Baseline

Baseline is a Next.js + Supabase app for daily operations, triage, implementations, and secure calendar-derived briefing signals.

Local Setup

1. Install dependencies:

   npm install
   

2. Create .env.local with your Supabase keys and optional provider keys (see Environment Variables below).
3. Run the app:

   npm run dev
   

Environment Variables

Required for the app: NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY. Optional: DEFAULT_USER_ID, LLM_ADMIN_USER_ID, and the settings below.

If you want MCP access outside the browser app, or you still maintain the legacy ChatGPT custom GPT fallback, also configure:

• MISSION_CONTROL_API_KEY for legacy or internal machine-auth MCP callers
• MISSION_CONTROL_ACTIONS_API_KEY only if you still use the private ChatGPT custom GPT Actions fallback
• MISSION_CONTROL_FAMCAL_API_KEY only if the FamCal Sunday brief reads your personal tasks
• MISSION_CONTROL_USER_ID for the user that machine-auth requests run as
• DEPLOYMENT_ROLE=main|mcp to split the protected main app from the public MCP deployment
• MCP_CANONICAL_APP_URL for stable user-facing URLs returned from search / fetch
• MCP_UPSTREAM_API_URL for the public MCP deployment to call the protected main app upstream routes

Calendar settings

• CALENDAR_SOURCE=local|ical|none
  • local: load events from local ICS file path (CALENDAR_LOCAL_ICS_PATH)
  • ical: load events from WORK_ICAL_URL
  • none: skip ingest and only serve already-stored sanitized rows
• WORK_ICAL_URL= optional remote Outlook/M365 ICS feed URL
• CALENDAR_LOCAL_ICS_PATH=data/calendar/work-calendar.ics
• CALENDAR_RETENTION_DAYS=183
• CALENDAR_FUTURE_HORIZON_DAYS=30
• CALENDAR_BODY_MAX_CHARS=4000
• CALENDAR_STORE_BODY=false (default privacy mode)
• ENABLE_CALENDAR_ENRICHMENT=false (reserved)

Planner settings

• PLANNER_ENABLE_CRITICAL_EXCEPTION=false
• PLANNER_CRITICAL_EXCEPTION_THRESHOLD=90
• NEXT_PUBLIC_SPRINT_HOLIDAYS=2026-01-01,2026-05-25,2026-07-03 optional comma-separated YYYY-MM-DD list used by Today sprint pacing; excluded from workday counts in addition to weekends

Machine auth settings

• MISSION_CONTROL_API_KEY: backend key used by legacy or internal machine-auth MCP callers and other existing external callers
• MISSION_CONTROL_ACTIONS_API_KEY: optional backend key used only by the legacy private ChatGPT custom GPT Actions integration
• MISSION_CONTROL_USER_ID: required when either machine key is enabled
• MISSION_CONTROL_ACTIONS_API_KEY is a Mission Control backend key, not an OpenAI API key
• Use a distinct value for MISSION_CONTROL_ACTIONS_API_KEY so /api/mcp remains reserved for the primary MCP machine-auth key
• The actions key is restricted to the allowlisted non-delete routes used by the legacy ChatGPT Actions surface
• Delete operations are intentionally excluded from the ChatGPT Actions v1 fallback surface
• MISSION_CONTROL_FAMCAL_API_KEY: optional backend key for the FamCal Sunday brief (FamCal stores the same value as BASELINE_API_KEY). Give it its own value. It sees only personal to-dos: tasks that carry the personal tag themselves, have no sprint, and are either standalone (no project, application or section) or in a project that is itself tagged personal. Anything else, e.g. a personal task inside a work or build project, is left out of the list and is a plain 404:
  • Routes: GET /api/tasks (always filtered to tag=personal; only tag, limit, offset, status, owner, needs_review, due_soon, include_done, include_parked and include_missed are accepted, each at most once; any other parameter is a 400, even when empty), GET|PATCH /api/tasks/{id}, GET|POST /api/tasks/{id}/comments, POST /api/tasks/park/{id}. Everything else, and every DELETE, is 403.
  • Any other task is a 404 on the per-task routes. Dependency rows in the list keep unresolved but their title ("Another task" / "A commitment"), status and linked ids are redacted.
  • Tasks come back in a fixed FamCal shape: the task's own fields (not source_url), recurrence as its scheduling fields only (never the template's id), its project and section only when the project is itself tagged personal, and never an application or sprint (those are null), so a personal task filed under a work project shows nothing about it.
  • PATCH accepts only status (Done, Parked, or back to an open status from Done/Parked), and (a date; clearing is a 400); any other field is a 400. Writes re-check the tag at write time; a status change that races another change is a 409.

