Twenty Shifts of the Pipeline: The Day-and-Night of the TVMOD Three-App De-Ad Campaign, the Front Line Shifting Between Emulator and Box, and the Full Falsification of the Skip Route

A full day of pipeline duty — twenty-odd cron shifts, a dozen subagents, and a test front that changed three times in one day — turned the TVMOD three-app de-ad campaign into a war of movement.

I. The Daytime Falsifications: Two Roads Sealed at Once

The noon and afternoon A/B experiments were the day’s hardest yield. First, the kernel loadVrsAdData branch was forced to the no-parts exit (patch at 0x22f42c, md5 c0d54dd1, loaded confirmed via /proc/maps) — the pre-roll still played. Second, the entire ad-parse function was blanked (patch at 0x222234) — QYPROBE still received seven events. Together they settle a new conclusion: pre-roll AdItems come from neither the kernel loadVrsAdData nor VendorAdParser::parse*; and since libcupid likewise contains creativeObject (0x12f701, eight code references), the suspicion for ad-response parsing formally moves to libcupid. The evening’s third proof completed the map: the ads are pulled and cached at play time (seven creatives under qcache/data/ad_cache with qd_idx/qd_tc tokens, landing on the same CDN domain as the feature film) — the domain-blocking road is also closed.

II. Plugin C2: First Chain-Through, Still Short of the Bar

Victory and shortfall shared a frame: with the plugin C2 build substituted directly, the C2 patch triggered its first full chain — TVMOD: C2PATCH: l() entered, auto-skip now and k() CALLED on the same millisecond, responding within four seconds (the main-package C2 had never fired in the emulator). Four skips across the evening, the feature film finally reached. But the reel length did not shorten: the skip intervals of 16/47/55 seconds approximated the creatives’ natural lengths, and click-to-feature still cost two minutes twenty-seven — the skip request does not advance the native slot. The merged verdict with the parallel shift: the C1/C2 skip route is falsified on both main-package and plugin sides; zeroing the reel length requires a new stake at the data-source layer (the FetchFeedAdDataTask/adlist chain, or the native slot layer). Occasional ad-free rounds (zero CUPID exec) came from server-side randomness — multi-round discipline pays.

III. Two Reversals of the Front and the Tuition of Discipline

The test front changed three times in a day: at dawn the master set emulator-testing plus scheduled long-shift mode; at 13:00, upon learning the emulator had been closed by the master himself, the order became adb box 192.168.5.42 in unbroken mode; at 21:00, after memory exhaustion and the discovery of touches on other software inside the box, the order reversed again — emulator only, never the TV box again. The tuition was recorded honestly: the model unification order (GLM-5.3-Flash only, subagents included), failure takeovers (the emulator dropped twice and was revived via MuMuManager), and a new methodology entry — confirm mCurrentFocus before declaring a deadlock. The evening’s urgent alarm — free sources playing ads forever, the feature never arriving — triggered an all-speed sweep that centered the v16 slot-wait chain.

IV. The Governance of a Pipeline

Between the twenty shifts ran another kind of order: disk governance freed 94GB in a day (C: from 12GB to 106GB, with .zcode at 21GB and the NVIDIA DXCache named item-by-item as a historical no-touch zone); a Chrome diagnosis used a clean headless instance to prove the fault lived in the user profile’s extension layer; and the Friendly Greeting #2 veteran window relayed the no-login standing order while verifying that the Hermes iOS TestFlight invitation had gone 404. The secret of an unbroken pipeline is not tirelessness — it is handover between shifts, iron rules above handovers, and above the rules, a channel that admits its own mistakes.