ddkc://swarm

2026-07-22 · BUILD LOG · EXECUTION CONTRACTS · ~9 MIN READ

Execution Contracts.
Not More Scripts.

A ChatAudit said the quiet part out loud: platform friction—not implementation difficulty—consumed the session. So Dream.OS gained a pre-slice contract board instead of another pile of one-off fixes.

Framing
Contracts
Identity
Resolved
Shell
Abstracted
Bus FSM
Restored

The audit did not ask “which script is missing?”

It asked: what permanent runtime capability would have prevented this class of failure?

That is a stronger design question. Scripts accumulate. Contracts reduce repeated operator interventions. The session that triggered this post had real product outcomes — Bot Lab ops scaffolding, a shared blog CSS SSOT, a live deploy — wrapped in preventable friction: false agent identity, PowerShell/bash dialect collisions, a missing message-bus module, green tests that did not prove the new invariant, and notification storms that burned turns on success chatter.

Four platform principles

1. Identity is never inferred

A worktree folder named agent4_… is not Agent-4. Identity must resolve from explicit argument, environment, or agent_workspaces/Agent-N — never from delivery-folder tokens, and never by defaulting to Agent-1 or Agent-4 when unresolved.

identity
IDENTITY_UNRESOLVED → BLOCK_SLICE
NO_DEFAULT_AGENT_1
NO_DEFAULT_AGENT_4
NO_CODING_AGENT_FOLDER_GUESS

2. Shells are an operating-system abstraction

Agents should emit an ExecutionRequest — an argv plan — not bash heredocs or PowerShell-fragile JSON on the command line. A Shell Contract Layer renders that plan for PowerShell, bash, cmd, or runs it directly via subprocess without a shell string.

shell contract
Execution Request
        ↓
Shell Contract Layer
        ↓
PowerShell | Bash | Termux | SSH | Docker

3. Verification must prove the intended invariant

“Tests passed” is not the same as “the property we introduced is proven.” After unifying dadudekc blogs onto blog-swarm.css, a soft editorial suite can stay green while posts silently drop the shared stylesheet. The invariant framework now treats that as a verification failure.

4. Notifications are governed events

Success belongs on a ledger. Background completion belongs on a dashboard. Human approval interrupts. Failure notifies immediately. Success chatter must not invent a new agent turn.

The contract board

Before a slice begins, Dream.OS evaluates a fixed board. Execution does not start until every contract passes or is explicitly waived.

execution contract
Identity ........ PASS|FAIL|WAIVED
Shell ........... PASS|FAIL|WAIVED
Workspace ....... PASS|FAIL|WAIVED
SSOT ............ PASS|FAIL|WAIVED
Verification .... PASS|FAIL|WAIVED
Bus ............. PASS|FAIL|WAIVED
Policy .......... PASS|FAIL|WAIVED

Many “random” session failures were the same pattern wearing different clothes: contract violations.

What shipped today

Verification that proves intent

The proof commands are intentionally boring:

verify
python runtime/scripts/verify_execution_contract_001.py --agent Agent-1
python -m pytest tests/test_execution_contract_engine_001.py -q
cd D:\websites && python -m pytest tests/websites/test_dadudekc_blog_swarm_css_ssot_001.py -q

If identity is unresolved, the board fails. If thread_fsm disappears again, bus fails and live dispatch stays blocked. If a blog post drops the shared CSS link, the SSOT invariant fails even when older soft tests stay green.

North star workflow

Hard onboard → contract board PASS → product slice → invariant proof → live handoff only if bus complete → closeout pack → silent success ledger.

The objective is not a prettier retrospective. The objective is a runtime that makes the same class of operator intervention illegal the second time.

Dream.OS does not need more scripts that remember today’s mistake.

It needs contracts that refuse to start tomorrow’s slice without the prerequisites already true.