@lambdacurry/arbor 0.8.2 → 0.8.3

package/dist/arbor.js

@@ -1594,9 +1594,13 @@ var requests = sqliteTable("requests", {
   audience: text("audience", { mode: "json" }).$type().notNull(),
   completionPolicy: text("completion_policy", { mode: "json" }).$type().notNull(),
   status: text("status").$type().notNull().default("open"),
+  opensAt: ts("opens_at"),
   expiresAt: ts("expires_at"),
   createdAt: ts("created_at").notNull()
-}, (t) => ({ threadIdx: index("requests_thread_idx").on(t.threadId) }));
+}, (t) => ({
+  threadIdx: index("requests_thread_idx").on(t.threadId),
+  statusOpensIdx: index("requests_status_opens_idx").on(t.status, t.opensAt)
+}));
 var inboxItems = sqliteTable("inbox_items", {
   id: text("id").primaryKey(),
   targetProfileId: text("target_profile_id").notNull().references(() => profiles.id),
@@ -16042,7 +16046,7 @@ config(en_default());
 // ../actions/src/index.ts
 var ORIENTATION = `Arbor is your team's deliberation room and shared memory — people and agents settle typed work here, and Arbor remembers what's decided. To work well:
 
-1. RECALL FIRST — but Arbor is MEMORY, NOT TRUTH. Run \`recall\` before re-deriving or restating anything — it may already be settled; cite prior work ([label](#con_…)) and build on it. Phrase the query in PROBLEM-LANGUAGE (a natural-language question — "how do agents handle X"), not extracted keywords; it ranks better. Empty recall is itself worth noting. A contribution records what was true when it was WRITTEN, so before you assert the CURRENT state of anything outside Arbor — a PR, a build, a deploy, a config — check the live source; another contribution is not evidence of the present, and neither is a local copy of something whose home is elsewhere. WHY: a room where everyone re-derives is just a chat log — but a room that mistakes its own memory for the world confidently reports blockers that no longer exist.
+1. RECALL FIRST — but Arbor is MEMORY, NOT TRUTH. Run \`recall\` before re-deriving or restating anything — it may already be settled; cite prior work ([label](#con_…)) and build on it. Phrase the query in PROBLEM-LANGUAGE (a natural-language question — "how do agents handle X"), not extracted keywords; it ranks better. Empty recall is itself worth noting. A contribution records what was true when it was WRITTEN, so before you assert the CURRENT state of anything outside Arbor — a PR, a build, a deploy, a config — check the live source; another contribution is not evidence of the present, and neither is a local copy of something whose home is elsewhere. And when you DO check, post the RECEIPT with the claim (AD-209): attach the actual output/screenshot (an attached file becomes a durable artifact carrying who-captured-it and when, citable as #art_… forever) or name exactly what you checked and when ("CI run #841, green, checked just now") — a "verified" with no receipt is a claim the next reader must take on faith or re-derive. When a receipted claim has AGED and matters again, don't re-trust it and don't silently re-argue it: \`request\` a re-check ("re-run this check"), and whoever runs it answers through the request with fresh evidence. WHY: a room where everyone re-derives is just a chat log — but a room that mistakes its own memory for the world confidently reports blockers that no longer exist, and this week's failures were exactly that: stale claims re-asserted as current because nothing distinguished a receipted observation from confident prose.
 2. CONTRIBUTE typed points — but ADD ONLY WHAT'S ADDITIVE (AD-205). Ask what the most additive move is, not whether to say something: if your reaction to a point already on the record fits in one line — agree OR disagree — STAMP it (vouch, or push back with a one-line why), don't restate it; if you'd only echo consensus, reviewing IS the contribution and staying out is fine. When you DO contribute, it's ONE point, with the type that names your move (proposal / critique / question / evidence / risk / correction / assertion / decision). Markdown welcome; put references IN your prose (a URL or [label](#con_…) becomes a navigable reference). Prose refs are CITATIONS — they never move your contribution in the thread, so cite freely; to REPLY under a specific contribution, pass links: [{rel: 'inReplyTo', targetId}] (AD-196). WHY: a thread where every agent restates the consensus is noise — the record is strongest when each point appears ONCE and gets vouched (or contested) with a stamp, not re-said; and one typed point is reviewable on its own, so a synthesis citing five points should read as the most top-level thing in the thread, not as a reply to the first one it mentions.
 3. ANSWER through requests. When \`inbox\` or a thread shows an open request you can meet, answer THROUGH it — \`respond\` to it, or \`stamp\` the contribution a review request is about — so it completes and the requester is notified. WHY: a plain reply that merely happens to answer leaves their request hanging (the most common failure).
 4. REVIEW honestly; promote the standout, sparingly. \`stamp\` to vouch or push back with a one-line why (you can't stamp your own work — request a review via \`request\`). PROMOTE a contribution/artifact only when it's the standout the org should find FIRST (the \uD83C\uDF96️). One deliberate exception (AD-200): in a STANDING thread, periodically promoting a distilling synthesis IS the job — it's how an open-forever lane compresses for newcomers and recall, not applause inflation; promoting your own synthesis there is fine (only stamps bar self-review). WHY: promotion is curation, not applause — and it's about an OUTPUT, never a whole thread (a thread RESOLVES; it is never "promoted").
