Merged
Conversation
fix cherry-pick fallout for LLM prompt integration
📝 WalkthroughWalkthroughAdds OpenAI as an alternative AI provider to Gemini, implements provider resolution (Gemini preferred, OpenAI fallback), refactors embedder and LLM clients for multi-provider support, infers embedding dimensions from provider/model, and updates docs/workflows/examples to accept Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant Config as Config (cfg.Embedding)
participant Init as Init
participant Env as Env (GEMINI/OPENAI)
participant Resolver as ResolveProvider
participant Embedder as Embedder
participant LLM as LLMClient
participant VectorDB as VectorDBPrep
Init->>Config: Load API key, model, dimensions
Init->>Env: Read GEMINI_API_KEY / OPENAI_API_KEY
Init->>Resolver: ResolveProvider(cfg key or env)
Resolver-->>Init: Provider + Resolved API Key
Init->>Embedder: NewEmbedder(apiKey, model)
Embedder->>Embedder: init provider client (Gemini/OpenAI)
Embedder->>Embedder: infer Dimensions()
Embedder-->>Init: Embedder(provider, model, dimensions)
Init->>LLM: NewLLMClient(apiKey, model...)
LLM->>LLM: resolve provider, init client
Init->>VectorDB: NewVectorDBPrep(..., embedder)
VectorDB->>Embedder: embedder.Dimensions()
VectorDB-->>Init: use inferred dimensions for collection ops
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Suggested labels
Poem
🚥 Pre-merge checks | ✅ 5 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing touches
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Fix all issues with AI agents
Verify each finding against the current code and only fix it if needed.
In `@DOCS/examples/multi-repo/simili.yaml`:
- Around line 12-15: Add an inline comment to the llm block to show OpenAI
alternatives: annotate the provider key (provider: "gemini" -> add comment like
# or "openai"), the api_key key (api_key: "${GEMINI_API_KEY}" -> note OpenAI env
var alternative), and the model key (model: "gemini-2.5-flash" -> add a hint
like # or "gpt-4o" or "gpt-4.1") so users see equivalent OpenAI values when
choosing provider, referencing the llm keys provider, api_key, and model in
simili.yaml.
In `@internal/integrations/gemini/embedder.go`:
- Around line 143-151: The code in embedOpenAI mutates the receiver field
e.dimensions (setting it to len(embedding)), which causes a data race if Embed
or EmbedBatch are used concurrently; remove the runtime assignment in
embedOpenAI and instead rely on the existing inferEmbeddingDimensions logic (or
initialize dimensions at construction), leaving e.dimensions immutable at
runtime; alternatively, if runtime detection is required, wrap the one-time
dimension detection in a sync.Once or use atomic operations—update embedOpenAI
to stop writing e.dimensions and adjust any callers that expected it to be set
there (e.g., EmbedBatch/Embed) to use inferEmbeddingDimensions or the new
one-time detection helper.
In `@internal/integrations/gemini/llm.go`:
- Around line 137-178: The NewLLMClient function sets unsafe default model names
that will cause runtime failures and soon-deprecated behavior: change the
default for ProviderOpenAI (where client.model is set under the ProviderOpenAI
case) from "gpt-5.2" to a valid OpenAI model such as "gpt-5" or "gpt-5-mini",
and update the ProviderGemini default (where client.model is set under the
ProviderGemini case) from "gemini-2.0-flash-lite" to a non-deprecated Gemini
model such as "gemini-2.5-flash-lite" or "gemini-3-flash-preview"; keep the
selectedModel override logic intact so passing a model still takes precedence.
In `@internal/integrations/gemini/prompts.go`:
- Around line 317-364: buildPRDuplicateDetectionPrompt currently dereferences
input.PullRequest.Title and Body without a nil check; add a guard inside
buildPRDuplicateDetectionPrompt to handle a nil PullRequest by reading Title and
Body into local variables (e.g., title, body) with safe defaults ("") if
input.PullRequest is nil, then use those locals in the fmt.Sprintf call and
truncate calls; ensure you still build candidateList as before and return the
prompt using the safe title/body values to avoid a panic when PullRequest is
nil.
