From 6c50308e1299e45feea6d363d384ae56955b42d5 Mon Sep 17 00:00:00 2001 From: antianqi <75944423+antianqi@users.noreply.github.com> Date: Sat, 7 Nov 2026 15:08:22 +0800 Subject: [PATCH] =?UTF-8?q?fix(ai-node-canvas):=20=E5=87=BA=E5=9B=BE?= =?UTF-8?q?=E8=B6=85=E8=BF=87=2030=20=E7=A7=92=E6=97=B6=E6=94=B9=E8=B5=B0?= =?UTF-8?q?=20Agent=20=E9=80=9A=E9=81=93=EF=BC=8C=E4=B8=8D=E5=86=8D?= =?UTF-8?q?=E7=9B=B4=E6=8E=A5=E5=A4=B1=E8=B4=A5?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 宿主对一次连接器调用有 30 秒上限，超时会被中止并报「结果未知」。图片没有任务号可查， provider 那边做出来的那张也就找不回来，所以带参考图的慢速出图过去会直接失败一次。 现在出图超时不再失败：直连被掐断后，把直连本来要发的那份请求原样交给 Agent 跑一轮 （宿主不限制 Agent 回合的时长），图从那条路回来，并落一份本地副本。页面上不弹失败， 运行日志里写明换了路。 只有 CONNECTOR_TIMEOUT 和 CONNECTOR_OUTCOME_UNKNOWN 走这条路；provider 明确报错不降级。 版本 0.86.0 -> 0.87.0。 --- .../.minimax-plugin/plugin.json | 2 +- plugins/antianqi/ai-node-canvas/README.md | 15 ++-- .../antianqi/ai-node-canvas/README.zh-CN.md | 8 +- .../miniapp/node/agent-runner.mjs | 70 +++++++++++++++++ .../miniapp/node/connector-runner.mjs | 10 ++- plugins/antianqi/ai-node-canvas/package.json | 2 +- .../ai-node-canvas/tests/smoke.test.mjs | 77 +++++++++++++++++-- 7 files changed, 167 insertions(+), 17 deletions(-) diff --git a/plugins/antianqi/ai-node-canvas/.minimax-plugin/plugin.json b/plugins/antianqi/ai-node-canvas/.minimax-plugin/plugin.json index 08fecab..1abe9a3 100644 --- a/plugins/antianqi/ai-node-canvas/.minimax-plugin/plugin.json +++ b/plugins/antianqi/ai-node-canvas/.minimax-plugin/plugin.json @@ -2,7 +2,7 @@ "schemaVersion": 1, "name": "ai-node-canvas", "displayName": "节点画布", - "version": "0.86.0", + "version": "0.87.0", "description": "无限画布节点编排：把 MiniMax 模型生成的文本、图片、视频串成工作流，连线即数据流，上游产物可直接被下游引用。", "author": "antianqi", "icon": "icon.png", diff --git a/plugins/antianqi/ai-node-canvas/README.md b/plugins/antianqi/ai-node-canvas/README.md index c1fcfc5..68870a1 100644 --- a/plugins/antianqi/ai-node-canvas/README.md +++ b/plugins/antianqi/ai-node-canvas/README.md @@ -5,7 +5,7 @@ English | [简体中文](README.zh-CN.md) An infinite-canvas node graph for MiniMax Code, modelled on the canvas of `work.batvid.com`. -Author: [antianqi](https://github.com/antianqi) · Version: `0.86.0` +Author: [antianqi](https://github.com/antianqi) · Version: `0.87.0` ![The canvas with the 图片复刻 template on it: a reference picture goes in, the prompt is read back out of that picture, and a new picture is made from the prompt](docs/preview.png) @@ -668,10 +668,15 @@ separate pictures. - **How long a connector call may take** — the Host puts a timeout on one connector call (30 seconds by default). A call that runs over it is aborted and reported as an unknown outcome: it may have been dispatched, and it may - have been charged, so it must never be retried automatically. A slow - generation (an image generation with a reference is the one that hits it - most often) can therefore fail once; the message names the tool and how long - it ran. Video is submitted and then polled, so it is not affected. + have been charged, so it must never be retried automatically. A picture has no + task number to ask about, so the one the provider makes cannot be found + afterwards — and a slow generation (an image generation with a reference is + the one that hits the limit most often) would be lost with it. So a picture + that runs over the limit does not fail: the run goes on by the other road. + The Agent is handed the very request the direct call would have sent and runs + it as a turn, which the Host does not time out, and the picture comes back + from there. The page is shown no failure; the run log says the road