Fix release docs and Chrysalis YAC smoke

CHANGELOG.md

@@ -23,8 +23,9 @@ under their entry.
   `endpoints add/list/remove`, `doctor`, `install-claude`.
 - Multi-endpoint config schema (`hpc.endpoints.<name>` with
   `endpoint_id`, `path_prefixes`, `timeout_seconds`) and config
-  discovery order: `$UXARRAY_MCP_CONFIG` → `~/.config/uxarray-mcp/config.yaml`
-  → `./config.yaml`.
+  discovery order: `$UXARRAY_MCP_CONFIG` → `./config.yaml` in the current
+  working directory → `~/.config/uxarray-mcp/config.yaml` → editable-install
+  repo fallback.
 - YAC remote build script (`scripts/hpc_build_yac.py`) and runtime
   fallback that loads `yac.core` via `importlib.machinery` when the
   upstream `__init__.py` is unconditional.

README.md

@@ -82,7 +82,10 @@ The ``uxarray-mcp`` CLI exposes:
 | ``install-claude``  | print or merge the Claude Desktop ``mcpServers`` block   |
 
 Config is discovered in this order: ``$UXARRAY_MCP_CONFIG`` →
-``~/.config/uxarray-mcp/config.yaml`` → ``./config.yaml`` (repo root).
+``./config.yaml`` in the current working directory →
+``~/.config/uxarray-mcp/config.yaml`` → the editable-install repo config
+fallback. The project-local file wins inside a checkout so development
+endpoints are not shadowed by an empty user config.
 
 ## Most Users Should Read These in Order
 

conda/recipe/meta.yaml

@@ -23,6 +23,9 @@ requirements:
     - pip
     - uv-build >=0.9.26,<0.10.0
   run:
+    # Core package only for the initial feedstock. Add optional HPC support as
+    # a second output/variant after globus-compute-sdk and academy-py solver
+    # behavior is validated on conda-forge.
     - python >={{ python_min }}
     - fastmcp >=3.4.0
     - holoviews >=1.19.0

docs/architecture.html

@@ -100,7 +100,7 @@
     <div class="logo-box">⬡</div>
     <div class="hdr-text">
       <h1>UXarray MCP Server — Architecture Diagram</h1>
-      <p>Mesh-aware assistant · provenance on every output · local or Argonne Improv HPC via Globus Compute · dynamic tool registration</p>
+      <p>Mesh-aware assistant · provenance on every output · local or named HPC endpoints via Globus Compute · unified tool surface</p>
     </div>
     <div class="branch">
       <svg width="12" height="12" viewBox="0 0 16 16" fill="#58a6ff"><path d="M11.75 2.5a.75.75 0 1 0 0 1.5.75.75 0 0 0 0-1.5zm-2.25.75a2.25 2.25 0 1 1 3 2.122V6A2.5 2.5 0 0 1 10 8.5H6a1 1 0 0 0-1 1v1.128a2.251 2.251 0 1 1-1.5 0V5.372a2.25 2.25 0 1 1 1.5 0v1.836A2.492 2.492 0 0 1 6 7h4a1 1 0 0 0 1-1v-.628A2.25 2.25 0 0 1 9.5 3.25z"/></svg>
@@ -205,7 +205,7 @@ <h1>UXarray MCP Server — Architecture Diagram</h1>
   <text x="540" y="96"  text-anchor="middle" font-size="15" font-weight="700" fill="#58a6ff" font-family="Inter,sans-serif">Receive tool call</text>
   <line x1="400" y1="103" x2="682" y2="103" stroke="#112238" stroke-width="1"/>
   <text x="540" y="121" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">run_scientific_agent(path)</text>
-  <text x="540" y="139" text-anchor="middle" font-size="11.5" fill="#484f58" font-family="Inter,sans-serif">HPC tools only if endpoint configured</text>
+  <text x="540" y="139" text-anchor="middle" font-size="11.5" fill="#484f58" font-family="Inter,sans-serif">Remote execution selected per call</text>
 
   <!-- Agent: Begin analysis -->
   <rect x="747" y="76" width="306" height="74" rx="8" fill="#0d0a20" stroke="#6e40c9" stroke-width="2"/>
@@ -261,13 +261,13 @@ <h1>UXarray MCP Server — Architecture Diagram</h1>
   <text x="1260" y="383" text-anchor="middle" font-size="11.5" fill="#8b949e" font-family="Inter,sans-serif">MPAS · UGRID · SCRIP · HEALPix</text>
   <text x="1260" y="398" text-anchor="middle" font-size="11"   fill="#2ea043" font-family="Inter,sans-serif" font-style="italic">→ n_face · n_node · n_edge · format</text>
 
