EvoX 文件
  • 產品
  • 價格
  • 文件
  • 能力市集
返回 EvoMap
登入註冊
概览
实践案例
功能
使用蜂群持续蜂群递归任务恢复与追加设备协作
配置
开发者
安全
管理
概览
实践案例
功能
使用蜂群持续蜂群递归任务恢复与追加设备协作
配置
开发者
安全
管理
EvoX 文件/Product/递归任务

Recursive tasks

Some goals cannot be fully enumerated upfront. A project onboarding guide needs verified features and entry points before its chapters can be planned, written, checked, and combined. Recursive tasks support work that must be refined in stages.

If you already know that two files need independent reviews, use ordinary swarm. Recursive mode is not simply a way to increase concurrency.

Define the deliverable first

Give one complete goal, its inputs, boundaries, and acceptance criteria:

Create an onboarding guide from the current project. Verify installation, actual entry points, and required configuration before organizing chapters, writing instructions, and checking links. Edit only docs/guide. Include the expected result of every step. List unsupported claims as unresolved and do not publish the site. Deliver the guide, verification results, and outstanding questions.

This defines what to deliver, where changes are allowed, and how to judge completion without prescribing a fixed number of steps or agents.

Submit a goal and identify recursive execution

The current desktop has no separate Recursive tasks switch. Ordinary swarm controls the conversation’s swarm posture; continuous swarm submits additional background instructions. Neither is a recursive-mode selector.

Enable swarm in the conversation, then submit one complete goal with facts to verify, allowed changes, and acceptance criteria. The connected runtime creates recursive execution. When the desktop receives its run facts, the shared swarm card displays hierarchy, dependencies, and run state. Identify execution from the actual records, not from asking for recursion, seeing an enabled badge, or counting nodes.

New continuous-swarm instructions omit the recursive option. Do not enable continuous swarm and look for the removed recursive selector. Use Continuous swarm to add independent work while the main AI is busy. Observation and recovery remain available for existing recursive runs.

If no recursive run is recorded, check the session connection, runtime capability, and execution actually used. Ordinary parallel work can still complete clearly bounded tasks; do not resubmit the same goal solely to obtain a task tree.

Understand the hierarchy

For example, organize an onboarding guide as follows. Use the example to define responsibilities rather than as a fixed role template:

StageWorkUsed for
Verify materialInspect installation scripts, configuration, and entry pointsEstablish facts for the guide
Organize chaptersArrange a reading order around verified featuresDefine chapter boundaries
Write contentDraft independent chapters in parallelProduce text for review
Check and integrateReview steps, links, terminology, and omissionsDeliver one consistent guide

Chapters depend on verified facts but independent chapters do not require each other’s outputs. Recursive execution refines tasks and dependencies as needed; it does not allow every node to create unlimited branches.

Follow execution

Identify completed, running, and dependency-blocked work. A failed prerequisite blocks downstream tasks that depend on it; it does not mean each downstream task failed during execution.

Check whether completed output supports the next stage. A finished outline does not make the guide ready for delivery while installation facts remain unresolved.

Tree size, management turns, tool iterations, and elapsed time are bounded. The current v1 contract defaults include depth 4 with the root at 0, 64 nodes per tree, 8 children per expansion, and a root deadline of at most 20 minutes. Per-run configuration can only lower these limits. An ordinary turn’s own deadline can shorten this further. Queue waiting counts toward that deadline; resuming does not grant another 20 minutes.

These are execution limits, not an exact credit budget in the client. Reserved turns and tool iterations differ from observed model usage. Unknown token usage must not be read as zero. When a limit is reached, inspect completed coverage and missing work before defining another bounded goal. A partial report is not the complete deliverable.

Continue after interruption

Distinguish waiting for prerequisites, requiring recovery, and being stopped. A dependency-blocked node is not an execution failure, and terminal states are not all successful. stopping means shutdown is still in progress; treat it as stopped only after the record reaches stopped.

For failed, blocked, or recovery-required runs, follow Recovery and follow-up to refresh, select eligible nodes, and confirm. Recovery retains completed output rather than rerunning the entire tree. It does not undo file changes or external actions. Expand retained results in the run card to inspect completed nodes’ summaries and evidence records, including run, node, and revision identifiers. These records display snapshot content and verification status; they do not reread files or validate digests for you. Check the original material when a record is marked truncated. A stopped run or one past its original deadline cannot resume; carry valid output and the remaining scope into a new task.

Accept the whole result

Return to the original request: can someone follow the guide, does every step state its expected result, do links work, and are uncertainties explicit?

Completing individual branches only establishes that those branches produced output. The final guide still needs duplicate sections, inconsistent terminology, and contradictions resolved. Publishing or deploying requires the corresponding authorization; recursive mode does not grant it.

上一篇持续蜂群下一篇恢复与追加