⏰ Copilot Scheduler

Schedule recurring or one-time AI prompts in VS Code.

📥 Install from VS Code Marketplace

Japanese / 日本語版はこちら

🎬 Demo

✨ Features

🗓️ Cron and One-Time Scheduling - Run recurring prompts with cron or dispatch a task once at a specified date and time

🤖 Agent & Model Selection - Choose from built-in agents (@workspace, @terminal) and AI models (GPT-4o, Claude Sonnet 4), including runtime quality or experimental quality variants when available

🌐 Multi-language Support - English and Japanese UI with auto-detection

📊 Sidebar TreeView - Manage all scheduled tasks with human-readable schedule summaries

🖥️ Webview GUI - Easy-to-use graphical interface for creating and editing tasks

🛠️ Copilot Chat Tools - Query, create, update, delete, enable/disable, and run scheduled tasks once from agent mode using Language Model Tools

📎 Attachments - Attach instructions, prompts, skills, or any workspace file to a task so they are sent with the prompt

In the task panel, Tab focuses the selected tab. Left/Right arrows switch tabs, Home/End select the first/last tab, and Tab moves into the selected panel. Enter/Space retain normal button activation.

⏰ Cron Expression Examples

The friendly cron builder applies your selected frequency, interval, time, weekday, or day-of-month to the cron expression as soon as you change the helper controls. The Generate button remains available as an explicit re-apply action, but you do not need to press it before saving.

For a single execution, select Run once at in the task form and choose a date and time. The task is disabled after dispatch by default; choose Delete task (keep history) to remove it after its history is saved. Missed one-time schedules follow the configured catch-up/skip policy. A completed disabled task must be assigned a new runAt before it can be enabled again.

The friendly cron builder only offers interval choices that can be represented exactly with standard cron. Intervals such as 40 or 90 minutes are generated as multiple cron lines instead of inaccurate expressions like */40 * * * *. All generated lines belong to the same task, and the scheduler runs the task at the earliest matching next time across those lines.

Monthly friendly schedules default to days 1-28 so the task can run every month. Use a custom cron expression if you intentionally want a schedule such as the 31st that only runs in months where that day exists.

📋 Commands

Execution history shows scheduled time, delay, attachment count, prompt source, path, hash, resolution time, and fallback reason when available. A successful entry is labeled Dispatched because it confirms that the prompt was sent to Chat, not that the model finished producing a response.

🛠️ Copilot Chat Tools

In Copilot Chat agent mode, use the scheduler tools with # references:

For a one-time manual run, the tool reports enabledStateChanged and taskDeleted after dispatch. A completed one-time task returns oneTimeCompleted with instructions to set a new run time; no second dispatch is attempted.

For kind=history, the response includes total, returned count, hasMore, statusSemantics, and newest-first entries. status: "success" confirms prompt dispatch rather than model response completion. Legacy malformed timestamps are preserved or omitted without inventing audit times and are marked with executedAtInvalid / nextRunAtInvalid when applicable.

Create a one-time task by sending runAt as an ISO 8601 date-time with an explicit offset (for example, 2030-09-26T21:00:00+09:00) instead of cronExpression, and optionally afterRun: "disable" | "delete". Update accepts the same fields; to switch back to a recurring schedule, set runAt: "" and provide cronExpression. kind=list / kind=get return these settings, and kind=history retains the task name and runAt even after automatic deletion. The Webview time picker uses the machine's local timezone; tools can specify any explicit offset.

kind=list returns task metadata with a short promptPreview and promptLength instead of the prompt body, because a local or global task stores a snapshot of the whole prompt file. The write tools return the same shape in their success payloads. Use kind=get when the full prompt is needed; a preview must never be written back to a task.

kind=list_models returns the selectable model ids together with supportedReasoningEfforts, and kind=list_agents returns the selectable agent ids without exposing file paths. The model list is the same one the Copilot Scheduler view offers, so Chat can never pin a model you cannot see or change in the UI. Create/update accept model (plus optional modelReasoningEffort), agent, and the execution controls autoMode, jitterSeconds, maxExecutionsPerDay, allowedTimeStart, and allowedTimeEnd. A model id that is not in that list is rejected with the list of valid ids instead of silently falling back to the default model; passing an empty model clears the selection and returns the task to the default model. When the Language Model API is unavailable and only the built-in fallback catalog is known, the requested model is saved with a warning instead of being rejected.

Auto and internal utility model ids remain selectable but do not inherit reasoning-effort options from the model named in their metadata.

Create/update also accept attachments: up to 10 entries of { source: "local" | "global", path }, where the path is relative to the task's workspace folder or to the global prompts folder. Absolute paths, .., NUL characters, denied files, and local attachments on a global task are rejected instead of being saved.

In agent mode, Copilot can also choose these tools from natural-language requests. Examples:

• "Schedule a workspace task every weekday at 9:00 to summarize this repository."
• "Change the daily summary task to run at 10:30."
• "Use Claude Sonnet for the daily summary task."
• "Pause the release reminder task until I turn it back on."
• "Run the disabled release reminder task once now without enabling it."
• "Show my scheduled Copilot tasks before changing anything."

