Feature Discussion: P2P Federation, Realtime Ephemeral Keys, Reasoning Parser, and Distributed Cache

Summary

I am planning to contribute several interconnected features to LocalAI that enhance P2P federation, realtime API security, function calling reasoning support, and distributed model caching. This issue serves as a discussion starter before breaking the work into smaller, reviewable PRs.

Proposed Changes

1. P2P Node Snapshot and Federation Routing (core/p2p/)

• NodeConfig struct for declarative node configuration
• discoveryTunnels full-node snapshot for cluster state synchronization
• ReplaceNodes for safe cluster membership updates
• HMAC-signed node advertisements for tamper-evident federation

2. Reasoning Content Parser (pkg/functions/)

• XMLToolCallFormat extended with three new fields for reasoning content
• ParseMsgWithXMLToolCalls now supports thinking reasoning extraction
• Enables models that emit chain-of-thought before tool calls

3. Realtime API Ephemeral Keys (core/http/endpoints/openai/realtime.go)

• 60-second HMAC-signed ephemeral keys for secure client authentication
• Optional transcription model per-session
• Any-to-any modality detection for unified audio/text sessions

4. Distributed Replica Cache (pkg/model/loader.go)

• replicaCache layer to reduce redundant FindAndLockNodeWithModel DB calls
• Per-model-ID caching with configurable TTL

5. Config YAML Endpoints (core/http/endpoints/localai/config_meta.go)

• New REST endpoints for retrieving and validating model configuration
• Supports dynamic config reloading without restart

6. MCP HTTP API Improvements (pkg/mcp/localaitools/httpapi/)

• Enhanced client with better error handling
• Additional routes for tool discovery

7. Metrics and Monitoring Enhancements (core/services/monitoring/)

• Additional backend monitoring metrics
• Improved worker file staging telemetry

8. Template Context Pipeline (core/templates/)

• Extended template loader with context-aware evaluation
• Support for dynamic template parameters

Breaking Down the Work

Per maintainer feedback, I will open separate PRs for each non-controversial subset:

1. PR 1 - Config YAML endpoints (non-controversial, self-contained)
2. PR 2 - Context pipeline and template loader improvements
3. PR 3 - Metrics and monitoring additions
4. PR 4 - MCP HTTP API routes and client hardening
5. PR 5 - Reasoning parser (with tests, proper implementation)
6. PR 6 - Distributed cache (with tests, document cache invalidation behavior)
7. PR 7 - P2P node snapshot and ReplaceNodes (with design doc, cancellation safety)
8. PR 8 - Realtime ephemeral key (with tests for HMAC round-trip)

Blocker Items to Resolve

• Dead code: GetConfigEndpointShutdown, StartReplicaCache, LoadModel - remove or implement with tests
• Reasoning parser no-op branches: proper implementation needed
• P2P cancellation safety: needs test under churn
• HMAC ephemeral key details: sign userID, fix response fields
• Tests: HMAC handshake, P2P snapshot/cancel, flag/validation round-trips
• AI-assisted code review: add Assisted-by trailer per project guidelines

Request for Feedback

I would appreciate early feedback on:

1. The overall feature set - are all of these wanted in LocalAI?
2. The P2P federation design - is a full-node snapshot the right approach?
3. The ephemeral key TTL (60s) - is this appropriate?
4. The distributed cache invalidation strategy - TTL vs event-driven?

Once we align on the design, I will open the individual PRs.

---

Related: This is a follow-up to the previously closed PR that combined all changes into one.

This issue is a discussion umbrella rather than a single task; split it into the proposed PRs before starting. For the planned first subset, read core/http/endpoints/localai/config_meta.go and the surrounding configuration endpoint code. Done means the config retrieval and validation endpoints, dynamic reload behavior, and their round-trip tests are defined in a focused PR.