Refactor implementation with focus on performance improvement

.github/copilot-instructions.md

@@ -1,5 +1,28 @@
-Copilot instructions
+## Copilot Usage Notes
 
-Look carefully through [AGENTS.md](../AGENTS.md) for a description of the project and how to contribute.
+Always skim [AGENTS.md](../AGENTS.md) before making changes—the document is the single source of truth for architecture, performance targets, and workflow expectations.
 
-Follow instructions carefully.
+### Hot-path rules
+
+- Reuse the helpers in `src/http/utils.ts` (`writeUnauthorized`, `writeNotFound`, `writeRateLimit`, `writeErrorResponse`) instead of hand-written JSON responses.
+- Preserve the SSE contract in `src/http/routes/chat.ts`: emit role chunk first, follow with `data: { ... }` payloads, and terminate with `data: [DONE]`.
+- When streaming, keep `socket.setNoDelay(true)` on the response socket to avoid latency regressions.
+- Honor `state.activeRequests` concurrency guard and return early 429s via `writeRateLimit`.
+
+### Tool calling compatibility
+
+- `mergeTools` already merges deprecated `functions`; prefer extending it over new code paths.
+- The bridge treats `tool_choice: "required"` like `"auto"` and ignores `parallel_tool_calls`—reflect this limitation in docs if behavior changes.
+- Stream tool call deltas using `delta.tool_calls` chunks containing JSON-encoded argument strings. Downstream clients should replace, not append, argument fragments.
+
+### Scope & contracts
+
+- Public endpoints are `/health`, `/v1/models`, `/v1/chat/completions`. Changing contracts requires README updates and a version bump.
+- Keep the bridge loopback-only unless a new configuration knob is explicitly approved.
+- Update configuration docs when introducing new `bridge.*` settings and run `npm run compile` before handing off changes.
+
+### Workflow
+
+- Plan with the todo-list tool, keep diffs minimal, and avoid formatting unrelated regions.
+- Capture limitations or behavior differences (e.g., missing OpenAI response fields) in comments or docs so clients aren’t surprised.
+- Summarize reality after each change: what was touched, how it was verified, and any follow-ups.

.github/instructions/ts.instructions.md

@@ -13,6 +13,7 @@ applyTo: '**/*.ts'
 - Prefer readable, explicit solutions over clever shortcuts.
 - Extend current abstractions before inventing new ones.
 - Prioritize maintainability and clarity, short methods and classes, clean code.
+- Keep edits aligned with [AGENTS.md](../../AGENTS.md) and `.github/copilot-instructions.md`.
 
 ## Programming Language: TypeScript
 
@@ -40,6 +41,11 @@ applyTo: '**/*.ts'
 - Use pure ES modules; never emit `require`, `module.exports`, or CommonJS helpers.
 - Rely on the project's build, lint, and test scripts unless asked otherwise.
 - Note design trade-offs when intent is not obvious.
+- Reuse the HTTP helpers in `src/http/utils.ts` (`writeUnauthorized`, `writeNotFound`, `writeRateLimit`, `writeErrorResponse`) instead of writing ad-hoc JSON responses.
+- Preserve the SSE contract in `src/http/routes/chat.ts`: send the role chunk first, follow with `data: { ... }` payloads, and always terminate with `data: [DONE]`.
+- When streaming, call `res.socket?.setNoDelay(true)` before emitting chunks to avoid latency regressions.
+- Honor the concurrency guard (`state.activeRequests`) and return early 429 responses via `writeRateLimit` when limits are exceeded.
+- Communicate limitations of the VS Code LM API, e.g., `tool_choice: "required"` behaving like `"auto"` and lack of `parallel_tool_calls` support.
 
 ## Project Organization
 
@@ -75,6 +81,7 @@ applyTo: '**/*.ts'
 - Send errors through the project's logging/telemetry utilities.
 - Surface user-facing errors via the repository's notification pattern.
 - Debounce configuration-driven updates and dispose resources deterministically.
+- Prefer the pre-serialized error helpers for fast paths and document any new reason codes in README + status handlers.
 
 ## Architecture & Patterns
 
@@ -126,6 +133,8 @@ applyTo: '**/*.ts'
 - Defer expensive work until users need it.
 - Batch or debounce high-frequency events to reduce thrash.
 - Track resource lifetimes to prevent leaks.
+- Avoid repeated configuration reads in hot paths; cache settings when practical.
+- Maintain streaming code paths without buffering entire responses; only accumulate when `stream: false`.
 
 ## Documentation & Comments