Module: Pikuri::RubyLlmPatches

Defined in:

lib/pikuri/ruby_llm_patches.rb

Overview

Local corrections to ruby_llm, applied at load time, carried until the fix ships upstream. Each patch pins us to a ruby_llm internal's private shape, so it's a liability: small, self-contained, tied to a tracked upstream issue so we know when to delete it. Reopening a third-party class is last resort — here only because the defect is on the live chat path and the release cadence isn't ours. Patching rather than dropping the dependency is the standing policy — DECISIONS.md D_keep_ruby_llm.

The streaming error-parse crash (ruby_llm #837)

https://github.com/crmne/ruby_llm/issues/837

On a streaming turn, a non-200 response whose body parses to a non-Hash JSON value crashes ruby_llm 1.16.0 with +TypeError: String does not have #dig+ at providers/openai/streaming.rb:43, instead of surfacing the endpoint's error. Two bodies trigger it, both because the method assumes a nested object to .dig:

1. A bare JSON string. The guard return unless error_data['error'] is meant as key-presence, but on a String that's substring matching, so a body containing "error" slips through into .dig.
2. A {"error": "<message>"} body where error is a string value (xAI/Grok), not an object: the guard passes, then error_data.dig('error', 'type') calls String#dig — same TypeError.

The non-streaming path already handles both (cases on Hash/Array/else, short-circuits a string error). This mirrors that: never .dig a non-Hash, hand a non-typed error up as the message with a nil status so ruby_llm's build_stream_error_response falls back to the real HTTP status.

Redefined whole (not prepend + super) because it's a small self-contained module_function with no state — faithful reproduction plus the guard is clearer than wrapping the buggy original. Re-running module_function replaces both the module method and the private copy the provider gets via include.

When to delete this file

ruby_llm main already carries an equivalent fix, in a module renamed to RubyLLM::Protocols::ChatCompletions::Streaming — so the next release past 1.16.0 obsoletes the patch and moves its target. The load-time guard below turns that into a boot failure naming this file; delete it and its spec then.

Class Method Summary

• .applied? ⇒ Boolean

  true once the patches below have been applied; a literal, kept honest by the load-time guard that raises if the target module has moved out from under it.

Class Method Details

.applied? ⇒ Boolean

Returns true once the patches below have been applied; a literal, kept honest by the load-time guard that raises if the target module has moved out from under it.

Returns:

• (Boolean) —

  true once the patches below have been applied; a literal, kept honest by the load-time guard that raises if the target module has moved out from under it

# File 'lib/pikuri/ruby_llm_patches.rb', line 54

def applied? = true