diff --git a/.cursor-plugin/marketplace.json b/.cursor-plugin/marketplace.json
index d7d2a96dd..99d6336b6 100644
--- a/.cursor-plugin/marketplace.json
+++ b/.cursor-plugin/marketplace.json
@@ -687,6 +687,11 @@
         "grokbot": "0.49.0",
         "sand": "0.49.0"
       }
+    },
+    {
+      "name": "tensorviz",
+      "source": "third_party/tensorviz",
+      "description": "Explore local PyTorch model graphs and measured tensor shapes (Apple Silicon preview)."
     }
   ]
 }
diff --git a/README.md b/README.md
index d2ac28d9d..3e7c0cb5f 100644
--- a/README.md
+++ b/README.md
@@ -125,6 +125,7 @@ Official Cursor plugins for popular developer tools, frameworks, and SaaS produc
 | `rippling` | [Rippling](third_party/rippling/) | Cursor | Integrations | Look up people, time off, and company policy in Rippling, and file time-off requests and draft hires. |
 | `onepay` | [OnePay](third_party/onepay/) | Cursor | Integrations | Access your OnePay accounts: balances, transactions, credit, Pay Later, investments, crypto, and rewards. |
 | `questrade` | [Questrade](third_party/questrade/) | Cursor | Integrations | Read your Questrade portfolio, get live market data, and discuss potential trades. |
+| `tensorviz` | [tensorViz](third_party/tensorviz/) | TensorViz | Developer Tools | Explore local PyTorch model graphs and measured tensor shapes (Apple Silicon preview). |
 Author values match each plugin’s `plugin.json` `author.name` (Cursor lists `plugins@cursor.com` in the manifest).
 
 ## Repository structure
diff --git a/third_party/tensorviz/.cursor-plugin/plugin.json b/third_party/tensorviz/.cursor-plugin/plugin.json
new file mode 100644
index 000000000..00b252f77
--- /dev/null
+++ b/third_party/tensorviz/.cursor-plugin/plugin.json
@@ -0,0 +1,31 @@
+{
+  "name": "tensorviz",
+  "displayName": "tensorViz",
+  "version": "0.3.61",
+  "description": "Explore local PyTorch model graphs and measured tensor shapes (Apple Silicon preview).",
+  "author": {
+    "name": "TensorViz",
+    "email": "tensorviz.app@gmail.com"
+  },
+  "homepage": "https://tensorviz.dev",
+  "repository": "https://github.com/tensorviz-org/tensorviz-plugins",
+  "license": "MIT",
+  "logo": "assets/logo.png",
+  "keywords": [
+    "pytorch",
+    "tensorviz",
+    "model-graph",
+    "tensor-shapes",
+    "machine-learning",
+    "mcp"
+  ],
+  "category": "developer-tools",
+  "tags": [
+    "pytorch",
+    "visualization",
+    "debugging",
+    "mcp"
+  ],
+  "skills": "./skills/",
+  "mcpServers": "./mcp.json"
+}
diff --git a/third_party/tensorviz/CHANGELOG.md b/third_party/tensorviz/CHANGELOG.md
new file mode 100644
index 000000000..f4b30a943
--- /dev/null
+++ b/third_party/tensorviz/CHANGELOG.md
@@ -0,0 +1,16 @@
+# Changelog
+
+## 0.3.61
+
+- Update the submission to the compiled runtime built from current source main
+  `d4e2865a3c43149907dd06840745eb037d55dffa`.
+- Include the current 27-tool host, with `find_computation`, `query_computation`,
+  and `analyze_computation_change`, and pin its matching runtime checksum.
+
+## 0.3.60
+
+- Submit the existing public tensorViz 0.3.60 integration to the Cursor catalog.
+- Include the local MCP launcher, pinned compiled-runtime lock, workflow skill,
+  logo, license, privacy, terms, support, and Cursor installation guide.
+- Target the macOS Apple Silicon / CPython 3.12 CPU preview; marketplace approval
+  and Cursor client acceptance remain separate.
diff --git a/third_party/tensorviz/LICENSE b/third_party/tensorviz/LICENSE
new file mode 100644
index 000000000..63a494992
--- /dev/null
+++ b/third_party/tensorviz/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 TensorViz contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/third_party/tensorviz/PRIVACY.md b/third_party/tensorviz/PRIVACY.md
new file mode 100644
index 000000000..c2ba31b90
--- /dev/null
+++ b/third_party/tensorviz/PRIVACY.md
@@ -0,0 +1,28 @@
+# Privacy and data handling
+
+TensorViz's agent plugin runs a local MCP host and Python workers. Discovery reads
+model source. Capturing executes the selected model with the selected environment
+and example inputs. Approved setup can write recipes and download Python runtimes
+or packages from their providers. Imported model code and dependencies can have
+their own filesystem and network behavior.
