[HarshRajSinghania]

User description

Summary

Updates the GitHub Action installation documentation so the rate-limit guidance points to the correct setting.

Motivation

Fixes #4193.

The Rate Limiting section and the "Rate limit exceeded" troubleshooting entry currently recommend raising config.ai_timeout (or adding a fallback model). A 429 is returned immediately; ai_timeout only bounds the duration of each individual request and does not change rate-limit behavior. The setting that controls per-call retries inside the completion client is config.num_retries (commented out by default in pr_agent/settings/configuration.toml).

Implementation

• In the Rate Limiting example, document config.num_retries and explain that each model in a fallback chain receives its own retry budget. Note that quota/hard-limit errors are retried but will not recover, and that ai_timeout is for timeouts only.
• In the Common Error Messages section, update the "Rate limit exceeded" solution to recommend config.num_retries and mention pr_reviewer.publish_error_details (settable only via workflow env) as a way to surface a sanitized failure reason for a manual /review.

No code or configuration defaults were changed.

Testing

• Manually verified the updated sections against the issue description and the source of num_retries in configuration.toml and litellm_ai_handler.py.
• Confirmed the Markdown renders the new guidance clearly and that links to the Review tool docs remain valid.

No automated tests are affected (documentation-only change).

---

PR Type

Documentation

---

Description

• Recommend config.num_retries for transient rate limits.

• Explain fallback retry budgets and quota limitations.

• Clarify timeouts do not resolve rate limits.

• Document workflow-only diagnostics for manual review failures.

---

File Walkthrough

Relevant files

Documentation

github.md Correct rate-limit retries and review troubleshooting guidance docs/docs/installation/github.md Add config.num_retries to the workflow example for transient provider errors. Explain client defaults, independent fallback retry budgets, and unrecoverable quota errors. Clarify that config.ai_timeout addresses timeouts rather than HTTP 429 responses. Document workflow-only pr_reviewer.publish_error_details for sanitized manual /review failure details and link to review documentation. +9/-3

---

---

⚙️ Agent run details

• Model: openai/gpt-6.1-sol
• Tokens: 5,520 in / 205 out / 5,725 total
• Time cost: 6.7s
• AI calls: 1

[the-pr-agent-bot]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis ✅ 4193 - Fully compliant Compliant requirements: Recommend config.num_retries in both rate-limit guidance sections. Explain independent retries for fallback models. Explain that retries do not resolve quota errors. Distinguish request timeouts from rate-limit responses. Document workflow-only error diagnostics for manual /review failures and link to the Review tool documentation.

⏱️ Estimated effort to review: 1 🔵⚪⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ No major issues detected

---

⚙️ Agent run details

• Model: openai/gpt-6.1-sol
• Tokens: 6,725 in / 352 out / 7,077 total
• Time cost: 9.9s
• AI calls: 1

[IsmaelMartinez]

Thanks @HarshRajSinghania for turning #4193 around so quickly; the num_retries recommendation, the fallback note and the publish_error_details pointer all match the code. Two small wording fixes inline: num_retries is not a plain retry count once LiteLLM adds its own retries, and the workflow env takes YAML, so the error-details example needs the YAML form.

[IsmaelMartinez]

Thanks @HarshRajSinghania. #4425 has since made the handler the only retry layer, so I pushed the matching num_retries comment and the YAML form of publish_error_details. Merging.