Review follow-up on 2bc624f:
1. finalSubmitTargets' payment-word block only covered PAYMENT_MARKERS
(立即支付/确认支付/输入支付密码). The ticket also requires blocking 去支付
and 付款. These cannot join the shared PAYMENT_MARKERS list: it also
gates the global pageProblem() payment guard, and UNPAID_MARKERS' "待付款"
contains "付款", so adding it there would misclassify every ordinary
unpaid-order page as a forbidden payment page. Added a separate
SUBMIT_TARGET_BLOCKED_MARKERS list (立即支付/确认支付/去支付/付款/输入支付密码)
used only by finalSubmitTargets' bottom-node subtree check.
2. Candidates were previously scanned across the whole snapshot, so an
unrelated clickable node outside the recognized panel (e.g. an
underlying goods-detail-page bottom bar still in the accessibility tree
behind the sheet) could out-rank the real button by sitting lower on
screen. Added panelContainerBounds(): climbs from the panel's unique
quantity input to the largest ancestor that still does not cover the
whole screen (the same "does not cover the whole screen" bounded notion
#331 already uses for sharesBoundedPanelContainer/boundedScrollables),
giving the whole bottom-sheet container. ParsedPddScreen.specPanelContainer
(PddScreenParser's `panelScrollable`) was considered but is the wrong
notion here: it is only the inner *scrollable dimension list* used for
heading/option parsing, and in a real PDD sheet the address/payment/
submit rows sit outside it as structural siblings, not descendants -
using it would incorrectly exclude the real submit row in most panels.
A candidate belongs to the panel when its center point falls inside the
container's bounds (geometric containment, matching the existing
`inside()` convention in this file, not path prefix, since PDD's own
tree can place the submit bar as a structural sibling that is still
visually part of the sheet). When the quantity input is missing or not
unique, the container cannot be determined and finalSubmitTargets fails
explicitly (no target) instead of guessing.
Tests added to SpecPanelRecognitionTest: bottom node labeled "去支付" and
"付款" are blocked (distinct from the existing PAYMENT_MARKERS-triggered
PURCHASE_PAYMENT_FORBIDDEN cases, since neither word reaches the global
guard); a clickable node outside the recognized panel container is never
chosen even though it is bottom-most on screen; a recognized panel with no
determinable quantity input/container fails explicitly. All existing
readOrderResult tests (which legitimately show "待付款"/"去支付" on the
order-result page, unrelated to finalSubmitTargets) stay green unchanged,
confirming SUBMIT_TARGET_BLOCKED_MARKERS did not leak into that path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
Task 570 (goods 8580) timed out because the order button's text kept
changing (促销价文案 instead of "提交订单"), so FINAL_SUBMIT_MARKERS never
matched and hasFinalSavedAddressEvidence stayed false forever. Per the
user's confirmed strategy, finalSubmitTargets now ignores button text and
instead picks the bottom-most clickable/enabled/non-zero-size node of an
already-recognized spec panel (reusing #331's PddScreenParser recognition,
including the REQUIRED_EVIDENCE fallback). Ties on the bottom edge pick the
rightmost node and are flagged in diagnostics. A node whose own label or
subtree contains a payment word (PAYMENT_MARKERS) is never a click target.
A candidate taller than 30% of screen height is excluded so a full-sheet/
full-body wrapper container can never win the tie against the real bottom
bar. The old FINAL_SUBMIT_MARKERS text match is kept only as non-required
diagnostic evidence. hasFinalSavedAddressEvidence, finalConfirmation and
submitOrderOnce all consume the same finalSubmitTargets, so the address-save
evidence check and the final click use one consistent rule.
Tests: PurchaseLiveAutomationTest's LiveDriver confirmation-page fixture now
carries a full #331 REQUIRED_EVIDENCE structure (address/payment rows nested
under the shared panel container) since finalSubmitTargets depends on panel
recognition; one assertion that encoded the old bare-tap address activation
is updated to reflect the now-wrapped clickable row. SpecPanelRecognitionTest
replaces the old text-matching submit-target test with cases for the 570
promo-price bottom row, legacy "提交订单" panels, side-by-side tie/rightmost,
payment-word blocking (visible and invisible-subtree), zero-size exclusion,
and unrecognized pages.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
Review follow-up on 56f24e1:
1. moveSpecPanelToTop's collapse detection was too sensitive: a normal
scroll-to-top on a multi-dimension panel can legitimately drop the
heading count (a lower heading scrolls out of view) without losing
any spec values. SPEC_PANEL_TOP_COLLAPSED is now only raised when
every collected spec dimension value vanishes (count goes from >0
to 0) or the spec panel is no longer recognized as open
(!specPanelOpen or specPanelType == UNKNOWN). Heading count alone
no longer triggers it.
2. A collector failure during the purchase spec probe (e.g.
SPEC_PANEL_TOP_COLLAPSED) previously vanished into
collectPurchaseProbe returning null, so probeOutcome() reported
the generic PURCHASE_SPEC_NOT_MATCHED "商品规格探测失败" — the same
as an ordinary spec mismatch. collectPurchaseProbe now encodes a
failed collect() result (code + message) into the opaque probe
JSON via PurchaseSpecProbePolicy.encodeCollectorFailure, and
PurchaseSpecProbePolicy.demote (replacing demoteIfEmpty, kept as a
deprecated alias) surfaces it as an explicit
PURCHASE_SPEC_PROBE_FAILED failure with a purchaser-readable
message ("规格探测时规格面板被拖动，规格标题消失" for
SPEC_PANEL_TOP_COLLAPSED, a generic message naming the code
otherwise). This keeps PurchaseRehearsalExecutor's probeSpecs
callback contract opaque, so its existing unit tests are untouched.
Tests: PddProductDetailCollectorTest (heading count dropping while
dimensions remain present must not fail collection, still collects
colors/sizes), PurchaseSpecProbePolicyTest (encode/extract collector
failure, demote surfaces SPEC_PANEL_TOP_COLLAPSED and unrecognized
codes explicitly).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
PddProductDetailCollector.moveSpecPanelToTop always swiped DOWN at
least once even when the panel already showed its topmost color
heading, and required two identical viewport signatures to stop. On
a real device that extra swipe could drag the bottom sheet and make
the color/size headings disappear, after which the purchase spec
probe silently reported spec_probe_completed with zero dimensions
and the server reported the generic PURCHASE_SPEC_NOT_MATCHED,
hiding the real cause (goods 8580, tasks 551/552).
- moveSpecPanelToTop now skips the restore swipe when the panel is
already at top (the first parsed dimension is "color" with visible
values), and stops and fails explicitly (SPEC_PANEL_TOP_COLLAPSED)
if a restore swipe makes headings/dimensions vanish, instead of
swiping further or returning an empty success.
- New AgentDiagnosticReason.SPEC_PANEL_TOP_ALREADY /
SPEC_PANEL_TOP_COLLAPSED record swipe count and heading/dimension
counts before/after (booleans/counts only, no page text).
- New PurchaseSpecProbePolicy demotes an Agent spec_probe_completed
outcome with zero collected dimensions into an explicit
PURCHASE_SPEC_PROBE_EMPTY failure ("规格探测未读取到任何颜色或尺码")
before it is persisted/reported, instead of reaching the server as
a normal empty probe.
- Server resolveProbedSpecs uses the same explicit
PURCHASE_SPEC_PROBE_EMPTY code/message when a probe result has zero
colors and zero sizes, as defense in depth for older Agent builds.
Tests: PddProductDetailCollectorTest (already-at-top skips the
restore swipe; not-at-top restores and still collects; vanishing
headings stop swiping and fail), PurchaseSpecProbePolicyTest, and
service_test.go TestLiveProbeWithNoDimensionsFailsWithExplicitEmptyProbeCode.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
- PddScreenParser: add REQUIRED_EVIDENCE fallback type after all existing
branches; requires address entry (masked phone in a nearby clickable row),
payment entry (specPanel.paymentAreaAliases in a nearby clickable row) and
one enabled quantity EditText, plus at least one auxiliary signal
(options, summary, close, +/- or submit action). +/- no longer required
on this path; existing branches unchanged.
- The three required items must share one panel container: their lowest
common ancestor must not be a window root and must not cover the whole
screen. Items from an address dialog, a payment dialog and an unrelated
input box are not recognized.