Remote MCP deployment settings

• DEPLOYMENT_ROLE=main on the normal app deployment
• DEPLOYMENT_ROLE=mcp on the public MCP-only deployment
• MCP_CANONICAL_APP_URL=https://your-main-app.example.com on both deployments
• MCP_UPSTREAM_API_URL=https://your-main-app.example.com on the public mcp deployment
• The public mcp deployment serves only /api/mcp, /oauth/*, /.well-known/*, /login, and /auth/callback
• The protected main deployment exposes /api/mcp-upstream and /api/mcp-upstream/* for bearer-token upstream calls only

Focus directive API

• GET /api/focus: returns current active focus directive (active) and optional fallback note
• POST /api/focus: creates a focus directive (supports implementation, stakeholder, task_type, query scopes)
• PATCH /api/focus/:id: updates an existing directive (including activation/deactivation)
• POST /api/focus/clear: clears active directive(s) for the current user

Calendar Privacy Model

• Calendar exports in data/calendar/ are ignored by git.
• Runtime calendar ingestion reads explicit source config only (no XML sharing URL discovery at runtime).
• Stored data is sanitized and scoped by user with Supabase RLS.
• /api/calendar returns only approved fields for brief generation and UI cards.
• User-authored meeting context is stored separately (calendar_event_context) and merged into /api/calendar for planner use.
• PATCH /api/calendar supports saving or clearing per-event meeting_context notes.

Authentication

• App API routes require an authenticated Supabase session.
• Use /login to request a magic-link sign-in email.
• Callback route (/auth/callback) exchanges the auth code and returns to app routes.
• MCP is the primary LLM interface for both ChatGPT and Claude.
• Internal or legacy MCP callers can still use /api/mcp with MISSION_CONTROL_API_KEY.
• The private ChatGPT custom GPT Actions flow remains available only as a fallback and sends X-Mission-Control-Key: <MISSION_CONTROL_ACTIONS_API_KEY>.
• Public remote MCP clients use OAuth authorization code + PKCE with dynamic client registration.
• Public remote MCP auth is limited to /api/mcp and /api/mcp-upstream/*; general app APIs do not accept MCP OAuth bearer tokens.

Remote MCP

Mission Control now supports a public remote MCP deployment for hosted MCP-compatible LLM products.

ChatGPT and Claude should both prefer this MCP path unless you have a specific reason to keep the legacy ChatGPT Actions fallback.

• Public MCP endpoint: https://<your-mcp-domain>/api/mcp
• OAuth authorize endpoint: https://<your-mcp-domain>/oauth/authorize
• OAuth token endpoint: https://<your-mcp-domain>/oauth/token
• Dynamic client registration endpoint: https://<your-mcp-domain>/oauth/register
• Authorization server metadata: https://<your-mcp-domain>/.well-known/oauth-authorization-server
• Protected resource metadata: https://<your-mcp-domain>/.well-known/oauth-protected-resource/api/mcp

Scope model:

• mcp.read: read/list/search/fetch/briefing/calendar style operations
• mcp.write: non-destructive mutations
• mcp.delete: destructive deletes only; request this separately from the default read/write consent

Tool behavior:

• Existing domain-specific tools remain available.
• Generic search(query) and fetch(id) are additive retrieval tools.
• search returns stable URLs pointing to MCP_CANONICAL_APP_URL.

Deployment outline:

1. Deploy the existing app normally with DEPLOYMENT_ROLE=main.
2. Deploy the same repo a second time with DEPLOYMENT_ROLE=mcp.
3. Point MCP_UPSTREAM_API_URL on the public mcp deployment at the protected main app domain.
4. Keep MISSION_CONTROL_API_KEY and MISSION_CONTROL_ACTIONS_API_KEY only on main while migrating legacy clients.

