Principal WP Agentic Pipeline Overview

The full pipeline as it runs today: thirty-one steps, driven from a Jira board. Dragging a ticket across the board runs the phase that owns the column, and the run pauses at human gates until someone answers or drags the ticket forward.

The board

Backlog

waiting; nothing runs

Definition

drag runs steps 1–6

Requirements

steps 7–9

Architecture

steps 10–14

Coding

steps 15–27

Code Review (human)

human review

UAT

human review

Done

manual wrap-up, steps 28–31

Drag rules:

One column forward runs the phase that owns it. If the previous phase stopped at a human gate, dragging on without answering adopts the suggested defaults; the adopted answers are recorded on the ticket. Answers left as a ticket comment are applied instead.

More than one column forward runs nothing. The ticket gets a comment naming the expected next column.

Done never auto-runs. Compounding (steps 28–31) is run manually from the workstation, and the ticket gets a comment explaining how.

New tickets assigned to the pipeline account in the Definition column start on their own; no drag needed.

Human review columns (Code Review, QA, UAT) run nothing. Comments left there are picked up and routed back to the pipeline as PR feedback.

The phases

DEFINITION

Steps 1–6

1 Environment Pre-Flight git state, build tools, test harness

2 Resume Check every phase re-runs steps 1–2 before its own work

3 Initialize Work Unit spec folder and state file on the feature branch

4 Classify Complexity and Route Trivial · Standard · Complex

5 Scope Discovery clarifying questions when the request is incomplete CPLX ONLY

6 Definition Phase stakeholder input in parallel (Standard · Complex); open questions go to the human ☺

Ends waiting on definition answers. Answer by comment, or drag to Requirements to adopt the defaults.

REQUIREMENTS

Steps 7–9

7 Research research runs in parallel; an independent audit checks it STD · CPLXAUDIT

8 Requirements Analysis a critic challenges the inputs first; the analyst writes the requirements from its findings STD · CPLX

9 Requirements Checkpoint independent audit, then human sign-off CPLX ONLYAUDIT ☺

Standard approves itself and moves on. Complex waits for human sign-off. Trivial skips steps 7–9.

ARCHITECTURE

Steps 10–14

10 Synthesis and Spec Writing the planner writes the spec from research and requirements STD · CPLX

11 Human-Input Reconciliation every human answer lands in the spec or is flagged STD · CPLX

12 Spec Audit structural failures are fixed before the checkpoint STD · CPLXAUDIT

13 Spec Architecture Review critical findings are fixed before the human sees the spec STD · CPLXAUDIT

14 Spec Checkpoint the human approves the spec before coding starts ☺

Ends waiting on spec approval. Answer by comment, or drag to Coding to adopt the defaults.

CODING

Steps 15–16

15 Post-Approval Setup spec artifacts verified; the coding lead takes over

16 Coding

Per-Task Loop ↻

1Slice the spec into tasks

2One coder per task, in order

3Each task lands as its own commit

4A stuck coder pauses with a report a human answers before the run resumes

VERIFY

Steps 17–20

17 Post-Coding Verification one lead runs steps 17–20 end to end

18 Post-Coding Audit the diff is checked against the spec; deviations go to the human STD · CPLXAUDIT

19 E2E Verification Gate the full end-to-end suite runs; failing tests get a capped number of fix rounds ↻

20 Build Verification Artifacts test results and static-analysis reports are saved with the run

REVIEW

Steps 21–22

21 Sequential Pre-Reviews STD · CPLX

Pre-Review Gates ↻

1Acceptance edge cases and boundaries · up to 5 rounds

2Architecture Complex only · up to 5 rounds

3Test quality single pass

Findings are fixed, then re-reviewed, round by round.

22 Parallel Main Cycles the main review, run in cycles

Review Cycles ↻

1Parallel reviewers read the diff

2Findings merge into one verdict

3A fix coder applies the batch

4The next cycle re-reviews which reviewers return depends on the findings

Up to 10 cycles. Conflicting findings stop the run for a human.

DELIVER

Steps 23–27

23 Demo Recording the change on video, or a written reason there is none

24 Performance Comparison before and after measurements; regressions are reported on the PR

25 Final E2E Re-Gate the suite re-runs after review fixes, with a capped number of fix rounds ↻

26 Pre-Completion Gate nine final checks, a docs sync, and a portability bundle for WordPress VIP before the PR opens

27 Delivery Verification & Phase Wrap-Up the PR is verified; the ticket moves to Code Review (human)

Ends at the PR. Human review continues on the board: Code Review (human), QA, UAT.

COMPOUNDING

Steps 28–31 · manual

Run manually from the workstation. Dragging the ticket to Done posts a how-to comment; nothing runs on its own.

28 Compound Learning observations from the run turn into proposed pipeline improvements

29 Post-Mortem what was asked versus what was built, across the whole run

30 Feedback Loop failures traced to their source in the pipeline itself

31 Final State Update terminal status recorded: completed, completed with gaps, or failed

Ends with the final status on record. Recommendations stay pending until a human approves them.

ICON LEGEND: Human Gate Loop AUDITAuditor — independent sub-agents that quality-check work before the human sees it.

Sizing: Every change is classified at Step 4 before work starts; the route decides which steps run.

Trivial covers a typo, a config value, or a small copy edit: about twenty lines, no new files. It skips scope discovery (5), research and the requirements steps (7–9), the reconciliation and spec audits (11–13), the post-coding audit (18), and the pre-reviews (21). It still gets a minimal spec, the spec checkpoint, the full end-to-end gate, and a lightweight code review.

Standard covers a multi-file change to something that already exists, with no new architectural patterns. It runs the full pipeline except the Complex-only steps: scope discovery (5), the requirements checkpoint (9), and the architecture pre-review inside step 21.

Complex covers a new feature, integration, or data model. It runs every step, records every major design decision, and is the default whenever scope is unclear.