QA System — Before / After

The robust QA platform: page-by-page coverage, machine runs ahead, human always adjudicates. The AFTER diagram keeps every BEFORE box in place — new components are green, upgraded ones amber, so the overlay reads at a glance.

exists today — kept exists — upgraded newly forged

Before — what runs today

flowchart LR
  T["Tester (human)
runs scripted packs manually"]
  TR["Tracker app
persona task packs · ticks"]
  APP["App under test (staging)
version: unknown"]
  W["Bug widget
screenshot · console · network
rrweb replay · voice"]
  BW["Bug worker
sync Jira filing (8s, 502s)
hash auto-merge dk-/tp-"]
  J["Jira board
report+bug+fix in one ticket"]
  D["Developers"]
  RL["Railway backend logs
(disconnected island)"]
  RP["Retest page
hand-written JSON"]

  TR --> T
  T --> APP
  APP --> W
  W --> BW
  BW --> J
  J --> D
  D -."manual copy of steps".- RP
  RP -.-> T

  classDef keep fill:#E4EBF0,stroke:#8FA3B3,color:#33475B
  classDef island fill:#E4EBF0,stroke:#8FA3B3,color:#33475B,stroke-dasharray:5 4
  class T,TR,APP,W,BW,J,D,RP keep
  class RL island

Pain lives in the grey: auto-merge buries reports, backend logs never reach the ticket, Jira progress never reaches the tester, and the human does the machine's job (scripted sweeps).

After — same skeleton, new muscle

flowchart LR
  subgraph TESTSIDE["Test side"]
    AGENT["Agent executor
runs packs page-by-page
span-per-action + video"]
    CON["Reconcile console
findings only: issue / pass / discuss
Take-Over button"]
    T2["Tester (human)
exploratory + adjudication"]
    TR2["Tracker app
packs + custom tasks (inline +)
+ retest tasks + live status bars"]
  end

  APP2["App under test
+ build stamp + update toast"]
  W2["Bug widget
(as is — already strong)"]

  subgraph PIPE["Report pipeline"]
    BW2["Bug worker
instant ack · queue · retries
auto-merge OFF → flags only"]
    LOGS["Log fetcher
±2 min around failure timestamp
identity-filtered"]
    LLM["LLM reconciler
%-similarity per prior report
what's-NEW · close/check notes"]
  end

  J2["Jira board
one ticket per report + links"]
  HG["Human gate
severity + merge approvals"]
  D2["Developers
ticket arrives armed"]
  RL2["Railway backend logs"]
  WH["Jira webhooks
status timeline stream"]

  TR2 --> AGENT
  AGENT --> APP2
  AGENT --> CON
  CON --> T2
  T2 -->|exploratory| APP2
  T2 -->|verdicts| TR2
  APP2 --> W2
  W2 --> BW2
  BW2 --> LOGS
  RL2 --> LOGS
  LOGS --> LLM
  LLM --> J2
  LLM --> HG
  HG --> J2
  J2 --> D2
  J2 --> WH
  WH --> TR2

  classDef keep fill:#E4EBF0,stroke:#8FA3B3,color:#33475B
  classDef new fill:#17A67D,stroke:#0E7D5D,color:#FFFFFF
  classDef changed fill:#FBEED2,stroke:#D99A2B,color:#5B4A1E
  class W2,D2,RL2 keep
  class APP2,BW2,TR2,J2 changed
  class AGENT,CON,LLM,LOGS,WH,HG,T2 new

Read it as the overlay: tracker, widget, worker, Jira, devs all stay where they were. Green is the forge-work; amber is surgery on what we own.

Build map — extend vs forge

Component	Action	What happens	Forged from
App build stamp + toast	extend	Version in every report; "new version — refresh" prompt kills ghost bugs	env var + existing service worker
Bug worker	extend	Instant ack, queued Jira filing with retries; auto-merge → duplicate flags	our worker (zero-dep) + CF Queues
Log fetcher	forge	±2 min of backend logs around the failure's own timestamp, filtered to the tester	Railway API + widget's network capture
LLM reconciler	forge	Per-report notes: % match vs prior reports (incl. fixed), what's-NEW, close/check recommendation with trail	pgvector (our Postgres) + one Claude call; trIAge (OSS) as reference
Agent executor	forge	Runs the persona packs page-by-page, evidence per step, video per session	Playwright (ours) + browser-use (OSS) + computer-use for free-form steps
Reconcile console	forge	Human watches live, verdicts on findings only (issue / pass / discuss), Take-Over	our tracker UI + span-per-action pattern
Tracker task system	extend	Inline "+" custom tasks · auto retest tasks · live Jira status bars per task	our D1 schema + Jira webhooks
Bug widget	extend	Adds build number; replay/console/network capture already battle-ready	ours (rrweb replay already recording)

Later layer (production phase, parked by design): machine-first discovery (error tracking, canary) and AI fix-to-merge flows. Nothing above depends on them.