Morgana: Skill for Claude Code

HarnessExecution

Runs the PromptHarness (PromptHarness/PromptHarness.csproj) on demand against whichever LLM providers are currently configured in User Secrets (Morgana:LanguageModels:Tiers), against a user-chosen scope and a user-chosen global run/pass threshold. Live LLM calls, real cost — never runs without the two questions below being answered first.

Trigger

Activated when the user says things like:

• "run the harness"
• "execute the prompt harness"
• "launch the harness"
• Any request to run PromptHarness scenarios against the current provider

Procedure

1. Confirm the active providers before anything else: read Morgana:LanguageModels:Tiers from the shared User Secrets store (UserSecretsId 374228be-4f26-4382-a3ef-7500a0b829dd, same as Morgana.Web) without printing the ApiKey/Endpoint values and tell the user which provider and model each tier (Economy / Efficiency / Performance: Provider and Options.ModelId) and Morgana:ActorSystem:Tier this run will hit. This is the harness's own design (PromptHarness/README.md): it never has its own Morgana: config, it inherits the host's. The same file is where Harness:HarnessDirectory lives when the user has overridden it (see step 7), so read both in one pass. Note the store is written with a UTF-8 BOM: parse it as utf-8-sig, or a plain JSON read fails on the first character.

2. Ask the target scope with AskUserQuestion, multi-select. Enumerate PromptHarness/Tests/ first and offer what is actually there — the class list below is a description of a moving directory, not a contract, and a --filter naming a class that no longer exists runs zero tests and exits 0: a green nobody asked for. Eighteen classes at the time of writing, in four families:

   Deterministic — no model, no cost. Run them first: they are the cheapest way to learn the topology under test is sane before any billed turn.

   • Startup validation (StartupValidationTests — an incoherent partner declaration must stop the boot)
   • Agent card (AgentCardTests — the published card and how far the gate behind it reaches)
   • Peer federation (PeerFederationTests — the outbound half: which cards this side accepts, what it signs, where a credential may go)
   • Conversation API (ConversationApiTests — who the channel API lets in, what a start must announce, how an unknown conversation is answered. The conversations it needs on record are synthesised into the run's storage; its one model call is the presentation of its single real start)