changed. + Video is submitted and then polled, so it is not affected. - **Processes spawned** — `ffmpeg` and `ffprobe`, and nothing else. Both are looked up on the machine and run with an argument array, never through a shell, so no part of a path or a parameter can be read as shell syntax. diff --git a/plugins/antianqi/ai-node-canvas/README.zh-CN.md b/plugins/antianqi/ai-node-canvas/README.zh-CN.md index fe41b4f..8c4ba69 100644 --- a/plugins/antianqi/ai-node-canvas/README.zh-CN.md +++ b/plugins/antianqi/ai-node-canvas/README.zh-CN.md @@ -5,7 +5,7 @@ MiniMax Code 里的无限画布节点图。每个节点是一次模型运行，每条连线是一条数据通路—— 把文本节点连到图片节点，上游文本会自动注入下游提示词。 -作者：[antianqi](https://github.com/antianqi) · 版本：`0.86.0` +作者：[antianqi](https://github.com/antianqi) · 版本：`0.87.0` ![画布：铺了「图片复刻」模板——参考图进去，提示词从这张图里反推出来，再照着它重新生成一张](docs/preview.png) @@ -419,8 +419,10 @@ provider 有**几种不同的请求**把素材变成视频，不是同一个请 （回环地址除外），请求由运行时自己发出，**不走**宿主连接器。 - **连接器调用的时限** — 宿主对一次连接器调用有超时上限（默认 30 秒），超时的调用会被 中止并报「结果未知」：调用可能已经发出去了，也可能已经扣费，所以**不能自动重试**。 - 慢的生成（带参考图的图片生成最容易碰到）因此可能失败一次，失败信息里会写明是哪一个 - 工具、跑了多少秒。视频走「提交 + 轮询」，不受这个时限影响。 + 图片没有任务号可查，provider 那边做出来的那张也就找不回来——出图超限**不会失败**： + 这次运行改走另一条路，把直连本来要发的那份请求原样交给 Agent 跑一轮（宿主不限制 Agent + 回合的时长），图从那条路回来。页面上不弹失败，运行日志里写明换了路。 + 视频走「提交 + 轮询」，不受这个时限影响。 - **启动的进程** — 只有 `ffmpeg` 与 `ffprobe`。两者都在本机查找，调用时用参数数组、 绝不经过 shell，所以路径或参数里的任何字符都不会被当成 shell 语法。具体命令： - `ffmpeg -hide_banner -loglevel error -y <输入> …` — 每个媒体工具一次调用： diff --git a/plugins/antianqi/ai-node-canvas/miniapp/node/agent-runner.mjs b/plugins/antianqi/ai-node-canvas/miniapp/node/agent-runner.mjs index 5ba4dc7..161b113 100644 --- a/plugins/antianqi/ai-node-canvas/miniapp/node/agent-runner.mjs +++ b/plugins/antianqi/ai-node-canvas/miniapp/node/agent-runner.mjs @@ -42,6 +42,11 @@ const MODES = new Set(['image', 'video', 'audio', 'music', 'text']); // so it keeps the Agent turn. const DIRECT_MODES = new Set(['image', 'video', 'audio', 'music']); +// The codes that mean the Host stopped a connector call it had already sent: +// the provider may still be working on it, and a picture has no task number +// to ask about afterwards. +const STOPPED_CODES = new Set(['CONNECTOR_TIMEOUT', 'CONNECTOR_OUTCOME_UNKNOWN']); + const ERROR_TEXT = { NOT_PERMITTED: '没有 Agent 访问权限（需要 MiniMax Code 3.1.2 或更新版本）', BUSY: 'Agent 正忙，请稍后重试', @@ -383,6 +388,35 @@ function buildPrompt({ mode, segments, prompt, references, params, model, path = return { agentPrompt: sections.join('\n\n'), compiled, providerPrompt, assets }; } +/** + * What the Agent is told when the direct call was stopped. + * + * The request is the one the direct path would have sent, written out as the + * tool call to make. The Agent decides nothing here: it is handed the same + * request and asked to run it by the other road, because the Host does not + * time a turn out the way it times a connector call out. + */ +function agentImageInstruction({ prompt, params, assets, count = 1 }) { + const request = { prompt: String(prompt ?? '') }; + if (params?.ratio) request.aspect_ratio = params.ratio; + if (params?.resolution) request.resolution = params.resolution; + const urls = (assets?.images ?? []) + .map((item) => String(item?.url ?? item ?? '')) + .filter(Boolean); + if (urls.length) request.input_urls = urls; + // A batch is one call with one request per asset: the direct path gives + // every asset a window of its own, but a turn has no such window. + const total = Math.max(1, Number(count) || 1); + const requests = []; + for (let index = 0; index < total; index += 1) requests.push({ ...request }); + return [ + '刚才那条直连生成通道超时了，现在改走另一条路：请调用图片生成工具重新生成，只调用这一次。', + '参数严格照抄下面这一份，不要改动，也不要额外生成别的：', + JSON.stringify({ requests }), + '生成之后，对每一张图片调用 get_asset_url 换成能访问的链接，并把每条链接都写在回复里。', + ].join('\n\n'); +} + export function createAgentRunner(context) { const tasks = new Map(); const nodeTasks = new Map(); @@ -696,6 +730,9 @@ export function createAgentRunner(context) { assets: { images: [], videos: [], audios: [] }, // Which path owns this run: the connector (direct) or an Agent turn. runner: 'agent', + // Set when the direct call was stopped and the run went on by the + // Agent road instead. + degraded: false, providerTaskId: null, providerModel: null, providerStatus: null, @@ -836,6 +873,9 @@ export function createAgentRunner(context) { // plan quota. The Host owns the credentials; no key is handled here. async function executeConnector(task, { mode, prompt, params, assets, modelId, count = 1, path, frames }) { const signal = armController(task); + // Set when the direct call was stopped and the run goes on by the Agent + // road instead. + let degrade = false; try { const result = await connector.submit({ mode, prompt, params, assets, modelId, signal, count, path, frames }); if (task.cancelled) { @@ -876,6 +916,15 @@ export function createAgentRunner(context) { task.status = 'aborted'; task.error = ''; task.log.push('已取消'); + } else if (mode === 'image' && STOPPED_CODES.has(error?.code)) { + // The Host stopped the call and the outcome is unknown: the provider + // may still be working on it. A picture has no task number to ask + // about, so the one it makes cannot be found afterwards. + // + // The page is not told the run failed. The run goes on by the other + // road: the Host does not time an Agent turn out, so the same + // request can run there as long as it needs. + degrade = true; } else { task.status = 'error'; task.error = describeError(error?.code ?? 'UNKNOWN', error?.message); @@ -884,6 +933,27 @@ export function createAgentRunner(context) { } finally { releaseController(task); } + if (degrade) { + task.degraded = true; + task.runner = 'agent'; + task.log.push('直连超过 30 秒，改走 Agent 通道重新生成'); + await execute(task, { text: agentImageInstruction({ prompt, params, assets, count }) }); + // The Agent's links are the provider's own and they expire, so the + // canvas keeps a local copy: the node still has its picture later. + if (task.status === 'done' && task.media?.length) { + task.media = await Promise.all(task.media.map(async (item) => { + const local = await store.publishRemote(item.url, item.kind).catch(() => null); + return local ? { ...item, url: local } : item; + })); + } + // A turn that ended without a picture is not a picture: the run says + // so, instead of finishing with nothing to show. + if (task.status === 'done' && !(task.media ?? []).length) { + task.status = 'error'; + task.error = 'Agent 通道没有返回图片，请再试一次'; + task.log.push(`失败：${task.error}`); + } + } return task; } diff --git a/plugins/antianqi/ai-node-canvas/miniapp/node/connector-runner.mjs b/plugins/antianqi/ai-node-canvas/miniapp/node/connector-runner.mjs index 62df2fe..366d987 100644 --- a/plugins/antianqi/ai-node-canvas/miniapp/node/connector-runner.mjs +++ b/plugins/antianqi/ai-node-canvas/miniapp/node/connector-runner.mjs @@ -332,15 +332,23 @@ export function createConnectorRunner(context, options = {}) { // calls themselves took. The two are told apart, because only one of // them is the provider's answer. const timing = { waitedMs: 0, callMs: 0 }; + // The first failure keeps its own code: whether the Host stopped the call + // or the provider refused it decides what the run does next, so the code + // is not flattened into one thing. + let firstError = null; for (let index = 0; index < wanted; index += 1) { try { media.push(...