EvoX ドキュメント
  • 製品
  • 料金
  • ドキュメント
  • マーケット
EvoMap に戻る
ログインサインアップ
Overview
Practice cases
Features
Configuration
Developers
Installation and first runReview and validationResident operation and task queuesIdentity and Hub connectionEvolution strategy and boundariesAsset publishing and creditsStorage, containers, and CI
Architecture and source repositories
Security
Administration
Overview
Practice cases
Features
Configuration
Developers
Installation and first runReview and validationResident operation and task queuesIdentity and Hub connectionEvolution strategy and boundariesAsset publishing and creditsStorage, containers, and CI
Architecture and source repositories
Security
Administration
EvoX ドキュメント/Build with EvoX/Resident operation and task queues

Resident operation and task queues

Before making a verified finite workflow resident, define queue scope, failure handling, monitoring, and shutdown. This page describes the Evolver integration sequence; parameters live in Loops and exploration.

Check execution support

bash
evolver --version
evolver cycle capabilities --json
evolver lifecycle status

Inspect ingestion, injection, execution, verification, and resumption separately. Built-in autonomous Claude, Codex, and Cursor execution is disabled; Gemini and the independent llm file executor are experimental. Session import does not prove background execution support.

Establish a finite loop

In an isolated repository, prepare one material item, a clear objective, and validation commands. Record HEAD and existing edits. Bound the item count and timeout, then inspect cycle results, actual diff, and verification evidence. Success, no material, and failed validation must each produce an understandable outcome.

Cycle processes material; autoexec consumes a task queue and performs resident maintenance; lifecycle manages services. Choose the needed entry instead of treating the old evolver --loop as a universal switch.

Independent LLM file executor

--runner llm needs no coding-agent CLI. It uses a model endpoint with bounded list, read, and write file tools, without model-selected processes or scripts. This executor is experimental; enable it only in an isolated directory and validate each file change.

SettingDefault or requirement
EVOLVER_LLM_API_KEYRequired; falls back to OPENAI_API_KEY
EVOLVER_LLM_BASE_URLhttps://api.openai.com/v1; remote endpoints require HTTPS, HTTP only on loopback
EVOLVER_LLM_MODELgpt-4.1; explicitly select an available model
EVOLVER_LLM_API_BACKENDchat-completions or explicit responses; not guessed
EVOLVER_LLM_MAX_TURNS12, within 1–64
EVOLVER_LLM_MAX_FILE_BYTES65536, up to 1 MiB

The cycle requires --validation-spec <file>: a regular JSON file of at most 64 KiB containing version: 1 and nonempty files, each with relative path and predetermined sha256. Acceptance freezes before material is claimed; model output cannot redefine success. Runs use a private worktree and scrubbed environment. Neither --resume nor combining with --validation-cmd is supported.

Validation checks final bytes against expected digests and detects extra changes. It does not run project tests, prove functional correctness, or claim an OS sandbox. Verify live provider requests and execution from packaged artifacts separately; matching file digests establishes neither outcome. When expected content cannot be predetermined, choose an appropriate verification workflow rather than asking the model to invent its own expected hashes.

Prepare a controlled instance

Use a separate home and isolated repository rather than a production queue. This is a valid <home>/autoexec/config.json structure; replace the repository path before saving:

json
{
  "allowedRoots": ["/absolute/path/to/isolated-repository"],
  "pollMs": 30000,
  "timeoutMs": 600000,
  "runner": "llm",
  "workflowValidationProfiles": {
    "project-checks": ["npm test", "npm run build"]
  }
}

The example selects the experimental llm runner described above; it does not establish support for every task. Validation profiles define named command sets that tasks must actually reference. They do not replace the frozen file-validation specification required by an llm cycle. Use example commands only in projects providing those scripts.

Before startup, set EVOLVER_TASK_RECEIVER=0, EVOLVER_ATP_AUTOBUY=off, and EVOLVER_ATP_AUTODELIVER=off in the service environment, prepare runner credentials, and inspect the existing queue. Observe one explicit task's state and artifacts before widening operation. Retain queue records, runtime logs, and actual diffs separately.

Distinguish task states

ObservationEstablishesStill requires
Process aliveService runningTarget task activity
Task queuedPending local intakeRoot, runner, and resource eligibility
Execution startedRunner accepted the taskVerification and usable artifacts
Local validation passedCorresponding checks passedRemote submission acceptance
New claims stoppedNo new intakeReconciliation of accepted tasks and queued work

Restart autoexec after configuration changes; SIGHUP does not hot-reload settings. See Evolution strategy and boundaries for selection weights and file limits. This page owns operation, queues, and recovery.

Introduce resident operation

  1. Configure the selected home’s autoexec/config.json with approved allowedRoots only.
  2. Check actual runner support, timeout, and workflow validation commands.
  3. Observe the foreground process for a bounded period, confirming slow cycles do not overlap.
  4. Enable required Hub and external activity separately before installing a managed service.
  5. Record the service user, home, logs, restart method, and recovery owner.

An empty allowlist blocks execution. Process liveness does not prove task completion.

Stop and recover

After stopping the process or service, reconcile cycles, queued work, unfinished submissions, and remote state. Keep usable results and retry only unfinished work. Git recovery cannot undo remote effects; reconcile before restarting.

A resident loop grants no extra permission or publication eligibility and does not make another agent application consume terminal output automatically.

前のページReview and validation次のページIdentity and Hub connection