feat(engine): retry transient draft failures with jittered backoff - #20
Open
jonasbrami wants to merge 1 commit into
Open
feat(engine): retry transient draft failures with jittered backoff#20jonasbrami wants to merge 1 commit into
jonasbrami wants to merge 1 commit into
Conversation
The agentic draft is the one expensive call in generation and the one that fails transiently — rate limits, provider overload, a dropped connection mid-stream. Today any failure goes straight to the error page. Wrap `llm.draft_quiz` in a capped retry loop (`_draft_with_retry`): - Retry only transient failures. The claude-agent adapter collapses every SDK failure into a `RuntimeError` carrying the upstream text, so retryability is classified on the message (rate limit / overload / dropped connection), not the type. A `ValidationError` (malformed submission) and fatal markers (missing `claude` binary, exhausted turn budget) surface immediately. - Exponential backoff with full jitter (`_backoff_delay`): a random wait in `[0, min(30s, 2s * 2**attempt)]`. The jitter decorrelates concurrent clients so a burst of retries doesn't re-collide on every wave. - Up to 3 attempts; the final attempt re-raises the original exception unchanged, so the caller's error handling sees the real cause. - Retry notices forward to the activity sink, so a streamed run shows the wait instead of going silent. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
jonasbrami
added a commit
that referenced
this pull request
May 24, 2026
) The old demo used a hand-authored stand-in quiz. This replaces it with a quiz cognit actually generated — captured by running the live engine against a real PR (#20, the retry-with-backoff change) — so the GIF shows genuine, model-authored questions, including a true before/after mermaid flow comparison of the retry exception path. - `scripts/capture_demo_quiz.py`: runs the real generation engine against a PR and records the generated quiz + activity feed into `scripts/demo_data/`. - `scripts/record_demo.py`: replays the captured quiz and feed. The Playwright driver is now data-driven — it computes and clicks the correct answer for each question from the quiz itself, so it adapts to whatever mix/order the model produced (no hard-coded option indices). Recording is higher resolution: Playwright records at the CSS-viewport resolution (device_scale_factor and an oversized record size don't add pixels), so the viewport is bumped to 1440x900 and the GIF is emitted 1:1 at 1440px, 16fps — no downscale softening. - `scripts/demo_data/{quiz.json,feed.json}`: the captured artifacts, so the demo stays offline and deterministic on re-record while the content is genuine. - README alt-text updated to match (five questions: three MCQ, a mermaid-pick flow diagram, true/false). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The agentic draft (
llm.draft_quiz) is the one expensive call in quiz generation, and the one that fails transiently: rate limits, provider overload, a connection dropped mid-stream. Today any exception goes straight to the error page — a momentary 429 means the whole run is lost.This wraps the draft in a capped, jittered retry loop (
_draft_with_retry).How it decides what to retry
The claude-agent adapter collapses every SDK failure into a
RuntimeErrorcarrying the upstream text, so retryability is classified on the message, not the exception type:rate limit/429/overloaded/529/503/timed out/connection reset.ValidationError(the model submitted a malformed quiz; deterministic), a missingclaudebinary, or an exhausted turn budget (maximum number of turns). Retrying those just burns the same tokens to the same dead end.Backoff
Exponential with full jitter: each wait is a random value in
[0, min(30s, 2s · 2**attempt)]. The jitter is load-bearing — fixed-schedule retries from a burst of clients re-collide on every wave (thundering herd); spreading each wait across the whole window decorrelates them.Up to 3 attempts. The final attempt re-raises the original exception unchanged so the caller's error handling sees the real cause. Retry notices forward to the activity sink, so a streamed run shows the wait instead of going silent.
Tests
tests/engine/test_generate.py: retryable/fatal classification, jittered-cap bounds, retry-then-succeed, immediate raise on non-retryable, and attempt exhaustion.🤖 Generated with Claude Code