- Purchaser hints on SPEC_PANEL_EVIDENCE_NOT_MATCHED, the spec-panel wait
diagnostic and the address-save timeout when a required item is missing
(default WeChat Pay / default address / quantity input). No Alipay alias.
- Purchase spec-entry wait and post address-save wait: bounded 5000 ms,
fallback panels need two identical structure samples, no Back while
waiting, explicit failure on timeout; boolean/count diagnostics only.
- Final submit target additionally rejects zero-size labels/containers.
- Diagnostics: evidence flags (incl. same-container) in panel evidence and
spec-panel entry event.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
- PddScreenParser: nested clickable nodes of one labelled option block
(outer block, image, inner text with a badge such as "零差评") are
collapsed into one spec value named by the outer block (task 535). The
click node is still chosen by safeOptionRank; selected/checked flags are
merged from the block members; no badge-text stripping.
- PurchaseRehearsalExecutor: on SPEC_SELECTION_UNCONFIRMED emit a
specSelectionUnconfirmed diagnostic with the dimension, the target value
and the parsed option texts (each truncated to 40 chars) with selected
flags. Spec values are product attributes only.
- Tests: SpecOptionDedupTest with a sanitized task 535 fixture.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
Address review findings on 01510a8:
1. BLOCKER: sessionRetryBackoff summed to 30min, shorter than the up-to-
~60min gap between a session dying and the next hourly SYB sync
refreshing it. Changed to 5m/10m/15m/30m/30m (total 90min across
maxSessionRetryAttempts=6), updated the code comment to state the
~90min > one hourly sync period rationale, and added
TestSessionRetryBackoffTotalExceedsHourlySyncWindow to guard it.
2. Test gap: the CheckSession probe added inside
restoreOrderWritebackClient was only exercised through a fake
Factory, never through a real sybclient.Client. Added
httptest-backed tests that run restoreOrderWritebackClient against
an emulated /am/user/get (matching the envelope shape in
sybclient/client.go's `envelope` type): valid session returns a
client, mismatched username maps to ErrSessionInvalid, 5xx/timeout
map to a non-invalid error — each asserting the syb_session row is
left untouched. Added an end-to-end worker test using the real
Factory against the invalid-session server, asserting
failed/SYB_SESSION_UNAVAILABLE with a scheduled backoff and an
intact session row.
3. sessionUnavailableMessage: renamed the default category to
"会话恢复失败（网络/其他）" and wrapped every category in an
actionable template ("SYB会话不可用（<类别>），将自动重试；如持续
失败请恢复登录后重试"), still well under the 300-char column limit
and free of raw error text/credentials.
Tests: go vet ./app/goauto/purchase/... (clean); go test
./app/goauto/purchase/... (ok, 3.4s, includes the new httptest-backed
CheckSession coverage and the backoff-window guard).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
SYB order-number writeback silently gave up on session-class failures
(SYB_SESSION_UNAVAILABLE), requiring manual resubmit even though the
hourly sync job refreshes the session on its own. This adds a bounded,
backoff-scheduled auto-retry for that error code only:
- restoreOrderWritebackClient now actively probes the cached cookie
jar with sybclient.CheckSession after import, so a remotely-expired
session is classified as retryable up front instead of surfacing
later as SYB_READ_FAILED. It never logs in, never triggers OCR and
never deletes the cached session.
- The dropped Factory error is now categorized into a safe message
(no cookies/tokens) and recorded in error_message.
- The worker's claim query additionally picks up failed rows with
error_code=SYB_SESSION_UNAVAILABLE once their backoff
(lease_expires_at) has elapsed and attempt_count is below
maxSessionRetryAttempts=6 (1m/2m/4m/8m/15m growing backoff, chosen
to span the hourly sync window); other failure codes are unchanged.
- CanSubmit no longer hides manual resubmit during that backoff
window; manual resubmit resets attempt_count to 0 and clears the
lease so the worker cannot double-claim the same row.
Diff is limited to the purchase package; sybimport/sybclient/
sybinnercode are untouched.
Tests: go test ./app/goauto/purchase/... (new
order_writeback_session_retry_test.go covers backoff scheduling,
reclaim timing, max-attempt cutoff, CheckSession invalid/network
classification with no session deletion, CanSubmit during backoff,
manual resubmit reset, and non-session codes being excluded).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTDbDcwbDw1TSAcE6wfh2F
