Shapes turns action-first: action leads, steps numbered, state restated. Use for ADHD-friendly output. Do not use to trim tokens; use response-compression.
The English summary is being prepared. Showing the original description.
Reports which delegation CLIs are installed and authenticated, installs missing ones, and stores the result. Use before delegating or when a provider errors.
The English summary is being prepared. Showing the original description.
Run score Not tested
See how to installReview the instructions the agent will receive before you install.
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.
Skill(conjure:delegation-core) routes)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.
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.
plugins/conjure/scripts/provider_ledger.py writes
~/.claude/hooks/delegation/provider-state.json. What it stores is
asymmetric on purpose:
| Fact | Cached | Why |
|---|---|---|
| Installed, version | Yes, under a 6-hour TTL | Cheap to re-derive, stable between runs |
| Confirmed credential | As a timestamp, never a boolean | A token expires without touching the binary |
| A recorded failure | Clears the confirmation immediately | A stale success must not outlive the credential |
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.
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.
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": 1UnverifiedBinaryError instead of running a guessed commandauth_confirmed_at, so a
provider that starts refusing work leaves --availableSkill(conjure:delegation-core): the routing this setup feedsSkill(leyline:service-registry): health checks for external
services generallyNo reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Shapes turns action-first: action leads, steps numbered, state restated. Use for ADHD-friendly output. Do not use to trim tokens; use response-compression.
The English summary is being prepared. Showing the original description.
Inverts burden of proof for code additions. Use when reviewing PRs, planning refactors, or running unbloat to challenge every addition's necessity.
The English summary is being prepared. Showing the original description.
Tracks per-agent token usage and flags waste in parallel dispatch. Use when evaluating parallel agent efficiency or after a multi-agent run.
The English summary is being prepared. Showing the original description.
Coordinates Claude agent teams via filesystem protocol. Use when orchestrating parallel agents with task dependencies. Do not use for single-agent tasks.
The English summary is being prepared. Showing the original description.
Evaluates API surface design, consistency, and exemplar alignment. Use when reviewing public API changes or before releasing a new API surface.
The English summary is being prepared. Showing the original description.
Selects architecture paradigm via research before scaffolding. Use when architecture is undecided and the choice needs justification and documentation.
The English summary is being prepared. Showing the original description.