@immediately-run/grove 0.1.5 → 0.1.8

package/src/lib/reachCard.ts

@@ -1,20 +1,64 @@
 // The reach card (GROVE_AGENT_SPEC §6) — the agent's envelope, rendered as rows in
 // the two-word vocabulary with a cause for every ✗ (R-SP-3). R-GA-1: every row is
-// COMPUTED from the session's envelope (provider three-state, chat grant, mount
-// writability, source trust); no capability claim on any pixel of this surface is
-// hand-written copy. The four old banners collapse into these rows; chips render
-// only for rows that are ✓ — derived, not curated.
+// COMPUTED from the session's envelope (provider four-state, chat grant, mount
+// writability, packaging, source trust); no capability claim on any pixel of this
+// surface is hand-written copy. The four old banners collapse into these rows; chips
+// render only for rows that are ✓ — derived, not curated.
+//
+// R3-752: a row whose outcome is a destination is not a denial — the apply row
+// renders `→ elsewhere` (R-GA-3 still holds: never from this panel), the packaging
+// substrate is its own neutral row, the source-trust sentence is its own line keyed
+// on WHY the source reads as shared (R-SP-3/R-SP-6), and the tool-less degrade is a
+// visible qualifier on the Q&A row instead of a silent downgrade (SPEC_AUDIT §2.8u).
 
-import type { ChatProviderState } from '@immediately-run/sdk';
+import type { AgentContextBlock, ChatProviderState } from '@immediately-run/sdk';
+
+/** Why `sourceShared` reads as it does — the SDK's own union, indexed off the
+ *  exported block so the two can never drift (R6: one home, the SDK's). */
+export type SourceSharedBasis = AgentContextBlock['sourceSharedBasis'];
+
+/** The visually-hidden word for a row's state — one home for the map the panel
+ *  renders per row and announces through its live region (R6). */
+export function stateWord(state: ReachRow['state']): string {
+  return state === 'ok' ? 'available' : state === 'blocked' ? 'unavailable' : state === 'elsewhere' ? 'opens elsewhere' : 'not applicable';
+}
+
+/** One reach-card row.
+ *
+ *  `state: 'neutral'` renders neither ✓ nor ✗. Four producers, all of them "we are not
+ *  claiming anything here":
+ *   1. the `unknown` provider state — rendering a cause there re-creates the false banner
+ *      R3-300 fixed: `unknown` means unanswered, not ungranted;
+ *   2. since R3-752, the packaging row — substrate context, not a capability claim;
+ *   3. since R3-688, the Q&A row while the CATALOG is unanswered — the same "not told
+ *      yet" as (1), one channel over;
+ *   4. since R3-688, the exhaustiveness fallback at the end of the Q&A chain. Unreachable
+ *      while `ChatProviderState` has four members, and deliberately neutral rather than ✓
+ *      so that a fifth member added without updating this file degrades to claiming
+ *      nothing instead of claiming everything.
+ *
+ *  `state: 'elsewhere'` is the apply row: the outcome happens at another surface, which is
+ *  where to go — never a ✗. */
+/** R3-790 — the one spelling of the ungranted-chat cause: the row renders it and
+ *  the composer's refusal toast composes from it, so a reword cannot drift the
+ *  two surfaces R-GA-1 holds to agreement. The tail is its own constant so the
+ *  cancelled toast composes the SAME clause without splitting a literal at
+ *  runtime (a reword of the cause must never render 'undefined' into copy). */
+export const READING_WORKS_NORMAL = 'reading works as normal';
+export const UNGRANTED_CHAT_CAUSE = `this Grove wasn't granted chat — ${READING_WORKS_NORMAL}`;
 
