@agent-native/core 0.84.21 → 0.84.23

package/corpus/templates/brain/README.md

@@ -1,564 +1,32 @@
-# Agent-Native Brain
+# Brain
 
-Brain is clean company chat backed by cited institutional knowledge. People ask
-questions in the Ask route; Brain answers from approved company knowledge with
-links back to the Slack thread, meeting, transcript, issue, or webhook capture
-that supports the answer. Sources, Review, and Knowledge are the admin/support
-surfaces for connecting data, approving proposals, and inspecting cited knowledge.
+An open-source, agent-native alternative to Glean — clean company chat backed by
+cited institutional knowledge. Ask a plain-English question and get an answer
+from approved company knowledge, with links back to the source.
 
-Brain ingests approved Slack channels, Clips recordings, Granola Team-space
-notes, GitHub issues/PRs, and generic transcript/webhook payloads. Transcript
-style captures pass through a pre-storage privacy filter first, so Brain stores
-the company-relevant capture text rather than the full raw call transcript or
-provider payload. It then distills imported captures into reviewable SQL-backed
-knowledge with source links.
+**Live app: [brain.agent-native.com](https://brain.agent-native.com)**
 
-The product direction is intentionally Glean-shaped, but the shipped V1 is not a
-full enterprise search replacement. Brain starts with open-source company knowledge
-over reviewed knowledge, then expands toward broader, permission-aware
-workspace search.
+Brain ingests approved Slack channels, meetings, transcripts, GitHub issues/PRs,
+and webhook captures, distills them into reviewable knowledge, and answers with
+exact evidence quotes and source links instead of guesses.
 
-## Version Direction
+## Features
 
-- **V1 Company Brain:** clean company chat over SQL-backed, reviewed,
-  distilled knowledge with exact evidence quotes and source links.
-- **V1.5 Search:** a Search route and `search-everything` action for searching
-  distilled knowledge, raw captures, and source records together. Agents should
-  use it as the broad first pass, then open records with `get-knowledge` or
-  `get-capture`. The response also includes `federatedCoverage` metadata that
-  names Brain source/provider coverage, workspace connection readiness, compact
-  discovered agent metadata when available, and deterministic delegation hints.
-  It does not directly search sibling apps.
-- **V1.5 shared integrations:** reusable workspace connections let Brain source
-  sync use provider credentials granted from Dispatch or the workspace layer.
-  The provider-reader runtime gives shared contracts for provider search/get
-  shapes, but live provider API calls stay template-owned unless a reader is
-  explicitly promoted to shared.
-- **V2 platform direction:** federated search across apps and sources,
-  permission-aware result ranking, and an expertise graph as a future/platform
-  layer.
-- **Portability:** V1 uses portable SQL text search and agentic query expansion.
-  There is no vector database requirement.
+- Company chat that answers from cited, reviewed knowledge — not hallucinations.
+- Source connectors for Slack, Granola, GitHub, Clips, and generic webhooks.
+- Pre-storage privacy filter and review queue before knowledge becomes durable.
+- Portable SQL text search and agentic query expansion — no vector database required.
+- Read-only, citation-backed retrieval exposed to other apps over A2A.
 