If multiple tasks could match the same name across scopes, ask Copilot to show the scheduled tasks first so it can confirm the exact task before updating, disabling, or deleting it.

scheduler_run_task accepts only { "id": "..." } and reuses Run Now, including prompt/attachment resolution and history. Disabled tasks stay disabled; enabled tasks advance nextRun according to manualRunNextRunPolicy. Like Run Now, it bypasses jitter, allowed time windows and daily limits. A workspace task must belong to the current workspace.

ok: true with executionSemantics: "prompt_dispatched" confirms dispatch, not model response completion. saveFailed also reports prompt_dispatched and retrySafe: false: only bookkeeping failed, so do not retry automatically. Unexpected failures report executionSemantics: "unknown" and retrySafe: false; check Chat and history first. Cancellation is checked before execution starts, not during dispatch, and cannot undo a sent prompt. Every invocation is a new manual request, not an idempotent operation; duplicate requests are rejected while dispatch and history recording are in progress in the current window.

Write tools are enabled by default and require a trusted workspace. Set copilotScheduler.lmTools.enableWriteTools to false to block create/update/delete/enable-disable/run operations. They remain visible, but only read-only tools can be used.

copilotScheduler.lmTools.confirmationMode controls only the extension-provided custom confirmation messages for write tools. VS Code or Copilot Chat may still show a generic approval dialog for extension tools, and users can use the built-in Always Allow flow when available.

Task snapshots and revision metadata are written through same-directory temporary files and atomically replaced. Empty/corrupt files and meta-less revision-zero [] snapshots do not override valid legacy global-state tasks, while revision-backed empty arrays remain valid deletes. Foreground saves and mirrors share one queue per destination. An atomic directory lock with heartbeat/stale recovery (via MIT-licensed proper-lockfile) plus a revision recheck prevents stale VS Code windows from overwriting newer tasks; stale windows reload the winning snapshot and ask the user to retry. Payload/mirror writes complete before metadata advances or the lock is released.

⚙️ Settings

To automatically keep AI-applied edits after review delay, configure VS Code setting chat.editing.autoAcceptDelay (0 = off, 1-100 = seconds, recommended: 5).

Missed-run policy uses the exact scheduler start/restart time: only an earlier nextRun is missed. skip advances it without an execution history entry or a daily-count increment; it cannot restore skipped occurrences later. A sleep delay after startup is not a missed startup run. Settings apply on the next check; active dispatches are not cancelled. Waiting tasks are selected oldest-due-first on a later tick, subject to time windows and daily limits. A daily slot reserved by an unfinished dispatch remains unavailable until that dispatch succeeds or fails. The default concurrency is now 1 (previously unrestricted); model responses may still overlap and other VS Code windows have independent dispatch limits.

Task-level controls (Chat Session, Max Runs/Day, Allowed Time Window) are configured per task in the Webview create/edit form. Max Runs/Day and Allowed Time Window are evaluated on the same clock as the schedule: copilotScheduler.timezone when it is set, otherwise the machine's local time.

The Webview previews Copilot Chat-like thinking effort options for supported model families. If it fails, choose Default.

Claude Opus/Sonnet are adaptive-thinking models. The extension writes the selected effort to the same per-model setting Copilot itself uses, but Copilot Chat governs Claude's effective thinking through adaptive thinking and may still apply Medium. GPT-5 family models honor the selected effort directly.

The selected custom agent is passed through the mode field of VS Code's workbench.action.chat.open, and reasoning effort is applied by writing Copilot Chat's per-model settings — both are sent identically for every model. Whether a Claude model actually honors the custom agent and reasoning depth is decided by VS Code / Copilot Chat. If they do not seem to take effect, set copilotScheduler.logLevel to debug and open the "Copilot Scheduler" output channel: when the Agent set: and Experimental model quality sync: lines show the expected mode and effective values, the scheduler did its part and the gap is on the Copilot Chat side.

If execution feels sluggish when a task is triggered, try:

• copilotScheduler.chatSession = continue
• copilotScheduler.commandDelayFactor = 0.6 (or 0.5)
• copilotScheduler.notificationMode = silentStatus
• copilotScheduler.logLevel = error (or none)

📝 Prompt Placeholders

Use these placeholders in your prompts:

📂 Task Scope

• Global: Task runs in all workspaces
• Workspace: Task runs only in the specific workspace where it was created

📄 Prompt Templates

Store prompt templates for reuse:

• Local: .github/prompts/*.md in your workspace
• Global: VS Code user prompts folder (or the folder set in copilotScheduler.globalPromptsPath)
• The edit form shows the selected file and Open prompt file action while Local/Global is selected; the prompt body is hidden. Use that action to edit the source file. Switch the source to Inline to show and edit the text on the task itself. Changing the model or other task settings preserves the file reference.
• The panel preview reads saved disk content only. At execution time, an open editor buffer is preferred; otherwise the latest saved file is read.
• A task only becomes Inline when you select the Inline source. Selecting a template keeps the task following that file.
• Task creation/updates, template loading and prompt path resolution reject paths containing NUL characters, even when a path is cached or lexically inside an allowed root.