(await callOneAsset(tool, argsFor(index), kind, signal, timing))); } catch (error) { + if (!firstError) firstError = error; warnings.push(`第 ${index + 1} 个没有生成出来：${error?.message ?? error}`); } } if (!media.length && warnings.length) { - throw Object.assign(new Error(warnings[0].replace(/^第 \d+ 个没有生成出来：/, '')), { code: 'CONNECTOR_ERROR' }); + throw Object.assign( + new Error(warnings[0].replace(/^第 \d+ 个没有生成出来：/, '')), + { code: firstError?.code ?? 'CONNECTOR_ERROR' }, + ); } return { media, warnings, waitedMs: timing.waitedMs, callMs: timing.callMs }; } diff --git a/plugins/antianqi/ai-node-canvas/package.json b/plugins/antianqi/ai-node-canvas/package.json index 902eca9..411f1c2 100644 --- a/plugins/antianqi/ai-node-canvas/package.json +++ b/plugins/antianqi/ai-node-canvas/package.json @@ -1,6 +1,6 @@ { "name": "ai-node-canvas", - "version": "0.86.0", + "version": "0.87.0", "private": true, "description": "Node-graph canvas MiniApp for MiniMax Code: wire MiniMax model runs into a visual workflow.", "mcode": { diff --git a/plugins/antianqi/ai-node-canvas/tests/smoke.test.mjs b/plugins/antianqi/ai-node-canvas/tests/smoke.test.mjs index 5c90a7f..641d1c4 100644 --- a/plugins/antianqi/ai-node-canvas/tests/smoke.test.mjs +++ b/plugins/antianqi/ai-node-canvas/tests/smoke.test.mjs @@ -2994,7 +2994,7 @@ test('a call that waited for the machine says how long it waited when it fails', // call, how long it ran, and how long it stood behind another call. const connector = createMockHostConnector(undefined, { callDelayMs: 60, - timeoutCall: { tool: 'generate_image', nth: 2 }, + timeoutCall: { tool: 'synthesize_speech', nth: 2 }, }); const h = await createHarness({ hostConnector: connector }); t.after(() => h.cleanup()); @@ -3002,14 +3002,14 @@ test('a call that waited for the machine says how long it waited when it fails', const first = await fetch(`${h.base}/api/runs`, { method: 'POST', headers: { 'content-type': 'application/json' }, - body: JSON.stringify({ nodeId: 'n_ok', mode: 'image', prompt: '第一张' }), + body: JSON.stringify({ nodeId: 'n_ok', mode: 'audio', prompt: '第一段' }), }); const firstRun = (await first.json()).run; - await waitForCall(connector.calls, (call) => call.toolRef.endsWith('generate-image')); + await waitForCall(connector.calls, (call) => call.toolRef.endsWith('synthesize-speech')); const second = await fetch(`${h.base}/api/runs`, { method: 'POST', headers: { 'content-type': 'application/json' }, - body: JSON.stringify({ nodeId: 'n_cut', mode: 'image', prompt: '第二张' }), + body: JSON.stringify({ nodeId: 'n_cut', mode: 'audio', prompt: '第二段' }), }); const secondRun = (await second.json()).run; @@ -3017,7 +3017,7 @@ test('a call that waited for the machine says how long it waited when it fails', const failed = await waitFor(h.base, secondRun.id, (item) => item.status === 'done' || item.status === 'error'); assert.equal(doneFirst.status, 'done'); assert.equal(failed.status, 'error'); - assert.match(failed.error, /generate_image/, 'the error does not say which call failed'); + assert.match(failed.error, /synthesize_speech/, 'the error does not say which call failed'); assert.match(failed.error, /前面排队/, 'the error does not say how long the call waited'); assert.match(failed.error, /耗时/, 'the error does not say how long the call itself ran'); // A call that the Host stopped leaves nothing behind: the outcome is @@ -3028,13 +3028,78 @@ test('a call that waited for the machine says how long it waited when it fails', const third = await fetch(`${h.base}/api/runs`, { method: 'POST', headers: { 'content-type': 'application/json' }, - body: JSON.stringify({ nodeId: 'n_after', mode: 'image', prompt: '第三张' }), + body: JSON.stringify({ nodeId: 'n_after', mode: 'audio', prompt: '第三段' }), }); const thirdRun = (await third.json()).run; const doneThird = await waitFor(h.base, thirdRun.id, (item) => item.status === 'done' || item.status === 'error'); assert.equal(doneThird.status, 'done', 'a failed call blocked the one behind it'); }); +test('a stopped image call goes on by the Agent road instead of failing', async (t) => { + // The Host stops a call that runs over its window, and a picture has no + // task number to ask about, so the one the provider makes cannot be found + // afterwards. The page is therefore not told the run failed: the run goes + // on by the other road, because the Host does not time a turn out. + const connector = createMockHostConnector(undefined, { + timeoutCall: { tool: 'generate_image', nth: 1 }, + }); + const h = await createHarness({ hostConnector: connector }); + t.after(() => h.cleanup()); + + const started = await fetch(`${h.base}/api/runs`, { + method: 'POST', + headers: { 'content-type': 'application/json' }, + body: JSON.stringify({ + nodeId: 'n_degrade', + mode: 'image', + prompt: '把参考图里的字搬到牌匾上', + params: { ratio: '9:16', resolution: '1K' }, + references: [ + { + id: 'n_ref', + title: '门头照', + kind: 'image', + text: '', + media: [{ url: 'https://cdn.example.com/ref.png', kind: 'image' }], + }, + ], + }), + }); + assert.equal(started.status, 202); + const runId = (await started.json()).run.id; + + // The Agent turn asks for its tool call to be approved before it makes it. + const waiting = await waitFor(h.base, runId, (run) => Boolean(run.pendingPermission)); + assert.ok(waiting.pendingPermission, 'the run never reached the Agent'); + assert.equal(waiting.degraded, true, 'the run did not say it changed road'); + assert.equal(waiting.status, 'running', 'the page was told the run failed'); + assert.ok(!waiting.error, 'the page was shown a failure'); + assert.ok( + waiting.log.some((line) => line.includes('改走 Agent 通道')), + 'the run did not say it changed road', + ); + + // The Agent is handed the very request the direct path would have sent. + const runCall = h.agent.calls.find((call) => call.kind === 'run'); + assert.ok(runCall, 'the Agent was not asked to run'); + assert.match(runCall.payload.text, /把参考图里的字搬到牌匾上/, 'the request lost the prompt'); + assert.match(runCall.payload.text, /"aspect_ratio":"9:16"/, 'the request lost the ratio'); + assert.match(runCall.payload.text, /"resolution":"1K"/, 'the request lost the resolution'); + assert.match(runCall.payload.text, /cdn\.example\.com\/ref\.png/, 'the request lost the reference'); + + const approved = await getJson(h.base, `/api/runs/${runId}/permission`, { + method: 'POST', + headers: { 'content-type': 'application/json' }, + body: JSON.stringify({ decision: 'allowOnce' }), + }); + assert.equal(approved.status, 200); + + const done = await waitFor(h.base, runId, (run) => run.status === 'done' || run.status === 'error'); + assert.equal(done.status, 'done', `the degraded run died: ${done.error}`); + assert.ok(!done.error, 'the degraded run ended with an error'); + assert.equal(done.media.length, 1, 'the degraded run produced no picture'); +}); + test('a finished video that cannot be fetched says so instead of finishing empty', async (t) => { // The task is done, but the answer carries no link and the video has no // drive asset. Saying "完成" with nothing behind it is a lie the user would