-  <!-- HPC: inspect_mesh_hpc() -->
+  <!-- HPC: inspect_mesh(..., use_remote=True) -->
   <rect x="1467" y="320" width="306" height="86" rx="8" fill="#180e00" stroke="#d29922" stroke-width="2"/>
-  <text x="1620" y="342" text-anchor="middle" font-size="15" font-weight="700" fill="#e3b341" font-family="Inter,sans-serif">inspect_mesh_hpc( )</text>
+  <text x="1620" y="342" text-anchor="middle" font-size="15" font-weight="700" fill="#e3b341" font-family="Inter,sans-serif">inspect_mesh(use_remote)</text>
   <line x1="1480" y1="350" x2="1762" y2="350" stroke="#201400" stroke-width="1"/>
   <text x="1620" y="362" text-anchor="middle" font-size="11.5" fill="#39d0d0" font-family="Inter,sans-serif" font-style="italic">✓ _endpoint_is_ready( ) pre-flight</text>
   <text x="1620" y="377" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">remote/compute_functions.py</text>
-  <text x="1620" y="392" text-anchor="middle" font-size="11.5" fill="#8b949e" font-family="Inter,sans-serif">Globus Executor → Improv cluster</text>
+  <text x="1620" y="392" text-anchor="middle" font-size="11.5" fill="#8b949e" font-family="Inter,sans-serif">Globus Executor → named endpoint</text>
   <text x="1620" y="406" text-anchor="middle" font-size="11"   fill="#d29922" font-family="Inter,sans-serif" font-style="italic">File stays on HPC · timeout 300 s</text>
 
   <!-- validate_dataset box -->
@@ -348,13 +348,13 @@ <h1>UXarray MCP Server — Architecture Diagram</h1>
   <text x="1260" y="772" text-anchor="middle" font-size="11.5" fill="#2ea043" font-family="Inter,sans-serif" font-style="italic">Runs entirely on local machine</text>
   <text x="1260" y="787" text-anchor="middle" font-size="11"   fill="#484f58" font-family="Inter,sans-serif">_provenance: venue=local</text>
 
-  <!-- HPC: calc_area_hpc + zonal_mean_hpc -->
+  <!-- HPC: calculate_area/calculate_zonal_mean with use_remote=True -->
   <rect x="1467" y="688" width="306" height="108" rx="8" fill="#180e00" stroke="#d29922" stroke-width="2"/>
   <text x="1620" y="710" text-anchor="middle" font-size="15" font-weight="700" fill="#e3b341" font-family="Inter,sans-serif">HPC Execution</text>
   <line x1="1480" y1="718" x2="1762" y2="718" stroke="#201400" stroke-width="1"/>
   <text x="1620" y="733" text-anchor="middle" font-size="11.5" fill="#39d0d0" font-family="Inter,sans-serif" font-style="italic">✓ _endpoint_is_ready( ) pre-flight</text>
-  <text x="1620" y="749" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">calculate_area_hpc( )</text>
-  <text x="1620" y="765" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">calculate_zonal_mean_hpc( )</text>
+  <text x="1620" y="749" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">calculate_area(use_remote)</text>
+  <text x="1620" y="765" text-anchor="middle" font-size="13"   fill="#8b949e" font-family="Inter,sans-serif">calculate_zonal_mean(use_remote)</text>
   <line x1="1480" y1="776" x2="1762" y2="776" stroke="#3d1400" stroke-width="1" stroke-dasharray="4,3"/>
   <text x="1620" y="790" text-anchor="middle" font-size="12" fill="#f85149" font-family="Inter,sans-serif" font-style="italic">⚠ auto-fallback if endpoint unreachable</text>
 