-## Product Shape
+## Develop locally
 
-- **Full-page company chat:** the Ask route is the main surface. It runs
-  `AgentChatSurface` in page mode and only surfaces setup/review attention when
-  action is needed. Keep this route on `AgentChatSurface` so Brain uses the
-  same composer UX as the agent sidebar and Agent-Native Code.
-- **Demo and evals stay out of the product UI:** seed product-decision sources
-  and run trust evals from actions/CLI/docs, not as prominent Ask-page controls.
-- **Search and drill-in:** the hidden Search route uses `search-everything` across
-  knowledge, raw captures, and source records, then agents can open exact
-  records with `get-knowledge` or `get-capture`. Cross-app expansion is exposed
-  as delegation guidance, not as hidden reads from other apps.
-- **Review queue:** the Review route lists pending/approved/rejected proposals,
-  lets reviewers edit proposed knowledge text, inspect evidence/source links, and
-  approve or reject. Reviewers can publish approved proposals as shared
-  `context/company-brain/...` workspace context when the knowledge should be
-  ambient for Dispatch and other apps.
-- **Source setup:** the Sources route leads with configured sources and one
-  clear Add source action. Provider catalog, source filtering, health checks,
-  and maintenance syncs live under Advanced. Sources are org-shared by default
-  so approved knowledge benefits the whole workspace.
-- **Ops and settings:** Ops stays available as an advanced/debug route for
-  queued, processing, stale, failed, and done distillation work. Settings
-  controls assistant identity, source posture, default publish tier,
-  company-knowledge approval, citations, redaction, and connector notifications.
-
-## Brain vs Dispatch
-
-Brain is the company-knowledge specialist. It ingests approved sources, reviews
-captures, distills durable facts and decisions, and answers from citations.
-
-Dispatch is the workspace control plane. It owns central messaging, the shared
-secrets vault, cross-app A2A routing, recurring jobs, approvals, and the
-distribution and approval of workspace-wide resources. In a workspace, Dispatch
-can route questions to Brain and grant Brain shared provider credentials, but
-Brain remains the place where company knowledge is ingested, reviewed, searched,
-and cited. Brain exposes
-read-only, citation-backed retrieval as its public A2A capability so Dispatch
-and sibling apps can ask company-knowledge questions. That is not anonymous data
-access: Brain's A2A agent card can be discovered publicly, but actual retrieval
-stays behind the authenticated A2A/action boundary and uses the same review,
-redaction, citation, and source-access rules as in-app Brain searches.
-
-## Start
+Scaffold your own copy and run it:
 
 ```bash
+npx @agent-native/core@latest create my-brain --standalone --template brain
+cd my-brain
 pnpm install
-pnpm --filter brain dev
-```
-
-Useful checks:
-
-```bash
-pnpm --filter brain typecheck
-pnpm --filter brain build
-```
-
-## Core Flow
-
-1. Check `get-brain-health` for the next first-run step, source freshness,
-   pending proposals, queue issues, and the last retrieval eval score.
-2. Create a source with `create-source`.
-3. Run `sync-source` for Slack/Granola/GitHub sources, import a transcript with
-   `import-transcript`, import raw text with `import-capture`, or POST a signed
-   `RawCapturePayload` to `/api/_agent-native/brain/ingest`.
-4. Review the raw capture inventory with `list-captures` or the Sources page,
-   then queue durable-company-context captures with `enqueue-distillation`.
-5. Distillation runs through the app agent: the open-tab bridge claims queued
-   work immediately, and the `brain-distillation` background sweep handles
-   queued or stale work when `RUN_BACKGROUND_JOBS` is enabled. The agent reads
-   the capture, writes cited knowledge or proposals with `write-knowledge`, and
-   closes the queue with `mark-capture-distilled`.
-6. Monitor failed or stale handoffs in the Ops route, or with
-   `list-distillation-queue` and `retry-distillation`.
-7. Review queued proposals in the Review route. Reviewers can edit a pending
-   proposal with `update-proposal`, then approve it with `approve-proposal` or
-   reject it with `reject-proposal`. Use `--publishCanonical true` only for
-   approved memories that should also be mirrored into workspace context. Brain
-   previews the exact Markdown first with `preview-canonical-resource`; the
-   Knowledge route and `set-knowledge-canonical` action can publish or unpublish
-   canonical context later without deleting the Brain knowledge row.
-8. Ask Brain or another workspace agent to search broadly with
-   `search-everything` when the V1.5 search surface is available, then drill
