wangjia/maestro
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(agent): 模型档位升 fable-5（复审/hard拆解/解冲突）+ 回退链 fable 置首

Browse Source 
按 plan ①②③④ 定稿（用户确认）：
- reviewer 三档统一 fable-5（不被 project.model 降档，保复审严格度）
- planner hard 拆解 opus→fable-5（最烧脑用最强档）；medium=opus/easy=sonnet 不变
- conflict 固定 fable-5（解冲突恒用最强档）
- MODEL_FALLBACK_CHAIN: [fable-5, opus-4-8, sonnet-4-6]（fable 置链首）
- executor 档不变（per plan ②）；CLAUDE.md 档位表同步更正

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
wangjia
2026-06-24 22:24:30 +08:00
parent 73f80d27c9
commit ff2d7c22e7
3 changed files with 36 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CLAUDE.md
+5 -3
View File
@@ -79,10 +79,12 @@ score = P(self) + Σ P(completed deps) + Σ P(tasks blocked by self)
| Role | Easy | Medium | Hard |
|---|---|---|---|
| executor | claude-sonnet-4-6 | claude-opus-4-8 | claude-fable-5 |
| reviewer | claude-sonnet-4-6 | claude-opus-4-8 | claude-opus-4-8 |
| executor | claude-sonnet-4-6 | claude-opus-4-8 | claude-opus-4-8 |
| planner | claude-sonnet-4-6 | claude-opus-4-8 | claude-fable-5 |
| reviewer | claude-fable-5 | claude-fable-5 | claude-fable-5 |
| conflict | claude-fable-5 | claude-fable-5 | claude-fable-5 |
Project-level `model` setting overrides env vars; env vars override built-in defaults.
Reviewer/conflict use the strongest tier (fable-5) and are NOT downgraded by project-level `model` (preserves self-review/conflict-resolution rigor). Project-level `model` overrides executor/planner env vars; env vars override built-in defaults. Model fallback chain on unavailability: `fable-5 → opus-4-8 → sonnet-4-6`.
### Data Directory
src/executor/models.ts
+13 -12
View File
@@ -4,8 +4,8 @@ import type { Complexity } from '../model/complexity.js';
/** 角色：executor（执行改动）/ reviewer（执行后复审）/ planner（拆解 Hard / 写方案 Medium）/ conflict（冲突裁决） */
export type ModelRole = 'executor' | 'reviewer' | 'planner' | 'conflict';
/** 模型不可用时的回退链（按序取下一档，同一 run 内只重试一次） */
export const MODEL_FALLBACK_CHAIN = ['claude-opus-4-8', 'claude-sonnet-4-6'] as const;
/** 模型不可用时的回退链（按序取下一档，同一 run 内只重试一次；fable 置链首=优先用最强档） */
export const MODEL_FALLBACK_CHAIN = ['claude-fable-5', 'claude-opus-4-8', 'claude-sonnet-4-6'] as const;
/** 复杂度 → [env 覆盖变量名, 默认模型] */
const MODEL_TABLE: Record<ModelRole, Record<Complexity, [env: string, fallback: string]>> = {
@@ -14,22 +14,23 @@ const MODEL_TABLE: Record<ModelRole, Record<Complexity, [env: string, fallback:
medium: ['MAESTRO_MODEL_MEDIUM', 'claude-opus-4-8'],
hard: ['MAESTRO_MODEL_HARD', 'claude-opus-4-8'],
},
// reviewer：复审是质量关口，统一用最强 fable-5（不受 project.model 降档，保自审严格度）
reviewer: {
easy: ['MAESTRO_MODEL_REVIEW_EASY', 'claude-opus-4-8'],
medium: ['MAESTRO_MODEL_REVIEW_MEDIUM', 'claude-opus-4-8'],
hard: ['MAESTRO_MODEL_REVIEW_HARD', 'claude-opus-4-8'],
easy: ['MAESTRO_MODEL_REVIEW_EASY', 'claude-fable-5'],
medium: ['MAESTRO_MODEL_REVIEW_MEDIUM', 'claude-fable-5'],
hard: ['MAESTRO_MODEL_REVIEW_HARD', 'claude-fable-5'],
},
// planner：拆解/写方案按复杂度分化（不受 project.model 覆盖）
// planner：拆解/写方案按复杂度分化（不受 project.model 覆盖）；hard 拆解最烧脑→fable-5
planner: {
easy: ['MAESTRO_MODEL_PLAN_EASY', 'claude-sonnet-4-6'],
medium: ['MAESTRO_MODEL_PLAN_MEDIUM', 'claude-opus-4-8'],
hard: ['MAESTRO_MODEL_PLAN_HARD', 'claude-opus-4-8'],
hard: ['MAESTRO_MODEL_PLAN_HARD', 'claude-fable-5'],
},
// conflict：固定 opus-4-8，TABLE 条目仅占位（pickModel 对此角色直接 return 不走此处）
// conflict：固定 fable-5（解冲突最难，恒用最强档），TABLE 条目仅占位（pickModel 对此角色直接 return）
conflict: {
easy: ['', 'claude-opus-4-8'],
medium: ['', 'claude-opus-4-8'],
hard: ['', 'claude-opus-4-8'],
easy: ['', 'claude-fable-5'],
medium: ['', 'claude-fable-5'],
hard: ['', 'claude-fable-5'],
},
};
