fix: retry premature close and connection errors in core

[RomneyDa]

Summary

• Add "premature close", "premature end", "connection reset", and "socket hang up" as retryable error patterns in core/util/withExponentialBackoff.ts and core/llm/utils/retry.ts
• These patterns were already handled in the CLI (extensions/cli/src/util/exponentialBackoff.ts) but missing from core, which is used by the VS Code extension
• Add 4 new test cases covering the new connection error patterns

Context

Five open issues report "premature close" errors across different providers (Ollama, OpenAI-compatible, xAI). All are from VS Code users. The CLI already retries these transient errors, but the core retry logic did not — so users saw unrecoverable "Unknown error" messages instead of automatic recovery.

Fixes #11108, #11102, #10948, #9667, #9999

Test plan

• All 29 retry tests pass (including 4 new ones for premature close, premature end, socket hang up, connection reset)
• Manual: trigger a premature close with a local Ollama instance and verify retry behavior

---

Summary by cubic

Core now retries transient connection/stream errors (“premature close”, “premature end”, “connection reset”, “socket hang up”) to match the CLI. This prevents unrecoverable “Unknown error” in the VS Code extension.

• Bug Fixes
  • Added retry patterns to core/util/withExponentialBackoff.ts and core/llm/utils/retry.ts.
  • Added 4 tests covering the new connection error cases. Fixes Error: DSllmO-Coder - Unknown error #11108, Error: DeepSeek Coder - Unknown error #11102, Error: DeepSeek Local - Unknown error #10948, Error: DeepSeek R1 - Unknown error #9667, Error: Grok 4 Fast Reasoning - Unknown error #9999.

^{Written for commit e9fccbc. Summary will update on new commits.}

[continue]

Docs Review

No documentation update needed for this PR.

This change adds retry logic for transient connection errors (premature close, socket hang up, connection reset, etc.) in core—aligning it with the existing CLI behavior. This is an internal reliability improvement that:

• Is completely transparent to users (no configuration required)
• Doesn't change any APIs, commands, or user-facing workflows
• Simply makes the extension more resilient to network hiccups

Users benefit automatically without needing to know or do anything differently. 👍

[continue]

Test Coverage Gap

The changes to core/llm/utils/retry.ts have excellent test coverage with 4 new test cases covering each connection error pattern.

However, the same patterns were also added to core/util/withExponentialBackoff.ts without corresponding tests. While a test file exists (withExponentialBackoff.test.ts), the entire test suite is currently skipped (describe.skip).

Suggestion: Consider adding test coverage for the new connection error patterns in withExponentialBackoff.ts:

• premature close
• premature end
• connection reset
• socket hang up

These tests would verify that the exponential backoff utility correctly retries on these connection errors, similar to the tests added for retry.ts.

[cubic-dev-ai]

1 issue found across 3 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="core/util/withExponentialBackoff.ts">

<violation number="1" location="core/util/withExponentialBackoff.ts:26">
P3: The new retryable connection errors are logged as "Hit rate limit," which is inaccurate and can mislead troubleshooting. Use a generic retry message (or include the actual error type) for this shared retry branch.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P3: The new retryable connection errors are logged as "Hit rate limit," which is inaccurate and can mislead troubleshooting. Use a generic retry message (or include the actual error type) for this shared retry branch.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At core/util/withExponentialBackoff.ts, line 26:

<comment>The new retryable connection errors are logged as "Hit rate limit," which is inaccurate and can mislead troubleshooting. Use a generic retry message (or include the actual error type) for this shared retry branch.</comment>

<file context>
@@ -19,7 +19,11 @@ const withExponentialBackoff = async <T>(
+        lowerMessage.includes("premature close") ||
+        lowerMessage.includes("premature end") ||
+        lowerMessage.includes("connection reset") ||
+        lowerMessage.includes("socket hang up")
       ) {
         const retryAfter = (error as APIError).response?.headers.get(
</file context>

[RomneyDa]

Actually nvm premature can happen mid-stream.