- Add an `npm run package` script that builds the webview + extension bundle
  and writes a shareable `.vsix` into `vsix/` (via @vscode/vsce, packaged with
  --no-dependencies since esbuild already bundles runtime deps).
- Commit the packaged preview build so testers/customers can side-load it
  before the Marketplace launch; un-ignore `vsix/*.vsix` in .gitignore.
- Tighten .vscodeignore so the vsix only ships dist/ + media/ (excludes
  src, webview-ui, maps, configs) — 1.3 MB, 7 files.
- README: add an "Install (Preview Build)" section (download link + UI/CLI
  install steps + how to share it for preview access), a "Build from Source"
  section documenting `npm run package`, and clarify launch for installed
  vs. source users.

Note: the preview vsix targets the Bedrock/Copilot providers; the on-device
local-model option isn't bundled.

…e vsix via releases

Local model support that works from the packaged .vsix (Option B):
- New `local-server` provider talks to an OpenAI-compatible /chat/completions
  endpoint you run yourself (Ollama, LM Studio, llama.cpp) — no native module,
  so a single portable vsix works on every platform. Configured via
  workflowInsight.localServerUrl (default http://localhost:11434/v1) and
  workflowInsight.localServerModel (default llama3.1); wired through
  generateQuery, verifyResult, and completeText (analyze + chart spec) via a
  setLocalServer() module setting, mirroring setLocalModel.
- Settings UI: provider option + Server URL / Model fields with Ollama guidance.
- The embedded on-device `local` provider is unchanged (source-only; not in the
  vsix since node-llama-cpp is a native module).

Distribution:
- Stop committing the .vsix; ignore vsix/ again. Preview builds are shared via
  GitHub Releases instead. README updated: download from Releases, and a
  `gh release create ... vsix/*.vsix` note for publishing.
- README: generalized the provider setup step (Bedrock / Copilot / Local
  server / on-device) and added the new settings to the reference table.

Typecheck (ext + webview) + build + package clean, 123 tests pass.

The "Favorites" and "Send · <mode>" text buttons took too much width. Favorites
is now an icon-only star dropdown (variant="icon"); Send is an icon-only primary
button (send icon) with the mode in its aria-label. The active mode is marked
with a check in the mode dropdown, and the mode-aware placeholder still signals
which mode is active.

…h, agent=gen-ai)

Adds an AI-usage disclosure + consent gate. Before the first use of any AI
feature — the Ask and Agent modes and the Visualize page — the user is shown a
notice explaining that an LLM is used and that their request text and, in some
cases, limited portions of their data (result column names and a small sample
of rows) are sent to the configured provider. The user must tick "I have read
and agree" and click "I agree" to proceed; the deferred action then runs.

- Consent is persisted in settings (workflowInsight.aiDisclosureAcceptedVersion)
  and versioned (AI_DISCLOSURE_VERSION) so a wording change re-prompts.
- Query mode is unaffected (it never calls an LLM) and stays usable without
  consent.
- Gating is enforced in the webview before any generate/visualize message is
  posted; the host persists the decision (onSetConsent).

LEGAL: the disclosure wording in AiConsentModal is a placeholder pending Legal
review; bump AI_DISCLOSURE_VERSION when the approved text lands to re-prompt.

Typecheck (ext + webview) + build clean, 123 tests pass.

…otice

Expands the disclosure to spell out where data goes per configured provider —
Amazon Bedrock (to AWS), GitHub Copilot (to GitHub/Microsoft), local server
(to a self-hosted endpoint you control), and on-device (never leaves the
machine) — so consent is informed for each option. Bumped AI_DISCLOSURE_VERSION
to 2 so previously-consented users re-accept the updated notice.

Expanded the settings.json descriptions so the disclosure is visible where
users configure the extension, not just in the modal:
- llmProvider: clarifies AI features (ask/agent/visualize) send request text and
  limited data (result column names + a small row sample) to the selected
  provider, that query mode uses no AI, and that data is subject to the
  provider's terms. Each enum option states where its data goes (Bedrock→AWS,
  Copilot→GitHub, local-server→your endpoint, local→on-device).
- bedrockModelId / localServerUrl / localServerModel / localModel: per-provider
  data-handling notes.
- queryMode: notes ask/agent use the LLM while query does not.
- aiDisclosureAcceptedVersion: documents that clearing it withdraws consent.

Wording remains pending Legal review.

…tting

- handleRunFavorite now runs saved queries verbatim in query mode without
  persisting query as the composer's default — re-running a favorite no longer
  silently changes the user's chosen default mode for future sessions.
- Reflow App.tsx/VisualizePage.tsx to satisfy Prettier (CI --check).
- Note why AI_DISCLOSURE_VERSION starts at "2".

Adds a host-side consent check (defense in depth) so a replayed or malformed
message can't reach the LLM without accepted consent. onGenerate (ask/agent)
and onVisualize now refuse when the stored aiDisclosureAcceptedVersion doesn't
match the current disclosure version; query mode stays exempt (no LLM).

- completeViaLocalServer now uses an AbortController timeout so a hung/slow
  local server can't hang the request forever, and requests response_format
  json_object so compatible servers constrain output to JSON.
- Replace the greedy /\{.*"query".*\}/ extraction with a balanced-brace parser
  that ignores braces inside string literals, so prose-wrapped or trailing
  braces don't capture the wrong span.
- Extract the parse logic into a dependency-free localServerParse module and
  add unit tests (valid, prose-wrapped, nested, string-brace, empty, invalid,
  missing/blank query).