Add Mastra Cloud MCP deployment with Gemini support#22
Conversation
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughAdded HTTP and SSE transport endpoints to the MCP manifest, extended LM Studio synthesizer and runtime types to support an OpenAI-style Changes
Sequence Diagram(s)sequenceDiagram
participant Client
participant MCP_Server
participant Synthesizer
participant LLM_API
Client->>MCP_Server: HTTP/SSE generation request
MCP_Server->>Synthesizer: forward request (apiStyle: chat_completions)
Synthesizer->>LLM_API: GET /v1/models (discover)
LLM_API-->>Synthesizer: models list
Synthesizer->>LLM_API: POST /v1/chat/completions (model, messages, temperature)
LLM_API-->>Synthesizer: choices[0].message.content
Synthesizer-->>MCP_Server: normalized generation result
MCP_Server-->>Client: HTTP response / SSE event
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| if (apiStyle === 'chat_completions') { | ||
| if (url.pathname === '/' || url.pathname === '') { | ||
| return `${url.origin}/v1/models`; | ||
| } | ||
| return `${trimmed}/models`; | ||
| } |
There was a problem hiding this comment.
🟡 buildModelsEndpoint produces invalid URL when chat_completions baseUrl includes /chat/completions path
When apiStyle is 'chat_completions' and the user-supplied baseUrl already includes the full /chat/completions path (e.g. http://localhost:1234/v1/chat/completions), buildModelsEndpoint returns http://localhost:1234/v1/chat/completions/models instead of the correct http://localhost:1234/v1/models.
The sibling function buildGenerationEndpoint correctly handles this case at src/runtime/lm-studio-synthesizer.ts:602-603 with if (trimmed.endsWith('/chat/completions')) { return trimmed; }, and the responses style has an equivalent guard in buildModelsEndpoint at line 674 (url.pathname.endsWith('/responses')), and the lmstudio_chat style handles /api/v1/chat at line 666. The new chat_completions branch (lines 658-663) is missing equivalent logic to strip /chat/completions and derive the models path. This causes health check model discovery to hit an incorrect endpoint.
| if (apiStyle === 'chat_completions') { | |
| if (url.pathname === '/' || url.pathname === '') { | |
| return `${url.origin}/v1/models`; | |
| } | |
| return `${trimmed}/models`; | |
| } | |
| if (apiStyle === 'chat_completions') { | |
| if (url.pathname.endsWith('/chat/completions')) { | |
| return `${url.origin}${url.pathname.replace(/\/chat\/completions$/, '/models')}`; | |
| } | |
| if (url.pathname === '/' || url.pathname === '') { | |
| return `${url.origin}/v1/models`; | |
| } | |
| return `${trimmed}/models`; | |
| } |
Was this helpful? React with 👍 or 👎 to provide feedback.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/runtime/lm-studio-synthesizer.ts (1)
295-300:⚠️ Potential issue | 🔴 CriticalGuard the Gemini key fallback by target host.
