mirror of
https://github.com/hansjone/dsh-im-ops.git
synced 2026-10-09 00:33:20 +08:00
feat: add reasoning effort commands
This commit is contained in:
xmanrui
parent
36cbbe48e7
commit
9e39333ff6
24 changed files with 1346 additions and 240 deletions
|
|
@ -25,6 +25,54 @@ const CATALOG = Object.freeze({
|
|||
failures: [],
|
||||
});
|
||||
|
||||
const REASONING_CATALOG = Object.freeze({
|
||||
groups: [
|
||||
{
|
||||
id: 'deepseek-official',
|
||||
name: 'DeepSeek',
|
||||
models: [
|
||||
{
|
||||
id: 'deepseek-v4-flash',
|
||||
name: 'DeepSeek V4 Flash',
|
||||
reasoning: {
|
||||
efforts: [
|
||||
{ id: 'off', name: 'Off' },
|
||||
{ id: 'high', name: 'High', description: '复杂任务' },
|
||||
{ id: 'max', name: 'Max' },
|
||||
],
|
||||
defaultEffort: 'high',
|
||||
},
|
||||
},
|
||||
{
|
||||
id: 'deepseek-v4-pro',
|
||||
name: 'DeepSeek V4 Pro',
|
||||
reasoning: {
|
||||
efforts: [
|
||||
{ id: 'medium', name: 'Medium' },
|
||||
{ id: 'xhigh', name: 'XHigh' },
|
||||
],
|
||||
defaultEffort: 'medium',
|
||||
},
|
||||
},
|
||||
],
|
||||
},
|
||||
CATALOG.groups[1],
|
||||
],
|
||||
failures: [],
|
||||
});
|
||||
|
||||
function reasoningSessionCatalog(current = {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'high',
|
||||
}) {
|
||||
return {
|
||||
...REASONING_CATALOG,
|
||||
current,
|
||||
routable: true,
|
||||
};
|
||||
}
|
||||
|
||||
function deferred() {
|
||||
let resolve;
|
||||
const promise = new Promise((settle) => { resolve = settle; });
|
||||
|
|
@ -113,13 +161,17 @@ function fixture({
|
|||
return { calls, harness, state, boundId: () => boundId };
|
||||
}
|
||||
|
||||
test('isModelCommand recognizes only /models and /model command prefixes', () => {
|
||||
test('isModelCommand recognizes model and reasoning command prefixes', () => {
|
||||
for (const command of [
|
||||
'/models', ' /MODELS ', '/models ignored', '/model', '/MoDeL openai/gpt-5',
|
||||
'/reasoning', '/REASONING high', '/reasoninglist', '/ReAsOnInGs',
|
||||
]) {
|
||||
assert.equal(isModelCommand(command), true, command);
|
||||
}
|
||||
for (const value of [null, '', 'model', '/modelx', '/modelsx', 'hello /models']) {
|
||||
for (const value of [
|
||||
null, '', 'model', '/modelx', '/modelsx', 'hello /models',
|
||||
'/reasoningx', '/reasoninglists', 'hello /reasoning',
|
||||
]) {
|
||||
assert.equal(isModelCommand(value), false, String(value));
|
||||
}
|
||||
});
|
||||
|
|
@ -209,6 +261,310 @@ test('/model reports current state without creating or selecting', async () => {
|
|||
assert.equal(existingFixture.calls.some(([name]) => name === 'selectModel'), false);
|
||||
});
|
||||
|
||||
test('/reasoninglist and /reasonings are identical read-only aliases', async () => {
|
||||
const first = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const second = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const listed = await runModelCommand(
|
||||
'/reasoninglist', first.harness, first.state, 'direct:one',
|
||||
);
|
||||
const aliased = await runModelCommand(
|
||||
'/REASONINGS', second.harness, second.state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.equal(aliased.message, listed.message);
|
||||
assert.match(listed.message, /deepseek-official\/deepseek-v4-flash/);
|
||||
assert.match(listed.message, /2\. High \(high\)（当前、默认）/);
|
||||
assert.match(listed.message, /复杂任务/);
|
||||
assert.match(listed.message, /\/reasoning --default/);
|
||||
assert.equal(first.calls.some(([name]) => name === 'selectModel'), false);
|
||||
assert.equal(second.calls.some(([name]) => name === 'selectModel'), false);
|
||||
});
|
||||
|
||||
test('/reasoninglist keeps default reset available without current-model metadata', async () => {
|
||||
for (const sessionCatalog of [
|
||||
{
|
||||
...CATALOG,
|
||||
