- 316shell scripts in ActiveKV, most of them validation and release harnesses
- 32probes against the AgentOS work graph
- 11checks in the integration platform's conformance suite, about 1,100 lines
Harnesses whose results cannot be lost
ActiveKV is judged by A/B runs on a quiet host, and a harness that misreports one cell can promote a regression. The main validation driver runs a matrix of serving cells, each named by prompt length, output length and concurrency.
I wanted the result impossible to lose, whatever happened to the run after the client returned.
Each row keeps HTTP status, exit code and health-wait time.
The release payload's entrypoint is shell too: it selects serve and diagnostic modes and redacts credentials from its logs.
The edge of a policy engine
AgentOS enforces policy through the coding CLI's hook system, and shell is its edge. Thin shims connect hook events to the Rust hook binary.
63 hooks
When shell hooks became policy, I ported them to Rust, because policy needs types and tests.
The contract suites for quality assurance and for the gatekeeper are shell, as are the probes against the PostgreSQL work graph. A probe that guards a fix is written to fail before the fix and pass after it.
Provisioning that stops instead of half-finishing
Helix needed a Microsoft 365 tenant provisioned repeatably. Tenant provisioning is portable bash, curl and Microsoft Graph, and each script stops with a specific exit code when a precondition is missing.
The test script for the meeting tools joins a real meeting and uses an EXIT trap so it always tries to leave. The Azure side of the media service is installed with PowerShell.
Loaders and installers
FAFO Core and Function Forge runs any function in a generic container, so updating a function needs no image rebuild. The API integration platform's conformance suite is shell too, covering the key surface, profile schema, wire format, error classification and build outputs. Agent Swarm member hosts are installed by scripts.
Each image is used only when its digest matches the pinned manifest; orange is a mismatch, refused. Illustrative.
No core system logic lives in shell.