Line 299 and Line 315 fall back to
GEMINI_AI_API_KEYwheneverFPF_LOCAL_LLM_API_KEYis unset. IfbaseUrlpoints at LM Studio or any other OpenAI-compatible backend, this forwards the Gemini bearer token to the wrong server. Restrict that fallback to Gemini’s OpenAI-compatible host, or require an explicitapiKeyfor non-Gemini targets.🔒 Proposed fix
+function resolveApiKey( + explicitApiKey: string | undefined, + env: NodeJS.ProcessEnv, + baseUrl: string, + apiStyle: LmStudioApiStyle, +): string | undefined { + const trimmedExplicit = explicitApiKey?.trim(); + if (trimmedExplicit) { + return trimmedExplicit; + } + + const configuredKey = env.FPF_LOCAL_LLM_API_KEY?.trim(); + if (configuredKey) { + return configuredKey; + } + + const url = safeUrl(baseUrl); + if ( + apiStyle === 'chat_completions' && + url?.hostname === 'generativelanguage.googleapis.com' + ) { + return env.GEMINI_AI_API_KEY?.trim(); + } + + return undefined; +} + export function createSynthesizerFromEnv( env: NodeJS.ProcessEnv = process.env, fetchImpl?: typeof fetch, ): LocalAnswerSynthesizer | undefined { @@ return new LmStudioSynthesizer({ baseUrl, model, apiStyle, - apiKey: env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim(), + apiKey: resolveApiKey(undefined, env, baseUrl, apiStyle), timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : DEFAULT_LM_STUDIO_TIMEOUT_MS, fetchImpl, env, }); } @@ - const apiKey = options.apiKey?.trim() || env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim(); + const apiKey = resolveApiKey(options.apiKey, env, baseUrl, apiStyle);Also applies to: 309-315
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/runtime/lm-studio-synthesizer.ts` around lines 295 - 300, The code is currently falling back to env.GEMINI_AI_API_KEY whenever env.FPF_LOCAL_LLM_API_KEY is unset, which may leak a Gemini bearer token to non-Gemini backends; update the LmStudioSynthesizer instantiation to only use GEMINI_AI_API_KEY when the target host is Gemini (guard by inspecting baseUrl for Gemini’s OpenAI-compatible host(s) or a gemini-specific apiStyle), otherwise do not fall back and require an explicit apiKey (or leave undefined) and surface a clear error; apply the same guarded fallback logic to the other occurrence in the same file (the second LmStudioSynthesizer/config block).
🧹 Nitpick comments (1)
src/mastra/index.ts (1)
30-33: Avoid eager duplicate runtime initialization.
export const mastra = createMastraRuntime().mastrainitializes one runtime at import time, andcreateHonoMastraRuntime()initializes another one for normal startup. Reusing a shared default runtime avoids duplicated setup work and side effects.♻️ Proposed refactor
export function createMastraRuntime(env: NodeJS.ProcessEnv = process.env) { @@ return { logger, observability, mastra, }; } +const defaultRuntime = createMastraRuntime(); + /** * Direct Mastra instance export required by `mastra build` / `mastra deploy`. */ -export const mastra = createMastraRuntime().mastra; +export const mastra = defaultRuntime.mastra; export async function createHonoMastraRuntime(env: NodeJS.ProcessEnv = process.env) { - const runtime = createMastraRuntime(env); + const runtime = env === process.env ? defaultRuntime : createMastraRuntime(env); const app = new Hono<{ Bindings: HonoBindings; Variables: HonoVariables; }>();🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/mastra/index.ts` around lines 30 - 33, The code eagerly constructs a runtime by calling createMastraRuntime() at module import, causing a second runtime when createHonoMastraRuntime() runs; change to a single shared singleton runtime: create and export a single default runtime instance (e.g., defaultMastraRuntime) and export its .mastra (export const mastra = defaultMastraRuntime.mastra), then have createHonoMastraRuntime() reuse or accept that default runtime instead of creating a new one so createMastraRuntime() is only invoked once.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/runtime/lm-studio-synthesizer.ts`:
- Around line 658-663: buildGenerationEndpoint currently maps an explicit
chat-completions baseUrl into ".../chat/completions/models", breaking model
discovery; update the apiStyle === 'chat_completions' branch in
buildGenerationEndpoint so that if url.pathname equals '/chat/completions' (with
or without trailing slash) or is empty it returns `${url.origin}/v1/models` (the
same normalization used for '/responses'), otherwise return `${trimmed}/models`;
refer to the buildGenerationEndpoint function and ensure runLmStudioHealthCheck
will receive the normalized `/v1/models` URL.