+
+Graph evidence, source excerpts, paths and errors returned through MCP are visible
+to the coding client and can enter its model-provider conversation. The client's
+privacy, retention and organization policies apply. Do not provide sensitive models
+or data unless those policies permit it.
+
+The plugin does not operate a TensorViz cloud ingestion service, and its browser
+and panel builds exclude TensorViz's hosted authentication and analytics modules.
+The browser viewer uses a local connection. Keep private viewer links private.
+The tldraw renderer has separate license-validation behavior; trial licenses can
+send license-usage information to tldraw. See
+[tldraw's licensing and data collection documentation](https://tldraw.dev/community/license).
+
+Model recipes and captures, local session resources, runtime installations and
+permission records can remain on disk. Removing the plugin does not delete model
+source, recipes or Python environments. Run the installed plugin's
+`scripts/launch.sh --revoke-permission` to revoke TensorViz permission, or disable
+the plugin in your coding client. See [installation and removal](docs/installation.md).
+
+Do not post private model code, credentials, full environment dumps or private
+viewer URLs in public issues. See [SUPPORT.md](SUPPORT.md) for reporting problems.
diff --git a/third_party/tensorviz/README.md b/third_party/tensorviz/README.md
new file mode 100644
index 000000000..c0c631b87
--- /dev/null
+++ b/third_party/tensorviz/README.md
@@ -0,0 +1,63 @@
+# tensorViz
+
+Explore local PyTorch model graphs and measured tensor shapes (Apple Silicon preview).
+
+The plugin connects Cursor to a local tensorViz MCP host and supplies the
+`model-workflow` skill. Capture a PyTorch `nn.Module`, inspect the measured input
+and output shapes, and refresh the graph after a saved source edit. Use the local
+browser graph when the client does not support the embedded Model graph panel.
+
+## Requirements
+
+- macOS on Apple Silicon, with CPU capture.
+- CPython 3.12 with PyTorch and your model's dependencies, or supported managed
+  Python setup after permission is granted.
+- Network access for the first runtime download and any approved dependency setup.
+- Cursor with plugins enabled. Account, organization, and tool permissions apply.
+
+This compiled preview does not support Windows, Linux, Intel Macs, or other Python
+ABIs. CUDA is unverified. VS Code, a separate Node.js installation, a tensorViz
+account, and a provider API key are not required by tensorViz.
+
+## Install and use
+
+Once approved and listed, open **Customize**, find **tensorViz**, and select
+**Install**. Until then, use the [local evaluation instructions](docs/installation.md).
+Start a new Agent chat in your model project and ask:
+
+> Use tensorViz to open my PyTorch model, capture its saved example input, and
+> explain the tensor shapes. Show the graph, then refresh it after my saved edit.
+
+Provide the model file, class, constructor arguments, and example inputs when
+needed. Review tensorViz's first-use permission prompt: capture executes your
+local model code, and approved setup can install Python and dependencies.
+Select a compatible project environment when requested. Saved `tensorviz.toml`
+recipes can be reused; model-specific packages and assets may still need setup.
+
+A successful capture contains measured shapes for the selected input and
+execution. A source preview alone does not establish a successful capture.
+Graph display and capture are separate results.
+
+## Runtime and data
+
+The readable launcher downloads the public **0.3.61** compiled runtime from GitHub
+Releases, verifies the SHA-256 digest pinned in `runtime-lock.json`, and caches the
+archive for later connections. It checks the archive on every connection and
+extracts separate session files. A checksum mismatch stops startup. This directory
+contains the integration; the runtime is downloaded separately.
+
+Graph evidence, source excerpts, and paths returned through MCP enter Cursor and
+may enter its model-provider conversation. Recipes, captures, environments,
+runtime caches, and permission records stay on the local machine until removed.
+Approved Python setup contacts package providers; imported model code can use its
+own services. The tldraw renderer's production trial can send license-usage
+information to tldraw, excluding canvas content. See [privacy](PRIVACY.md).