-   into `get-knowledge` / `get-capture` for cited answers. In V1-only
-   workspaces, use `search-knowledge` and `get-knowledge`.
-
-## Agent Retrieval Pattern
-
-Agents should treat Brain as cited company knowledge, not a guess engine:
-
-- Start with `search-everything` for broad questions so knowledge, raw captures,
-  and sources can all be considered. Inspect `federatedCoverage` before
-  claiming "everything": it shows what Brain actually searched, which provider
-  connections are visible/granted, and whether the agent should next use
-  `call-agent` for Analytics dashboards, Mail/Gmail mailbox search, or Dispatch
-  workspace resources/provider grants.
-- Use `get-knowledge` for reviewed facts, decisions, policies, and durable
-  summaries.
-- Use `get-capture` when the answer needs source context, exact quote checking,
-  or a direct link back to a meeting/transcript/message. Capture content is
-  redacted by default; pass `includeRawContent: true` only for
-  editor-authorized distillation or exact quote validation.
-- Cite links from evidence or capture metadata whenever available.
-- If Brain does not contain supporting results, say that the answer was not
-  found instead of filling in from general knowledge.
-- Cross-app delegation happens in the agent loop through `call-agent`. Brain
-  actions stay deterministic and read-only; they report coverage and hints but
-  do not call sibling app agents or read sibling app databases.
-
-## Privacy And Gating
-
-Brain is scoped to company knowledge, not personal surveillance:
-
-- Slack sync reads only configured channels and rejects DMs/MPIMs.
-- Granola sync reads Team-space notes exposed by Granola's API, not private
-  notes or private folders.
-- Source setup should use allow-lists, exclusions, redaction, and review gates
-  before broad sync is enabled.
-- Raw capture bodies are omitted from list/search responses by default. Use
-  previews for intentional human review and `includeRawContent` only for
-  distillation or exact quote validation.
-- Source configs default to review-required, and Settings can require approval
-  for company-tier knowledge before publishing.
-- Settings can require citations, auto-redact emails, and notify reviewers when
-  connectors degrade.
-- `run-demo-eval` covers proposal gating, PII redaction, personal-content
-  exclusion, citation presence, and honest not-found behavior.
-- `run-retrieval-eval` covers offline real-channel-style retrieval, using
-  existing workspace data when #dev-fusion/stale Fusion branch support is
-  already present and seeding fallback Slack-style data when absent.
-
-## Slack Source Config
-
-Slack resolves `SLACK_BOT_TOKEN` from a granted workspace connection first,
-then from backward-compatible Brain-local or registered vault credentials. It
-only scans configured channels and rejects DMs/MPIMs.
-
-```bash
-pnpm --filter brain action create-source \
-  --title "Slack product channels" \
-  --provider slack \
-  --visibility org \
-  --config '{"channelIds":["C0123456789"],"historyLimit":15}'
-```
-
-Useful config keys:
-
-- `channelIds`, `channels`, or `allowedChannels`: Slack channel IDs or names to
-  scan.
-- `historyLimit`: page size per channel. Keep this small for non-Marketplace
-  Slack apps because `conversations.history` can be heavily rate limited.
-- `oldest` / `updatedAfter`: optional timestamp boundary for initial backfill.
-- `autoSync` and `pollMinutes`: opt the source into background polling and set
-  the cadence. Background polling runs when `RUN_BACKGROUND_JOBS=1` in dev, and
-  by default in production unless `RUN_BACKGROUND_JOBS=0`.
-
-Slack scopes for the bot token should be the smallest set that supports the
-configured source:
-
-- `auth.test`: validate the token before any history read.
-- `conversations.info`: verify allow-listed conversations and reject DMs/MPIMs.
-- `conversations.history`: read message history from allow-listed channels.
-- `chat.getPermalink`: store durable citation links for each capture.
-- `conversations.list`: optional, only when a setup flow resolves channel names
-  instead of using channel IDs.
-
-Private channels require inviting the bot to the channel. Public channels may
-also require an explicit join/invite depending on the Slack app posture.
-
-For CLI/action-runner tests, store `SLACK_BOT_TOKEN` through a workspace
-connection or Brain/app credential first. The source credential resolver does
-not read process environment variables directly; a token in `.env.local` alone
-is only useful after it has been copied into the local credential store for the
-test user.
-
-Before reading real Slack history, run a credential/channel smoke test:
-
-```bash
-pnpm --filter brain action test-slack-connection \