In `@internal/integrations/gemini/provider.go`:
- Around line 50-56: The heuristic in inferProviderFromKey only checks for "sk-"
and misses other OpenAI key formats like project-scoped "sk-proj-" (and similar
variants); update inferProviderFromKey to trim and lowercase the apiKey, then
check for additional known OpenAI prefixes (e.g., "sk-proj-" plus any other
documented OpenAI prefixes you find) before returning ProviderOpenAI, falling
back to ProviderGemini otherwise, and add a brief comment above the function
documenting which prefixes are treated as OpenAI keys; reference the
inferProviderFromKey function and ProviderOpenAI/ProviderGemini constants when
making the change.
🧹 Nitpick comments (2)
🤖 Fix all nitpicks with AI agents
Verify each finding against the current code and only fix it if needed. In `@DOCS/examples/multi-repo/simili.yaml`: - Around line 12-15: Add an inline comment to the llm block to show OpenAI alternatives: annotate the provider key (provider: "gemini" -> add comment like # or "openai"), the api_key key (api_key: "${GEMINI_API_KEY}" -> note OpenAI env var alternative), and the model key (model: "gemini-2.5-flash" -> add a hint like # or "gpt-4o" or "gpt-4.1") so users see equivalent OpenAI values when choosing provider, referencing the llm keys provider, api_key, and model in simili.yaml. In `@internal/integrations/gemini/provider.go`: - Around line 50-56: The heuristic in inferProviderFromKey only checks for "sk-" and misses other OpenAI key formats like project-scoped "sk-proj-" (and similar variants); update inferProviderFromKey to trim and lowercase the apiKey, then check for additional known OpenAI prefixes (e.g., "sk-proj-" plus any other documented OpenAI prefixes you find) before returning ProviderOpenAI, falling back to ProviderGemini otherwise, and add a brief comment above the function documenting which prefixes are treated as OpenAI keys; reference the inferProviderFromKey function and ProviderOpenAI/ProviderGemini constants when making the change.DOCS/examples/multi-repo/simili.yaml (1)
12-15: Consider adding OpenAI alternative comments to the LLM section for consistency.The embedding section includes helpful comments showing OpenAI alternatives, but the LLM section doesn't have similar guidance. For users who want to use OpenAI throughout, adding inline comments (e.g.,
provider: "gemini" # or "openai") would improve consistency.📝 Suggested enhancement
llm: - provider: "gemini" - api_key: "${GEMINI_API_KEY}" - model: "gemini-2.5-flash" + provider: "gemini" # or "openai" + api_key: "${GEMINI_API_KEY}" # or "${OPENAI_API_KEY}" + model: "gemini-2.5-flash" # OpenAI example: "gpt-4o-mini"🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@DOCS/examples/multi-repo/simili.yaml` around lines 12 - 15, Add an inline comment to the llm block to show OpenAI alternatives: annotate the provider key (provider: "gemini" -> add comment like # or "openai"), the api_key key (api_key: "${GEMINI_API_KEY}" -> note OpenAI env var alternative), and the model key (model: "gemini-2.5-flash" -> add a hint like # or "gpt-4o" or "gpt-4.1") so users see equivalent OpenAI values when choosing provider, referencing the llm keys provider, api_key, and model in simili.yaml.internal/integrations/gemini/provider.go (1)
50-56: Thesk-prefix heuristic may not cover all OpenAI key formats.OpenAI has introduced project-scoped keys with prefixes like
sk-proj-and organization keys. While the current heuristic works for common cases and falls back to Gemini safely, consider documenting this behavior or extending the check.📝 Suggested documentation or extended check
func inferProviderFromKey(apiKey string) Provider { - // OpenAI keys commonly use sk-* prefixes. Fall back to Gemini for compatibility. - if strings.HasPrefix(strings.TrimSpace(apiKey), "sk-") { + // OpenAI keys use sk-* prefixes (including sk-proj-*, sk-svcacct-*). + // Fall back to Gemini for compatibility when prefix is unrecognized. + trimmed := strings.TrimSpace(apiKey) + if strings.HasPrefix(trimmed, "sk-") { return ProviderOpenAI } return ProviderGemini }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/provider.go` around lines 50 - 56, The heuristic in inferProviderFromKey only checks for "sk-" and misses other OpenAI key formats like project-scoped "sk-proj-" (and similar variants); update inferProviderFromKey to trim and lowercase the apiKey, then check for additional known OpenAI prefixes (e.g., "sk-proj-" plus any other documented OpenAI prefixes you find) before returning ProviderOpenAI, falling back to ProviderGemini otherwise, and add a brief comment above the function documenting which prefixes are treated as OpenAI keys; reference the inferProviderFromKey function and ProviderOpenAI/ProviderGemini constants when making the change.