+
+The bundled renderer trial expires **26 December 2026**. Continued renderer use
+requires a renewed license and an updated runtime; existing downloads retain
+their original embedded key. See [terms](TERMS.md).
+
+See [installation, updates, and removal](docs/installation.md),
+[support](SUPPORT.md), [the public integration repository](https://github.com/tensorviz-org/tensorviz-plugins),
+and [tensorviz.dev](https://tensorviz.dev). The integration is [MIT licensed](LICENSE);
+runtime components retain their applicable terms and notices.
diff --git a/third_party/tensorviz/SUPPORT.md b/third_party/tensorviz/SUPPORT.md
new file mode 100644
index 000000000..3d0669ee0
--- /dev/null
+++ b/third_party/tensorviz/SUPPORT.md
@@ -0,0 +1,17 @@
+# Support and compatibility
+
+This compiled candidate targets macOS Apple Silicon and CPython 3.12 CPU capture.
+It does not require VS Code or Node.js. Explicitly selected incompatible Python
+versions produce an error; TensorViz does not silently replace that environment.
+Windows, Linux, Intel Macs, CUDA and other Python ABIs remain unverified targets.
+
+The integration works with clients that support its local MCP transport and
+plugin format. Directory publication is distinct from local installation, capture
+and rendering. No official marketplace acceptance is claimed for this candidate.
+
+When reporting a problem, include the integration version, OS/client version,
+whether it failed during download, startup or capture, and a small synthetic
+model. Do not share credentials, private viewer links, model secrets or unredacted
+environment dumps. Use
+[the integration repository's issues](https://github.com/tensorviz-org/tensorviz-plugins/issues)
+for public, non-sensitive bug reports.
diff --git a/third_party/tensorviz/TERMS.md b/third_party/tensorviz/TERMS.md
new file mode 100644
index 000000000..ff2ab6b26
--- /dev/null
+++ b/third_party/tensorviz/TERMS.md
@@ -0,0 +1,22 @@
+# Terms and third-party components
+
+The integration files in this repository carry the accompanying MIT license.
+The core runtime is distributed separately as compiled artifacts; this repository
+does not publish its source or change the licensing of existing core code.
+Runtime distribution terms and third-party notices must accompany its release.
+
+Public runtime releases require the owner's confirmation that the tldraw license
+covers the bundled renderer supplied to end users. Private build candidates are
+not permission to redistribute the renderer. The integration's MIT license does not
+grant a tldraw production or redistribution license. Dependencies retain their
+original licenses and the binary includes their applicable notices.
+
+The initial preview uses a production trial of tldraw that expires on
+26 December 2026. Continued renderer use after the trial requires a renewed license
+and an updated TensorViz runtime. Downloaded versions retain their embedded key;
+renewal does not retroactively update old installations.
+
+Your coding client, model provider, downloaded packages and model repositories
+have their own terms. Capturing runs selected model code locally. A graph records
+one selected execution or an explicitly identified source preview; it does not
+guarantee model correctness or behavior for other inputs.
diff --git a/third_party/tensorviz/assets/logo.png b/third_party/tensorviz/assets/logo.png
new file mode 100644
index 000000000..14de08be7
Binary files /dev/null and b/third_party/tensorviz/assets/logo.png differ
diff --git a/third_party/tensorviz/docs/installation.md b/third_party/tensorviz/docs/installation.md
new file mode 100644
index 000000000..b06af7f58
--- /dev/null
+++ b/third_party/tensorviz/docs/installation.md
@@ -0,0 +1,82 @@
+# Cursor installation and maintenance
+
+This entry packages the public
+[tensorViz 0.3.61 preview](https://github.com/tensorviz-org/tensorviz-plugins/releases/tag/v0.3.61)
+for Cursor. It requires macOS Apple Silicon and CPython 3.12 CPU capture. Local
+installation, package validation, runtime checks, and marketplace approval are
+separate results. This submission does not claim completed Cursor client
+acceptance or an approved marketplace listing.
+
+## Local evaluation
+
+1. Copy this complete `third_party/tensorviz` directory to
+   `~/.cursor/plugins/local/tensorviz`. The destination must directly contain
+   `.cursor-plugin/plugin.json`, `mcp.json`, `runtime-lock.json`, `scripts/`,
+   `skills/`, and `assets/`. Copy regular files rather than a symlink.
