Bug: OMO internal prompts can be labeled user-initiated in Copilot, causing extra premium request usage

[maximharizanov]

🐞 Bug Description

Some internally generated OMO prompts/messages are being attributed as user-initiated in GitHub Copilot contexts, instead of agent-initiated.

This appears to affect flows where OMO injects prompts into the main session (continuation/recovery/background notifications/babysitter patterns). As a result, Copilot users may see unexpected premium request consumption.

✅ Expected Behavior

OMO-internal/system-generated prompts should be attributed as agent-initiated (or otherwise avoid being counted as user-origin prompts), while real user prompts remain user-initiated.

❌ Actual Behavior

At least some OMO-generated prompts are treated as user-initiated by downstream Copilot attribution heuristics, increasing premium usage for Copilot users.

🔁 Reproduction (high-level)

1. Use OpenCode + GitHub Copilot provider with OMO enabled.
2. Trigger OMO internal prompt flows (e.g. background-task notifications, continuation injections, recovery flows).
3. Observe that some of these prompts are classified as user-initiated (and can contribute to premium request counting).

📌 Notes / Context

• This is similar in nature to the 2026 opencode-dynamic-context-pruning Copilot misattribution incident/fixes.
• OMO currently emits many internal prompts via session.promptAsync(...) in main sessions, which can hit user-role boundary heuristics.
• We should keep OMO messages model-visible where needed (not blanket-ignore them), while preventing misattribution.

💡 Suggested direction (minimal)

• Add an OMO chat.headers safeguard (Copilot-only) to force x-initiator=agent for clearly OMO-internal emissions.
• Mark OMO-internal injected prompts with a stable internal marker + synthetic metadata so override conditions are explicit and narrow.
• Preserve normal user prompt attribution and existing noReply behavior semantics for background notifications.

Environment

• OMO: current main
• OpenCode: current versions where Copilot plugin applies role/content heuristics for initiator
• Provider: GitHub Copilot

Related

• #8721 fix: Prevent excessive Copilot premium request consumption

[sisyphus-dev-ai]

Thanks for this detailed bug report. This is a valid concern about Copilot attribution for OMO-internal prompts.

The issue of system/agent prompts being misattributed as user-initiated is something that needs to be addressed to prevent unexpected premium usage for Copilot users. Your suggested direction around using chat.headers with x-initiator=agent marking for OMO-internal emissions makes sense.

This will need maintainer review to implement the proper attribution handling without breaking existing functionality. Marking this for attention.

[code-yeongyu]

[sisyphus-bot] 🔍 Investigation Complete: Copilot Labeling Bug

ACTION

Searched codebase for Copilot integration, internal prompt labeling, and message flow.

FINDINGS

Root Cause: OMO internal prompts (background tasks, sync delegation, parent notifications) are sent via parts: [{ type: "text", text: ... }] without any internal/system marker. Copilot treats these as user-initiated messages, incurring premium request charges.

Locations of Relevant Code:

1. Background Task Launch (src/features/background-agent/spawner.ts:136-150)

promptWithModelSuggestionRetry(client, {
  path: { id: sessionID },
  body: {
    agent: input.agent,
    system: input.skillContent,
    parts: [{ type: "text", text: input.prompt }],  // ← Sent as plain text
  },
})

2. Sync Task Delegation (src/tools/delegate-task/sync-prompt-sender.ts:53-65)

const promptArgs = {
  body: {
    agent: input.agentToUse,
    system: input.systemContent,
    parts: [{ type: "text", text: input.args.prompt }],  // ← Same issue
  },
}

3. Parent Session Notifications (src/features/background-agent/parent-session-notifier.ts:68-77)

await client.session.promptAsync({
  path: { id: task.parentSessionID },
  body: {
    parts: [{ type: "text", text: notification }],  // ← Background completion notification
  },
})

4. Resume Task (src/features/background-agent/spawner.ts:221-234)

Same pattern for task continuation.

Key Insight

All internal prompts use the same parts: [{ type: "text", text: ... }] structure without metadata to distinguish them as internal/system messages. Copilot's billing differentiates between user-initiated vs system-initiated based on message origin indicators.

RECOMMENDATION

Consider adding a marker/metadata to distinguish OMO internal messages from actual user messages when sending prompts via OpenCode SDK. This could involve:

• Adding an internal: true flag to the prompt body
• Using a specific header or metadata field
• Leveraging OpenCode's system message injection capabilities

This is a billing/cost issue for Copilot users - internal OMO orchestration is being charged as user requests.

[code-yeongyu]

[sisyphus-bot]

Great news! This issue has been fixed in PR #1953 which was merged to the dev branch. The fix adds an internal marker to all OMO-initiated messages and sets the x-initiator: agent header for Copilot providers, ensuring internal prompts are correctly attributed as agent-initiated. This will be included in the next release. Thank you for the detailed report, @maximharizanov!