-  --channelRefs '["C0123456789"]'
-```
-
-This calls Slack `auth.test` and optional channel metadata checks only. It never
-calls `conversations.history`.
-
-For a fuller rollout report, use the Slack pilot workflow:
-
-```bash
-pnpm --filter brain action run-slack-pilot \
-  --sourceId <source-id>
-```
-
-The default pilot validates credentials and allow-listed channels, summarizes
-guardrails, privacy exclusions, current knowledge/proposal counts, and next
-steps, and still reads no history. Only run a tiny sample sync when explicitly
-requested:
-
-```bash
-pnpm --filter brain action run-slack-pilot \
-  --sourceId <source-id> \
-  --readHistory true
-```
-
-Pilot sync caps reads to two validated channels, one history page per channel,
-ten messages per page, ten permalinks, `autoSync: false`, and a recent default
-history window.
-
-For a production rollout, keep the first pass deliberately narrow:
-
-1. Create one Slack source for one or two high-signal channels, using channel
-   IDs where possible.
-2. Keep `autoSync: false` while testing and reviewing the first imported
-   captures.
-3. Run `test-slack-connection`, then `run-slack-pilot` without `readHistory`.
-4. If the report is clean, run one tiny `run-slack-pilot --readHistory true`
-   sample.
-5. Review imported captures with previews only when needed; mark social,
-   personal, or thin messages ignored.
-6. Distill only durable company context, approve proposal-gated memories, and
-   confirm `ask-brain` cites the expected Slack permalinks.
-7. Expand the page/window manually with bounded `sync-source` runs before
-   turning on `autoSync`.
-
-After the first sample succeeds, review capture inventory before distillation:
-
-```bash
-pnpm --filter brain action list-captures \
-  --sourceId <source-id> \
-  --status queued
-```
-
-`list-captures` omits raw message bodies by default and includes the latest
-distillation queue state for each capture. Pass `--includePreview true` only
-when a human is intentionally reviewing snippets. Open individual records with
-`get-capture`; use `--includeRawContent true` only for distillation or exact
-quote validation. Distill durable company context into `write-knowledge`, and
-keep `autoSync` disabled until the source rules and review behavior look right.
-
-After any pilot sync, generate the source-level quality report:
-
-```bash
-pnpm --filter brain action get-pilot-report \
-  --sourceId <source-id>
-```
-
-The report summarizes sync health, capture counts, distillation queue state,
-published knowledge, pending proposals, privacy notes, and recommended next
-steps without returning raw capture bodies. It also includes a compact
-`pilotTrustLane` for the real `#dev-fusion` pilot: a status, four checks,
-minimal next actions, and eval questions to run before broadening sync.
-
-The Sources page exposes the same review inventory from each source card. Open
-**Captures** to inspect queued records, enable short previews only when needed,
-queue distillation for durable context, see whether a capture is waiting on the
-distillation worker, or mark non-company material ignored.
-
-For Slack sources, the source card presents the recommended pilot path as a
-minimal four-step flow: **Test** validates credentials and the allow-list
-without history reads, **Safe pilot** imports a tiny capped sample,
-**Review captures** opens the raw inventory, and **Review queue** takes
-reviewers to proposal approval before memories become durable company
-knowledge. Opening **Report** shows the same trust lane inline so operators can
-see whether `#dev-fusion` is blocked, needs distillation, needs review, needs
-retrieval eval, or is ready for a narrow expansion.
-
-Distillation has two worker paths. When a Brain tab is open, the app shell
-claims queued items with `claim-distillation` and hands them to the app agent in
-the background. When no tab is open, the `brain-distillation` server sweep runs
-with `RUN_BACKGROUND_JOBS`, claims due queued rows, reclaims stale `processing`
-rows, and invokes the same agent loop headlessly. Re-running
-`enqueue-distillation` for an active queue item refreshes the handoff instead
-of duplicating queue rows. The agent reads the capture, writes cited knowledge
-or review proposals, then calls `mark-capture-distilled`, which marks the
-active queue row done. If the agent does not close the queue, the worker requeues
-the item with a short delay and eventually fails it after repeated attempts.
-
-The Ops route is the operator surface for that pipeline. It shows queued,
-processing, failed, done, stale, and retryable distillation work. The matching
-actions are `list-distillation-queue` and `retry-distillation`; retries are
-allowed only for failed or stale processing items the current user can edit.