---
Outside diff comments:
In `@src/runtime/lm-studio-synthesizer.ts`:
- Around line 295-300: The code is currently falling back to
env.GEMINI_AI_API_KEY whenever env.FPF_LOCAL_LLM_API_KEY is unset, which may
leak a Gemini bearer token to non-Gemini backends; update the
LmStudioSynthesizer instantiation to only use GEMINI_AI_API_KEY when the target
host is Gemini (guard by inspecting baseUrl for Gemini’s OpenAI-compatible
host(s) or a gemini-specific apiStyle), otherwise do not fall back and require
an explicit apiKey (or leave undefined) and surface a clear error; apply the
same guarded fallback logic to the other occurrence in the same file (the second
LmStudioSynthesizer/config block).
---
Nitpick comments:
In `@src/mastra/index.ts`:
- Around line 30-33: The code eagerly constructs a runtime by calling
createMastraRuntime() at module import, causing a second runtime when
createHonoMastraRuntime() runs; change to a single shared singleton runtime:
create and export a single default runtime instance (e.g., defaultMastraRuntime)
and export its .mastra (export const mastra = defaultMastraRuntime.mastra), then
have createHonoMastraRuntime() reuse or accept that default runtime instead of
creating a new one so createMastraRuntime() is only invoked once.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: cbd85547-4450-4aed-8e84-8cdeb580299b
⛔ Files ignored due to path filters (2)
bun.lockis excluded by!**/*.lockpackage-lock.jsonis excluded by!**/package-lock.json
📒 Files selected for processing (9)
.mastra-project.jsonmanifest.jsonpackage.jsonsrc/mastra/index.tssrc/mcp/tool-contracts.tssrc/runtime/lm-studio-synthesizer.tssrc/runtime/types.tssrc/server.tstests/mcp-server.test.ts
There was a problem hiding this comment.
Pull request overview
This PR adds Mastra Cloud deployment support for the FPF Memory MCP server and extends the local LLM synthesizer to support Google Gemini’s OpenAI-compatible chat_completions API style, while aligning the Mastra entrypoint/export conventions required for mastra build/deploy.
Changes:
- Add
chat_completionssupport to the LM Studio synthesizer (request/response parsing, endpoint building, env key fallback for Gemini). - Update Mastra runtime wiring to support serverless/stateless HTTP MCP deployments and adjust imports to the
src/mastra/index.tsentrypoint. - Update MCP manifest to include HTTP transport configuration and add Mastra project metadata + dependency.
Reviewed changes
Copilot reviewed 9 out of 11 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| tests/mcp-server.test.ts | Updates Mastra runtime import path to new entrypoint layout. |
| src/server.ts | Updates server to import Mastra runtime from src/mastra/index.ts. |
| src/runtime/types.ts | Extends runtime type unions to include chat_completions. |
| src/runtime/lm-studio-synthesizer.ts | Implements chat_completions API style, Gemini API key fallback, and endpoint/payload parsing updates. |
| src/mcp/tool-contracts.ts | Extends Zod schema enum to allow chat_completions. |
| src/mastra/index.ts | Adds serverless MCP options and exports a direct Mastra instance for Mastra build/deploy conventions. |
| package.json | Adds mastra CLI as a dev dependency. |
| manifest.json | Adds HTTP transport + paths for MCP endpoints. |
| bun.lock | Lockfile updates for the new mastra dependency tree. |
| .mastra-project.json | Adds Mastra Cloud project metadata for deployment. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| if (apiStyle === 'chat_completions') { | ||
| if (url.pathname === '/' || url.pathname === '') { | ||
| return `${url.origin}/v1/models`; | ||
| } | ||
| return `${trimmed}/models`; |
There was a problem hiding this comment.
buildGenerationEndpoint explicitly supports passing a baseUrl that already ends with /chat/completions, but buildModelsEndpoint doesn’t mirror that behavior. With apiStyle === 'chat_completions' and baseUrl set to a full generation endpoint, this will produce an invalid models URL like .../chat/completions/models. Consider stripping /chat/completions (similar to the /responses handling) when computing the models endpoint.
| if (apiStyle === 'chat_completions') { | ||
| return { | ||
| model, | ||
| temperature: 0.1, | ||
| messages: [ |
There was a problem hiding this comment.