current: { provider: 'deepseek-official', model: 'deepseek-v4-flash' },
|
||||
routable: true,
|
||||
},
|
||||
{
|
||||
...CATALOG,
|
||||
current: {
|
||||
provider: 'private-provider',
|
||||
model: 'private-model',
|
||||
reasoningEffort: 'custom-high',
|
||||
},
|
||||
routable: true,
|
||||
},
|
||||
]) {
|
||||
const current = fixture({ initialSessionId: 'session-one', sessionCatalog });
|
||||
const listed = await runModelCommand(
|
||||
'/reasoninglist', current.harness, current.state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(listed.message, /不提供可切换的推理等级/);
|
||||
assert.match(listed.message, /恢复默认等级：\/reasoning --default/);
|
||||
assert.equal(current.calls.some(([name]) => name === 'selectModel'), false);
|
||||
}
|
||||
});
|
||||
|
||||
test('reasoning read commands require an existing Session and validate syntax', async () => {
|
||||
const missing = fixture();
|
||||
for (const command of ['/reasoning', '/reasoninglist', '/reasonings']) {
|
||||
const result = await runModelCommand(command, missing.harness, missing.state, 'direct:one');
|
||||
assert.match(result.message, /还没有会话/, command);
|
||||
}
|
||||
assert.equal(missing.calls.some(([name]) => name === 'createSession'), false);
|
||||
|
||||
const existing = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
assert.match(
|
||||
(await runModelCommand(
|
||||
'/reasoninglist extra', existing.harness, existing.state, 'direct:one',
|
||||
)).message,
|
||||
/不带参数/,
|
||||
);
|
||||
assert.match(
|
||||
(await runModelCommand(
|
||||
'/reasoning high extra', existing.harness, existing.state, 'direct:one',
|
||||
)).message,
|
||||
/用法/,
|
||||
);
|
||||
});
|
||||
|
||||
test('/reasoning reports the explicit or model-default effort without selecting', async () => {
|
||||
const explicit = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const explicitResult = await runModelCommand(
|
||||
'/reasoning', explicit.harness, explicit.state, 'direct:one',
|
||||
);
|
||||
assert.match(explicitResult.message, /当前推理等级：High \(high\)/);
|
||||
|
||||
const inherited = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog({
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
}),
|
||||
});
|
||||
const inheritedResult = await runModelCommand(
|
||||
'/reasoning', inherited.harness, inherited.state, 'direct:one',
|
||||
);
|
||||
assert.match(inheritedResult.message, /当前推理等级：High \(high\)/);
|
||||
assert.equal(explicit.calls.some(([name]) => name === 'selectModel'), false);
|
||||
assert.equal(inherited.calls.some(([name]) => name === 'selectModel'), false);
|
||||
});
|
||||
|
||||
test('/reasoning preserves an unknown current effort and represents Provider Default', async () => {
|
||||
const groups = structuredClone(REASONING_CATALOG.groups);
|
||||
delete groups[0].models[0].reasoning.defaultEffort;
|
||||
const raw = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: {
|
||||
groups,
|
||||
failures: [],
|
||||
current: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'gateway-ultra',
|
||||
},
|
||||
routable: true,
|
||||
},
|
||||
});
|
||||
const listed = await runModelCommand(
|
||||
'/reasoninglist', raw.harness, raw.state, 'direct:one',
|
||||
);
|
||||
assert.match(listed.message, /当前推理等级：gateway-ultra/);
|
||||
assert.doesNotMatch(listed.message, /（当前/);
|
||||
|
||||
const providerDefault = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: {
|
||||
groups,
|
||||
failures: [],
|
||||
current: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
},
|
||||
routable: true,
|
||||
},
|
||||
});
|
||||
const current = await runModelCommand(
|
||||
'/reasoning', providerDefault.harness, providerDefault.state, 'direct:one',
|
||||
);
|
||||
assert.match(current.message, /Default（由模型或 Provider 决定）/);
|
||||
});
|
||||
|
||||
