菜单
Configure custom Spell agent and model profiles. Use when creating or modifying .agent.edn files, choosing namespaces, changing system prompts, wiring sub-agents, selecting provider-prefixed models, or doing agent smoke tests.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | spell-custom-agents |
| description | Configure custom Spell agent and model profiles. Use when creating or modifying .agent.edn files, choosing namespaces, changing system prompts, wiring sub-agents, selecting provider-prefixed models, or doing agent smoke tests. |
Use this skill for runtime configuration, not evaluator changes. Keep docs/api.md as the stable public reference and config/AGENTS.md as the local directory guide.
Spell separates model-call policy from the runtime profile exposed to the model:
config/model-profiles/*.edn. They say how to call a provider: :provider, credentials, :default-model, reasoning effort, token/time limits, retries, pricing overrides, and transport flags like OpenAI :force-tool-call.config/agent-profiles/*.agent.edn. They say what the model sees and can do: system prompt, exposed namespaces, worker/sub-agent topology, default model profile, budget, output format, and recovery/evaluation behavior.If a setting changes the HTTP/provider request, put it in a model profile. If it changes prompt-visible capabilities or Spell runtime behavior, put it in an agent profile.
Start from a nearby existing profile in config/agent-profiles/:
base-pf.agent.edn, base-msg.agent.edn, base-tc.agent.edn: transport bases with the right system prompt and no effect namespaces.cli.agent.edn: interactive default; exposes io, web, patterns, agents, and globals, plus explore as a worker.io-pf.agent.edn, io-msg.agent.edn, io-tc.agent.edn: I/O-capable profiles without web by default.Use :base inheritance and add only the differences:
{:base cli.agent.edn
:agent-name my-agent
:agent-description "Short purpose of this profile."
:default-model-profile "../model-profiles/openai-tc.edn"
:namespaces
{io stdlib/io
patterns stdlib/patterns
agents stdlib/agents
globals stdlib/globals}}
Paths in :base, :system-prompt {:file ...}, and :default-model-profile are resolved relative to the file that declares them. Child scalar values override parent scalar values; :namespaces merge by key.
Expose only the capabilities the task needs. Use io for file/shell work, web for search/fetch, patterns for reusable Spell programs, agents for agent communication, and globals for shared state. Use :available-agents to expose workers through the workers/ namespace; do not expose .agent.edn files as namespace values for new public config.
Start from config/model-profiles/*.edn and keep provider details there. Common public provider paths are:
codex-tc:<model>
openai-tc:<model>
anthropic-tc:<model>
anthropic-pf:<model>
fireworks:<model>
fireworks-tc:<model>
ollama:<model>
Profile files use provider keys such as :provider, :api-key-env, :default-model, :default-reasoning-effort, :max-tokens, :request-timeout-sec, and :default-agent-profile. Keep model names exact and update data/pricing.edn when adding priced models.
Credential expectations are:
codex once so ~/.codex/auth.json exists.OPENAI_API_KEY.ANTHROPIC_API_KEY.FIREWORKS_API_KEY.ollama:<model> model spec.Model names must be exact. Do not guess model IDs; check provider docs before adding or recommending new model strings.
Keep model transport and agent base aligned:
base-pf.agent.edn or a profile inheriting from it.base-msg.agent.edn or a profile inheriting from it.base-tc.agent.edn or a profile inheriting from it.OpenAI tool-call configs still use :provider :openai plus :force-tool-call true; do not invent a separate OpenAI provider type.
Run the smallest useful checks:
bin/spell -h
bin/spell -t "Return a short greeting"
bin/spell -a config/agent-profiles/my-agent.agent.edn -t "Return a short greeting"
Then run one live task with the intended provider:
bin/spell -a config/agent-profiles/my-agent.agent.edn -m openai-tc:gpt-5.4 "Return the number 42."
For Clojure API validation, use explicit profiles:
(spell/run {:prompt "Return 42."
:model-profile "config/model-profiles/openai-tc.edn"
:agent-profile "config/agent-profiles/my-agent.agent.edn"})
For debugging, add -v, --log FILE, or -T. Run bin/spell -h for the authoritative option list in the current checkout.
Run clj -M:test-fast after changing loader behavior, profile semantics, or public examples.
Navigate and develop the Spell source code. Use when modifying Spell internals, investigating evaluator/runtime/provider behavior, finding the right source files, updating tests, or explaining how the implementation fits together.
Set up Spell from a fresh checkout and run the first smoke tests or examples. Use when a user asks to install Spell, get started, verify the CLI, configure provider credentials, choose a first example, or try Spell for the first time.
Use Spell from the CLI, the Clojure API, and the Python benchmark adapter. Use when writing commands, embedding Spell in Clojure, calling Spell from benchmark Python code, or checking current API vocabulary.
Navigate and develop the Spell source code. Use when modifying Spell internals, investigating evaluator/runtime/provider behavior, finding the right source files, updating tests, or explaining how the implementation fits together.
Set up Spell from a fresh checkout and verify prerequisites, CLI smoke tests, and available model provider authorization. Use when helping a user get to a runnable checkout, configure API provider access, or try the first Spell example.
Create, inspect, or modify Spell model profiles and agent profile files. Use when the user asks to configure providers, add or change `.agent.edn` files, expose namespaces, define sub-agents, adjust budgets, or use `spell.api/run`.