New chat_completions behavior (request payload shape, endpoint construction, and response text extraction from choices[].message.content) is not covered by existing tests. Since this file already has dedicated tests, please add cases validating chat_completions for both endpoint building and parsing/extraction to prevent regressions across providers (e.g., Gemini OpenAI-compatible endpoints).
| "projectId": "52e5f16d-40b5-4f27-9896-5537b883c4b4", | ||
| "projectName": "fpf-memory", | ||
| "projectSlug": "fpf-memory", | ||
| "organizationId": "org_O1KNSZHGZ4FXT065PGP9BSJEXT" |
There was a problem hiding this comment.
This file hard-codes a specific Mastra Cloud projectId / organizationId. If this is environment- or account-specific (similar to other deployment tool metadata files), committing it can unintentionally couple the repo to one cloud project and may expose internal identifiers. Consider documenting why it must be in-repo or ignoring it and configuring these values via deployment-time configuration instead.
| "projectId": "52e5f16d-40b5-4f27-9896-5537b883c4b4", | |
| "projectName": "fpf-memory", | |
| "projectSlug": "fpf-memory", | |
| "organizationId": "org_O1KNSZHGZ4FXT065PGP9BSJEXT" | |
| "projectId": "REPLACE_WITH_MASTRA_PROJECT_ID", | |
| "projectName": "fpf-memory", | |
| "projectSlug": "fpf-memory", | |
| "organizationId": "REPLACE_WITH_MASTRA_ORGANIZATION_ID" |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for the chat_completions API style within the LM Studio synthesizer, encompassing payload extraction, endpoint generation, and expanded environment variable support for API keys. Additionally, it configures the Mastra runtime for serverless MCP deployment and updates project metadata. Feedback was provided to simplify the text extraction logic in extractGenerationText to improve readability and reduce redundancy.
| 'output' in payload ? (payload.output ?? []) : [], | ||
| ); | ||
| if (preferredMessage) { | ||
| return preferredMessage; | ||
| } | ||
|
|
||
| return findAnyGenerationText(payload.output ?? []); | ||
| return findAnyGenerationText('output' in payload ? (payload.output ?? []) : []); |
There was a problem hiding this comment.
The logic for extracting generation text from the output property is repeated twice using ternary operators. This can be simplified by extracting the output array once, which also improves readability and makes the type narrowing clearer for TypeScript.
const output = 'output' in payload ? (payload.output ?? []) : [];
const preferredMessage = findStructuredMessageText(output);
if (preferredMessage) {
return preferredMessage;
}
return findAnyGenerationText(output);![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ff8fbcb574
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| model, | ||
| apiStyle, | ||
| apiKey: env.FPF_LOCAL_LLM_API_KEY?.trim(), | ||
| apiKey: env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim(), |
There was a problem hiding this comment.
This fallback unconditionally reuses GEMINI_AI_API_KEY whenever FPF_LOCAL_LLM_API_KEY is unset, regardless of which baseUrl is configured. In mixed environments, requests to non-Gemini OpenAI-compatible hosts will carry the Gemini bearer token, leaking that credential to unrelated endpoints and potentially causing auth failures (401/403) on providers that expect a different token. Restrict this fallback to explicit Gemini-targeted configurations (for example, compatible API style + Google endpoint) or require opt-in.
Useful? React with 👍 / 👎.
| if (url.pathname === '/' || url.pathname === '') { | ||
| return `${url.origin}/v1/models`; | ||
| } | ||
| return `${trimmed}/models`; |
There was a problem hiding this comment.
For chat_completions, this branch appends /models to any non-root path, so a valid full generation URL like .../v1/chat/completions becomes .../v1/chat/completions/models. runLmStudioHealthCheck then probes an invalid models endpoint and can report discovery failure even when generation is correctly configured. Handle /chat/completions similarly to the /responses rewrite so model discovery targets .../v1/models.