test('/reasoning switches the current model effort by index or exact ID', async () => {
|
||||
for (const [requested, expected] of [['3', 'max'], ['off', 'off']]) {
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const result = await runModelCommand(
|
||||
`/reasoning ${requested}`, harness, state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(result.message, new RegExp(`推理等级已切换为：[\\s\\S]*${expected}`));
|
||||
assert.deepEqual(calls.find(([name]) => name === 'selectModel'), [
|
||||
'selectModel',
|
||||
'session-one',
|
||||
{
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: expected,
|
||||
},
|
||||
{},
|
||||
]);
|
||||
}
|
||||
});
|
||||
|
||||
test('/reasoning prefers an exact opaque numeric effort ID before an index', async () => {
|
||||
const groups = structuredClone(REASONING_CATALOG.groups);
|
||||
groups[0].models[0].reasoning = {
|
||||
efforts: [
|
||||
{ id: '3', name: 'Literal Three' },
|
||||
{ id: '0', name: 'Literal Zero' },
|
||||
{ id: '9007199254740992', name: 'Huge Numeric' },
|
||||
],
|
||||
defaultEffort: '3',
|
||||
};
|
||||
|
||||
for (const requested of ['3', '0', '9007199254740992']) {
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: {
|
||||
groups,
|
||||
failures: [],
|
||||
current: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: '3',
|
||||
},
|
||||
routable: true,
|
||||
},
|
||||
});
|
||||
const result = await runModelCommand(
|
||||
`/reasoning ${requested}`, harness, state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(result.message, /推理等级已切换为/);
|
||||
assert.equal(
|
||||
calls.find(([name]) => name === 'selectModel')?.[2]?.reasoningEffort,
|
||||
requested,
|
||||
);
|
||||
}
|
||||
});
|
||||
|
||||
test('/reasoning --default restores the current model default effort', async () => {
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog({
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'max',
|
||||
}),
|
||||
selectionResult: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'high',
|
||||
},
|
||||
selectionReadback: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'high',
|
||||
},
|
||||
});
|
||||
const result = await runModelCommand(
|
||||
'/reasoning --default', harness, state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(result.message, /推理等级已切换为：[\s\S]*High \(high\)/);
|
||||
assert.deepEqual(calls.find(([name]) => name === 'selectModel')?.[2], {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
});
|
||||
});
|
||||
|
||||
test('/reasoning --default can reset an advisory-unlisted current model', async () => {
|
||||
const sessionCatalog = {
|
||||
...CATALOG,
|
||||
current: {
|
||||
provider: 'private-provider',
|
||||
model: 'private-model',
|
||||
reasoningEffort: 'custom-high',
|
||||
},
|
||||
routable: true,
|
||||
};
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog,
|
||||
});
|
||||
const result = await runModelCommand(
|
||||
'/reasoning --default', harness, state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(result.message, /推理等级已切换为/);
|
||||
assert.match(result.message, /Default/);
|
||||
assert.deepEqual(calls.find(([name]) => name === 'selectModel')?.[2], {
|
||||
provider: 'private-provider',
|
||||
model: 'private-model',
|
||||
});
|
||||
});
|
||||
|
||||
test('/reasoning rejects unsupported levels without changing the selection', async () => {
|
||||
for (const requested of ['0', '4', '9007199254740992', 'medium']) {
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const result = await runModelCommand(
|
||||
`/reasoning ${requested}`, harness, state, 'direct:one',
|
||||
);
|
||||
assert.match(result.message, /序号无效|不支持推理等级/, requested);
|
||||
assert.equal(calls.some(([name]) => name === 'selectModel'), false, requested);
|
||||
}
|
||||
|
||||
const unsupported = fixture({ initialSessionId: 'session-one' });
|
||||