Comment on lines
+137
to
178
| // NewLLMClient creates a new LLM client. | ||
| func NewLLMClient(apiKey string, model ...string) (*LLMClient, error) { | ||
| provider, resolvedKey, err := ResolveProvider(apiKey) | ||
| if err != nil { | ||
| return nil, fmt.Errorf("failed to create Gemini client: %w", err) | ||
| return nil, err | ||
| } | ||
|
|
||
| if model == "" { | ||
| model = "gemini-2.5-flash" // Fast and cost-effective | ||
| client := &LLMClient{ | ||
| provider: provider, | ||
| apiKey: resolvedKey, | ||
| } | ||
| selectedModel := "" | ||
| if len(model) > 0 { | ||
| selectedModel = strings.TrimSpace(model[0]) | ||
| } | ||
|
|
||
| return &LLMClient{ | ||
| client: client, | ||
| model: model, | ||
| }, nil | ||
| switch provider { | ||
| case ProviderGemini: | ||
| ctx := context.Background() | ||
| geminiClient, err := genai.NewClient(ctx, option.WithAPIKey(resolvedKey)) | ||
| if err != nil { | ||
| return nil, fmt.Errorf("failed to create Gemini client: %w", err) | ||
| } | ||
| client.gemini = geminiClient | ||
| if selectedModel != "" { | ||
| client.model = selectedModel | ||
| } else { | ||
| client.model = "gemini-2.0-flash-lite" | ||
| } | ||
| case ProviderOpenAI: | ||
| client.openAI = &http.Client{Timeout: 60 * time.Second} | ||
| if selectedModel != "" { | ||
| client.model = selectedModel | ||
| } else { | ||
| client.model = "gpt-5.2" | ||
| } | ||
| default: | ||
| return nil, fmt.Errorf("unsupported provider: %s", provider) | ||
| } | ||
|
|
||
| return client, nil | ||
| } |
There was a problem hiding this comment.
