Summary
184 lines are copy-pasted identically between OpenAiLlmClient.kt (319 lines) and GeminiLlmClient.kt (359 lines). Only ~77-98 lines per client are genuinely unique. Any change to shared logic must be manually applied in both files.
Duplicated methods (identical, copy-paste)
| Method |
OpenAI lines |
Gemini lines |
executeRequest() |
90-110 |
95-115 |
buildSystemPrompt() |
151-159 |
148-156 |
buildUserContent() |
162-175 |
159-172 |
formatAction() |
177-185 |
174-182 |
scaleBitmap() |
197-207 |
193-203 |
parseActionSchema() |
216-224 |
215-223 |
toDecision() |
226-267 |
225-266 |
requireOnly() |
269-282 |
268-281 |
buildActionSchema() |
284-301 |
283-300 |
validateNoUnknownKeys() |
307-312 |
302-307 |
logFailure() |
303-305 |
309-311 |
| Error-handling shell |
71-87 |
76-92 |
Parts that genuinely differ
- URL construction and auth — Bearer header vs query param
- Request body structure (
buildRequestJson()) — different JSON schemas
- Response parsing (
parseAssistantContent()) — choices[0].message.content vs candidates[0].content.parts[].text
- Image encoding (
encodeImage()) — data URL wrapper vs raw base64
- Data model classes —
OpenAiModels.kt vs inline Gemini classes
Drift already detected
LOG_TAG is "LlmClient" in OpenAI but "GeminiLlmClient" in Gemini — the OpenAI value was never updated.
Suggestion
Extract shared logic into an abstract base class or shared utility, leaving only the API-specific parts (request building, response parsing, auth) in the concrete implementations.
Files
action-executor/src/main/java/com/scroller/agent/executor/OpenAiLlmClient.ktaction-executor/src/main/java/com/scroller/agent/executor/GeminiLlmClient.kt
Summary
184 lines are copy-pasted identically between
OpenAiLlmClient.kt(319 lines) andGeminiLlmClient.kt(359 lines). Only ~77-98 lines per client are genuinely unique. Any change to shared logic must be manually applied in both files.Duplicated methods (identical, copy-paste)
executeRequest()buildSystemPrompt()buildUserContent()formatAction()scaleBitmap()parseActionSchema()toDecision()requireOnly()buildActionSchema()validateNoUnknownKeys()logFailure()Parts that genuinely differ
buildRequestJson()) — different JSON schemasparseAssistantContent()) —choices[0].message.contentvscandidates[0].content.parts[].textencodeImage()) — data URL wrapper vs raw base64OpenAiModels.ktvs inline Gemini classesDrift already detected
LOG_TAGis"LlmClient"in OpenAI but"GeminiLlmClient"in Gemini — the OpenAI value was never updated.Suggestion
Extract shared logic into an abstract base class or shared utility, leaving only the API-specific parts (request building, response parsing, auth) in the concrete implementations.
Files
action-executor/src/main/java/com/scroller/agent/executor/OpenAiLlmClient.ktaction-executor/src/main/java/com/scroller/agent/executor/GeminiLlmClient.kt