@@ -44,7 +45,7 @@ const MODEL_TABLE: Record<ModelRole, Record<Complexity, [env: string, fallback:
export function pickModel(task: Task, project: Project, role: ModelRole): string {
// reviewer/conflict 用最强档，不受 project.model 覆盖（保自审独立性）
if (role !== 'reviewer' && role !== 'conflict' && project.model) return project.model;
if (role === 'conflict') return 'claude-opus-4-8';
if (role === 'conflict') return 'claude-fable-5';
const [env, fallback] = MODEL_TABLE[role][task.complexity];
const override = process.env[env]?.trim();
return override || fallback;
test/models.test.ts
+18 -13
View File
@@ -40,10 +40,14 @@ test('pickModel：复杂度映射默认值（executor / reviewer）', () => {
assert.equal(pickModel(fakeTask('easy'), p, 'executor'), 'claude-sonnet-4-6');
assert.equal(pickModel(fakeTask('medium'), p, 'executor'), 'claude-opus-4-8');
assert.equal(pickModel(fakeTask('hard'), p, 'executor'), 'claude-opus-4-8');
// reviewer 三档统一 opus-4-8（保复审独立性）
assert.equal(pickModel(fakeTask('easy'), p, 'reviewer'), 'claude-opus-4-8');
assert.equal(pickModel(fakeTask('medium'), p, 'reviewer'), 'claude-opus-4-8');
assert.equal(pickModel(fakeTask('hard'), p, 'reviewer'), 'claude-opus-4-8');
// reviewer 三档统一 fable-5（复审是质量关口，用最强档保严格度）
assert.equal(pickModel(fakeTask('easy'), p, 'reviewer'), 'claude-fable-5');
assert.equal(pickModel(fakeTask('medium'), p, 'reviewer'), 'claude-fable-5');
assert.equal(pickModel(fakeTask('hard'), p, 'reviewer'), 'claude-fable-5');
// planner hard 拆解最烧脑 → fable-5；medium=opus；easy=sonnet
assert.equal(pickModel(fakeTask('hard'), p, 'planner'), 'claude-fable-5');
assert.equal(pickModel(fakeTask('medium'), p, 'planner'), 'claude-opus-4-8');
assert.equal(pickModel(fakeTask('easy'), p, 'planner'), 'claude-sonnet-4-6');
});
});
@@ -53,10 +57,10 @@ test('pickModel：project.model executor 最优先，但 reviewer/conflict 不
for (const c of ['easy', 'medium', 'hard'] as const) {
// executor 受 project.model 覆盖
assert.equal(pickModel(fakeTask(c), p, 'executor'), 'claude-opus-4-6');
// reviewer 不受 project.model 覆盖，始终用 opus-4-8（保自审独立性）
assert.equal(pickModel(fakeTask(c), p, 'reviewer'), 'claude-opus-4-8');
// conflict 固定 opus-4-8
assert.equal(pickModel(fakeTask(c), p, 'conflict'), 'claude-opus-4-8');
// reviewer 不受 project.model 覆盖，始终用 fable-5（保自审独立性）
assert.equal(pickModel(fakeTask(c), p, 'reviewer'), 'claude-fable-5');
// conflict 固定 fable-5（解冲突恒用最强档）
assert.equal(pickModel(fakeTask(c), p, 'conflict'), 'claude-fable-5');
}
});
});
@@ -70,8 +74,8 @@ test('pickModel：env 覆盖默认值，executor/reviewer 各用各的 env', ()
assert.equal(pickModel(fakeTask('easy'), p, 'executor'), 'claude-haiku-4-5');
// reviewer easy 被 MAESTRO_MODEL_REVIEW_EASY 覆盖
assert.equal(pickModel(fakeTask('easy'), p, 'reviewer'), 'claude-sonnet-4-6');
// reviewer hard 无 env → 默认 opus-4-8
assert.equal(pickModel(fakeTask('hard'), p, 'reviewer'), 'claude-opus-4-8');
// reviewer hard 无 env → 默认 fable-5
assert.equal(pickModel(fakeTask('hard'), p, 'reviewer'), 'claude-fable-5');
assert.equal(pickModel(fakeTask('hard'), p, 'executor'), 'claude-opus-4-8');
});
});
@@ -98,8 +102,9 @@ test('isModelError：模型不可用类错误才触发回退', () => {
});
test('pickFallbackModel：回退链取下一档（循环）', () => {
// 链：opus → sonnet → opus（循环）
assert.deepEqual([...MODEL_FALLBACK_CHAIN], ['claude-opus-4-8', 'claude-sonnet-4-6']);
// 链：fable → opus → sonnet → fable（循环，fable 置链首=优先最强档）
assert.deepEqual([...MODEL_FALLBACK_CHAIN], ['claude-fable-5', 'claude-opus-4-8', 'claude-sonnet-4-6']);
assert.equal(pickFallbackModel('claude-fable-5'), 'claude-opus-4-8');
assert.equal(pickFallbackModel('claude-opus-4-8'), 'claude-sonnet-4-6');
assert.equal(pickFallbackModel('claude-sonnet-4-6'), 'claude-opus-4-8');
assert.equal(pickFallbackModel('claude-sonnet-4-6'), 'claude-fable-5');
});