-/** One reach-card row. `state: 'neutral'` renders neither ✓ nor ✗ — used only for
- *  the unknown provider state (rendering a cause there re-creates the false banner
- *  R3-300 fixed: `unknown` means unanswered, not ungranted). */
 export interface ReachRow {
-  key: 'answer' | 'read' | 'draft' | 'apply';
+  key: 'packaging' | 'answer' | 'read' | 'draft' | 'apply';
   label: string;
-  state: 'ok' | 'blocked' | 'neutral';
+  state: 'ok' | 'blocked' | 'neutral' | 'elsewhere';
   cause?: string;
+  /** R3-790 — the row's earning affordance: on the ✗ chat row, an 'Enable chat'
+   *  control whose CLICK invokes (the host's lazy consent gesture mints on that
+   *  same activation, site-main#612). Declared here so the card stays the one
+   *  computed surface; the panel renders it, never invents it. */
+  action?: 'enable-chat';
+  /** Where a `'elsewhere'` row's outcome happens — rendered after `→`, never a ✗. */
+  destination?: string;
   /** Chips this row contributes when ✓ — the panel renders exactly these. */
   chips?: string[];
 }
@@ -22,42 +66,149 @@ export interface ReachRow {
 export interface ReachInputs {
   providerState: ChatProviderState;
   /** Whether the grant-filtered catalog advertises `llm:chat` (the consented
-   *  capability — absent on an ungranted fork, a distinct cause from "no key"). */
+   *  capability — absent on an ungranted fork, a distinct cause from "no key").
+   *
+   *  Since R3-688 this is the BELT, not the primary discriminator: a 0.72.0 host marks
+   *  the grantless answer on the provider channel itself, and that mark is read first.
+   *  But it is still checked BEFORE `not-configured`, not after, because a host
+   *  predating the mark answers an ungranted fork with `{ provider: null }` — the same
+   *  payload as keyless — so not-configured would otherwise win and send a user who HAS
+   *  a key to Settings. See the ordering note in `computeReachRows` for the one
+   *  imprecision that buys, and how `catalogAnswered` bounds it. */
   chatGranted: boolean;
+  /** Whether the host has actually ANSWERED the catalog (`useCatalogAnswered()`).
+   *
+   *  `!chatGranted` is true both for "not granted" and for "has not replied yet", and
+   *  the value cannot tell them apart because an empty catalog is a legitimate answer.
+   *  This does, because the push channel has no value-equality check: a second
+   *  notification means the host spoke. While it is `false`, the card declines to name
+   *  a cause rather than guessing one. */
+  catalogAnswered: boolean;
   writable: boolean;
-  /** Fail-closed source trust (git ⇒ indeterminate ⇒ treated as shared). */
+  /** Fail-closed source trust (git ⇒ indeterminate ⇒ treated as shared). Not
+   *  rendered on any row since R3-752 — it is the source-trust LINE's input (see
+   *  `sourceTrustLine`) — but part of the envelope, so the cross-product over the
+   *  card's inputs can assert no input re-blocks the apply row. */
   sourceShared: boolean;
+  /** The corpus mount id (`getCorpusMountId()`): `null` is a fork reading its own
+   *  bundled corpus, an id is a wiki mounted into it. One fact, decided once at
+   *  boot by the same delegation as the root — never a second source of truth. */
+  mountId: string | null;
+  /** Whether the configured provider advertises `features.tools`. `false` with a
+   *  configured provider degrades the agent to context-stuffing (G-GA-8) — a real
+   *  capability change the card must show, not hide (SPEC_AUDIT §2.8u). */
+  toolsSupported: boolean;
 }
 