-
-## Granola Source Config
-
-Granola resolves `GRANOLA_API_KEY` from a granted workspace connection first,
-then from backward-compatible Brain-local or registered vault credentials, and
-polls `https://public-api.granola.ai/v1/notes`. Enterprise API keys expose
-Team-space notes; private notes are not included by Granola's API.
-
-```bash
-pnpm --filter brain action create-source \
-  --title "Granola team notes" \
-  --provider granola \
-  --visibility org \
-  --config '{"pageSize":10,"updatedAfter":"2026-05-01T00:00:00.000Z"}'
-```
-
-Brain persists Granola cursors in the source cursor JSON and normalizes note
-summary, transcript, attendees, calendar metadata, and `web_url` into imported
-captures. Before SQL insert, the sanitizer filters the title/body down to
-company-relevant content and strips raw transcript segments, attendees, owner,
-and calendar objects from metadata. Safe linkage such as `sourceUrl`, connector,
-sync run id, and note timestamps remains available for review/citations.
-
-## GitHub Source Config
-
-GitHub is Brain's first reusable connector proof. It resolves `GITHUB_TOKEN`
-from a granted workspace connection first, then from backward-compatible
-Brain-local or registered vault credentials, and imports bounded issue/PR
-context from configured repositories through GitHub's REST API. This is company
-context for Brain ingestion, not full GitHub analytics.
-
-```bash
-pnpm --filter brain action create-source \
-  --title "GitHub product repos" \
-  --provider github \
-  --visibility org \
-  --config '{"repositories":["owner/repo"],"state":"all","limit":25}'
-```
-
-Useful config keys:
-
-- `repositories` or `repos`: repository slugs like `owner/repo`.
-- `state`: `open`, `closed`, or `all`; defaults to `all`.
-- `limit`: bounded page size per repository, capped by the connector.
-- `includeIssues` / `includePullRequests`: disable either side when a source
-  should capture only issues or only PRs.
-- `linkedSlackSourceIds`, `slackSourceIds`, or `linkedSourceIds`: import GitHub
-  issue and PR URLs found in accessible Slack Brain captures.
-- `linkedCaptureLimit`, `linkedRefLimit`, `linkedDetailLimit`, `commentLimit`,
-  `reviewLimit`, and `repoDetailLimit`: keep linked imports bounded.
-
-## Workspace Connections
-
-`list-connection-providers` returns the Brain provider catalog plus
-`workspaceConnection`, `credentialHealth`, and `providerHealth` summaries for
-`appId=brain`. Use those summaries before asking for duplicate provider
-credentials:
-
-- `grantState: "connected"` means Brain already has a granted workspace
-  connection for that provider.
-- `grantState: "granted"` means Brain has a grant, but the connection is not
-  currently active.
-- `grantState: "needs_grant"` means a workspace connection exists but still
-  needs a Brain grant.
-- `grantState: "not_connected"` means there is no shared connection for Brain
-  yet, though Brain-local or registered vault credentials may still exist.
-
-Source sync resolves credentials in this order:
-
-1. A source's configured `workspaceConnectionId`, when present.
-2. Granted `workspace_connections` / `workspace_connection_grants` credential
-   refs for `appId=brain`.
-3. Brain-local SQL credentials.
-4. Registered vault secrets for the same user/org/workspace scope.
-
-`workspaceConnectionId` is non-secret source config. Use it when a workspace has
-multiple Slack, Granola, or GitHub connections and a source must use one exact
-shared integration. Bound sources are strict: if the selected connection is
-missing, not granted to Brain, unhealthy, or missing the required vault-backed
-credential ref, sync fails with that specific message instead of silently using
-another shared connection or Brain-local credential. Leave it unset for the
-legacy automatic fallback behavior.
-
-It does not fall back to deploy-level environment variables for source
-credentials. Connection and grant refs point at vault secret names; they never
-contain raw credential values.
-
-The Sources route shows the same shared integration state in the provider
-catalog, including readiness labels such as ready, grant needed, missing keys,
-needs repair, or metadata only. Use Dispatch to connect or grant reusable
-workspace credentials, then create Brain sources against those providers
-without copying secret values into the Brain app.
-
-The boundary is intentional:
-
-- Reusable workspace integrations own provider identity, account metadata,
-  credential ref names, and per-app grants.
-- Dispatch is the workspace control plane where admins usually connect, repair,
-  and grant those integrations.
-- The vault owns the secret values.
-- Brain owns source-specific choices: Slack channel IDs, GitHub repositories,
