Technology deep dive

"Why did it do that?" Now the app can answer.

Every app you use does mysterious things — a message vanishes into a folder, a rule fires, something gets archived — and the honest answer to "why?" is usually a shrug. SwarmMarshal makes that question answerable by construction: every consequential action runs one mandatory lifecycle and leaves a durable receipt, and the assistant answers "why did you do that?" from those receipts — never from a guess.

The user-facing payoff

Evidence-first answers, or an honest "I don't know."

Ask the assistant why a message moved, why a rule fired, why something was filed — and it queries the decision ledger, collapses the action's lifecycle to its causal receipt, and tells you what actually happened.

The whole story

Who or what acted, what triggered it, the reason and the policy behind it, the before-and-after state, exactly which items were touched — and how to undo it.

Never a guess

The answer comes only from durable receipts. If no receipt exists, the assistant says it doesn't know — it will not reverse-engineer a plausible cause from your current settings, because a plausible cause and the real cause are not the same thing.

Undo travels with the receipt

Because receipts record the before-state and the touched items, "put it back" is a data operation, not an archaeology project.

By construction

Actions can't skip the paper trail.

This isn't logging bolted onto features — it's a single action framework every consequential operation must run through. The base lifecycle owns every ledger transition; a feature module only supplies the factual plan and the domain work.

  1. Plan The module states, factually, what it intends to do — the targets, the change, the reason, the policy that authorizes it.
  2. Approve (when required) Consequential plans wait for approval. Nothing physical happens before the approval entry exists, which is what makes a pending approval safely resumable after a restart.
  3. Apply The domain work runs — a move, a send, a rule change — through the module, which is the sole physical boundary for that action type. There is no side door that mutates state without a plan.
  4. Terminal receipt The outcome — success, failure, or skipped — is written as the action's final, durable word. That receipt is what "why did you do that?" reads later.
One lifecycle for every consequential action Modules can't own their own bookkeeping Append-only — receipts are never edited
Failure discipline

A crash never repeats a side effect — and never hides one.

Idempotent by claim

Each action takes a durable single-owner claim before applying, so a retried or duplicated request can't perform the same physical action twice.

Interrupted means "unknown"

If the process dies mid-action, the lifecycle is closed as unknown for reconciliation — the system checks what actually happened rather than blindly re-applying. An email is never moved twice because a write timed out.

Honest outcomes

Failures and skips get receipts too. The ledger records what the app tried and didn't do — which is often exactly what you're asking about.

Trust chain

Approval evidence is earned, never manufactured.

When a plan says "the user approved this," that claim must come from a real authority — the approval that you actually clicked flows through the agent's turn context into the action as evidence. A feature module cannot fabricate its own permission, so the receipt trail is trustworthy end to end: the same chain that gates the action is the chain that explains it later.

Consolidation

One ledger replaced three dashboards.

SwarmMarshal used to answer "what has the app been doing?" with separate diagnostic surfaces — an autonomy log, an auto-email management view, an agent diagnostics page — each reading its own store. Those are gone. The decision ledger is now the single source for plans, authorizations, outcomes, explanations, and undo, and the conversational explain tool is the front door. You don't hunt through admin pages; you ask.

Ask "why did that email move?" Ask "what did you do overnight?" Ask "who approved that?"
Keep reading

The rest of the stack.

Receipts explain actions; provenance explains answers. Together they make the whole system accountable.