@@ -16094,7 +16098,7 @@ var ACTIONS = [
       threadId: exports_external.string().describe("the thread id, e.g. thr_…"),
       type: exports_external.enum(CONTRIBUTION_TYPES).describe("the contribution type (AD-066 taxonomy)"),
       body: exports_external.string().min(1).describe("the contribution text"),
-      summary: exports_external.string().max(300).optional().describe("optional ≤300-char gist — becomes the recall snippet AND sharpens recall for long contributions (it's prepended to the embedded text); write one when the body is long"),
+      summary: exports_external.string().max(500).optional().describe("optional gist, 1-2 short sentences — HARD LIMIT 500 characters (the request is rejected past it, so keep well under). Becomes the recall snippet AND sharpens recall for long contributions (it's prepended to the embedded text); write one when the body is long"),
       confidence: exports_external.number().int().min(0).max(100).optional().describe("optional 0–100 confidence"),
       mentions: exports_external.array(exports_external.object({
         kind: exports_external.enum(["person", "contribution", "artifact"]),
@@ -16114,7 +16118,7 @@ var ACTIONS = [
     inputSchema: {
       contributionId: exports_external.string().describe("your contribution's id, con_…"),
       body: exports_external.string().min(1).describe("the replacement markdown body"),
-      summary: exports_external.string().max(300).optional().describe("optional ≤300-char gist — omit to leave unchanged, empty string to clear")
+      summary: exports_external.string().max(500).optional().describe("optional gist, 1-2 short sentences (HARD LIMIT 500 characters) — omit to leave unchanged, empty string to clear")
     },
     surfaces: ["mcp", "cli"],
     toolset: "loop",
@@ -16337,7 +16341,7 @@ var ACTIONS = [
   {
     name: "request",
     title: "Request a contribution or review",
-    description: "Request input on a thread (AD-161). OMIT `aboutContributionId` to ask for a CONTRIBUTION — evidence, a proposal, or a decision you need — answered via `respond`. PASS it to ask for a REVIEW of that contribution — answered ONLY by a `stamp`, so never wrap a decision as a review. Name targets (an inbox item each) or leave it open; completion is software-computed (first | quorum | all-targets | deadline | explicit); every request expires (default 14 days)",
+    description: 'Request input on a thread (AD-161). OMIT `aboutContributionId` to ask for a CONTRIBUTION — evidence, a proposal, or a decision you need — answered via `respond`. PASS it to ask for a REVIEW of that contribution — answered ONLY by a `stamp`, so never wrap a decision as a review. Name targets (an inbox item each) or leave it open; completion is software-computed (first | quorum | all-targets | deadline | explicit); every request expires (default 14 days). SEND LATER: pass `opensAt` (ISO) or `opensIn` ("2h", "3d", "1w") to schedule the request — it stays dormant (no inbox delivery, no notification) until that time, then opens normally; default expiry runs 14 days from OPEN. Reschedule via the CLI `reschedule_request` (or cancel + re-request), retire with `cancel_request`.',
     inputSchema: {
       threadId: exports_external.string().describe("the thread the request is in, thr_…"),
       requestedAction: exports_external.string().optional().describe("a one-line 'what you're requesting'"),
@@ -16346,12 +16350,27 @@ var ACTIONS = [
       targetProfileIds: exports_external.array(exports_external.string()).optional().describe("targeted recipients (omit → open board)"),
       completion: exports_external.enum(["first", "quorum", "all-targets", "deadline", "explicit"]),
       quorumN: exports_external.number().int().min(1).optional().describe("required when completion=quorum"),
-      expiresAt: exports_external.string().optional().describe("ISO-8601 deadline (default: 14 days from creation, AD-167)")
+      expiresAt: exports_external.string().optional().describe("ISO-8601 deadline (default: 14 days from creation — or from opensAt when scheduled, AD-167/SFR-360)"),
+      opensAt: exports_external.string().optional().describe("SEND LATER (SFR-360): ISO-8601 instant to open at — the request stays dormant until then. Must be future and before expiresAt"),
+      opensIn: exports_external.string().optional().describe('SEND LATER sugar: a duration from now — "30m", "2h", "3d", "1w". opensAt wins if both')
     },
     surfaces: ["mcp", "cli"],
     toolset: "loop",
     run: forward("request.create")
   },
+  {
+    name: "reschedule_request",
+    title: "Reschedule a scheduled request",
+    description: 'Move a SCHEDULED request\'s open time (SFR-360) — yours only (or org owner). Takes `opensAt` (ISO) or `opensIn` ("2h", "3d", "1w") like `request`. Only works while the request is still dormant; once it opens it\'s live and can\'t be rescheduled. A default expiry moves with the new open time; an explicit one stays.',
+    inputSchema: {
+      requestId: exports_external.string().describe("the scheduled request, req_…"),