Global custom agents are auto-discovered from the VS Code user prompts/customization folder and ~/.copilot/agents when copilotScheduler.globalAgentsPath is empty.

This follows current Copilot custom agent and Copilot CLI file locations, but this extension only discovers agent files. Prompt templates still use the VS Code user prompts folder or copilotScheduler.globalPromptsPath, not ~/.copilot/prompts, and the extension does not manage Copilot CLI sessions.

Agent definitions are refreshed automatically when you create, edit, or delete workspace *.agent.md / AGENTS.md files (or the global agent files). You can also reload them on demand with the refresh button next to the agent picker.

Only user-invocable agents appear in the picker. Agents with user-invocable: false in their frontmatter are subagent-only — they cannot be selected as a chat mode, so they are intentionally hidden. Agents without the field stay listed.

📎 Attachments

A task can carry up to 10 attachment files that are sent with the prompt, so instructions, prompt files, skills, or any other workspace file can be attached explicitly instead of being mentioned inside the prompt text.

• Add attachment opens a quick pick with a Recommended group (AGENTS.md, .github/copilot-instructions.md, .github/prompts/, .github/instructions/, .github/skills/) followed by the rest of the workspace. Browse... opens the regular file dialog.
• Attachments are stored relative to the workspace folder (local) or to the global prompts folder (global), never as absolute paths.
• Only workspace-scoped tasks can attach workspace files. A global task runs in any window, where the same relative path could point at a different file, so workspace attachments are rejected for them.
• Files such as .env*, *.pem, *.key, id_rsa* and anything under secrets/ or .ssh/ are never attached.
• If an attachment cannot be found when the task runs, the task is skipped instead of running without it. The run is recorded as blocked in the execution history and reported once per attachment set, so a renamed file does not silently stop the schedule and changing the attachments makes the next failure visible again.
• Trust boundary: the checks above are a path boundary, not a content-provenance boundary. Anything that can write inside the workspace folder can also place or link a file there, and the file is read at run time, so it is the content at that moment that is sent. Attach files only from workspaces you trust, the same way you would trust a prompt file that runs unattended.

📋 Requirements

• VS Code 1.95.0 or higher
• GitHub Copilot extension

Development Tests

npm test compiles and runs the complete suite in an isolated VS Code profile. No Copilot sign-in is required. To select tests by their full suite/test title, pass a JavaScript regular expression with --grep (or -g). On Windows PowerShell, use npm.cmd to preserve option forwarding:

npm.cmd test -- --grep "Test Runner Arguments"

For shell-sensitive expressions (for example, alternation with |), compile first and invoke Node directly:

npm.cmd run pretest
node ./out/test/runTest.js --grep 'tabs|claim'

Patterns are preserved verbatim. Unknown arguments, empty/invalid patterns and zero matching tests fail the run. Without arguments the full suite runs even if a filter environment variable was inherited. A focused test pass is not a substitute for npm test before release.

🛠️ Release Automation

Maintainers can publish from GitHub Actions instead of running vsce publish locally.

• Push a tag in the form vX.Y.Z after updating package.json to the same version.
• GitHub Actions runs npm ci, npm run compile, npm test, packages a .vsix, publishes to VS Code Marketplace, and attaches the .vsix to the GitHub release.
• For a pre-release dry run, dispatch Publish Extension on the branch with publish=false (the default). Wait for all gates to pass before pushing the version tag. Runs are serialized across branches and tags. Explicit publish=true publishes to Marketplace only; a matching tag is required for the GitHub Release.
• Add the repository secret VSCE_PAT before using the workflow.

⚠️ Known Issues

• Copilot Chat API is still evolving; some features may require updates as the API stabilizes
• Model selection may not work in all configurations
• Experimental model quality relies on evolving VS Code/Copilot internals and may not work in all configurations

Disclaimer: This extension automates Copilot Chat. GitHub's Acceptable Use Policies prohibit "excessive automated bulk activity", the Terms of Service § H (API Terms) allow account suspension for excessive API usage, and the GitHub Copilot Additional Product Terms apply these policies directly to Copilot. Use at your own risk; your account could be rate-limited or restricted. Configure jitter/daily limits/longer intervals to reduce risk, but there is no guarantee.

Note: There are reports of Copilot access being restricted even without using automation tools. These mitigations reduce obvious automation patterns but cannot eliminate that risk.

🐛 Report a bug

📦 Release Notes

0.1.0

Initial release:

• Cron-based task scheduling
• Agent and model selection
• English/Japanese localization
• Sidebar TreeView
• Webview GUI for task management
• Prompt template support

📄 License

CC-BY-NC-SA-4.0 © aktsmm

Enjoy scheduling your Copilot prompts! 🚀