-/** The rows, computed. Order is the card's display order. */
-export function computeReachRows({ providerState, chatGranted, writable, sourceShared }: ReachInputs): ReachRow[] {
-  // Row 1 — Q&A. Three provider states × the grant, with the two NOT-causes never
+/** The rows, computed. Order is the card's display order: the substrate first
+ *  (context for every row under it), then the capability rows. */
+export function computeReachRows({
+  providerState,
+  chatGranted,
+  catalogAnswered,
+  writable,
+  mountId,
+  toolsSupported,
+}: ReachInputs): ReachRow[] {
+  // Row 0 — packaging (R3-752). Context, not a claim: `neutral`, no chips, no
+  // capability words. A pinned-library consumer is indistinguishable from a fork at
+  // runtime and the distinction does not change reach, so the row states the
+  // substrate — never a guess at how the app was assembled.
+  const packaging: ReachRow = {
+    key: 'packaging',
+    label: mountId === null ? 'Reads its own entries' : 'Reads a wiki mounted into it',
+    state: 'neutral',
+  };
+
+  // Row 1 — Q&A. The provider states × the grant, with the two NOT-causes never
   // conflated (G-GA-10): "no key" is the user's to fix in Settings; "not granted"
-  // is this copy's consent state, and reading works either way.
+  // is this copy's consent state, and reading works either way. A configured
+  // provider without `features.tools` keeps the ✓ — asking still works — and
+  // carries the degrade as a qualifier (G-GA-8, SPEC_AUDIT §2.8u).
+  //
+  // R3-688: the host now marks the grantless answer on the provider channel itself
+  // (`ungranted`), so the consent cause is computable even though an ungranted frame is
+  // never told the provider — the host's grant decision IS the fact. The catalog check is
+  // the BELT for a host predating the mark, and it is checked BEFORE not-configured
+  // because a pre-mark host answers an ungranted fork with `{provider: null}` — the same
+  // payload as keyless — so not-configured would win and send a user who has a key to
+  // Settings.
+  //
+  // THE UNANSWERED CATALOG, and the two wrong answers before this one (grove#75 r1, r2).
+  //
+  // `!chatGranted` is true both for "not granted" and for "the host has not replied yet",
+  // and the catalog's VALUE cannot separate them because an empty catalog is a legitimate
+  // answer. Round 1 caught the belt firing on the unanswered case and claiming a consent
+  // state about a frame that may hold the grant.
+  //
+  // Wrong answer 1: gate on `catalog.length > 0`. That reads an empty ANSWER as silence,
+  // which turns the G-GA-10 case below (a configured provider beside an empty catalog)
+  // from a permanent correct ✗ into a permanent unbacked ✓ — the R-GA-1 violation this
+  // card exists to prevent. Measured; two tests catch it.
+  //
+  // Wrong answer 2: declare it unfixable without an SDK change. Round 2 found it is
+  // derivable here, because the push channel has no value-equality check — a second
+  // notification means the host spoke, even when the value equals the initial `[]`. That
+  // is `useCatalogAnswered()`.
+  //
+  // So: while the catalog is unanswered the row is NEUTRAL — no cause named. `catalogAnswered`
+  // alone is not enough, because falling through to the ✓ arm would trade an under-claim
+  // for an over-claim, and over-claiming is the one R-GA-1 forbids outright.
   let answer: ReachRow;
+  const degrade = toolsSupported ? '' : ' (reads a summary of this wiki, not entries on demand)';
   if (providerState.status === 'unknown') {
     answer = { key: 'answer', label: 'Answer questions about this wiki', state: 'neutral' };
-  } else if (providerState.status === 'not-configured') {
+  } else if (providerState.status === 'ungranted') {
     answer = {
       key: 'answer',
       label: 'Answer questions about this wiki',
       state: 'blocked',
-      cause: 'no model key connected — add one in Settings',
+      cause: UNGRANTED_CHAT_CAUSE,
+      // R3-790: the host's MARK says the grant is missing (provider and key state
+      // unknown — the mark does not tell); the invoke this row's own click makes is
+      // what the host's lazy gesture earns on that same activation.
+      action: 'enable-chat',
     };
   } else if (!chatGranted) {
+    answer = catalogAnswered
+      ? {
+          key: 'answer',
+          label: 'Answer questions about this wiki',
+          state: 'blocked',
+          cause: UNGRANTED_CHAT_CAUSE,
+          // R3-790: the affordance only when a provider RESOLVES (status
+          // 'configured' with the grant missing) — on a keyless host the invoke's
+          // earning path is the SP-7 connect flow, which the not-configured copy
+          // already names (Settings); offering both rows' affordances here would
+          // blur which half is missing.
+          ...(providerState.status === 'configured' ? { action: 'enable-chat' as const } : {}),
+        }
+      : // The host has not answered the catalog. We know nothing about the grant, so name
+        // nothing — the same shape `unknown` uses for an unanswered provider.
+        { key: 'answer', label: 'Answer questions about this wiki', state: 'neutral' };
+  } else if (providerState.status === 'not-configured') {
     answer = {
       key: 'answer',
       label: 'Answer questions about this wiki',
       state: 'blocked',
-      cause: "this Grove wasn't granted chat — reading works as normal",
+      cause: 'no model key connected — add one in Settings',
     };
-  } else {
+  } else if (providerState.status === 'configured') {
     answer = {
       key: 'answer',
-      label: 'Answer questions about this wiki',
+      label: `Answer questions about this wiki${degrade}`,
       state: 'ok',
       chips: ['Summarize this entry', 'What entries are tagged security?'],
     };
+  } else {
+    // The ✓ arm is NARROWED to `configured` and this is the exhaustiveness check. The open
+    // `else` it replaces dates from R3-489, the file's first commit, and survived R3-752
+    // untouched; R3-688 is what closes it. Left open, a FIFTH provider state would land on
+    // ✓ with chips — an unbacked capability claim (R-GA-1) that `tsc` would not mention.
+    // A fourth was just added; the fifth must be a compile error, not a silent grant.
+    const unreachable: never = providerState;
+    void unreachable;
+    answer = { key: 'answer', label: 'Answer questions about this wiki', state: 'neutral' };
   }
 