-  Granola polling windows, sync cursors, review posture, and distillation state.
-- Agents should inspect shared connection readiness first, then ask for a Brain
-  grant or source config instead of asking for another raw provider token.
-
-## Scheduled Sync
-
-Use `sync-source` to run one source immediately, or `sync-due-sources` to run
-accessible Slack/Granola sources whose `autoSync` cadence is due. The Nitro
-plugin in `server/plugins/brain-jobs.ts` registers the same due-source sweep for
-long-lived deployments.
-
-## Clips And Generic Webhook
-
-Create a Clips or generic source with `sourceKey` to receive a one-time ingest
-token:
-
-```bash
-pnpm --filter brain action create-source \
-  --title "Clips exports" \
-  --provider clips \
-  --sourceKey clips \
-  --visibility org
-```
-
-Then send:
-
-```json
-{
-  "sourceKey": "clips",
-  "externalId": "meeting-123",
-  "title": "Product decision review",
-  "participants": ["Ada", "Grace"],
-  "occurredAt": "2026-05-15T15:00:00.000Z",
-  "transcript": "We decided to...",
-  "sourceUrl": "https://example.com/share/meeting-123",
-  "tags": ["product", "pricing"],
-  "raw": {}
-}
-```
-
-Use `Authorization: Bearer <ingestToken>`.
-
-Clips exports use this endpoint without Brain reading the Clips database
-directly. Generic sources use the same payload shape for transcripts, customer
-research, meeting exports, or any bounded capture that should enter the review
-and distillation pipeline.
-
-Transcript payloads from Clips, generic webhooks, Granola, and
-`import-transcript` are sanitized before they are saved. The default model is
-the Brain agent model, but Settings can specify a cheaper model override for
-this filtering pass. The model sanitizer receives `AGENTS.md`, including the
-capture sanitization rules, so privacy policy tweaks can be made in instructions.
-Recruiting and candidate-evaluation content is always stripped. If no model is
-available, Brain falls back to a conservative deterministic filter that keeps
-only likely company-relevant lines and redacts obvious contact details, secrets,
-recruiting signals, and links.
-
-After enabling or tightening the filter, re-run it on already imported
-transcript captures:
-
-```bash
-pnpm --filter brain action resanitize-captures \
-  --sourceId <source-id> \
-  --limit 25 \
-  --dryRun true
-```
-
-Non-dry-run resanitization skips captures that already have knowledge or
-proposals citing their current text, so evidence quotes do not silently drift.
-After reviewing the dry-run preview, re-distill affected captures or pass
-`--allowCitationDrift true` only for an intentional repair.
-
-## Data
-
-Brain stores data in portable SQL through Drizzle:
-
-- `brain_sources`
-- `brain_raw_captures`
-- `brain_knowledge`
-- `brain_proposals`
-- `brain_sync_runs`
-- `brain_ingest_queue`
-
-JSON stays in text columns. V1 does not require a vector database.
-
-## Demo and Eval
-
-Load the product-decision demo corpus:
-
-```bash
-pnpm --filter brain action seed-demo-data
-```
-
-Then run the repeatable quality check:
-
-```bash
-pnpm --filter brain action run-demo-eval
-```
-
-The eval checks product-decision recall, citation presence, supersede links,
-proposal gating, PII redaction, and personal-content exclusion. Keep these
-demo/eval controls in actions and docs rather than the Ask-page product UI.
-
-Run the real-channel-style retrieval eval:
-
-```bash
-pnpm --filter brain action run-retrieval-eval
-```
-
-This eval checks #dev-fusion stale Fusion branch retrieval, Slack-style citation
-presence, branch-safety terms, and an unsupported cleanup-cron not-found case.
-It evaluates existing workspace data first; if the answer cases do not have
-citation-backed support and `seedIfMissing` is true, it seeds a tiny portable SQL
-fallback corpus and re-runs the checks. The same eval is available through
-`run-demo-eval` with `mode: "retrieval"`.
-
-CI and `pnpm prep` run the deterministic action evals through the repository
-script:
-
-```bash
-pnpm test:brain-evals
-```
-
-That command uses a disposable local SQLite database at
-`templates/brain/data/brain-evals-ci.db`, seeds any missing demo/fallback
-fixtures, and removes the database when it exits. It does not call Slack,
-Granola, Clips, or any external service.
-
-The Slack pilot regression set lives in
-`templates/brain/evals/slack-pilot-corpus.ts`. It contains redacted pilot
-questions for reasoning-effort controls, Fusion PR #13340 missing-branch
-handling, Figma Plugin JSON uploader feedback, non-English support, Slack
-history guardrails, citation requirements, personal-content exclusion, and
-honest not-found behavior.
-
-```bash
-pnpm --filter brain exec vitest --run --config vitest.config.ts evals/slack-pilot-corpus.test.ts
+pnpm dev
 ```
 
