Skip to content

ci: update renovate and fix version branch - #3

Merged
marcusrbrown merged 1 commit into
mainfrom
amaranth-written-wolverine
Jan 2, 2026
Merged

marcusrbrown merged 1 commit into
mainfrom
amaranth-written-wolverine

Conversation

@marcusrbrown

Copy link
Copy Markdown
Collaborator

No description provided.

@marcusrbrown marcusrbrown self-assigned this Jan 2, 2026
@marcusrbrown
marcusrbrown merged commit b17b8ab into main Jan 2, 2026
9 checks passed
@marcusrbrown
marcusrbrown deleted the amaranth-written-wolverine branch January 2, 2026 03:03
marcusrbrown added a commit that referenced this pull request Jan 21, 2026
- Only track tokens/cost/artifacts from successful attempts (Issue #1, #2)
- Check remaining timeout before retry attempts (Issue #3)
- Increase event stream grace period from 1s to 2s (Issue #4)
- Add comprehensive tests for retry logic:
  - Retries on LLM fetch error and succeeds on second attempt
  - Stops retrying after MAX_LLM_RETRIES attempts
  - Does not retry on non-LLM errors
  - Only tracks results from successful attempt
  - Sends continuation prompt on retry instead of initial prompt
marcusrbrown added a commit that referenced this pull request Jan 21, 2026
…ors (#111)

* fix(agent): wrap entire SDK execution in retry loop for LLM fetch errors

- Extract attemptExecution() helper to encapsulate full SDK lifecycle
- Retry up to 3 times with 5s delay on recoverable LLM fetch errors
- Detect errors from both event stream and caught exceptions
- Post error comment to PR/issue on final failure
- Exit 0 for recoverable LLM errors (don't fail CI)
- Fix PR number extraction using triggerResult.context.target

* chore: update build artifacts

* fix(agent): reuse session across retries and send continuation prompt

Session context is now preserved on LLM fetch error retry:
- Create server and session ONCE (outside retry loop)
- On retry, send continuation prompt to same session instead of recreating
- Agent can resume where it left off with full conversation context
- Accumulate token usage, cost, and artifacts across retry attempts

* fix(agent): address PR feedback on retry behavior

- Only track tokens/cost/artifacts from successful attempts (Issue #1, #2)
- Check remaining timeout before retry attempts (Issue #3)
- Increase event stream grace period from 1s to 2s (Issue #4)
- Add comprehensive tests for retry logic:
  - Retries on LLM fetch error and succeeds on second attempt
  - Stops retrying after MAX_LLM_RETRIES attempts
  - Does not retry on non-LLM errors
  - Only tracks results from successful attempt
  - Sends continuation prompt on retry instead of initial prompt
@fro-bot fro-bot mentioned this pull request Feb 24, 2026
86 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant