fix: retry dropped permission replies

[marius-kilocode]

What Problem This Solves

Auto-approve and manual permission replies travel over a pooled HTTP connection to the CLI backend. When that socket is closed before the response completes, the reply fails with a transient transport error and the permission request stays pending. With auto-approve on, the agent waits indefinitely. With auto-approve off, the prompt cannot be resolved.

Why This Change Was Made

The reply path had no retry for transport failures, so a single dropped connection stranded the request. The fix retries only transient transport failures (including undici's TypeError: terminated) and never retries a decisive server error, so a missing request or a client/server mismatch still fails fast. It covers manual replies, auto-approve replies, and the pending-permission recovery list.

User Impact

• Auto-approve no longer stalls when a reply hits a dropped connection.
• Manual "Allow once", "Deny", and "Always" approvals recover from the same drop.
• No behavior change for genuine server errors.

Evidence

Reproduced in isolated VS Code with the self-test harness and a TCP proxy that drops the first permission reply:

• Before: the reply was dropped and the chat stayed blocked at "Permission required".
• After: the same permission ID was retried 503 ms after the drop, forwarded successfully, and the command completed with exit status 0 without manual approval.

Focused checks: SDK permission tests 13 passed, extension permission recovery tests 33 passed, extension typecheck and lint passed.

This fixes the dropped-reply failure only. It does not address separate reports of "Unknown permission route" or an unclickable "Allow once".

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/sdk/js/src/kilocode/permission.ts
• packages/kilo-vscode/tests/unit/permission-recovery.test.ts

Previous Review Summary (commit 675ed4b)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 675ed4b)

Status: 3 Issues Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	3

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
packages/sdk/js/src/kilocode/permission.ts	14	Comment names packages/core/src/util/retry.ts as the sync source, but core lacks the exact "terminated" match and "fetch failed", so a future resync would regress this fix.
packages/sdk/js/src/kilocode/permission.ts	54	throw last is unreachable dead code; the loop always throws on its final iteration.
packages/kilo-vscode/tests/unit/permission-recovery.test.ts	376	The cancellation test would pass even if the retry loop did not stop, because reply is never re-invoked after shouldContinue returns false.

Files Reviewed (6 files)

• .changeset/retry-permission-transport-drops.md - 0 issues
• packages/kilo-vscode/src/commands/toggle-auto-approve.ts - 0 issues
• packages/kilo-vscode/src/kilo-provider/handlers/permission-handler.ts - 0 issues
• packages/kilo-vscode/tests/unit/permission-recovery.test.ts - 1 issue
• packages/sdk/js/src/kilocode/permission.ts - 2 issues
• packages/sdk/js/test/kilocode/permission.test.ts - 0 issues

No new memory leaks or unbounded resources were introduced: retry's backoff timer is awaited and reclaimed, replyOnce always settles (errors are caught, cancellation is intentionally suppressed), and void fetchAndSendPendingPermissions(ctx) cannot reject because its body is wrapped in try/catch. The shared deadline in respondToPermission is still respected because each retry recomputes AbortSignal.timeout(budget()) against the same absolute deadline.

Fix these issues in Kilo Cloud

---

_{Reviewed by deepseek-v4.1-flash · Input: 0 · Output: 0 · Cached: 0}

_{Review guidance: REVIEW.md from base branch main}