model-router

Preview · a replayed session in a sandbox

claude · ~/work/app · model-router

› fix the failing auth test and add an audit log call ⏺ Read(src/auth.ts) ⎿ Read 6 lines ⏺ Update(src/auth.ts) ⎿ Added 2 lines, removed 1 line ⏺ Bash(bun test) ⎿ 3 pass, 1 fail ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM › /route ⎿ model-router: model-router: the main loop keeps the model it asked for ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts

README

turn.step を使って、モデルへのリクエストを 1 回ずつ振り分ける Mod。

ルール

#	条件	書き換え
1	`e.agentId` がある（サブエージェントのループ）	`model` を `claude-haiku-4-5-20251001` に
2	メインループで `e.index === 0`、直前のプロンプトが 80 文字未満	`effort` を `low` に
3	`/route <model>` で上書きが設定済み	メインループの `model` をその値に

ルール 3 の上書きは $.store に残るので、次のセッションでも効く。/route auto で解除。

使い方

claude --plugin-dir C:/Users/lifes/kasegu/mods/model-router

グローバル設定の env に CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 が必要。

証拠ログ

turn.step が 1 回走るたびに PROOF-steps.log に 1 行追記される。

index=0 | loop=a16bed91c0e2520c2 | rule=subagent | model=claude-opus-5[1m] -> claude-haiku-4-5-20251001 | effort=high -> high | messages=10 | answered=claude-haiku-4-5-20251001 | stop=end_turn | tools=0 | trace=jev-model-router/user:returned > engine/core:returned

answered= は結果の usage.model、つまり 実際に答えたモデル。ここが書き換え後の id になっていれば、書き換えが API まで届いたということ。

注意（重要）

turn.step の model は そのまま API に送られる。'haiku' のようなエイリアスは解決されず、リクエストが 404 (not_found_error, model: haiku) で落ちる。だからこの Mod は MODEL_IDS で id に直してから渡している。詳細は NOTES.md。

テスト

claude plugin validate .
claude plugin test .

Source 1 files

hooks/register.ts 96 lines

1import type { Register } from 'claude-code'
2
3// `turn.step`'s `model` reaches the API verbatim: nothing resolves an alias
4// on the way down. `model: 'haiku'` makes the request itself 404 with
5// `not_found_error, model: haiku`, so a router writes ids, not family names.
6const MODEL_IDS: Record<string, string> = {
7  haiku: 'claude-haiku-4-5-20251001',
8  sonnet: 'claude-sonnet-5',
9  opus: 'claude-opus-5',
10}
11const modelId = (name: string): string => MODEL_IDS[name.trim().toLowerCase()] ?? name
12const SUBAGENT_MODEL = modelId('haiku')
13const SHORT_PROMPT = 80
14
15export const register: Register = on => {
16  let override: string | null = null
17  let lastPromptLength = 0
18  const log: string[] = []
19
20  on('session.start', async ($, e, next) => {
21    const r = await next(e)
22    const saved = await $.store.get('override')
23    override = typeof saved === 'string' ? saved : null
24    await $.command.register({
25      name: 'route',
26      description: 'Pin the main loop to a model, or hand it back (model-router)',
27      argumentHint: '<model | auto>',
28    })
29    return r
30  })
31
32  on('command.run', { command: 'route' }, async ($, e) => {
33    const arg = String(e.args ?? '').trim()
34    if (arg === 'auto') {
35      override = null
36      await $.store.delete('override')
37    } else if (arg !== '') {
38      // Resolved here, so what the person is told is what the API is sent.
39      override = modelId(arg)
40      await $.store.set('override', override)
41    }
42    return {
43      text: override
44        ? `model-router: the main loop runs on ${override}`
45        : 'model-router: the main loop keeps the model it asked for',
46    }
47  })
48
49  // The length of the last prompt is the only thing rule 2 needs, and a
50  // length is not the person's text, so nothing of the prompt is kept.
51  on('prompt.submit', async ($, e, next) => {
52    lastPromptLength = e.text.length
53    return next(e)
54  })
55
56  on('turn.step', async function* ($, e, next) {
57    let model = e.model
58    let effort = e.effort
59    let rule = 'none'
60    if (e.agentId !== undefined) {
61      if (model !== SUBAGENT_MODEL) {
62        model = SUBAGENT_MODEL
63        rule = 'subagent'
64      }
65    } else if (override !== null) {
66      model = override
67      rule = 'override'
68    } else if (e.index === 0 && lastPromptLength > 0 && lastPromptLength < SHORT_PROMPT) {
69      effort = 'low'
70      rule = 'short-prompt'
71    }
72
73    const r = yield* next({ ...e, model, effort })
74
75    const trace = next.trace
76      .map(t => `${t.plugin}/${t.tier}:${t.outcome}${t.chunks === undefined ? '' : `(${t.chunks}c)`}`)
77      .join(' > ')
78    log.push(
79      [
80        `index=${e.index}`,
81        `loop=${e.agentId ?? 'main'}`,
82        `rule=${rule}`,
83        `model=${e.model} -> ${model}`,
84        `effort=${String(e.effort)} -> ${String(effort)}`,
85        `messages=${e.messageCount}`,
86        `answered=${r.usage === null ? 'none' : r.usage.model}`,
87        `stop=${String(r.stopReason)}`,
88        `tools=${r.toolUses.length}`,
89        `trace=${trace}`,
90      ].join(' | '),
91    )
92    await $.fs.write(`${$.plugin.root}/PROOF-steps.log`, `${log.join('\n')}\n`)
93    return r
94  })
95}
96