hot reloading the BRUTAL way

.serena/memories/project_overview.md

@@ -0,0 +1,16 @@
+# Project Overview
+
+`lda-workflow-as-struct` is a Python 3.14 prototype for `lda.chat`: an AI-assisted workflow system where an LLM plans structured workflows and a deterministic executor validates/runs them.
+
+Main packages live under `src/`:
+- `wf_core`: workflow model, validation, runtime semantics, frames, trace, interrupts, foreach, async execution.
+- `wf_authoring`: ergonomic authoring layer including `@node`, `NodeSpec`, `WorkflowBuilder`, conditions, paths, and subgraph wrapping.
+- `wf_mcp`: MCP broker/proxy layer for managing multiple backend MCP connections, discovery/catalog snapshots, transparent FastMCP proxying, config/admin tools, and eventual workflow build/run integration.
+
+Important docs:
+- `readme.md`: running design notes and architecture.
+- `authoring_sketch.md`: authoring API direction.
+- `wf_mcp_plan.md`: MCP proxy/broker/workflow integration plan.
+- `scratchpad.md`: rough design history.
+
+Current MCP direction: transparent proxy mode is the main product path. Old broker tools remain useful for debugging/admin/catalog operations, but protocol-native FastMCP proxying exposes upstream tools/resources/prompts as first-class MCP capabilities.

.serena/memories/style_and_conventions.md

@@ -0,0 +1,20 @@
+# Style And Conventions
+
+General:
+- Python 3.14, `src/` layout, Pydantic v2 where external/boundary validation is useful.
+- Prefer explicit dataclasses for runtime/internal models and Pydantic for config/wire-ish boundary validation.
+- Async-first for MCP calls and workflow runtime interactions.
+- Keep MCP proxy concepts separate from workflow-specific concepts like `outcome`.
+- Do not leak workflow-only fields into MCP `tools/list`.
+
+Code style:
+- Use precise type hints and modern Python collection syntax (`list[str]`, `dict[str, Any]`).
+- Keep modules layered by responsibility; avoid stuffing everything into service/runtime files.
+- Prefer small helper modules when behavior becomes a boundary (`config_models.py`, `config_manager.py`, `proxy_validation.py`).
+- Validation should fail early with clear errors.
+- Tests should exercise behavior through public APIs/MCP calls where practical.
+
+Editing rules from repo collaboration:
+- Use `apply_patch` for manual code edits.
+- Do not revert user-owned changes.
+- Treat `wf_mcp.config.json` as user-owned live config unless explicitly asked.

.serena/memories/suggested_commands.md

@@ -0,0 +1,27 @@
+# Suggested Commands
+
+Use PowerShell on Windows from the repo root.
+
+Testing:
+- `uv run --with pytest pytest -q`
+- Focused example: `uv run --with pytest pytest tests/test_wf_mcp_transparent_proxy.py -q`
+
+Lint/type checks:
+- `uv run ruff check src/wf_mcp tests`
+- Focused basedpyright example: `uv run basedpyright src/wf_mcp/transparent_proxy.py --level error`
+
+Formatting:
+- `uv run ruff format`
+
+CLI / MCP server:
+- `uv run wf-mcp --config wf_mcp.config.json serve`
+- Transparent proxy mode is default.
+- Old broker mode: `uv run wf-mcp --config wf_mcp.config.json serve --mode broker`
+- Optional compatibility/search flags: `--resources-as-tools`, `--prompts-as-tools`, `--search-tools`
+
+Useful Windows shell commands:
+- Fast search: `rg "pattern" path`
+- List files: `Get-ChildItem -Force`
+- Read file: `Get-Content -Path path`
+- Git status: `git status --short`
+- Diff: `git diff -- path`

.serena/memories/task_completion_checklist.md

@@ -0,0 +1,12 @@
+# Task Completion Checklist
+
+Before considering a code task done:
+- Run focused tests for the touched area.
+- Run full tests when the change affects shared behavior: `uv run --with pytest pytest -q`.
+- Run ruff on touched source/tests, usually `uv run ruff check src/wf_mcp tests` for MCP work.
+- Run focused basedpyright at error level for new or heavily changed files when type issues are likely.
+- Check `git status --short` and distinguish user-owned config changes from code changes.
+- Summarize functional changes and verification results concisely.
+
+Known environment note:
+- Windows sandbox may block commands with `CreateProcessAsUserW failed: 5`; retry important commands with escalation rather than working around via unsafe shell tricks.