-The eval is offline and validates the real `searchEverythingRows` retrieval
-path plus `ask-brain` cited-answer behavior.
+Full docs: [agent-native.com/docs/template-brain](https://agent-native.com/docs/template-brain).

package/corpus/templates/calendar/.agents/skills/frontend-design/SKILL.md

@@ -53,6 +53,8 @@ Default to Apple/Linear-level restraint: make the primary workflow obvious, then
 - **Visual assets**: Websites, games, and object-focused pages need real or generated media when images help users understand the subject.
 - **Responsive fit**: Text must not overflow buttons, cards, tabs, sidebars, or fixed-format tools. Use stable dimensions for boards, grids, toolbars, and counters.
 
+**Beat convergence, not just defaults.** You sample toward the "on-distribution" center, so naming what to avoid is not enough: every "don't" needs a "do", or you converge on the next safe option (ban Inter and you reach for Roboto; ban purple gradients and you reach for Space Grotesk + a teal accent on every screen). Commit to one named direction, pair any reference with the reason it fits ("Linear: the quiet confidence of its spacing" — a bare "Linear" collapses back to the average), and match implementation effort to the vision: maximalist wants elaborate motion and effects, minimal wants restraint and precise spacing. When building on an existing app, inspect its tokens/type/components first and treat any drift back to a default as a missing token to pin, not something to re-prompt.
+
 ## Agent-Native UI Rules
 
 - Agent-native apps use React, Vite, Tailwind CSS, shadcn/ui, and `@tabler/icons-react`.

package/corpus/templates/calendar/README.md

@@ -1,41 +1,35 @@
 # Calendar
 
-Google Calendar integration with a Calendly-style public booking page.
+An open-source, agent-native alternative to the Google Calendar + Calendly combo.
+An agent-powered calendar with Google Calendar sync and Calendly-style public
+booking links — schedule, find slots, and manage availability in plain English.
 