+2. Restart Cursor or run **Developer: Reload Window**.
+3. Open **Customize** and confirm the `model-workflow` skill and tensorViz MCP
+   server. Your account or organization must allow local plugin imports. A
+   marketplace installation of the same name takes precedence over a local copy.
+4. Start a new Agent chat in your model project and use the prompt in the
+   [README](../README.md). Complete the permission and environment prompts.
+
+The MCP configuration uses `${CURSOR_PLUGIN_ROOT}/scripts/launch.sh`. Cursor
+expands that variable to this plugin's installed directory. No Node.js command
+or global Python executable is used to start the compiled host.
+
+## Updates, removal, and troubleshooting
+
+For marketplace installations, use Cursor's plugin management to update, disable,
+or uninstall tensorViz. Start a new chat after an update; the runtime remains
+pinned to this entry's version until its lock and integration are updated together.
+
+For a local evaluation update, close tensorViz chats, move the old complete plugin
+directory outside `~/.cursor/plugins/local`, and copy the new complete directory
+into place. Reload Cursor and start a new chat. Restore the old complete directory
+for rollback; do not combine files from different versions.
+
+If download fails, check access to the URL in `runtime-lock.json` and reconnect.
+The verified archive is cached under `~/Library/Caches/dev.tensorviz/runtime`.
+A checksum error names the damaged archive: remove only that archive and
+reconnect to download it again. Do not bypass checksum verification.
+
+To remove a local evaluation, close its chats, remove only
+`~/.cursor/plugins/local/tensorviz`, and reload Cursor. Removal preserves model
+source, recipes, Python environments, cached runtimes, and saved permission.
+Before removal, run the installed plugin's `scripts/launch.sh --revoke-permission`
+if you want to revoke tensorViz permission. See [privacy](../PRIVACY.md).
+
+## Provenance and review scope
+
+The launcher, runtime lock, workflow skill, logo, support, and terms were copied
+from public tag `v0.3.61`, commit
+`d01630d9e9d6e453b1ef6f7786b7dd77b2601bed` of
+[`tensorviz-org/tensorviz-plugins`](https://github.com/tensorviz-org/tensorviz-plugins/tree/v0.3.61/cursor/plugins/tensorviz).
+Their hashes were checked against that release's `PUBLICATION.json`. The MIT
+license is copied from the distribution repository. The privacy document changes
+only its installation link to this guide.
+
+The Cursor manifest uses this marketplace's supported author fields, adds display
+and discovery metadata, and keeps the release version and component paths. This
+entry also adds the catalog registration, README, changelog, and this guide. It
+does not include the other clients' manifests or a compiled runtime archive.
+
+Local submission checks on 3 November 2026 passed the repository's
+`node scripts/validate-plugins.mjs` validator, launcher shell syntax, catalog and
+manifest consistency, documentation links, and copied-file hash checks. An
+isolated macOS Apple Silicon profile launched the configured MCP command from a
+path containing spaces, downloaded the runtime without GitHub credentials, and
+verified its 35,773,473-byte archive against the pinned SHA-256 digest. MCP
+initialization reported tensorViz 0.3.61 and discovery returned 27 tools, including
+`open_graph`, `open_graph_view`, `find_computation`, `query_computation`, and
+`analyze_computation_change`. All three computation queries returned `NO_MODEL`
+for an unopened model without creating user state. No model ran and no tensorViz
+permission was granted. These checks did not run the Cursor application.
+
+For client acceptance, verify skill/tool discovery, a small CPU model capture,
+measured shapes, visible graph display, refresh after an edit, and an update with
+retained setup. A manifest check or a direct MCP connection alone cannot establish
+those Cursor behaviors.