Useful? React with 👍 / 👎.
|
@codex[agent] review |
|
@claude[agent] review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 12d55649cd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (apiStyle === 'chat_completions') { | ||
| if (url.pathname === '/' || url.pathname === '') { | ||
| return `${url.origin}/v1/models`; | ||
| } | ||
| return `${trimmed}/models`; |
There was a problem hiding this comment.
When apiStyle is chat_completions, this branch always returns ${trimmed}/models for non-root paths. If users provide a base URL that already includes /chat/completions (which buildGenerationEndpoint explicitly supports), runLmStudioHealthCheck will query .../chat/completions/models instead of the provider’s models endpoint (typically .../models). In that configuration, model discovery falsely fails even though generation works, so lm-check reports a broken setup.
Useful? React with 👍 / 👎.
Handled the review feedback: chat_completions models endpoint now strips full generation URLs and added coverage for payload parsing and endpoint building in tests (5c48d72). The Mastra project file is already removed/ignored (9d640ea). |
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/runtime/lm-studio-synthesizer.ts (1)
291-300:⚠️ Potential issue | 🟠 MajorDon't reuse
GEMINI_AI_API_KEYfor arbitrary backends.With this fallback, any non-empty
GEMINI_AI_API_KEYgets sent to whateverFPF_LOCAL_LLM_BASE_URLpoints at whenFPF_LOCAL_LLM_API_KEYis unset. That leaks a provider-specific secret to localhost or other OpenAI-compatible hosts. Gate the fallback on a Gemini base URL, or require the explicit local key for non-Gemini backends.🔐 Suggested fix
+function isGeminiOpenAiBaseUrl(baseUrl: string): boolean { + const url = safeUrl(baseUrl.trim()); + return url?.hostname === 'generativelanguage.googleapis.com' + && /\/openai(?:$|\/)/.test(url.pathname); +} + +function resolveApiKey( + baseUrl: string, + env: NodeJS.ProcessEnv, + explicitApiKey?: string, +): string | undefined { + const configuredApiKey = explicitApiKey?.trim() || env.FPF_LOCAL_LLM_API_KEY?.trim(); + if (configuredApiKey) { + return configuredApiKey; + } + + return isGeminiOpenAiBaseUrl(baseUrl) ? env.GEMINI_AI_API_KEY?.trim() : undefined; +} + export function createSynthesizerFromEnv( env: NodeJS.ProcessEnv = process.env, fetchImpl?: typeof fetch, ): LocalAnswerSynthesizer | undefined { @@ return new LmStudioSynthesizer({ baseUrl, model, apiStyle, - apiKey: env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim(), + apiKey: resolveApiKey(baseUrl, env), timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : DEFAULT_LM_STUDIO_TIMEOUT_MS, fetchImpl, env, }); } @@ - const apiKey = options.apiKey?.trim() || env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim(); + const apiKey = resolveApiKey(baseUrl, env, options.apiKey);Also applies to: 313-315
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/runtime/lm-studio-synthesizer.ts` around lines 291 - 300, The current fallback that sets apiKey to env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim() leaks a provider-specific secret to arbitrary LmStudioSynthesizer backends; change the logic so that GEMINI_AI_API_KEY is only used when the target is actually Gemini (e.g., when baseUrl equals the Gemini default or when apiStyle/normalizeLmStudioApiStyle indicates Gemini), otherwise require FPF_LOCAL_LLM_API_KEY and leave apiKey undefined; apply the same guarded-fallback change to the other occurrence(s) around the LmStudioSynthesizer construction at the later block (the repeated apiKey usage referenced in the comment).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Outside diff comments:
In `@src/runtime/lm-studio-synthesizer.ts`:
- Around line 291-300: The current fallback that sets apiKey to
env.FPF_LOCAL_LLM_API_KEY?.trim() || env.GEMINI_AI_API_KEY?.trim() leaks a
provider-specific secret to arbitrary LmStudioSynthesizer backends; change the
logic so that GEMINI_AI_API_KEY is only used when the target is actually Gemini
(e.g., when baseUrl equals the Gemini default or when
apiStyle/normalizeLmStudioApiStyle indicates Gemini), otherwise require
FPF_LOCAL_LLM_API_KEY and leave apiKey undefined; apply the same
guarded-fallback change to the other occurrence(s) around the
LmStudioSynthesizer construction at the later block (the repeated apiKey usage
referenced in the comment).
