database cannot duplicate them.
Keyword and NormalizedKeyword were typed text, which MySQL refuses to put in
a key specification without a prefix length (Error 1170), so the unique index
could not be built at all. They are sized 200 now, matching how SYBShop sizes
its own matching key. SQLite accepts text in an index, which is why the whole
test suite passed while the real migration failed.
The 只看启用 filter is removed rather than kept: hiding a disabled structural
rule would conceal the one state on this page that matters most, since a
disabled "#" means roughly 65% of 档口 rows start entering the system again.
Keyword search stays.
Verified against the local MySQL database: migration applies, eight rules
present with traditional Chinese preserved (印花大學T normalizes to 印花大學t),
menu row created under 采采管理.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
The samples are the ones the production database actually holds: 档口 codes
like "新寮-雅伊阁大码女装#301" and "DD#004" must be filtered, while own
codes like "PDD0582", "my238", "1" and "引流款" must not, and an empty
variationSku is a product without a supplier code rather than a 档口 item.
The test also asserts the two structural rules record different hit counts.
That is the property the previous aggregate-per-kind write could not hold,
and the disable-confirmation dialog depends on it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
UpdateHits wrote one aggregate per kind, so every structural rule stored
the same number. The disable-confirmation dialog quotes that number
("停用「#」后…约 N 条明细会恢复入库"), and on real data "#" matches 8498
rows while "-" matches 6392 — the dialog would have warned with a figure
2100 rows off for one of them.
Match now returns the rule that matched instead of a char/keyword pair,
so the caller can attribute each hit to one rule, and UpdateHits writes
every rule's own count. Rules that matched nothing are written as 0 so a
previous run's number is never left behind.
The management page is reformatted to the repo's usual Vue layout and
gains the search / 只看启用 / 重置 controls the approved prototype shows.
Cancelling the disable confirmation now just reverts the switch instead
of throwing an unhandled rejection.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
QiuSW