Exploration Lessons

The nine lessons of the DSH exploration plan. Work them in order; each assumes only the artifacts of the lessons before it. The ordering rationale is the learning path, and the capability inventory is the feature map.

Work them in order. If you have half an hour rather than a weekend, the learning path names the two that carry the mental model.

• L1 — Mount your first plugin - Install the kit's plugin bundle, write a real plugin into it, and observe the fiber lifecycle including a loud failure and a silent PENDING.
• L2 — Register a tool, compose with config - A model-facing tool with a validated schema, a deliberately broken config, and a config override from a second patch layer.
• L3 — Services, isolation, and hot reload - Provide ctx.yourService, force a consumer into PENDING, hot-edit a running plugin, and inspect the live loader tree.
• L4 — Build a policy gate - Waterfall interception at tools/pre-execute, monotonic guards, and mapping policy to the seams that already own it.
• L5 — Assemble context deliberately - Hook the turn flow, inject durable context, ship a discoverable skill, and add a command with no model turn.
• L6 — Give the session durable state - Fold a known first-party session event into a projection, prove the derived state survives a restart, and meet the trap that inventing an event type destroys the session.
• L7 — Operate the harness - Session query tools, trajectory replay, token accounting, telemetry modes, and runtime invariants.
• L8 — Orchestrate multiple agents - Subagents and forks, workflow fan-out with schema-validated results, and a measured monolith-versus-fan-out comparison.
• L9 — Automate the harness - Headless runs, the SDKs and ACP, schedules, webhooks, and the operating bar before unattended deployment.