+
+References: [Cursor plugin format and MCP paths](https://cursor.com/docs/reference/plugins),
+[local plugin imports](https://cursor.com/docs/plugins#test-plugins-locally).
diff --git a/third_party/tensorviz/mcp.json b/third_party/tensorviz/mcp.json
new file mode 100644
index 000000000..f25140e4f
--- /dev/null
+++ b/third_party/tensorviz/mcp.json
@@ -0,0 +1,12 @@
+{
+  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
+  "mcpServers": {
+    "tensorviz": {
+      "type": "stdio",
+      "command": "sh",
+      "args": [
+        "${CURSOR_PLUGIN_ROOT}/scripts/launch.sh"
+      ]
+    }
+  }
+}
diff --git a/third_party/tensorviz/runtime-lock.json b/third_party/tensorviz/runtime-lock.json
new file mode 100644
index 000000000..2637e54d3
--- /dev/null
+++ b/third_party/tensorviz/runtime-lock.json
@@ -0,0 +1,9 @@
+{
+  "formatVersion": 1,
+  "version": "0.3.61",
+  "platform": "darwin-arm64",
+  "pythonABI": "cpython-312-darwin",
+  "sha256": "c6d4a606f45d880ed2967a16059a9130cfc46111e913d91dc808c7ac32341430",
+  "bytes": 35773473,
+  "url": "https://github.com/tensorviz-org/tensorviz-plugins/releases/download/v0.3.61/tensorviz-runtime-0.3.61-darwin-arm64.tar.gz"
+}
diff --git a/third_party/tensorviz/scripts/launch.sh b/third_party/tensorviz/scripts/launch.sh
new file mode 100755
index 000000000..208871b5a
--- /dev/null
+++ b/third_party/tensorviz/scripts/launch.sh
@@ -0,0 +1,60 @@
+#!/bin/sh
+# Public integration only. Core code arrives as a version-pinned binary archive.
+set -eu
+umask 077
+fail() { echo "TensorViz: $*" >&2; exit 1; }
+plugin=$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd -P)
+if [ "$(uname -s)-$(uname -m)" != Darwin-arm64 ]; then
+  fail 'This preview supports macOS Apple Silicon. No runtime was downloaded.'
+fi
+lock="$plugin/runtime-lock.json"
+version=$(/usr/bin/plutil -extract version raw -o - "$lock")
+sha=$(/usr/bin/plutil -extract sha256 raw -o - "$lock")
+url=$(/usr/bin/plutil -extract url raw -o - "$lock")
+bytes=$(/usr/bin/plutil -extract bytes raw -o - "$lock")
+case "$sha" in *[!a-f0-9]*|'') fail 'Invalid runtime digest.';; esac
+[ "${#sha}" -eq 64 ] || fail 'Invalid runtime digest.'
+case "$bytes" in *[!0-9]*|'') fail 'Invalid runtime size.';; esac
+case "$url" in https://*) ;; *) fail 'Runtime downloads require HTTPS.';; esac
+cache="${HOME:?}/Library/Caches/dev.tensorviz/runtime"
+mkdir -p "$cache"
+[ ! -L "$cache" ] || fail 'The runtime cache must not be a symbolic link.'
+archive="$cache/$sha.tar.gz"
+download=''
+session=''
+child=''
+cleanup() {
+  if [ -n "$child" ]; then kill "$child" 2>/dev/null || true; wait "$child" 2>/dev/null || true; fi
+  [ -z "$download" ] || rm -f -- "$download"
+  [ -z "$session" ] || rm -rf -- "$session"
+}
+trap cleanup 0
+trap 'exit 130' INT
+trap 'exit 143' TERM HUP
+if [ ! -f "$archive" ]; then
+  download=$(mktemp "$cache/.download.XXXXXX")
+  echo "TensorViz: downloading runtime $version for reuse by later connections." >&2
+  /usr/bin/curl --fail --silent --show-error --location --proto '=https' --proto-redir '=https' \
+    --connect-timeout 15 --max-time 300 --max-filesize "$bytes" --output "$download" "$url" < /dev/null &
+  child=$!
+  wait "$child" || fail 'Runtime download failed. Retry the plugin connection.'
+  child=''
+  [ "$(/usr/bin/shasum -a 256 "$download" | cut -d ' ' -f 1)" = "$sha" ] \
+    || fail 'Runtime checksum mismatch. Nothing was executed.'
+  mv -f -- "$download" "$archive"
+  download=''
+fi
+[ ! -L "$archive" ] || fail 'The cached archive must not be a symbolic link.'
+[ "$(/usr/bin/shasum -a 256 "$archive" | cut -d ' ' -f 1)" = "$sha" ] \
+  || fail "Cached runtime is damaged. Remove $archive and reconnect. Nothing was executed."
+# Extract anew from the verified archive, so modified unpacked caches cannot run.