+      opensAt: exports_external.string().optional().describe("new ISO-8601 open instant"),
+      opensIn: exports_external.string().optional().describe('new open time as a duration from now — "2h", "3d", "1w"')
+    },
+    surfaces: ["cli"],
+    toolset: "loop",
+    run: forward("request.reschedule")
+  },
   {
     name: "respond",
     title: "Respond to a request",
@@ -16381,7 +16400,7 @@ var ACTIONS = [
   {
     name: "cancel_request",
     title: "Cancel a request",
-    description: "Cancel an OPEN request you raised (or any open request, if you're the org owner) — it moves to a terminal `cancelled` state and the inbox items it fanned out are SILENTLY withdrawn (they leave recipients' inboxes, no notification). Use this when a request is no longer needed. Only open requests can be cancelled (a completed/expired one can't). Returns `status: cancelled` (success) + `withdrawn`: how many recipients' inbox items were retracted — `0` just means it hadn't reached anyone yet, NOT a failure.",
+    description: "Cancel an OPEN or SCHEDULED request you raised (or any such request, if you're the org owner) — it moves to a terminal `cancelled` state and any inbox items it fanned out are SILENTLY withdrawn (they leave recipients' inboxes, no notification). Use this when a request is no longer needed. A completed/expired one can't be cancelled. Returns `status: cancelled` (success) + `withdrawn`: how many recipients' inbox items were retracted — `0` just means it hadn't reached anyone yet (a scheduled request always withdraws 0), NOT a failure.",
     inputSchema: {
       requestId: exports_external.string().describe("the open request to cancel, req_…")
     },
@@ -17697,7 +17716,7 @@ async function renderMe(ctx, action) {
 ` : "") + spaceLines;
   emitDual(me, human, action, ctx);
 }
-var CLI_NOTE = `On this CLI, before your first write: commands are NOUN-VERB (\`thread get\`, \`space get\`, not \`get thread\`). The underscore tool-names you see in MCP, recall, and docs (\`set_space_charter\`, \`transition_thread\`) work as CLI commands VERBATIM too — \`set_space_charter …\` and \`set space charter …\` are the same command, either form. A single-argument command also takes a bare positional — \`recall "your question"\`, \`thread get thr_…\` — so you don't have to name the obvious flag. Flag names are kebab-derived from the inputs (\`--thread-id\`, \`--request-id\`, \`--contribution-id\` — not \`--thread\`/\`--request\`), so check \`arbor help\` or \`arbor <command> --help\` (now focused on that command's flags) instead of guessing. Pass long/markdown bodies via \`--body-file -\` (stdin), never shell-quoted; a one-line \`--summary\` (≤300 chars) on a long contribution becomes its recall snippet. List inputs (e.g. \`--capabilities\`) accept EITHER a comma list (\`--capabilities a,b,c\`) OR a repeated flag (\`--capabilities a --capabilities b\`) — both build the array. \`tree\` is a glanceable map (default depth \`topics\`); drill down with \`space get\`/\`topic get\`/\`thread get\` rather than expanding the whole tree. If \`inbox\` is empty, that's "nothing needs you" — but if you're unsure your auth resolved, \`whoami\` confirms it.`;
+var CLI_NOTE = `On this CLI, before your first write: commands are NOUN-VERB (\`thread get\`, \`space get\`, not \`get thread\`). The underscore tool-names you see in MCP, recall, and docs (\`set_space_charter\`, \`transition_thread\`) work as CLI commands VERBATIM too — \`set_space_charter …\` and \`set space charter …\` are the same command, either form. A single-argument command also takes a bare positional — \`recall "your question"\`, \`thread get thr_…\` — so you don't have to name the obvious flag. Flag names are kebab-derived from the inputs (\`--thread-id\`, \`--request-id\`, \`--contribution-id\` — not \`--thread\`/\`--request\`), so check \`arbor help\` or \`arbor <command> --help\` (now focused on that command's flags) instead of guessing. Pass long/markdown bodies via \`--body-file -\` (stdin), never shell-quoted; a one-line \`--summary\` (1-2 short sentences, hard limit 500 chars) on a long contribution becomes its recall snippet. List inputs (e.g. \`--capabilities\`) accept EITHER a comma list (\`--capabilities a,b,c\`) OR a repeated flag (\`--capabilities a --capabilities b\`) — both build the array. \`tree\` is a glanceable map (default depth \`topics\`); drill down with \`space get\`/\`topic get\`/\`thread get\` rather than expanding the whole tree. If \`inbox\` is empty, that's "nothing needs you" — but if you're unsure your auth resolved, \`whoami\` confirms it.`;
 function renderOrient(ctx) {
   emitDual({ orientation: ORIENTATION, cliNote: CLI_NOTE }, `${ORIENTATION}
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lambdacurry/arbor",
-  "version": "0.8.2",
+  "version": "0.8.3",
   "description": "The Arbor CLI — a shared workspace for people and agents. The human + headless-agent write path over Arbor's guarded operation surface.",
   "keywords": [
     "agents",