🧩 Analysis chain
🌐 Web query:
OpenAI GPT models available API 2026 current
💡 Result:
As of February 16, 2026, these GPT-branded models are available via the OpenAI API (availability can depend on usage tier / org verification) [1]:
-
GPT‑5 series (general + reasoning-capable)
gpt-5.1,gpt-5,gpt-5-mini,gpt-5-nano[2][3]gpt-5-pro(Responses API only) [4]gpt-5.1-codex(optimized for agentic coding; Responses API only) [5]
-
GPT‑4.1 series (non-reasoning, long-context)
gpt-4.1,gpt-4.1-mini,gpt-4.1-nano[6][7][8]
-
GPT‑4o mini (small multimodal GPT)
gpt-4o-mini[9]
-
GPT‑4o mini audio models
gpt-4o-mini-transcribe(speech-to-text) [10]gpt-4o-mini-tts(text-to-speech) [11]
-
Legacy GPT
gpt-3.5-turbo(still available, but OpenAI recommends newer models) [12]
Note: OpenAI states that GPT‑4o and GPT‑4.1-family models were retired from ChatGPT on Feb 13, 2026 but “continue to be available in the API.” [13]
Sources: [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]
🌐 Web query:
Google Gemini models available API 2026
💡 Result:
As of Feb 16, 2026, these Gemini-family models are available in the Gemini API (Google AI for Developers), per Google’s current model docs:
Core Gemini (text output; multimodal inputs)
gemini-3-pro-preview(Preview) [1]gemini-3-flash-preview(Preview) [1]gemini-2.5-pro(Stable) [2]gemini-2.5-flash(Stable) [1]gemini-2.5-flash-lite(Stable) [3]gemini-2.0-flash-lite(plus stable-001) [4]gemini-2.0-flash(plus stable-001, and-exp) [4]
Gemini image model (image + text output)
gemini-3-pro-image-preview(Preview) [1]
Important lifecycle notes (API)
- Google changed the “latest” aliases in Jan 2026:
gemini-pro-latest→gemini-3-pro-previewgemini-flash-latest→gemini-3-flash-preview[5]
- Deprecation announced:
gemini-2.0-flash/gemini-2.0-flash-lite(and their-001variants) scheduled to shut down Mar 31, 2026 [5]
How to see the exact current list in your project
Use the API’s models list endpoint to programmatically enumerate what’s available to you right now. [6]
Sources: [1] [2] [3] [4] [5] [6]
The default OpenAI model "gpt-5.2" does not exist and will cause runtime failures.
As of February 2026, available OpenAI API models include gpt-5.1, gpt-5, gpt-5-mini, gpt-5-nano, gpt-4.1-mini, gpt-4o-mini, and others—but gpt-5.2 is not among them. Use gpt-5 or gpt-5-mini instead.
Additionally, gemini-2.0-flash-lite (line 164) is scheduled for deprecation on March 31, 2026. Replace with gemini-2.5-flash-lite or gemini-3-flash-preview for longer-term stability.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@internal/integrations/gemini/llm.go` around lines 137 - 178, The NewLLMClient
function sets unsafe default model names that will cause runtime failures and
soon-deprecated behavior: change the default for ProviderOpenAI (where
client.model is set under the ProviderOpenAI case) from "gpt-5.2" to a valid
OpenAI model such as "gpt-5" or "gpt-5-mini", and update the ProviderGemini
default (where client.model is set under the ProviderGemini case) from
"gemini-2.0-flash-lite" to a non-deprecated Gemini model such as
"gemini-2.5-flash-lite" or "gemini-3-flash-preview"; keep the selectedModel
override logic intact so passing a model still takes precedence.
Contributor
Author
There was a problem hiding this comment.
The bot is out-of-date: gpt-5.2 is real
There was a problem hiding this comment.
🧩 Analysis chain
🌐 Web query:
OpenAI gpt-5.2 model API availability 2026
💡 Result:
As of February 16, 2026, GPT‑5.2 is available in the OpenAI API.
-
API model IDs / endpoints
gpt-5.2(GPT‑5.2 “Thinking”) — available in the Responses API and Chat Completions API. [1]gpt-5.2-chat-latest(GPT‑5.2 “Instant”) — available in the Responses API and Chat Completions API. [1]gpt-5.2-pro(GPT‑5.2 “Pro”) — available in the Responses API. [1]
-
Who can access it
- OpenAI states GPT‑5 / GPT‑5.1 / GPT‑5.2 are accessible to API users on usage tiers 1–5, with some access/features subject to organization verification. [3]
-
Docs / snapshot
- The model documentation lists
gpt-5.2plus a snapshot such asgpt-5.2-2025-12-11. [2]
- The model documentation lists
Sources:
[1] OpenAI release post “Introducing GPT‑5.2” (Dec 11, 2025)
[2] OpenAI API docs: GPT‑5.2 model page
[3] OpenAI Help Center: “API Model Availability by Usage Tier and Verification Status”
@nick1udwig, you're absolutely right—thank you for the correction! My initial review was incorrect.