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 6689008f-5572-4528-8f75-de0efdf70223
📒 Files selected for processing (2)
src/runtime/lm-studio-synthesizer.tstests/lm-studio-synthesizer.test.ts
- Move src/mastra.ts to src/mastra/index.ts for mastra build convention - Export direct Mastra instance for mastra deploy - Add serverless mcpOptions for stateless MCP over HTTP - Add chat_completions API style to synthesizer for Google Gemini OpenAI-compatible endpoint support - Add GEMINI_AI_API_KEY env var fallback in synthesizer - Update manifest.json with HTTP transport endpoints - Add mastra CLI v1.5.0 as devDependency - Add .mastra-project.json for Mastra Cloud project linking Deployed to https://fpf-memory.server.mastra.cloud MCP endpoint: /api/mcp/fpf_memory/mcp Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contains environment-specific Mastra Cloud project/org IDs that should not be committed. Generated by `mastra server deploy`. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Project uses Bun as its package manager. The package-lock.json was a temporary workaround for a bun install resolution hang during mastra build — no longer needed since the deploy pipeline works. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: venikman <22530205+venikman@users.noreply.github.com>
Bun is the project package manager, but mastra build's bun install hangs on zod@4 resolution in the output directory. A local package-lock.json forces npm as fallback during deploy only. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
venikman
force-pushed
the
feat/mastra-cloud-mcp-deploy
branch
from
a0afb1a to
cfe983f
Compare
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/runtime/lm-studio-synthesizer.ts`:
- Around line 299-300: The code currently falls back to GEMINI_AI_API_KEY for
apiKey regardless of target host; change the apiKey selection so
GEMINI_AI_API_KEY is used only when the chosen baseUrl clearly targets Gemini
(check the baseUrl variable or FPF_LOCAL_LLM_BASE_URL for a Gemini/Google
hostname substring or known Gemini URL pattern) and otherwise avoid using the
Gemini secret; apply the same guarded logic to the other occurrence where apiKey
is set (the second apiKey selection in this file) so both places only use
GEMINI_AI_API_KEY when baseUrl indicates a Gemini endpoint.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 97a04ee4-dc3d-413f-b673-ce640faa8baa
⛔ Files ignored due to path filters (1)
bun.lockis excluded by!**/*.lock
📒 Files selected for processing (10)
.gitignoremanifest.jsonpackage.jsonsrc/mastra/index.tssrc/mcp/tool-contracts.tssrc/runtime/lm-studio-synthesizer.tssrc/runtime/types.tssrc/server.tstests/lm-studio-synthesizer.test.tstests/mcp-server.test.ts
✅ Files skipped from review due to trivial changes (4)
- package.json
- .gitignore
- tests/mcp-server.test.ts
- src/server.ts
🚧 Files skipped from review as they are similar to previous changes (3)
- src/mastra/index.ts
- tests/lm-studio-synthesizer.test.ts
- src/runtime/types.ts
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cfe983fe3a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| /** | ||
| * Direct Mastra instance export required by `mastra build` / `mastra deploy`. | ||
| */ | ||
| export const mastra = createMastraRuntime().mastra; |
There was a problem hiding this comment.