const result = await runModelCommand(
|
||||
'/reasoning high', unsupported.harness, unsupported.state, 'direct:one',
|
||||
);
|
||||
assert.match(result.message, /不提供可切换的推理等级/);
|
||||
assert.equal(unsupported.calls.some(([name]) => name === 'selectModel'), false);
|
||||
|
||||
const noSession = fixture({ globalCatalog: REASONING_CATALOG });
|
||||
const noSessionResult = await runModelCommand(
|
||||
'/reasoning high', noSession.harness, noSession.state, 'direct:one',
|
||||
);
|
||||
assert.match(noSessionResult.message, /还没有会话/);
|
||||
assert.equal(noSession.calls.some(([name]) => name === 'createSession'), false);
|
||||
});
|
||||
|
||||
test('reasoning commands reject image-bearing control messages', async () => {
|
||||
const current = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
for (const command of ['/reasoning', '/reasoninglist', '/reasonings', '/reasoning high']) {
|
||||
const result = await runModelCommand(
|
||||
command, current.harness, current.state, 'direct:one', { hasImages: true },
|
||||
);
|
||||
assert.match(result.message, /仅支持纯文字/, command);
|
||||
}
|
||||
assert.equal(current.calls.some(([name]) => name === 'models'), false);
|
||||
assert.equal(current.calls.some(([name]) => name === 'selectModel'), false);
|
||||
});
|
||||
|
||||
test('/model uses an exact catalog ID and preserves slashes inside the model ID', async () => {
|
||||
const { calls, harness, state } = fixture({ initialSessionId: 'session-one' });
|
||||
const control = Object.freeze({ route: 'direct:one' });
|
||||
|
|
@ -242,6 +598,71 @@ test('/model accepts the current catalog\'s global 1-based model number', async
|
|||
]);
|
||||
});
|
||||
|
||||
test('/model accepts an optional effort and otherwise applies the target model default', async () => {
|
||||
for (const [command, expectedEffort] of [
|
||||
['/model 2 xhigh', 'xhigh'],
|
||||
['/model deepseek-official/deepseek-v4-pro', 'medium'],
|
||||
]) {
|
||||
const resolvedDefault = command.includes(' xhigh') ? undefined : {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-pro',
|
||||
reasoningEffort: 'medium',
|
||||
};
|
||||
const { calls, harness, state } = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
globalCatalog: REASONING_CATALOG,
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
selectionResult: resolvedDefault,
|
||||
selectionReadback: resolvedDefault,
|
||||
});
|
||||
const result = await runModelCommand(command, harness, state, 'direct:one');
|
||||
|
||||
assert.match(result.message, /deepseek-official\/deepseek-v4-pro/);
|
||||
assert.match(result.message, new RegExp(`推理等级：.*${expectedEffort}`));
|
||||
assert.deepEqual(calls.find(([name]) => name === 'selectModel')?.[2], {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-pro',
|
||||
...(command.includes(' xhigh') ? { reasoningEffort: expectedEffort } : {}),
|
||||
});
|
||||
}
|
||||
});
|
||||
|
||||
test('/model validates an optional effort against the target model atomically', async () => {
|
||||
const invalid = fixture({ globalCatalog: REASONING_CATALOG });
|
||||
const failed = await runModelCommand(
|
||||
'/model 2 high', invalid.harness, invalid.state, 'direct:one',
|
||||
);
|
||||
|
||||
assert.match(failed.message, /不支持推理等级：high/);
|
||||
assert.match(failed.message, /medium, xhigh/);
|
||||
assert.equal(invalid.calls.some(([name]) => name === 'createSession'), false);
|
||||
assert.equal(invalid.calls.some(([name]) => name === 'selectModel'), false);
|
||||
|
||||
const valid = fixture({ globalCatalog: REASONING_CATALOG });
|
||||
const switched = await runModelCommand(
|
||||
'/model 2 xhigh', valid.harness, valid.state, 'direct:one',
|
||||
);
|
||||
assert.match(switched.message, /模型已切换为/);
|
||||
assert.equal(valid.boundId(), 'session-created');
|
||||
