Provider Setup

Provider Setup

Intent

An operator should be able to ask what this machine can delegate to, get a truthful answer, install what is missing, and have that answer remembered. Before this skill existed, an unconfigured provider surfaced only as a failed delegation, and every call re-derived the same facts by spawning up to sixteen subprocesses.

When To Use

• Before the first delegation on a new machine
• When a provider errors out and it is unclear whether the cause is the binary, the credential, or the model
• When deciding which provider to name in a delegation request

When NOT To Use

• Choosing between two providers that are both ready (Skill(conjure:delegation-core) routes)
• Debugging a provider's output rather than its availability

What it reports

scripts/delegation_setup.py probes each registered service and prints a table: binary, version, installed, authenticated, and the fix for each unhealthy row.

python3 scripts/delegation_setup.py --status
python3 scripts/delegation_setup.py --doctor
python3 scripts/delegation_setup.py --available

--available is the short answer: the providers confirmed ready to take work right now.

Installing

Installs come from the provenance map in delegation_executor, which is the only source of install commands. A binary with no recorded provenance raises instead of falling back to a guess, because #655 shipped a service naming a binary that an unaffiliated package publishes.

python3 scripts/delegation_setup.py --install gemini
python3 scripts/delegation_setup.py --all

Every install asks first. Declining is a recorded outcome, not a failure, so --all on a machine where two of eight CLIs are wanted still exits 0.

What is remembered, and what is not

plugins/conjure/scripts/provider_ledger.py writes ~/.claude/hooks/delegation/provider-state.json. What it stores is asymmetric on purpose:

A provider whose credentials live inside its CLI reports auth_checked: false, and that never becomes a confirmation. "We did not look" and "it works" are different claims, and the table keeps them apart.

The file is a cache. A truncated or unreadable ledger loads as empty and costs one round of probes.

What this does not change

Routing still probes live. Delegator.verify_service spawns the version and auth probes on the delegation path as it always did, and the ledger does not short-circuit it.

That is a deliberate boundary, not an unfinished edge. Reading a cached success into the router would send work to a provider whose token expired since the probe, and reading a cached failure into the router would disable a provider for the rest of the TTL after one transient error. The ledger answers an operator's question, "what is set up on this machine", where a six-hour-old answer is useful and a wrong one costs a re-run. The router answers "can this call succeed now", where a wrong answer costs the call.

Wiring the ledger into verify_service needs a failure taxonomy that separates a transient refusal from a revoked credential. That does not exist yet.

Exit Criteria

• python3 scripts/delegation_setup.py --status prints one row per registered service with its installed and authenticated state
• --available names only providers that are installed, carry a confirmed credential, are inside the TTL, and have no recorded failure
• ~/.claude/hooks/delegation/provider-state.json exists after a probe and parses as JSON with "version": 1
• An install request for a binary with no provenance entry raises UnverifiedBinaryError instead of running a guessed command
• A recorded failure clears any standing auth_confirmed_at, so a provider that starts refusing work leaves --available
• A corrupt ledger file loads as empty rather than raising

Related Skills

• Skill(conjure:delegation-core): the routing this setup feeds
• Skill(leyline:service-registry): health checks for external services generally