-## Setup
+**Live app: [calendar.agent-native.com](https://calendar.agent-native.com)**
 
-### 1. Install dependencies
+Connect your Google Calendar and the agent can read your schedule, find free
+slots, create events, and manage booking links. Anything you can do in the UI,
+the agent can do through the same actions.
 
-```bash
-pnpm install
-```
+## Features
 
-### 2. Google OAuth Setup (optional)
+- Day, week, and month views with multiple Google accounts overlayed.
+- Google Calendar sync and read-only ICS feed subscriptions.
+- Weekly availability with timezone support for slot-finding.
+- Calendly-style public booking links at `/book/{slug}` with custom fields.
+- Ask the agent anything schedule-related, from "am I free Thursday?" to
+  creating and rescheduling events.
+- Share booking links with teammates and required co-hosts.
 
-1. Go to [Google Cloud Console](https://console.cloud.google.com/)
-2. Create a project and enable the Google Calendar API
-3. Create OAuth 2.0 credentials (Web application type)
-4. Set the authorized redirect URI to `http://localhost:5173/_agent-native/google/callback`
-5. Add credentials in the app's Settings page, or set them only in a local
-   `.env` / deployment secret. Never commit real credential values:
-   ```
-   GOOGLE_CLIENT_ID=your-client-id
-   GOOGLE_CLIENT_SECRET=your-client-secret
-   ```
+## Develop locally
 
-### 3. Run
+Scaffold your own copy and run it:
 
 ```bash
+npx @agent-native/core@latest create my-calendar --standalone --template calendar
+cd my-calendar
+pnpm install
 pnpm dev
 ```
 
-Open http://localhost:5173
-
-## Features
-
-- Monthly/weekly/daily calendar views
-- Google Calendar sync (pull-based)
-- Event CRUD (local + Google)
-- Configurable availability schedule
-- Public booking page at `/book/meeting`
-- Real-time updates via SSE when agent modifies data
+Connecting Google Calendar in dev needs a Google OAuth client — see the docs for
+setup. Full docs: [agent-native.com/docs/template-calendar](https://agent-native.com/docs/template-calendar).

package/corpus/templates/chat/.agents/skills/frontend-design/SKILL.md

@@ -53,6 +53,8 @@ Default to Apple/Linear-level restraint: make the primary workflow obvious, then
 - **Visual assets**: Websites, games, and object-focused pages need real or generated media when images help users understand the subject.
 - **Responsive fit**: Text must not overflow buttons, cards, tabs, sidebars, or fixed-format tools. Use stable dimensions for boards, grids, toolbars, and counters.
 
+**Beat convergence, not just defaults.** You sample toward the "on-distribution" center, so naming what to avoid is not enough: every "don't" needs a "do", or you converge on the next safe option (ban Inter and you reach for Roboto; ban purple gradients and you reach for Space Grotesk + a teal accent on every screen). Commit to one named direction, pair any reference with the reason it fits ("Linear: the quiet confidence of its spacing" — a bare "Linear" collapses back to the average), and match implementation effort to the vision: maximalist wants elaborate motion and effects, minimal wants restraint and precise spacing. When building on an existing app, inspect its tokens/type/components first and treat any drift back to a default as a missing token to pin, not something to re-prompt.
+
 ## Agent-Native UI Rules
 
 - Agent-native apps use React, Vite, Tailwind CSS, shadcn/ui, and `@tabler/icons-react`.