frontpage-host 0.3.9 → 0.3.10

package/AGENTS.md

@@ -276,6 +276,8 @@ Currency is owned by Payments and read by Store, Offerings and invoices. Offerin
 
 AI Support Agent setup: enable website chat through Inbox settings; Frontpage public sites load it without a file edit or publish. Other websites use the explicit widget snippet. `get website` reports managed addresses/public availability. Website knowledge is default-on, read-only public retrieval; `websiteKnowledge.enabled:false` disables it. Source toggles, attention filtering and receipt cursors use the documented versioned resources.
 
+AI Support Agent recovery: CUSTOMER-INBOX.md explains bounded read continuations and the existing jobs/<id>/resume operation for context pauses. Read current updatedAt and reuse the exact request key/body after a lost response. Retry cannot bypass screening, stale work, unknown paid requests or uncertain email sends. Working context is separate from credits.
+
 AI Support Agent team guidance: administrators may set `guidanceEmail` through the versioned settings write. The owner toolbar has the same setting. Teammates reply as the configured mailbox; no teammate account or separate mailbox activation is needed. Do not set it to the public Inbox address. Conversation management uses close/reopen; do not use read/trash/restore.
 
 ## Emails workspace workflows

package/CUSTOMER-INBOX.md

@@ -30,7 +30,13 @@ Setup defaults private guidance to the verified administrator who requested it.
 
 Settings allow `enabled`, `instructions`, `tools`, `maxSteps` (1–12), `maxOutputTokens` (256–2400), and `maxInputBytes` (4000–64000). Supported tools: `knowledge_read`, `knowledge_update`, `customer_read`, `verify_customer`, `customer_update`, `notes_read`, `note_add`, `ask_owner`, `reply_customer`, `wait`, `finish`. Keep `wait` and `finish` enabled. Mailbox verification remains available whenever customer-data tools are enabled; it cannot be bypassed by settings. Defaults: 8 steps, 1200 output tokens per call, 32000 bytes of public text. Model and billing multiplier are fixed. Only an administrator can choose the private guidance mailbox; the support agent cannot choose recipients. The worker also enforces platform concurrency, deadlines and input limits.
 
-`maxContextTokens` (12000–96000, default 32000) sets a conservative per-call input plus output allowance. UTF-8 text bytes, tool definitions, instructions, transcript and image capacity count toward admission; exact provider tokens still determine billing. Optional older history and preloaded notes drop first. Essential evidence is never clipped: if it cannot fit, the job pauses for review. History stays stored without a lifetime message-count stop. The agent receives up to 40 recent messages plus required current input and owner guidance, and can retrieve older messages in bounded pages under its existing `knowledge_read` grant.
+`maxContextTokens` (12000–96000, default 32000) keeps its existing API name but is a conservative working-size allowance, not an exact provider-token count or AI credit balance. Admission measures UTF-8 request text and tools in the provider's wire representation once, with reply and image reserves. It never counts both canonical assistant content and provider replay. Exact provider receipts still determine billing. Optional older history, preloaded notes and expendable read results can leave the working window; current input, required owner guidance, safety instructions and completed action outcomes remain intact. Omitted reads are labeled and must be reread before use. Full history and sources remain saved. Screening uses complete required public input and removes only optional older context. Required input that cannot fit needs owner review.
+
+Retrieval calls share the remaining allowance within each tool batch. Website reads return smaller chunks when needed, with their existing cursors. Knowledge search previews return accurate `hasMore` flags. `knowledge_read({id,cursor?})` and `knowledge_read({sourceId,cursor?})` return bounded bodies with `offset`, `truncated` and `nextCursor`; continue the identical ID. Knowledge cursors bind the site, ID and content version. Changed or expired content requires restarting without a cursor. Management source/note reads keep their full-body contract.
+
+`retrieval_budget_exhausted` tells the support agent to use one focused read or answer from sufficient evidence still present, rather than repeat the same broad batch. Local preparation removes optional context before another paid call. An unresolved paid call or email send still requires reconciliation, never a blind inference retry.
+
+For a saved `input_budget` or `context_limit` pause, the workspace offers **Retry reply**. The existing `jobs/<id>/resume` operation rebuilds the working transcript for the current unanswered request and preserves paid-attempt identities and the remaining step allowance. Read the job, then POST its exact `updatedAt` with a stable request key. Retry is idempotent. New input, a completed answer, changed ownership/settings, unresolved attempts or uncertain sends prevent an unsafe restart. `essential_context_limit`, screening blocks and unknown paid outcomes do not gain a retry bypass. Take over to handle the customer when human review is required.
 
 `reviewAfterReplies` (0–200, default 0/off) and `reviewWindowHours` (1–168, default 24) request private owner guidance after that many recent automated replies since the last guidance. Pending requests deduplicate. This is a rolling work control, not a cap on returning customers or retained history.
 
@@ -204,7 +210,7 @@ Use a focused query such as `repair pricing`. Follow a search `nextCursor` with
 
 Expected tool failures return `error`, `message`, `nextAction` and `retryable`. `website_cursor_invalid` explains how to restart the request. Page failures identify whether another result or a bounded retry is useful. `retryable` refers to a transient failure, not permission to repeat invalid input. The agent corrects recoverable inputs before escalating; revoked permissions and stale jobs still stop execution. No source files, private pages, credentials or external-site override become available.
 
-`knowledge_read({query})` searches saved notes and enabled approved sources by keywords. Previews include matching passages and `hasMore:{notes,sources}`. When either flag is true, refine the subject to find the relevant item; do not treat the preview list as complete. Use `knowledge_read({id})` for a complete note or `knowledge_read({sourceId})` for a complete enabled source. Support source searches filter before bounding results, and direct source reads do not depend on the management list's page. Existing approval, freshness, locks and learning rules remain in force.
+`knowledge_read({query})` searches saved notes and enabled approved sources by keywords. Previews include matching passages and `hasMore:{notes,sources}`. When either flag is true, refine the subject to find the relevant item; do not treat the preview list as complete. Use `knowledge_read({id,cursor?})` or `knowledge_read({sourceId,cursor?})` to read bounded body chunks, continuing the returned cursor as needed. Support source searches filter before bounding results, and direct source reads do not depend on the management list's page. Existing approval, freshness, locks and learning rules remain in force.
 
 The widget shows sending/upload feedback and animated dots while an automatic response is queued or running. The existing chat session response adds `activity` with `processing`, `awaiting_team`, `paused` or `idle`; it exposes no private job details. Awaiting owner guidance replaces the dots with a team-wait message; paused work does not keep typing. These are customer-safe statuses, not private tool results. Private guidance requests include customer-visible conversation history, with a full Inbox link and a labeled recent excerpt for oversized threads. Reply privately with new guidance; quoted history is context, not new instructions. Recipients, takeover controls, verification and email delivery rules are unchanged.
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "frontpage-host",
-  "version": "0.3.9",
-  "description": "Frontpage CLI — manage your site from the terminal",
+  "version": "0.3.10",
+  "description": "Frontpage CLI \u2014 manage your site from the terminal",
   "type": "module",
   "bin": {
     "frontpage": "dist/bin.js"