+session=$(mktemp -d "${TMPDIR:-/tmp}/tensorviz-native.XXXXXX")
+/usr/bin/tar -xzf "$archive" -C "$session"
+[ -x "$session/tensorviz-host" ] || fail 'Runtime archive has no executable host.'
+"$session/tensorviz-host" "$@" <&0 &
+child=$!
+set +e
+wait "$child"
+status=$?
+child=''
+exit "$status"
diff --git a/third_party/tensorviz/skills/model-workflow/SKILL.md b/third_party/tensorviz/skills/model-workflow/SKILL.md
new file mode 100644
index 000000000..5d53dc1f3
--- /dev/null
+++ b/third_party/tensorviz/skills/model-workflow/SKILL.md
@@ -0,0 +1,68 @@
+---
+name: model-workflow
+description: Create, understand, modify or debug local PyTorch nn.Module models with a current TensorViz graph. Use for local model architecture, forward-pass tensor flow, shape errors and activation or layer changes. Exclude unrelated Python work, conceptual questions without local code and requests to avoid graphing or execution.
+---
+
+# TensorViz model workflow
+
+Read the user's request first. Do not run anything on startup. Explicit opt-outs,
+read-only instructions and the client's native execution/edit approvals take
+priority. Work only on projects relevant to the user's request.
+
+Discover the advertised TensorViz tools and host capabilities once. Use the exact
+workspace checkout, model file and class from the current task. Do not infer a
+workspace from the launcher's directory or silently use another worktree.
+
+Use `open_graph` when advertised. It waits briefly for completion and yields early
+for permission or setup. Poll the same requestId only while it remains pending;
+never start duplicate captures or sleep repeatedly in the agent. Continue its returned next action until a
+current capture succeeds, a model failure has evidence, the user cancels, or a
+concrete blocker requires their input. Inspect repository evidence for missing
+setup; ask about unresolved domain choices or assets. A source preview, connected
+server or tool call is not a successful capture. Do not change inputs to hide a
+model failure or retry an unchanged failed action.
+
+Use the environment reported by tensorViz. The tensorViz plugin remembers the
+saved/project environment and offers available previously used interpreters when
+there is no project choice. Start without `managedRuntime`; select an offered
+interpreter when the user or project identifies it. Do not search arbitrary
+directories or guess shell Python. With no available choices on a supported
+platform, the host prepares a managed environment under its existing permission.
+An explicit `managedRuntime: true` requests a separate environment.
+The tensorViz plugin remembers the
+user's plugin permission from installation or first use. Once granted, continue
+inspection, setup, dependency repair and capture without asking for additional
+TensorViz approvals. Honor a declined or revoked permission. Never approve a
+permission prompt, invoke the permission helper, edit permission records, change
+client permission settings or send an invented approval flag on the user's behalf.
+If the user explicitly asks to enable TensorViz after a declined or cancelled
+decision, use the stopped opening's `request_permission` recovery action. It
+shows the user's confirmation and resumes the same opening after acceptance;
+it does not grant permission itself. Never repeat it automatically after a decline.
+If the project defines example inputs in a Python function such as `graph_inputs`,
+reference that exact function with `inputs: { factory: { file, callable, kwargs } }`
+in the setup draft. Read its source first; do not guess a function or manually
+serialize its tensors. It must return named forward inputs or an explicit
+`{args, kwargs}` binding. TensorViz executes it once in the capture worker.
+Keep saved inputs
+and recapture after a coherent source change, not after every file write.
+
+In clients supporting MCP Apps, use `open_graph_panel` when advertised to show
+the native Model graph panel beside this chat. Open it once; it follows the current
+capture. Check `list_models` for a connected current view before claiming display,
+and use that viewId for `get_selection` and `focus_view`. Do not call the panel's
+app-only transport tools. The panel opening alone does not prove rendering.
+
+For clients without MCP Apps, use `open_graph_view`: open the returned private
+URL once in a browser and poll its handoff for the exact display receipt. Display
+failure is separate from capture failure. Keep private links out of logs and
+external messages.
+
+For plugin hosts, ordinary source edits use the coding agent's normal review
+workflow. Use TensorViz source proposals only when the host advertises and supports
+the complete review path. Never bypass a pending proposal. Use graph comparison
+only when advertised; focusing a changed node does not show a diff.
+
+Finish with the graph result, the model result and any evidenced limitation. If
+the installed host lacks a required capability, name it accurately rather than
+claiming completion or routing to a different host.