Creating mastra at module load (createMastraRuntime()) eagerly initializes the logger and observability stack even when callers only import createHonoMastraRuntime. In this codebase those initializers call resolveLogPath, which performs mkdirSync/openSync; in read-only build/deploy environments this can throw during import before the server starts, and it also bypasses any env object a caller intended to pass later. Keeping this export lazy (or isolating it to a deploy-only entrypoint) avoids import-time side effects and environment drift.
Useful? React with 👍 / 👎.
Public surface (deployed): ask_fpf, query_fpf_spec, get_fpf_index_status Expert surface (local only): refresh, trace, inspect, read_doc, anchor, citations Deployed server defaults to public. Set FPF_MCP_SURFACE=full for all tools. Stdio transport always exposes all tools for local Codex/Claude use. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 475eff4077
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| export function createMastraRuntime(env: NodeJS.ProcessEnv = process.env) { | ||
| const logger = getRuntimeLogger(env); | ||
| const observability = getRuntimeObservability(env); | ||
| const mcpServer = env.FPF_MCP_SURFACE === 'full' ? fpfMcpServer : fpfPublicMcpServer; |
There was a problem hiding this comment.
createMastraRuntime now selects fpfPublicMcpServer unless FPF_MCP_SURFACE is explicitly set to "full", which means the default HTTP server path (createHonoMastraRuntime() from src/server.ts) exposes only the 3 public tools. In normal local/server startup where that env var is unset, HTTP clients lose core debug/retrieval tools (refresh_fpf_index, trace_fpf_path, read_fpf_doc, etc.) even though the stdio path still exposes them, creating an unexpected capability regression. Make the restricted surface opt-in instead of the default.
Useful? React with 👍 / 👎.
Manifest, README, AGENTS.md, docs/mcp-interface.md, and Codex config now distinguish public tools (deployed HTTP) from expert tools (local stdio only). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…action - Only use GEMINI_AI_API_KEY when baseUrl points to *.googleapis.com, preventing credential leakage to non-Gemini endpoints - Extract output array once in extractGenerationText instead of repeating the ternary check Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Flips the default so local HTTP server exposes all 9 tools. Deployed server sets FPF_MCP_SURFACE=public to restrict to 3 tools. Addresses review feedback about local dev losing expert tools. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 228a1a6dc9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (trimmed.endsWith('/chat/completions')) { | ||
| return trimmed; |
There was a problem hiding this comment.
buildGenerationEndpoint checks trimmed.endsWith('/chat/completions') on the full URL string, so a valid base URL that includes query params (for example Azure-style .../chat/completions?api-version=...) is not recognized as already complete. The function then appends another /chat/completions (and buildModelsEndpoint similarly appends /models after the query), yielding malformed endpoints and failing both generation and lm-check in that configuration. Fresh evidence: with https://example.openai.azure.com/.../chat/completions?api-version=2024-05-01-preview, the code produces ...?api-version=.../chat/completions and ...?api-version=.../models.
Useful? React with 👍 / 👎.
Summary
https://fpf-memory.server.mastra.cloudchat_completionsAPI style to synthesizer for Google Gemini OpenAI-compatible endpointsrc/mastra.tstosrc/mastra/index.ts(required bymastra buildconvention)mcpOptionsfor stateless HTTP MCPGEMINI_AI_API_KEYenv var fallback in synthesizer.mastra-project.jsonandpackage-lock.json(deploy-only artifacts)MCP Endpoints (deployed)
/api/mcp/fpf_memory/mcp/api/mcp/fpf_memory/sse/api/mcp/fpf_memory/tools/api/mcp/v0/serversGemini Configuration
Test plan
bun run check— TypeScript passesbun run test— 36/36 tests passbun run build— bundles 3 entry pointsnpx mastra server deploy— deployed, health check passedhttps://fpf-memory.server.mastra.cloud/health,/api/mcp/v0/servers,/api/mcp/fpf_memory/tools🤖 Generated with Claude Code
Summary by CodeRabbit
New Features
Tests