The next chapter
Build workflows that know how to pause.
A proposed architecture for long-running WordPress work: saved checkpoints, bounded retries, exact review versions, and a person who can resolve uncertainty.
Design proposal · not an operational unattended service
The durable loop
1. Scope
Name the site, task, access, budget, and acceptance criteria.
2. Inspect & prepare
Read current state. Prepare a draft. Save the target IDs and before-state.
3. Review
Ask a named human to approve the exact proposed version.
4. Recheck & apply
Confirm the target has not changed. Perform the authorized step once.
5. Verify & record
Read saved state and inspect rendering. Save the evidence.
6. Resume or escalate
Reconcile uncertain writes before retrying. Pause when the result is unknown.
One task, one visible chain of evidence
Save progress, prepare an exact review, and verify the result before marking work complete.
View Mermaid workflow source
flowchart TD
A[Owner chooses one site and outcome] --> B[Record task scope and limits]
B --> C[Durable coordinator]
C <--> D[Checkpoints and attempt history]
C --> E[Read current site and available tools]
E --> F[Prepare draft or proposed change]
F --> G[Save evidence and exact review version]
G --> H{Human approves this version?}
H -->|No| I[Pause or revise]
I --> C
H -->|Yes| J[Recheck target and permission]
J --> K{Still matches review?}
K -->|No| I
K -->|Yes| L[Apply bounded change through MCPWP]
L --> M[Read saved result and inspect rendering]
M --> N{Evidence meets acceptance?}
N -->|Yes| O[Record completion and notify owner]
N -->|No or uncertain| P[Stop and escalate with evidence]
Check before trying again
Inspect an uncertain outcome before retrying a write. Pause when duplicate safety cannot be proved.
View Mermaid workflow source
sequenceDiagram
participant C as Coordinator
participant S as Durable task store
participant W as MCPWP and WordPress
participant H as Human owner
C->>S: Claim task lease and save unique action intent
C->>W: Read target and confirm expected version
W-->>C: Current state
C->>S: Save before state and attempt number
C->>W: Request one authorized change
alt Response and read-back confirm intended result
W-->>C: Target ID and result
C->>W: Read saved state
W-->>C: Matching saved state
C->>S: Save evidence and completed step
else Timeout or lost response
C->>S: Mark outcome unknown
C->>W: Inspect target for intended effect
alt Intended effect is proven
W-->>C: Matching target and result
C->>S: Save recovered evidence without another write
else Effect absent and retry is proven safe
C->>S: Check retry budget and record next attempt
C->>C: Wait bounded backoff then repeat validation
else Outcome cannot be established or budget is exhausted
C->>S: Pause task with unresolved evidence
C->>H: Request decision with target and attempt history
end
end
Pause clearly. Resume carefully.
Carry task state across sessions while preserving review, cancellation and human decisions.
View Mermaid workflow source
stateDiagram-v2
[*] --> Scoped
Scoped --> Inspecting: Owner authorizes bounded task
Inspecting --> Preparing: Site and permissions confirmed
Inspecting --> Paused: Access or evidence missing
Preparing --> AwaitingReview: Save draft and checkpoint
AwaitingReview --> Preparing: Reviewer requests revision
AwaitingReview --> Ready: Exact version approved
Ready --> Applying: Recheck target and approval
Ready --> AwaitingReview: Target or proposal changed
Applying --> Verifying: Response received
Applying --> Reconciling: Response uncertain
Reconciling --> Verifying: Intended effect found
Reconciling --> Ready: No effect and safe retry within limit
Reconciling --> Paused: Unknown outcome or retries exhausted
Verifying --> Complete: Saved state and render accepted
Verifying --> Paused: Mismatch or incomplete evidence
Paused --> Inspecting: Owner resolves cause and resumes
Scoped --> Cancelled: Owner cancels
AwaitingReview --> Cancelled: Owner cancels
Paused --> Cancelled: Owner cancels
Complete --> [*]
Cancelled --> [*]
What would make this ready to run?
- A durable task store, scheduler, and worker with explicit ownership.
- Per-site credentials, scoped access, and protected secret storage.
- Idempotency, leases, retry budgets, cancellation, and failure recovery.
- Human approval tied to a saved version, with expiry and drift checks.
- Monitoring that distinguishes saved, delivered, verified, and complete.
These diagrams specify the intended system. This website redesign does not install or arm that runtime.