gpt-5.2 is indeed a valid OpenAI model available in the API as of February 2026, with variants including gpt-5.2-chat-latest and gpt-5.2-pro. The default model choice in the code is correct. I apologize for the confusion caused by outdated search results in my earlier analysis.
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Fix all issues with AI agents
Verify each finding against the current code and only fix it if needed.
In `@internal/integrations/gemini/llm.go`:
- Around line 324-326: The code accesses resp.Candidates[0].Content.Parts
without checking that Content is non-nil, risking a nil pointer panic; update
the validation around resp.Candidates to also verify resp.Candidates[0].Content
!= nil before reading .Parts (and propagate an informative error like "empty or
blocked response from LLM"), e.g., add a nil check for
resp.Candidates[0].Content and handle that case the same as when Candidates or
Parts are empty so the function returns a clear error instead of panicking.
- Around line 445-452: The extractJSONObject function can return mismatched
brackets because it finds any opening char with IndexAny and any closing with
LastIndexAny; update extractJSONObject to locate the earliest opening delimiter
between '{' and '[' (use strings.Index for "{" and "["), determine which
delimiter was chosen, then use the corresponding closing delimiter ('}' for '{'
and ']' for '[') with LastIndex to find the matching end; if a valid pair is
found return the trimmed substring, otherwise return "" (keep function name
extractJSONObject and behavior otherwise).
🧹 Nitpick comments (1)
🤖 Fix all nitpicks with AI agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/llm.go`: - Around line 445-452: The extractJSONObject function can return mismatched brackets because it finds any opening char with IndexAny and any closing with LastIndexAny; update extractJSONObject to locate the earliest opening delimiter between '{' and '[' (use strings.Index for "{" and "["), determine which delimiter was chosen, then use the corresponding closing delimiter ('}' for '{' and ']' for '[') with LastIndex to find the matching end; if a valid pair is found return the trimmed substring, otherwise return "" (keep function name extractJSONObject and behavior otherwise).internal/integrations/gemini/llm.go (1)
445-452: Minor edge case: mismatched bracket extraction.
extractJSONObjectusesIndexAny("{[")andLastIndexAny("}]")independently, which could extract mismatched delimiters (e.g.,{...data...]if]appears after}). Since this is a fallback after primary parsing attempts and would still fail atjson.Unmarshal, the impact is low.♻️ Optional: match bracket types
func extractJSONObject(text string) string { - first := strings.IndexAny(text, "{[") - lastObj := strings.LastIndexAny(text, "}]") - if first >= 0 && lastObj > first { - return strings.TrimSpace(text[first : lastObj+1]) + firstObj := strings.Index(text, "{") + lastObj := strings.LastIndex(text, "}") + firstArr := strings.Index(text, "[") + lastArr := strings.LastIndex(text, "]") + + // Prefer object if it starts first, else array + if firstObj >= 0 && lastObj > firstObj && (firstArr < 0 || firstObj < firstArr) { + return strings.TrimSpace(text[firstObj : lastObj+1]) + } + if firstArr >= 0 && lastArr > firstArr { + return strings.TrimSpace(text[firstArr : lastArr+1]) } return "" }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/llm.go` around lines 445 - 452, The extractJSONObject function can return mismatched brackets because it finds any opening char with IndexAny and any closing with LastIndexAny; update extractJSONObject to locate the earliest opening delimiter between '{' and '[' (use strings.Index for "{" and "["), determine which delimiter was chosen, then use the corresponding closing delimiter ('}' for '{' and ']' for '[') with LastIndex to find the matching end; if a valid pair is found return the trimmed substring, otherwise return "" (keep function name extractJSONObject and behavior otherwise).