@@ -446,15 +446,15 @@ <h1>UXarray MCP Server — Architecture Diagram</h1>
     Flow runs <strong>top → bottom</strong>, arrows crossing lanes are handoffs.
     <strong style="color:#e3b341">Orange diamonds</strong> = routing decisions made automatically by the agent.
     <strong style="color:#56d364">Green boxes</strong> = compute runs on your local machine.
-    <strong style="color:#e3b341">Orange boxes</strong> = dispatched to Argonne Improv via Globus Compute — the file never leaves the cluster.
-    <strong style="color:#39d0d0">Teal dashed boxes</strong> = steps that only activate under certain conditions (data_path provided, endpoint configured).
+    <strong style="color:#e3b341">Orange boxes</strong> = dispatched to a named HPC endpoint via Globus Compute — the file never leaves the cluster.
+    <strong style="color:#39d0d0">Teal dashed boxes</strong> = steps that only activate under certain conditions (data_path provided, remote mode requested).
     <strong style="color:#f85149">Red dashed</strong> inside HPC boxes = automatic local fallback when the endpoint is unreachable.
-    <strong>Dynamic registration:</strong> HPC tools (rows 10–13) only appear in Claude's tool list when <code>endpoint_id</code> is set in <code>config.yaml</code>.
+    <strong>Unified registration:</strong> tools are registered once; remote execution is selected with <code>use_remote=True</code> and optional endpoint names.
   </div>
 
   <!-- Tool table -->
   <div class="card">
-    <div class="card-h">All MCP Tools &nbsp;<span style="font-weight:400;color:#6e7681;font-size:12px">— 9 always registered · 4 conditional on endpoint_id</span></div>
+    <div class="card-h">Representative MCP Tools &nbsp;<span style="font-weight:400;color:#6e7681;font-size:12px">— unified local / remote surface</span></div>
     <table>
       <thead>
         <tr><th>#</th><th>Tool Name</th><th>Source File</th><th>Runs On</th><th>Domain Module</th><th>What It Does</th></tr>
@@ -469,14 +469,14 @@ <h1>UXarray MCP Server — Architecture Diagram</h1>
         <tr><td>7</td><td><code>get_execution_mode</code></td><td><code>execution_control.py</code></td><td><span class="tag tl">LOCAL</span></td><td>—</td><td>Returns current execution mode (local / hpc / auto) and whether an HPC endpoint is configured.</td></tr>
         <tr><td>8</td><td><code>set_execution_mode</code></td><td><code>execution_control.py</code></td><td><span class="tag tl">LOCAL</span></td><td>—</td><td>Switch execution mode from the Claude UI without editing config.yaml directly.</td></tr>
         <tr><td>9</td><td><code>run_scientific_agent</code></td><td><code>scientific_agent.py</code></td><td><span class="tag ta">AUTO</span></td><td>All 4 modules</td><td>Autonomous 4-stage pipeline: Analyze → Plan → Execute → Verify. Validation-gated. Returns full reasoning trace + provenance + artifacts.</td></tr>
-        <tr><td>10</td><td><code>inspect_mesh_hpc</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>mesh.py</code> on HPC</td><td>Mesh inspection via Globus Compute on Improv. Pre-flight health check. Auto-fallback to local.</td></tr>
-        <tr><td>11</td><td><code>calculate_area_hpc</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>area.py</code> on HPC</td><td>Face area calculation via Globus Compute. Pre-flight health check. Auto-fallback to local.</td></tr>
-        <tr><td>12</td><td><code>inspect_variable_hpc</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>variable.py</code> on HPC</td><td>Variable inspection via Globus Compute. Pre-flight health check. Auto-fallback to local.</td></tr>
-        <tr><td>13</td><td><code>calculate_zonal_mean_hpc</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>zonal.py</code> on HPC</td><td>Zonal mean via Globus Compute. Pre-flight health check. Auto-fallback to local.</td></tr>
+        <tr><td>10</td><td><code>inspect_mesh</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>mesh.py</code> on HPC</td><td>Mesh inspection with <code>use_remote=True</code>. Pre-flight health check. Auto-fallback to local.</td></tr>
+        <tr><td>11</td><td><code>calculate_area</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>area.py</code> on HPC</td><td>Face area calculation with <code>use_remote=True</code>. Pre-flight health check. Auto-fallback to local.</td></tr>
+        <tr><td>12</td><td><code>inspect_variable</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>variable.py</code> on HPC</td><td>Variable inspection with <code>use_remote=True</code>. Pre-flight health check. Auto-fallback to local.</td></tr>
+        <tr><td>13</td><td><code>calculate_zonal_mean</code></td><td><code>remote_tools.py</code></td><td><span class="tag th">HPC*</span></td><td><code>zonal.py</code> on HPC</td><td>Zonal mean with <code>use_remote=True</code>. Pre-flight health check. Auto-fallback to local.</td></tr>
       </tbody>
     </table>
     <div style="padding:10px 15px;font-size:11.5px;color:#6e7681;border-top:1px solid #21262d;">
-      * HPC tools are only registered when <code>endpoint_id</code> is set in <code>config.yaml</code>. Without an endpoint, Claude's tool list shows 9 tools.
+      * Remote execution is selected per call with <code>use_remote=True</code>. Endpoint readiness controls remote dispatch and fallback.
     </div>
   </div>
 

docs/architecture.md

@@ -114,6 +114,21 @@ of exposing separate HPC-only tool names.
 
 **Validation gating** — The scientific agent runs dataset validation before zonal mean. If validation fails, zonal mean is skipped rather than producing unreliable results.
 
+## Maintenance Notes
+
+The current implementation favors a small number of broad tool modules while
+the MCP surface is still evolving. That keeps related behavior easy to audit
+for the first release, but the largest files should be split once the public
+contracts settle:
+
+- `remote/compute_functions.py` should be divided by remote capability family
+  (inspection, plotting, vector calculus, remapping, diagnostics).
+- `tools/advanced.py` should be divided into spatial, comparison, remapping,
+  temporal/ensemble, and export modules.
+
+Keep those refactors behavior-preserving and test-backed; they are polish and
+maintainability work, not blockers for the core conda package.
+
 ## Interactive Diagram
 
 An interactive architecture diagram is available at `docs/architecture.html` in the repository.

docs/chrysalis.md

@@ -19,69 +19,44 @@ which hosts the E3SM next-generation mesh library.
   are sent as source code via `AllCodeStrategies` and only need `uxarray` + deps
   in the worker environment.
 - Login nodes **kill compute processes** — always use the Slurm backend.
-- `unset PYTHONPATH` before every endpoint start — the conda `uxarray-yac` env
-  injects broken `pydantic_core` paths that crash workers.
+- YAC remapping needs the Python 3.12 `uxarray-yac` environment plus YAC, MKL,
+  MPICH, NetCDF, and local shim library paths. Use
+  `scripts/chrysalis_endpoint.sh` instead of hand-writing those paths.
+- If a remote probe times out after the endpoint is `registered`, inspect the
+  endpoint logs on Chrysalis with `scripts/chrysalis_endpoint.sh logs`.
 
 ## Worker Environment
 
 | Item | Value |
 |---|---|
-| Venv | `~/venvs/globus-compute-py313` (Python 3.13) |
+| UXarray/YAC env | `~/.conda/envs/uxarray-yac` (Python 3.12) |
+| Endpoint helper venv | `~/venvs/globus-compute-py313` |
 | Slurm partition | `debug` (4h walltime, 20 nodes) |
 | Compute nodes | 251 GB RAM, 128 CPUs |
 | Endpoint name | `uxarray-chrysalis` |
 
 ## First-Time Setup
 
+The checked-in helper script writes the endpoint profile, YAC runtime library
+paths, and small BLAS/LAPACK shims needed by the current YAC build:
+
 ```bash
-# 1. Create Python 3.13 conda env
-/gpfs/fs1/soft/chrysalis/manual/miniforge3/25.3.1/bin/conda create \
-    -n gc-py313 python=3.13 -y
-
-# 2. Build the globus-compute venv
-~/.conda/envs/gc-py313/bin/python -m venv ~/venvs/globus-compute-py313
-~/venvs/globus-compute-py313/bin/pip install \
-    "globus-compute-endpoint==4.12.0" \
-    uxarray xarray netCDF4 h5netcdf numpy matplotlib holoviews cartopy
-
-# 3. Configure the endpoint (Slurm-backed)
-unset PYTHONPATH
-~/venvs/globus-compute-py313/bin/globus-compute-endpoint configure uxarray-chrysalis
-
-cat > ~/.globus_compute/uxarray-chrysalis/user_config_template.yaml.j2 << 'EOF'
-endpoint_setup: ""
-engine:
-  type: GlobusComputeEngine
-  max_workers_per_node: 4
-  provider:
-    type: SlurmProvider
-    partition: debug
-    nodes_per_block: 1
-    init_blocks: 0
-    min_blocks: 0
-    max_blocks: 2
-    walltime: "04:00:00"
-    worker_init: |
-      unset PYTHONPATH
-    launcher:
-      type: SrunLauncher
-idle_heartbeats_soft: 10
-idle_heartbeats_hard: 5760
-EOF
-
-cat > ~/.globus_compute/uxarray-chrysalis/user_environment.yaml << 'EOF'
-PYTHONPATH: ""
-PATH: "/home/<username>/venvs/globus-compute-py313/bin:/usr/bin:/bin"
-EOF
+git clone https://github.com/UXARRAY/uxarray-mcp-server.git
+cd uxarray-mcp-server
+bash scripts/chrysalis_endpoint.sh configure slurm-debug
+bash scripts/chrysalis_endpoint.sh check-yac
 ```
 
+The `check-yac` command runs a tiny Slurm job that imports `yac.core`, imports
+UXarray's YAC helper, and remaps HEALPix zoom 2 to zoom 3. It should report
+`yac_core_ok: true` and `remap_ok: true` before the endpoint is used by MCP.
+
 ## Starting the Endpoint
 
 Run this every time you log in:
 
 ```bash
-unset PYTHONPATH
-~/venvs/globus-compute-py313/bin/globus-compute-endpoint start uxarray-chrysalis
+bash scripts/chrysalis_endpoint.sh start
 ```
 
 The endpoint prints its UUID. Add it to your private local config on your
@@ -98,6 +73,8 @@ From your laptop after the endpoint is running:
 
 ```bash
 uv run python scripts/hpc_doctor.py --endpoint chrysalis --timeout-seconds 120
+uv run --extra hpc python scripts/yac_smoke_test.py \
+  --endpoint chrysalis --timeout-seconds 300
 ```
 
 Or manually:
@@ -110,6 +87,14 @@ print(validate_hpc_setup(endpoint='chrysalis', run_remote_probe=True,
                           probe_timeout_seconds=120))
 ```
 
+If the manager reports `registered` but worker probes time out, inspect the
+remote side on Chrysalis:
+
+```bash
+bash scripts/chrysalis_endpoint.sh logs
+squeue -u "$USER"
+```
+
 ## E3SM Next-Generation Ocean Meshes
 
 Available at `/lcrc/group/e3sm/ac.xylar/polaris_1.0/chrysalis/test_20260520/unified-mesh-topo-cull2/`:
@@ -123,8 +108,23 @@ Available at `/lcrc/group/e3sm/ac.xylar/polaris_1.0/chrysalis/test_20260520/unif
 
 ## Troubleshooting
 
-**`ENDPOINT_NOT_ONLINE`** — the Slurm debug job timed out (4h limit). Restart with `unset PYTHONPATH && ~/venvs/globus-compute-py313/bin/globus-compute-endpoint start uxarray-chrysalis`.
+**`ENDPOINT_NOT_ONLINE`** — the Slurm debug job timed out (4h limit). Restart
+with `bash scripts/chrysalis_endpoint.sh restart`.
+
+**Worker probe timeout after `registered`** — the manager is connected, but a
+Slurm worker did not return. Run `bash scripts/chrysalis_endpoint.sh logs` on
+Chrysalis and inspect the latest submit script/log pair.
+
+**`pydantic_core` not found** — the worker is running from the wrong Python
+environment. Re-run `bash scripts/chrysalis_endpoint.sh configure slurm-debug`
+and restart the endpoint.
 
-**`WorkerLost` or `SystemError`** — PYTHONPATH is set. Always `unset PYTHONPATH` before starting the endpoint.
+**`libnetcdf.so.22`, `liblapack.so.3`, or `libblas.so.3` not found** — the YAC
+runtime paths or local MKL shims are missing. Re-run
+`bash scripts/chrysalis_endpoint.sh configure slurm-debug`, then
+`bash scripts/chrysalis_endpoint.sh check-yac`.
 
-**`pydantic_core` not found** — conda env leaked into the worker. Check `user_environment.yaml` has `PYTHONPATH: ""` and restart.
+**`PMI_Init failed` or `WorkerLost` during YAC import** — YAC initializes MPI.
+Inside a Globus Compute worker, run the YAC smoke/remap through the dedicated
+smoke path, which launches the native YAC child process with
+`srun --ntasks 1` when `SLURM_JOB_ID` is present.

docs/release.md

@@ -123,6 +123,13 @@ The conda package should install the core MCP server and CLI. HPC-specific
 Globus Compute dependencies can be added to the feedstock later if conda-forge
 availability and solver behavior are acceptable.
 
+The seed recipe intentionally targets the **core** package only. Keep
+`globus-compute-sdk` and `academy-py` out of the initial conda-forge recipe
+until those dependencies and their transitive solver behavior are validated on
+conda-forge. If conda-native HPC support becomes necessary, prefer a second
+output such as `uxarray-mcp-hpc` or a feedstock variant rather than making every
+local-only user solve the remote-execution stack.
+
 ## Privacy Check
 
 Before every release, verify endpoint UUIDs and local config did not re-enter