菜单
View and edit which Claude model each Cosmo/Inception surface runs on. Single entry point for OpenCosmos's inference-cost knobs — reads and edits apps/web/lib/ai-models.ts so model choices never have to be hunted down across route files.
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
The Companion path's wellbeing check-in — "how am I, really?" The opposite of a task list: attunes across the whole person (body, heart, mind, spirit, relationships, creativity), gently surfaces drift from purpose, names the current season of life and meets it, and ends by bridging to the body. Presence first; offers one small thing, not ten. Grounded in Cosmo's Attune rhythm. Triggers: /attune, "attune", "how am I", "where am I at", "check in with me", "how am I doing really".
The Companion path's discernment loop — clarity for a decision or a crossroads, Socratic in spirit. Surfaces untested assumptions, contingent beliefs ("what has to be true for this to make sense?"), and blind spots (who or what isn't in the room?). Checks options against the person's living values, and notices when a value itself is up for revision. Doesn't decide — clarifies so the person can. Ends by bridging to the felt sense of the choice. Triggers: /discern, "help me discern", "I have a decision", "I'm at a crossroads", "what's true here", "help me see clearly", "should I".
The Companion path's core loop — being with what's present, not fixing it. Attune to the moment, inquire gently, offer in the person's wisdom language, and always bridge insight back to the body. Makes room for resistance (what is it protecting?), polarities (hold both poles, don't collapse the tension), and perspective shifts. Guards against spiritual bypassing. Witnesses; doesn't fix. Grounded in Cosmo's constitution. Triggers: /reflect, "reflect with me", "I'm sitting with something", "help me process", "I'm stuck in my head", "I need to be with this".
The Companion path's teacher-synthesis loop — reconciles the many coaches, teachers, traditions, and sources a person draws from into one coherent through-line, and names the contradictions honestly rather than smoothing them. Renders the synthesis in the person's own wisdom language. Faithful attribution; says so when weaving across sources. Reads brain/teachers.md. Triggers: /synthesize, "synthesize my teachers", "what's the through-line", "where do my teachers disagree", "reconcile my teachers", "weave this together".
The Companion path's growth-witnessing loop — the counterweight to every mirror that only names what's lacking. Reads the witness journal and reflects back growth: what's shifted, what edge softened, what got integrated, what values evolved. Celebrates genuinely and specifically. Honors rest, play, and connection as real growth. Holds the anti-treadmill guard: if celebration curdles into "now do more," it names that and invites spaciousness. Triggers: /witness, "witness me", "celebrate", "how have I grown", "look back with me", "see my growth".
Your synthetic self's check-in — "what matters now, and why?" Renders the current plate (what's open, due, in progress, waiting) and ranks the single highest-impact thing to do right now, showing the reasoning. The "why" is the deliverable. Reads goals (priority order) + the recent log; traces every recommendation to a goal or flags drift. Honest mirror: no flattery. Triggers: /check-in, "check in", "what should I do today", "what matters now", "morning check-in", "evening check-in".
| name | inference-cost |
| description | View and edit which Claude model each Cosmo/Inception surface runs on. Single entry point for OpenCosmos's inference-cost knobs — reads and edits apps/web/lib/ai-models.ts so model choices never have to be hunted down across route files. |
| argument-hint | [tier] [model-id] |
| disable-model-invocation | true |
| user-invocable | true |
All of OpenCosmos's Claude model selection lives in one file:
apps/web/lib/ai-models.ts. Every API route imports its model from
there — nothing should hardcode a model ID inline. This skill reads and
edits that file so the current settings are never more than one command
away, and stay in one place instead of drifting across route files.
Read apps/web/lib/ai-models.ts in full (it's short — three exports with
comments explaining the cost/quality reasoning behind each). Also skim
apps/web/app/api/chat/route.ts and apps/web/app/api/inception/route.ts
for anything that imports a model constant, to confirm nothing has drifted
back to a local hardcoded model ID (grep for claude- in those two files —
should return nothing outside the ai-models.ts import line).
Show a table:
| Export | Model | Controls | Notes |
|---|---|---|---|
MODEL_GENERAL | (current value) | /dialog — general-audience sessions (free, subscriber, BYOK) | highest-volume surface |
MODEL_ADMIN | (current value) | /dialog — Shalom's admin sessions incl. creative mode | |
MODEL_INCEPTION | (current value) | /inception — personal-agent-brief interview | shares free-tier budget with /dialog |
$ARGUMENTS → this is a browse/edit session. Ask the user (via
AskUserQuestion) which tier they want to change, if any — offer the three
export names plus "just show me, no changes."$ARGUMENTS given as <tier> <model-id> (e.g. admin claude-opus-4-8
or general claude-sonnet-5) → apply that change directly, matching
tier loosely against general/admin/inception (case-insensitive,
ignoring a MODEL_ prefix if given).When the user names a target model loosely ("the cheaper one", "latest
Haiku", "Opus") rather than an exact ID, resolve it using current Claude
model IDs — check the claude-api skill or ask if you're unsure of an
exact alias, since model names/aliases change over time and this skill
should never guess a stale one.
Use the Edit tool on apps/web/lib/ai-models.ts only. Update the constant's
value on its export const MODEL_X = '...' line. If the change reflects a
deliberate cost/quality tradeoff decision (not just a typo fix), update or
append to that export's comment block with the same brevity as the existing
ones — this file is the durable record of why each tier is set the way it
is, and future-you (or future-Claude) will read those comments before
asking the user again.
Do not touch the two route files — they import from this module and need no changes when a model value changes.
Run cd apps/web && pnpm exec tsc --noEmit to confirm the edit didn't break
anything (a bad constant name would surface here). Report the before/after
value and which surface(s) it affects. Remind the user this is a source
change — it needs a deploy (PR → merge) to take effect, same as any other
code change; there's no separate runtime toggle.