There was a problem hiding this comment.
🤖 Fix all issues with AI agents
Verify each finding against the current code and only fix it if needed.
In `@internal/integrations/gemini/embedder.go`:
- Around line 4-10: The TODO in package gemini (file embedder.go) about renaming
the package to provider-neutral naming should be turned into a tracked issue:
create an issue in the repo tracker describing the rename (current path/package
"gemini", desired example name "internal/integrations/ai"), include rationale,
migration steps, and any API/compatibility considerations, then update the TODO
comment in embedder.go to reference the created issue ID (e.g.,
TODO(2026-02-16): see issue `#123`) so the TODO is tied to an actionable ticket;
ensure the issue number and short link are added to the comment and push the
updated comment.
🧹 Nitpick comments (1)
🤖 Fix all nitpicks with AI agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/embedder.go`: - Around line 4-10: The TODO in package gemini (file embedder.go) about renaming the package to provider-neutral naming should be turned into a tracked issue: create an issue in the repo tracker describing the rename (current path/package "gemini", desired example name "internal/integrations/ai"), include rationale, migration steps, and any API/compatibility considerations, then update the TODO comment in embedder.go to reference the created issue ID (e.g., TODO(2026-02-16): see issue `#123`) so the TODO is tied to an actionable ticket; ensure the issue number and short link are added to the comment and push the updated comment.internal/integrations/gemini/embedder.go (1)
4-10: Track the package rename TODO in an issue.Optional, but turning this TODO into a tracked issue helps avoid long-lived TODOs.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/embedder.go` around lines 4 - 10, The TODO in package gemini (file embedder.go) about renaming the package to provider-neutral naming should be turned into a tracked issue: create an issue in the repo tracker describing the rename (current path/package "gemini", desired example name "internal/integrations/ai"), include rationale, migration steps, and any API/compatibility considerations, then update the TODO comment in embedder.go to reference the created issue ID (e.g., TODO(2026-02-16): see issue `#123`) so the TODO is tied to an actionable ticket; ensure the issue number and short link are added to the comment and push the updated comment.
There was a problem hiding this comment.
🤖 Fix all issues with AI agents
Verify each finding against the current code and only fix it if needed.
In `@internal/integrations/gemini/llm.go`:
- Line 136: The default Gemini model in internal/integrations/gemini/llm.go is
set to "gemini-2.0-flash-lite" (see the client.model assignment); change that
default to a non-deprecated model such as "gemini-2.5-flash-lite" or
"gemini-3-flash-preview" by updating the client.model value, and optionally make
the default configurable (via env/config) so future model swaps don't require
code changes.
- Around line 445-452: The current extractJSONObject selects the first opening
brace from "{[" and the last closing from "}]" independently, which can yield
mismatched pairs; update extractJSONObject to determine which opening char was
found (e.g., open := text[first]) then search for the matching closing char
(close := '}' if open == '{' else ']' if open == '[') using a last-index search
for that specific close (e.g., LastIndexByte/LastIndex) and only return the
substring if last > first; otherwise return "" to avoid extracting invalid
mixed-bracket JSON. Ensure you update the logic inside extractJSONObject and
keep trimming behavior the same.
