A year of your work, one cell a day. A dashboard you arrange. A profile a visitor can only ever see less of.

One cell a day. Every cell is a count, nothing is decoration.

Hover or arrow through the specimen. Switch the tier to see the same year through a visitor's lens: the numbers get smaller, the chrome stays identical, and no title ever appears. Tick include private to watch the public total change without a single object being revealed.

Day one

The graph earns its place in onboarding by being nearly empty. A new account has one cell: today, the welcome relay they just claimed. Tomorrow adds a second. Nothing is celebrated; the cell simply fills, and the page says so once.

The mix toward text-bright is what makes one formula serve both modes: it darkens in light mode and brightens in dark mode, which is exactly the anchor flip a sequential ramp needs. The values shown are computed, not chosen; the tokens remain the only source.

A module is the smallest unit a screen can lend to another surface.

The shipped dashboard is five bands hand-wired in overview.ts. The Root Manifest already projects one substrate into three lenses (agent boot, the Overview, the Graph Map). The contract generalises that: every screen lends modules; every host arranges them; the lens decides the tier. One specimen, three widths, three tiers. Its chrome never changes; only the corpus behind it does.

// web/src/modules/contract.ts — proposed. A screen registers what it lends; a host arranges it.
interface RelayModule {
  id: "streams.active";            // <screen>.<module> — belongs to ONE screen, always
  screen: "stream";                // the route it projects; "Open Stream →" is the only way deeper
  title: "Active streams";
  hosts: ("dashboard" | "profile" | "public-profile" | "stream" | "org-home" | "agent-boot")[];
  sizes: (1 | 2 | 3)[];              // columns on a 3-column host; a phone is always 1 and keeps the order
  tiers: { owner: "full"; connection: "counts"; public: "counts" };   // or "none" → the module collapses
  empty: "collapse" | "teach";      // zero height, or an empty state that teaches (first run only)
  data: { source: "GET /api/user/v1/streams?lifecycle=open"; freshness: "live" | "polled" | "snapshot" };
  pinned?: true;                    // a host may move it, never remove it (Needs you)
  hat?: "org"; role?: "admin"; flag?: "files";   // presence conditions — hide, don't disable
  onboarding?: { step: 6 };          // a step it satisfies, or a teach moment it plays once
  render(ctx: { tier; size; corpus }): string;   // receives an ALREADY-FILTERED corpus; never reads raw data
}
interface HostLayout {             // GET/PUT /api/user/v1/layouts/{host} — per user, per host, versioned
  version: 1; host: "dashboard"; lens: { scope: "context" | "mine" };
  modules: { id: string; size: 1|2|3; lens?: "mine"; options?: Record<string, unknown> }[];
}

Six places a module can land. Two are drawn at full scale.

The agent-boot row is the point of naming this a contract rather than a widget system. The manifest already is a module host for agents (SPEC.manifest.md §3: streams, series, open sessions, recent relays). Making the human dashboard the same shape means a screen lends one module and both audiences get it.

Sixteen modules, each lent by a screen that already ships.

Nothing here invents a screen. The table is generated from the same registry the prototypes run on, so it cannot drift from them.

Hats are walls, lenses are windows, and a viewer is a lens too.

Four lenses touch a module, and they compose in a fixed order. The rulings already ratified for the Context Map carry over unchanged: filters never grant · permissions are evaluated once, before any lens renders · a lens is presentation over an already-filtered corpus · counts are per-viewer and honest.

Step 6 becomes “Make it yours”: arrange the dashboard, choose what the profile shows.

The First Relay series already ends on Profile (handle · bio · visibility → “You're set up”). The step keeps its number and its signal; it gains the two decisions this concept introduces, and it gets the graph's first cell as its quiet beat.

Sequence cost: zero new steps. The signals R5 (Graph opened) and R6 (Profile set) already exist as stubs in the initiative; 6a adds one more accepted signal for R6 (a layout written), so a user who arranges the dashboard but never opens the profile still completes.

The shipped form becomes the edit state. The profile becomes what you did.

Where a contribution graph lies if nobody decides.

Worked in full since this page shipped. The ten below are the shape of the problem; the complete register is 60 cases across ten families — what counts, time and timezones, privacy and the lens, layout and host, performance, accessibility, mobile, i18n, onboarding and degradation — each row carrying a decision and the id of the test that pins it: relay-board/docs/product/activity-modules-program-map.md §4. Four decisions came out of it (the agent split is owner-only; layouts are per workspace; “view as” is a server round-trip; Step 6’s signal is unchanged) and one live privacy finding: the shipped public badge already publishes the counts this page makes opt-in.

1. Agent activity. Counts as the user's (an agent works for them) but is marked ◆ and split out in the tooltip and the table. A public visitor never learns which agent.
2. Timezone. Day boundaries follow the viewer's profile timezone; the streak is computed the same way. A trip across the dateline cannot break a streak that was real.
3. Multiple hats. The owner's graph spans every hat (mine). A visitor's spans org-visible activity, plus personal only if opted in. An org can restrict its share; it cannot add to it.
4. Deleted or expired relays. The event happened; the count stays. The tombstone grammar applies to the object, not to history.
5. Backfill. relay_timeline reaches back to each object's creation; the rollup runs once over history, then incrementally. Accounts older than the timeline show honest blanks, not zeros dressed as quiet days.
6. The 53rd column. A year of Monday-start weeks needs up to 53 columns; the leftmost may be partial. Month labels drop the first when two would collide.
7. Empty year. Selecting a year before the account existed says so (“No activity in 2025 — you joined Apr 2026”), never an all-grey grid pretending to be a quiet year.
8. Small orgs. The admin lens carries an aggregation floor; a member's public card never carries org names beyond the org display the shipped card already shows.
9. Streamlined cohort. Under the streamlined preset the dashboard keeps its module host but the catalogue hides team/network modules, exactly as the rail does.
10. Mobile. One column, band order preserved. The graph shows the trailing weeks that fit, at full cell size, with Full year one tap away — and the totals stay the whole window’s, so a narrow screen never reports a smaller year. A 12px cell cannot be a 44px target, so on a coarse pointer a tap opens a day sheet naming the day, with prev/next: a mis-tap is visible and one tap from fixed. No hover tooltip, no drag grip and no resize control on touch.

Most of the substrate ships. The surface and one endpoint do not.

Phases

Decisions to ratify

1. M1 · What counts. The seven kinds above, reads never. One definition for the graph, the streak and the badge.
2. M2 · Counts, or nothing, on any non-owner surface. Titles, Codes and counterparts never leave the owner's tier. Rule 3 as law.
3. M3 · Private contributions are opt-in and numeric. Default off; org policy may restrict, never widen.
4. M4 · The dashboard is a host and Needs-you is pinned. Customization can reorder it, never remove it.
5. M5 · One registry, both audiences. A screen lends a module once; the human dashboard and the agent's manifest both draw from it.
6. M6 · Step 6 = Make it yours. The Profile step absorbs the two new decisions and the first-cell beat; no new step is added.