Shapes turns action-first: action leads, steps numbered, state restated. Use for ADHD-friendly output. Do not use to trim tokens; use response-compression.
日本語の概要は準備中です。原文の説明を表示しています。
Reports which delegation CLIs are installed and authenticated, installs missing ones, and stores the result. Use before delegating or when a provider errors.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
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 generallyまだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Shapes turns action-first: action leads, steps numbered, state restated. Use for ADHD-friendly output. Do not use to trim tokens; use response-compression.
日本語の概要は準備中です。原文の説明を表示しています。
Inverts burden of proof for code additions. Use when reviewing PRs, planning refactors, or running unbloat to challenge every addition's necessity.
日本語の概要は準備中です。原文の説明を表示しています。
Tracks per-agent token usage and flags waste in parallel dispatch. Use when evaluating parallel agent efficiency or after a multi-agent run.
日本語の概要は準備中です。原文の説明を表示しています。
Coordinates Claude agent teams via filesystem protocol. Use when orchestrating parallel agents with task dependencies. Do not use for single-agent tasks.
日本語の概要は準備中です。原文の説明を表示しています。
Evaluates API surface design, consistency, and exemplar alignment. Use when reviewing public API changes or before releasing a new API surface.
日本語の概要は準備中です。原文の説明を表示しています。
Selects architecture paradigm via research before scaffolding. Use when architecture is undecided and the choice needs justification and documentation.
日本語の概要は準備中です。原文の説明を表示しています。