assert.deepEqual(valid.calls.find(([name]) => name === 'selectModel')?.[2], {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-pro',
|
||||
reasoningEffort: 'xhigh',
|
||||
});
|
||||
});
|
||||
|
||||
test('/model reports the current reasoning effort when the Session exposes it', async () => {
|
||||
const current = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const result = await runModelCommand('/model', current.harness, current.state, 'direct:one');
|
||||
|
||||
assert.match(result.message, /当前推理等级：High \(high\)/);
|
||||
assert.match(result.message, /\/reasoninglist/);
|
||||
assert.equal(current.calls.some(([name]) => name === 'selectModel'), false);
|
||||
});
|
||||
|
||||
test('/model rejects invalid model numbers without creating or selecting a Session', async () => {
|
||||
for (const requested of ['0', '4', '9007199254740992']) {
|
||||
const { calls, harness, state } = fixture();
|
||||
|
|
@ -469,6 +890,78 @@ test('/model refuses pending interactions and active or running Sessions', async
|
|||
}
|
||||
});
|
||||
|
||||
test('/reasoning mutations refuse pending or busy Sessions while reads remain available', async () => {
|
||||
const pending = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
});
|
||||
const pendingResult = await runModelCommand(
|
||||
'/reasoning off', pending.harness, pending.state, 'direct:one',
|
||||
{ pendingInteraction: true },
|
||||
);
|
||||
assert.match(pendingResult.message, /等待你的回答或审批/);
|
||||
assert.equal(pending.calls.some(([name]) => name === 'selectModel'), false);
|
||||
|
||||
for (const runState of [{ running: true }, { activeTurn: true }]) {
|
||||
const active = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
...runState,
|
||||
});
|
||||
const failed = await runModelCommand(
|
||||
'/reasoning off', active.harness, active.state, 'direct:one',
|
||||
{ control: 'owner-one' },
|
||||
);
|
||||
assert.match(failed.message, /当前任务正在运行/);
|
||||
assert.equal(active.calls.some(([name]) => name === 'selectModel'), false);
|
||||
|
||||
const listed = await runModelCommand(
|
||||
'/reasoninglist', active.harness, active.state, 'direct:one',
|
||||
);
|
||||
assert.match(listed.message, /可用推理等级/);
|
||||
}
|
||||
});
|
||||
|
||||
test('/reasoning verifies both the selected effort and authoritative readback', async () => {
|
||||
const wrongSelection = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
selectionResult: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'high',
|
||||
},
|
||||
});
|
||||
const selectionFailure = await runModelCommand(
|
||||
'/reasoning max',
|
||||
wrongSelection.harness,
|
||||
wrongSelection.state,
|
||||
'direct:one',
|
||||
);
|
||||
assert.match(selectionFailure.message, /推理等级切换失败/);
|
||||
assert.match(selectionFailure.message, /reasoningEffort=max/);
|
||||
assert.match(selectionFailure.message, /reasoningEffort=high/);
|
||||
|
||||
const wrongReadback = fixture({
|
||||
initialSessionId: 'session-one',
|
||||
sessionCatalog: reasoningSessionCatalog(),
|
||||
selectionReadback: {
|
||||
provider: 'deepseek-official',
|
||||
model: 'deepseek-v4-flash',
|
||||
reasoningEffort: 'off',
|
||||
},
|
||||
});
|
||||
const readbackFailure = await runModelCommand(
|
||||
'/reasoning max',
|
||||
wrongReadback.harness,
|
||||
wrongReadback.state,
|
||||
'direct:one',
|
||||
);
|
||||
assert.match(readbackFailure.message, /推理等级切换失败/);
|
||||
assert.match(readbackFailure.message, /reasoningEffort=max/);
|
||||
assert.match(readbackFailure.message, /reasoningEffort=off/);
|
||||
});
|
||||
|
||||
test('a missing bound Session is cleared and /models falls back to the global catalog', async () => {
|
||||
const { calls, harness, state, boundId } = fixture({
|
||||
initialSessionId: 'session-missing',
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue