@@ -1,12 +1,12 @@
 {
   "name": "braina-plugins",
   "owner": {
-    "name": "Andrea Brovelli — BraiNets, Institut de Neurosciences de la Timone"
+    "name": "Andrea Brovelli \u2014 BraiNets, Institut de Neurosciences de la Timone"
   },
   "plugins": [
     {
       "name": "braina",
-      "source": "."
+      "source": "./plugin"
     }
   ]
 }
@@ -0,0 +1,25 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    timeout-minutes: 40
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+      - name: Check the environment
+        run: uv run check_env.py
+      - name: Library verification (Frites + HOI)
+        run: uv run plugin/mcp/verify_libs.py
+      - name: MCP tool tests
+        run: uv run tests/run_tests.py
+      - name: Plugin manifests are valid JSON
+        run: |
+          python3 -c "import json; json.load(open('.claude-plugin/marketplace.json')); json.load(open('plugin/.claude-plugin/plugin.json')); json.load(open('plugin/hooks/hooks.json')); print('manifests OK')"
@@ -6,3 +6,7 @@ tutorials/seeg_ebrains_frites/datasets/seeg/*.nc
 
 # Python bytecode
 __pycache__/
+
+# Eval runs and test artefacts
+plugin/evals/results/
+.pytest_cache/
@@ -0,0 +1,56 @@
+# Changelog
+
+## 1.1.0 - 2026-09-18
+
+### Fixed
+- MCP server is declared in `plugin/.claude-plugin/plugin.json` (`mcpServers`);
+  the root `.mcp.json`, which failed to resolve `${CLAUDE_PLUGIN_ROOT}` in a
+  plain clone, is gone.
+- The `mcp/` directory no longer shadows the `mcp` SDK package.
+- Frites connectivity tools now pass the `roi`/`times` coordinate names to
+  Frites, so `.nc` input really keeps ROI names, a time axis in seconds and
+  `sfreq` (previously silently replaced by `roi_0, roi_1, ...` and 1 Hz).
+- HOI tools convert `.nc` input to numpy (HOI rejects xarray) and keep the
+  multiplet metadata (`multiplets`, `order`, `multiplet_names`) in their
+  `.nc` output; `hoi_get_nbest_mult` resolves region names from it.
+- `frites_wf_stats` writes `.nc` with `times`/`roi` coords; `frites_conn_pid`
+  takes `dt` instead of a nonexistent `max_delay`; `frites_conn_te` forwards
+  `sfreq`; random-effect inference with fewer than two subjects is refused.
+- All tools catch exceptions and return `Error in <tool>: ...` strings, and
+  return a summary (shape, coords, min/mean/max) of what they wrote.
+
+### Added
+- Multi-subject workflows: `frites_wf_mi` / `frites_wf_conn_comod` accept one
+  file per subject or a `.nc` with a `subject` dimension, plus `mcp`,
+  `cluster_th`, `cluster_alpha`, `random_state`, `estimator`.
+- `frites_wf_mi_combine` (WfMiCombine) via `save_workflow=True` pickles.
+- `frites_conn_reshape` (pairs -> sources x targets matrix) and
+  `frites_conn_net` (A->B minus B->A).
+- Extra Frites parameters: `norm`, `n_jobs` (covgc), `estimator` (dfc),
+  `normalized` (ccf), `n_jobs` everywhere; `frites_sim_ar` gains `sf`, `dt`,
+  `n_std`, `stim_onset`, `random_state` and stores `sfreq`.
+- HOI: `method` (gc / gauss / binning / knn / kernel), `samples_path`, and
+  bootstrap confidence intervals (`n_boots`, `ci_percentiles`,
+  `random_state`, written to `<output>_ci.nc`) on every metric; new metrics
+  `hoi_tc`, `hoi_sinfo`, `hoi_infotot`, `hoi_transfer_entropy`, `hoi_dotot`,
+  `hoi_redundancy_phiid`, `hoi_atoms_phiid`.
+- `convert_to_nc` (MNE `.fif` epochs, `.mat`, `.csv`, relabelling `.npy`) and
+  `plot_result` (PNG of any braina output).
+- `/braina:check` skill with `scripts/smoke_test.py`; SessionStart hook that
+  pre-warms the `uv` environment; `uv` lockfile for the server script.
+- pytest suite (`tests/`), GitHub Actions CI, `claude plugin eval` cases.
+
+### Changed
+- Repository layout: everything the plugin installs now lives under
+  `plugin/` (manifest, MCP server, skills, examples, hooks, evals); papers,
+  tutorials and use cases stay at the root and are no longer copied on
+  install (68 MB -> ~2 MB).
+- Skills restructured for progressive disclosure: short `SKILL.md` plus
+  per-tool `references/` files; limitations fixed in 1.1.0 removed from
+  the text; papers cited by DOI/URL.
+- Tool count corrected (22 in 1.0.0, 34 in 1.1.0).
+
+## 1.0.0 - 2026-09-13
+
+Initial release: MCP server wrapping Frites and HOI (22 tools) and four
+skills (orientation, frites-connectivity, hoi-metrics, workflows).
@@ -15,32 +15,52 @@ The user may be a novice in computational neuroscience. Explain mathematical con
 uv run check_env.py
 
 # Run the verification test suite for Frites + HOI
-uv run mcp/verify_libs.py
+uv run plugin/mcp/verify_libs.py
+
+# Run the pytest suite over all MCP tools (synthetic data, ~30 s)
+uv run tests/run_tests.py
+
+# End-to-end smoke test of the installation (what /braina:check runs)
+uv run plugin/scripts/smoke_test.py
 
 # Launch the MCP server standalone
-uv run mcp/braina_mcp.py
+uv run plugin/mcp/braina_mcp.py
+
+# Refresh the server lockfile after editing its dependency block
+uv lock --script plugin/mcp/braina_mcp.py
+
+# Behavioural evals of the skills
+claude plugin eval plugin
 
 # Run any example script (all use uv inline dependencies)
-uv run examples/frites/conn/plot_covgc.py
-uv run examples/hoi/metrics/plot_oinfo.py
+uv run plugin/examples/frites/conn/plot_covgc.py
+uv run plugin/examples/hoi/metrics/plot_oinfo.py
 ```
 
 All scripts use `uv` with PEP 723 inline script metadata (`# /// script` blocks) for dependency management — no virtualenv setup needed.
 
 ## Architecture
 
-### MCP Server (`mcp/braina_mcp.py`)
+### Repository layout
 
-The central integration layer. A FastMCP server exposing 22 tools that wrap Frites and HOI library functions with standardized file-based I/O. Each tool takes file paths as input (`.npy` or `.nc`), calls the underlying library function, saves results, and returns a text summary (shape, coords, min/mean/max). Exceptions are caught and returned as `Error in <tool>: ...` strings. The server is declared in `.claude-plugin/plugin.json` under `mcpServers`. This is registered as a Claude Code MCP server (`braina`).
+Everything the plugin installs lives under `plugin/` (manifest, MCP server + lockfile, skills, examples, hooks, scripts, evals); the marketplace points at it (`.claude-plugin/marketplace.json`, `source: ./plugin`). Papers, tutorials, use cases, tests and CI stay at the root and are not installed.
+
+### MCP Server (`plugin/mcp/braina_mcp.py`)
+
+The central integration layer. A FastMCP server exposing 34 tools that wrap Frites and HOI library functions with standardized file-based I/O. Each tool takes file paths as input (`.npy` or `.nc`), calls the underlying library function, saves results, and returns a text summary (shape, coords, min/mean/max). Exceptions are caught and returned as `Error in <tool>: ...` strings. The server is declared in `plugin/.claude-plugin/plugin.json` under `mcpServers`; `plugin/hooks/hooks.json` pre-warms its uv environment at session start. This is registered as a Claude Code MCP server (`braina`).
 
 `load_data`/`save_data` are internal I/O helpers (not exposed as MCP tools) that handle the `.npy` vs `.nc` dispatch for every tool.
 
 Tool categories:
-- **Data I/O**: `inspect_data`, `read_pdf`
-- **Frites connectivity**: `frites_conn_covgc`, `frites_conn_dfc`, `frites_conn_pid`, `frites_conn_ii`, `frites_conn_te`, `frites_conn_fit`, `frites_conn_spec`, `frites_conn_ccf`
-- **Frites workflows**: `frites_wf_stats` (WfStats), `frites_wf_mi` (WfMi), `frites_wf_conn_comod` (WfConnComod)
-- **Frites simulation**: `frites_sim_ar` (StimSpecAR)
-- **HOI metrics**: `hoi_oinfo`, `hoi_gradient_oinfo`, `hoi_infotopo`, `hoi_redundancy_mmi`, `hoi_synergy_mmi`, `hoi_rsi`, `hoi_dtc`, `hoi_get_nbest_mult`
+- **Data I/O**: `inspect_data`, `convert_to_nc` (.fif/.mat/.csv/.npy → labelled .nc), `plot_result` (PNG), `read_pdf`
+- **Frites connectivity**: `frites_conn_covgc`, `frites_conn_dfc`, `frites_conn_pid`, `frites_conn_ii`, `frites_conn_te`, `frites_conn_fit`, `frites_conn_spec`, `frites_conn_ccf`; post-processing `frites_conn_reshape` (pairs → sources×targets), `frites_conn_net` (A→B minus B→A)
+- **Frites workflows**: `frites_wf_stats` (WfStats), `frites_wf_mi` (WfMi, multi-subject, `save_workflow`), `frites_wf_mi_combine` (WfMiCombine from pickles), `frites_wf_conn_comod` (WfConnComod)
+- **Frites simulation**: `frites_sim_ar` (StimSpecAR, `random_state`)
+- **HOI static metrics** (all with `method`, `samples_path`, bootstrap `n_boots`): `hoi_oinfo`, `hoi_gradient_oinfo`, `hoi_infotopo`, `hoi_redundancy_mmi`, `hoi_synergy_mmi`, `hoi_rsi`, `hoi_dtc`, `hoi_tc`, `hoi_sinfo`, `hoi_infotot`
+- **HOI dynamic metrics** (3D input, lag `tau`): `hoi_transfer_entropy`, `hoi_dotot`, `hoi_redundancy_phiid`, `hoi_atoms_phiid`
+- **HOI utilities**: `hoi_get_nbest_mult`
+
+Skills: `orientation`, `frites-connectivity`, `hoi-metrics`, `workflows`, `check` — each a short `SKILL.md` plus `references/*.md` read on demand. When adding or changing a tool, update the matching reference file, the pytest suite (`tests/test_mcp_tools.py`, which asserts the tool count) and `CHANGELOG.md`.
 
 ### Core Libraries
 
@@ -60,9 +80,10 @@ Tool categories:
 
 ## Working in this Repo
 
-- Consult `/examples` for usage patterns before writing new analysis code.
-- Read `mcp/braina_mcp.py` for canonical function signatures and parameter conventions.
+- Consult `plugin/examples` for usage patterns before writing new analysis code.
+- Read `plugin/mcp/braina_mcp.py` for canonical function signatures and parameter conventions.
 - For library internals, read source on GitHub via `gh api` or `gh browse` against `brainets/frites` and `brainets/hoi`.
-- When writing analysis scripts, test with small dummy data first (see `mcp/verify_libs.py` for patterns).
+- When writing analysis scripts, test with small dummy data first (see `plugin/mcp/verify_libs.py` for patterns).
 - Prefer JAX for high-performance math in HOI contexts; use xarray/MNE structures for Frites connectivity.
+- Papers are cited in the skills by their GitHub URL under `papers/`; the installed plugin does not ship them.
 - Frites must be `>=0.4.5`; earlier versions crash on import under NumPy 2. HOI has never restricted the NumPy version. With Frites `>=0.4.5`, NumPy 2 is supported.
@@ -22,10 +22,11 @@ at the Institut de Neurosciences de la Timone:
   (O-information, synergy, redundancy, RSI, DTC, InfoTopo), GPU-capable via
   JAX.
 
-It ships as an MCP server (22 tools wrapping Frites and HOI) plus 4 skills
-that know how to pick the right tool, explain what it actually computes, and
-run the statistics correctly — grounded in the real library source, not just
-tool docstrings.
+It ships as an MCP server (34 tools wrapping Frites and HOI, plus data
+conversion and plotting) and 5 skills that know how to pick the right tool,
+explain what it actually computes, run the statistics correctly and check
+the installation — grounded in the real library source, not just tool
+docstrings.
 
 ## Quick start
 
@@ -35,8 +36,11 @@ claude plugin marketplace add brainets/braina
 claude plugin install braina@braina-plugins
 ```
 
-That's it — no need to clone the repo or run from inside it. Then, in any
-Claude Code session:
+That's it — no need to clone the repo or run from inside it. The first
+session pre-warms the Python environment in the background (about 150 MB of
+wheels, once); if the `braina` MCP server shows "Connection closed" in that
+very first session, run `/braina:check` or simply start a new session. Then,
+in any Claude Code session:
 
 > **You:** I have LFP recordings from two regions during a working-memory
 > task (`data.nc`). Does region A drive region B?
@@ -61,26 +65,31 @@ Claude Code session:
 | `frites-connectivity` | Explains each Frites connectivity tool (Granger causality, transfer entropy, PID, coherence, cross-correlation, ...) once one's been picked. |
 | `hoi-metrics` | Explains HOI's higher-order metrics (O-info, synergy, redundancy, DTC, ...) — including sign-convention gotchas and CPU/GPU notes. |
 | `workflows` | Explains permutation testing, cluster correction, bootstrapping, and fixed/random-effect inference, for Frites *and* HOI output. |
+| `check` | `/braina:check` — verifies the installation: versions, JAX backend (CPU/GPU), MCP connection, end-to-end smoke test. |
+
+Each skill is a short `SKILL.md` plus `references/` files that Claude reads
+on demand, so only the relevant detail enters the context.
 
 ## MCP tools
 
 <details>
-<summary>22 tools wrapping Frites and HOI (click to expand)</summary>
+<summary>34 tools wrapping Frites and HOI (click to expand)</summary>
 
 | Category | Tools |
 |---|---|
-| Data I/O | `inspect_data`, `read_pdf` |
-| Frites connectivity | `frites_conn_covgc`, `frites_conn_dfc`, `frites_conn_pid`, `frites_conn_ii`, `frites_conn_te`, `frites_conn_fit`, `frites_conn_spec`, `frites_conn_ccf` |
-| Frites workflows | `frites_wf_mi`, `frites_wf_stats`, `frites_wf_conn_comod` |
+| Data I/O | `inspect_data`, `convert_to_nc` (MNE `.fif`, `.mat`, `.csv`, `.npy` → labelled `.nc`), `plot_result` (PNG), `read_pdf` |
+| Frites connectivity | `frites_conn_covgc`, `frites_conn_dfc`, `frites_conn_pid`, `frites_conn_ii`, `frites_conn_te`, `frites_conn_fit`, `frites_conn_spec`, `frites_conn_ccf`, `frites_conn_reshape`, `frites_conn_net` |
+| Frites workflows | `frites_wf_mi` (multi-subject), `frites_wf_mi_combine`, `frites_wf_stats`, `frites_wf_conn_comod` |
 | Frites simulation | `frites_sim_ar` |
-| HOI metrics | `hoi_oinfo`, `hoi_gradient_oinfo`, `hoi_infotopo`, `hoi_redundancy_mmi`, `hoi_synergy_mmi`, `hoi_rsi`, `hoi_dtc`, `hoi_get_nbest_mult` |
+| HOI metrics | `hoi_oinfo`, `hoi_gradient_oinfo`, `hoi_infotopo`, `hoi_redundancy_mmi`, `hoi_synergy_mmi`, `hoi_rsi`, `hoi_dtc`, `hoi_tc`, `hoi_sinfo`, `hoi_infotot`, `hoi_transfer_entropy`, `hoi_dotot`, `hoi_redundancy_phiid`, `hoi_atoms_phiid`, `hoi_get_nbest_mult` |
 
 Each tool wraps a Frites or HOI function with file-based I/O (`.npy` or
 `.nc`) and returns a summary of the result (shape, coordinates, min/mean/max).
 Use `.nc` input with `roi` and `times` coordinates so that outputs keep ROI
 names and a time axis in seconds; HOI outputs saved as `.nc` keep the
-multiplet metadata that `hoi_get_nbest_mult` needs. See `mcp/braina_mcp.py`
-for exact signatures.
+multiplet metadata that `hoi_get_nbest_mult` needs. Every HOI metric accepts
+`method` (estimator) and bootstrap confidence intervals (`n_boots`). See
+`plugin/mcp/braina_mcp.py` for exact signatures.
 
 </details>
 
@@ -89,8 +98,9 @@ for exact signatures.
 Beyond the plugin itself, this repo also carries the reference material the
 skills point to:
 
-- **`examples/`** — ~50 self-contained scripts (Frites + HOI), runnable with
-  `uv run examples/frites/conn/plot_covgc.py`.
+- **`plugin/examples/`** — ~50 self-contained scripts (Frites + HOI), runnable with
+  `uv run plugin/examples/frites/conn/plot_covgc.py` (shipped with the plugin,
+  the skills point to them).
 - **`tutorials/`** — longer walkthroughs, including a full SEEG analysis
   pipeline and a Frites+HOI+XGI integration notebook.
 - **`usecases/`** — end-to-end analysis scenarios (AR simulation, dynamic FC,
@@ -104,20 +114,29 @@ skills point to:
 git clone https://github.com/brainets/braina.git
 cd braina
 
-# Verify the environment
+# Verify the environment and run the test suite
 uv run check_env.py
-uv run mcp/verify_libs.py
+uv run plugin/mcp/verify_libs.py
+uv run tests/run_tests.py            # pytest over all 34 MCP tools
+uv run plugin/scripts/smoke_test.py  # what /braina:check runs
 
 # Register the MCP server (one-time setup)
-claude mcp add braina -- uv run mcp/braina_mcp.py
+claude mcp add braina -- uv run plugin/mcp/braina_mcp.py
 
 claude
 ```
 
-The MCP server is declared in `.claude-plugin/plugin.json` (`mcpServers`,
-using `${CLAUDE_PLUGIN_ROOT}`), which only applies to a plugin install — a
-plain clone does not auto-register the server, hence the manual
-`claude mcp add` step above.
+`plugin/mcp/braina_mcp.py.lock` pins the server's dependencies (`uv lock
+--script plugin/mcp/braina_mcp.py` to refresh after changing them).
+Behavioural checks of the skills live in `plugin/evals/` and run with
+`claude plugin eval plugin`. CI (`.github/workflows/ci.yml`) runs the
+environment check, the library verification and the test suite.
+
+The MCP server is declared in `plugin/.claude-plugin/plugin.json`
+(`mcpServers`, using `${CLAUDE_PLUGIN_ROOT}`), which only applies to a plugin
+install — a plain clone does not auto-register the server, hence the manual
+`claude mcp add` step above. The marketplace installs only the `plugin/`
+directory (~2 MB); papers, tutorials and use cases stay in the repo.
 
 Reads `CLAUDE.md` for project context.
 
@@ -128,19 +147,25 @@ Reads `CLAUDE.md` for project context.
 
 ```
 braina/
-├── .claude-plugin/
-│   ├── plugin.json         # Plugin manifest (incl. MCP server declaration)
-│   └── marketplace.json    # Self-hosted marketplace
-├── skills/                  # orientation, frites-connectivity, hoi-metrics, workflows
-├── mcp/
-│   ├── braina_mcp.py        # MCP server — 22 tools wrapping Frites & HOI
-│   └── verify_libs.py       # Test suite for all wrapped functions
-├── examples/                # ~50 example scripts (frites/, hoi/)
-├── tutorials/
-├── usecases/
-├── papers/
-├── CLAUDE.md                 # Project context for Claude Code
-└── check_env.py              # Environment verification
+├── .claude-plugin/marketplace.json   # Self-hosted marketplace → source: ./plugin
+├── plugin/                           # Everything the plugin installs
+│   ├── .claude-plugin/plugin.json    # Manifest (incl. MCP server declaration)
+│   ├── mcp/
+│   │   ├── braina_mcp.py             # MCP server — 34 tools wrapping Frites & HOI
+│   │   ├── braina_mcp.py.lock        # uv lockfile for the server script
+│   │   └── verify_libs.py            # Library-level verification
+│   ├── skills/                       # orientation, frites-connectivity, hoi-metrics, workflows, check
+│   │   └── <skill>/SKILL.md + references/
+│   ├── examples/                     # ~50 example scripts (frites/, hoi/)
+│   ├── hooks/hooks.json              # SessionStart: pre-warm the uv environment
+│   ├── scripts/                      # warmup.sh, smoke_test.py
+│   └── evals/                        # `claude plugin eval` cases
+├── tests/                            # pytest suite (uv run tests/run_tests.py)
+├── .github/workflows/ci.yml
+├── tutorials/  usecases/  papers/  docs/
+├── CHANGELOG.md
+├── CLAUDE.md                         # Project context for Claude Code
+└── check_env.py                      # Environment verification
 ```
 
 </details>
 
@@ -1,6 +1,11 @@
 # /// script
-# dependencies = ["frites", "hoi", "xgi", "numpy", "xarray", "jax", "jaxlib", "mne"]
+# dependencies = ["frites>=0.4.5", "hoi", "xgi", "numpy", "xarray", "jax", "jaxlib", "mne"]
 # ///
+"""Quick environment check: imports, versions and the JAX backend.
+
+For an end-to-end check of the MCP tools, run
+`uv run plugin/scripts/smoke_test.py` (or the /braina:check skill).
+"""
 import frites
 import hoi
 import xgi
@@ -11,4 +16,8 @@
 print(f"frites version: {frites.__version__}")
 print(f"hoi version: {hoi.__version__}")
 print(f"xgi version: {xgi.__version__}")
+print(f"numpy version: {np.__version__}")
+print(f"xarray version: {xr.__version__}")
+print(f"mne version: {mne.__version__}")
+print(f"jax version: {jax.__version__} | devices: {jax.devices()}")
 print("All packages imported successfully!")
Original file line number	Diff line number	Diff line change
`@@ -1,12 +1,12 @@`
`1`	`1`	`{`
`2`	`2`	`"name": "braina-plugins",`
`3`	`3`	`"owner": {`
`4`		`- "name": "Andrea Brovelli — BraiNets, Institut de Neurosciences de la Timone"`
	`4`	`+ "name": "Andrea Brovelli \u2014 BraiNets, Institut de Neurosciences de la Timone"`
`5`	`5`	`},`
`6`	`6`	`"plugins": [`
`7`	`7`	`{`
`8`	`8`	`"name": "braina",`
`9`		`- "source": "."`
	`9`	`+ "source": "./plugin"`
`10`	`10`	`}`
`11`	`11`	`]`
`12`	`12`	`}`