ChatGPT setup:

1. Open ChatGPT Developer Mode or MCP app setup.
2. Add the public MCP server URL: https://<your-mcp-domain>/api/mcp.
3. Let ChatGPT dynamically register as a public OAuth client.
4. Approve the default mcp.read mcp.write consent in the browser.
5. Re-run consent with mcp.delete only if you explicitly want destructive tools enabled.

Product note:

• As of March 8, 2026, OpenAI Help Center documentation says full write-capable MCP in ChatGPT beta is for Business and Enterprise/Edu, while Pro is read/fetch only in developer mode.
• This is a product-availability constraint, not a code limitation in this repo.

Legacy ChatGPT Custom GPT Fallback

Prefer MCP for both ChatGPT and Claude. Use this section only if you still want the older private custom GPT plus Actions integration as a fallback.

You do not need an OpenAI API key for normal use inside ChatGPT. The only extra credential for this integration is MISSION_CONTROL_ACTIONS_API_KEY.

1. Open the GPT builder in ChatGPT and create a private GPT.
2. Paste your adapted Mission Control instructions into the GPT Instructions field.
3. Import the tracked OpenAPI schema from /openapi/chatgpt-actions-v1.yaml.
4. Configure the action authentication header as X-Mission-Control-Key with MISSION_CONTROL_ACTIONS_API_KEY.
5. Add prompt starters such as morning brief, midday brief, eod brief, weekly review, monthly review, recommend my today list but do not sync yet, review these meeting notes and suggest updates, and apply these meeting notes to Mission Control.

Notes:

• The tracked schema file in this repo is /Users/owner/dev/Cooper Mission Control/mission-control/public/openapi/chatgpt-actions-v1.yaml.
• A paste-ready builder instruction draft lives at /Users/owner/dev/Cooper Mission Control/mission-control/chatgpt-gpt-instructions.md.
• Re-import the schema in the GPT builder after action-surface changes; the custom GPT does not auto-refresh from this repo.
• ChatGPT and Claude should both prefer MCP. This fallback exists only for the older private custom GPT setup.
• Deletes are intentionally excluded from the ChatGPT Actions v1 schema.

Database

Apply Supabase SQL migrations from supabase/migrations/:

• 001_initial_schema.sql
• 002_calendar_schema.sql
• 003_task_comments_dependencies.sql
• 004_add_task_type_task.sql
• 005_focus_planner_brain.sql
• 006_calendar_event_context.sql
• 007_llm_model_eval.sql — LLM model catalog, user preferences, and usage telemetry
• 008_task_status_workflow.sql — Renames task status enum values (Next->Backlog, Scheduled->Planned, Waiting->Blocked/Waiting) and adds In Progress
• 009_add_quick_capture_feature.sql — Adds quick_capture to LLM feature CHECK constraints
• 010_missing_rls_policies.sql — Adds missing RLS policies for inbox_items and status_updates
• 027_add_mcp_oauth.sql — Adds OAuth client, authorization code, access token, and refresh token storage for public remote MCP
• 028_add_project_focus_scope.sql — Adds project-scoped focus directives
• 029_add_task_tags.sql — Adds tags for tasks
• 030_add_project_review_history.sql — Adds durable project status history and persisted weekly/monthly review snapshots

Task status values

After migration 008, the valid task_status enum values are:

Backlog | Planned | In Progress | Blocked/Waiting | Done

The old values (Next, Scheduled, Waiting) no longer exist.

LLM features

Supported feature names for model preferences and usage tracking:

briefing_narrative | intake_extraction | quick_capture

Validation

Run checks:

npm run lint
npm run build
npm run test:planner-scoring
npm run test:calendar-sanitize
npm run test:mcp-oauth

Calendar API contract check (requires running app at http://localhost:3000 by default):

npm run dev
# in another terminal
npm run test:calendar-api

Optional env for contract test:

• CALENDAR_API_BASE_URL (default http://localhost:3000)
• CALENDAR_TEST_BEARER_TOKEN (if you want to provide an existing token)