🧹 Nitpick comments (2)
🤖 Fix all nitpicks with AI agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/llm.go`: - Line 136: The default Gemini model in internal/integrations/gemini/llm.go is set to "gemini-2.0-flash-lite" (see the client.model assignment); change that default to a non-deprecated model such as "gemini-2.5-flash-lite" or "gemini-3-flash-preview" by updating the client.model value, and optionally make the default configurable (via env/config) so future model swaps don't require code changes. - Around line 445-452: The current extractJSONObject selects the first opening brace from "{[" and the last closing from "}]" independently, which can yield mismatched pairs; update extractJSONObject to determine which opening char was found (e.g., open := text[first]) then search for the matching closing char (close := '}' if open == '{' else ']' if open == '[') using a last-index search for that specific close (e.g., LastIndexByte/LastIndex) and only return the substring if last > first; otherwise return "" to avoid extracting invalid mixed-bracket JSON. Ensure you update the logic inside extractJSONObject and keep trimming behavior the same.internal/integrations/gemini/llm.go (2)
136-136: Consider updating default Gemini model before deprecation.The default model
gemini-2.0-flash-liteis scheduled for deprecation on March 31, 2026 (approximately 6 weeks away). Consider updating togemini-2.5-flash-liteorgemini-3-flash-previewfor longer-term stability.♻️ Suggested update
if selectedModel != "" { client.model = selectedModel } else { - client.model = "gemini-2.0-flash-lite" + client.model = "gemini-2.5-flash-lite" }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/llm.go` at line 136, The default Gemini model in internal/integrations/gemini/llm.go is set to "gemini-2.0-flash-lite" (see the client.model assignment); change that default to a non-deprecated model such as "gemini-2.5-flash-lite" or "gemini-3-flash-preview" by updating the client.model value, and optionally make the default configurable (via env/config) so future model swaps don't require code changes.
445-452: Minor: Bracket matching heuristic may extract invalid JSON.The function finds the first
{or[and last}or]independently, which could extract invalid JSON for malformed responses (e.g.,{data]would extract{data]). Since this is a fallback for error recovery and LLMs typically produce valid JSON, the impact is minimal.♻️ Optional: verify bracket type matches
func extractJSONObject(text string) string { first := strings.IndexAny(text, "{[") - lastObj := strings.LastIndexAny(text, "}]") - if first >= 0 && lastObj > first { + if first < 0 { + return "" + } + openChar := text[first] + var closeChar byte + if openChar == '{' { + closeChar = '}' + } else { + closeChar = ']' + } + lastObj := strings.LastIndexByte(text, closeChar) + if lastObj > first { return strings.TrimSpace(text[first : lastObj+1]) } return "" }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@internal/integrations/gemini/llm.go` around lines 445 - 452, The current extractJSONObject selects the first opening brace from "{[" and the last closing from "}]" independently, which can yield mismatched pairs; update extractJSONObject to determine which opening char was found (e.g., open := text[first]) then search for the matching closing char (close := '}' if open == '{' else ']' if open == '[') using a last-index search for that specific close (e.g., LastIndexByte/LastIndex) and only return the substring if last > first; otherwise return "" to avoid extracting invalid mixed-bracket JSON. Ensure you update the logic inside extractJSONObject and keep trimming behavior the same.
Member
|
Hi @nick1udwig, I will solve the conflics and merge. Changes looks good but I need to do some minor changes to this which I will do in a later PR. Thank you. |
Kavirubc
added a commit
that referenced
this pull request
Feb 17, 2026
Adds OpenAI as an alternative AI provider alongside Gemini. Provider selection is automatic: Gemini is preferred when both keys are set; OpenAI is used when only OPENAI_API_KEY is present. Covers embeddings, LLM, config, docs, workflow examples, and CI. Conflict with exponential-backoff retry logic resolved: both embedGemini and embedOpenAI are now wrapped in withRetry. Closes #42 Co-authored-by: nick1udwig <nick1udwig@users.noreply.github.com>
This was referenced Feb 17, 2026
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
2 participants
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.
Description
Split out the openai provider portion of #40 to resolve #42
Type of Change
Related Issues
Fixes #42
Changes Made
Add OpenAI API support for LLM and embedding
Use Gemini by default (i.e. if both Gemini and OpenAI are provided)
Testing
go build ./...successfullygo test ./...successfullygo vet ./...successfullyScreenshots (if applicable)
N/A
Checklist
Additional Notes
Summary by CodeRabbit
New Features
Documentation