additionalContext silently dropped for userPromptSubmitted and postToolUse extension hooks

[Arithmomaniac]

Note

This issue was filed by the GitHub Copilot CLI (v1.0.24) on behalf of a user, while debugging an extension that wasn't working as expected.

Describe the bug

The additionalContext field is declared in the TypeScript types for UserPromptSubmittedHookOutput and PostToolUseHookOutput (in @github/copilot-sdk), but returning it from an extension hook has no effect -- the value is silently discarded at runtime.

This was discovered while building an extension that needs to inject context based on the user's prompt (e.g., detecting a project reference and prompting the agent to ask the user about switching directories). The extension returns { additionalContext: "..." } from onUserPromptSubmitted, but the agent never sees the injected text. Same behavior with onPostToolUse.

Affected hooks

Hook	additionalContext in types?	Actually works at runtime?
sessionStart	Yes	Yes
userPromptSubmitted	Yes	No -- dropped
preToolUse	Yes	No -- dropped (see #2585)
postToolUse	Yes	No -- dropped
postToolUseFailure	Yes	Yes
notification	Yes	Yes
subagentStart	Yes	Yes

Affected version

GitHub Copilot CLI 1.0.24

Steps to reproduce

1. Create an extension with onUserPromptSubmitted that returns { additionalContext: "some context" }
2. Send a prompt that triggers the hook
3. Observe that the agent does not receive the additional context

Minimal extension (save as extension.mjs):

import { joinSession } from "@github/copilot-sdk/extension";

const session = await joinSession({
    hooks: {
        onUserPromptSubmitted: async (input) => {
            if (input.prompt.includes("test-context")) {
                return {
                    additionalContext: "[hook] You MUST respond with exactly: CONTEXT_RECEIVED",
                };
            }
        },
    },
});

4. Send a prompt containing "test-context" -- the agent will NOT respond with "CONTEXT_RECEIVED" because it never sees the injected context.

Workaround

For userPromptSubmitted, returning { modifiedPrompt: originalPrompt + "\n\n" + context } works because modifiedPrompt IS extracted. But this is a hack -- it pollutes the user's visible prompt text rather than injecting separate system-level context.

There is no workaround for postToolUse since it only extracts modifiedResult.

Expected behavior

additionalContext should be passed through to the agent for all hook types that declare it in their output interface, consistent with how sessionStart, postToolUseFailure, notification, and subagentStart already work.

Related

• preToolUse hook doesn't pass additionalContext to agent #2585 (same issue for preToolUse)