   // Row 2 — the body source. Both packagings define one post-S2 (the fork's own
@@ -81,18 +232,18 @@ export function computeReachRows({ providerState, chatGranted, writable, sourceS
     : { key: 'draft', label: 'Draft changes', state: 'blocked', cause: 'you’re a reader here' };
 
   // Row 4 — applying. NEVER from this panel (R-GA-3): the widget renders content and
-  // is exactly the broker core_concepts §8a Axis D forbids. The cause is where
-  // changes go, plus the shared-source sentence when trust says others can write.
+  // is exactly the broker core_concepts §8a Axis D forbids. That is a DESTINATION,
+  // not a denial — where to go, stated as one — so it renders `→`, never ✗. The
+  // source-trust sentence does not ride here: it is a property of the source, and it
+  // has its own line under the card (`sourceTrustLine`).
   const apply: ReachRow = {
     key: 'apply',
     label: 'Apply changes',
-    state: 'blocked',
-    cause:
-      'changes open in the editor / workbench, where you confirm them' +
-      (sourceShared ? ' — this repo is treated as if others can write (sole authorship can’t be verified yet), so agent actions go past you first' : ''),
+    state: 'elsewhere',
+    destination: 'in the editor or workbench, where you confirm them',
   };
 
-  return [answer, read, draft, apply];
+  return [packaging, answer, read, draft, apply];
 }
 
 /** The chips the panel shows: exactly the ✓ rows' chips, in card order (R-GA-1 —
@@ -101,6 +252,20 @@ export function reachChips(rows: ReachRow[]): string[] {
   return rows.flatMap((r) => (r.state === 'ok' ? r.chips ?? [] : []));
 }
 
+/**
+ * The source-trust line under the card (R3-752). It is a property of the SOURCE,
+ * not of the apply row it used to be glued to, and its copy is chosen by WHY the
+ * source reads as shared — a cause the reader can act on (R-SP-3), never a
+ * classification (R-SP-6). `null` when the source is not shared: no line, no
+ * reassurance about a regime that is not running.
+ */
+export function sourceTrustLine(shared: boolean, basis: SourceSharedBasis): string | null {
+  if (!shared) return null;
+  return basis === 'git-indeterminate'
+    ? 'anyone who can push to this repo can change what the agent reads'
+    : 'others can change what the agent reads here';
+}
+
 /** R-GA-6's unconditional egress line — shown whenever a provider is bound,
  *  whatever the wiki's trust mode: Q&A composes read + provider egress, and the
  *  confidentiality axis is grant-based, not sharedness-based. */

package/src/lib/shell.ts

@@ -44,8 +44,9 @@ export interface GroveShell {
   /** True when the host REFUSED the last edit request — render it where the
    *  affordance was offered (3.3.1, R3-608); a cancelled request never sets it. */
   editRefused: boolean;
-  /** What a save actually does, for the affordance's title — under dispatch it says that
-   *  proposing a change back to the content repo is not wired yet (R3-266's residual). */
+  /** What a save actually does, for the affordance's title — under dispatch it says
+   *  the write lands in the mounted content and can be proposed back as a PR
+   *  (R3-643 / site-main #576 wired the contribute flow to the content repo). */
   editHint: string;
   siteTitle: string;
   /** Interpreter mode (TRUST_MODES §5): render this entry's body through the