@sellable/mcp 0.1.463 → 0.1.464

package/dist/tools/prompts.js

@@ -373,7 +373,7 @@ export function getPostFindLeadsScoutRegistry() {
             codex: 'After confirm_lead_list copies source rows and the initial campaign-table execution slice exists, ask the filter-choice question immediately. Do not spawn anything before that question. After the answer, launch only Message Drafting. The filter-choice answer is the post-import user gate for this single worker; do not ask another question about starting it in step-wise or YOLO mode. The registry lookup is not a launch: after get_post_find_leads_scout_registry, immediately invoke Task/spawn_agent or the host background-agent mechanism before loading filter-leads.md, before saving rubrics, and before treating skip-filters as ready for message review. Both choices route through this kickoff; do not let filters_skipped jump straight from filter-choice to message-generation. If filters are chosen, the parent stays on Filter Rules and drafts/saves rubrics with MCP tools while Message Drafting runs in the background. If filters are skipped, move to Messages/message review only after Message Drafting has started or is ready; update_campaign(currentStep=messages) is not proof of launch. If the named Message Drafting custom agent is unavailable, spawn a generic gpt-5.5 xhigh Message Drafting background agent with the same lean campaign/table basis. When the background worker starts, persist workerDetails.messageDraftBuilder with statusSource "branch", status "branch-running", runId, startedAt, updatedAt, basisToken when known, and basis containing campaignId, selectedLeadListId, workflowTableId, filterChoice, and reviewBatchRowHash or reviewBatchRowIds; workerStatuses.messageDraftBuilder may be "running" as a simple badge only. Never put rich proof under workerStatuses and never use workerStatuses.messageDrafting. If no background-agent tool is callable, start the same full message branch inline before filter drafting or before skip-filter message review, record workerDetails.messageDraftBuilder with statusSource "parent-thread-fallback" and status "fallback-active", and require the same live context, prompt, assets, and validation gate before message review; do not wait until filters are saved and then call the registry.',
             claude: "After confirm_lead_list copies source rows and the initial campaign-table execution slice exists, ask the filter-choice question immediately. Do not invoke any Task/Agent before that question. After the answer, invoke only Message Drafting. If filters are chosen, parent drafts/saves rubrics with MCP tools while Message Drafting runs, asks filter approval, then joins Message Drafting. If filters are skipped, invoke only Message Drafting and move to Messages/message review.",
             parentThreadRule: 'Named agents are optional acceleration, but message drafting is not optional. The only normal background worker is Message Drafting. The filter-choice answer is the campaign-scoped go-ahead for this single post-import worker; do not ask another question to start it in step-wise or YOLO mode. If a named agent is unavailable, use a generic gpt-5.5 xhigh Message Drafting background agent. source work and filter work stay in the parent thread with MCP tools. If post-find-leads-message-scout is available, run it as the background Message Draft Builder after the filter-choice answer. The registry lookup is not a launch: get_post_find_leads_scout_registry only identifies the worker, and Message Drafting counts as started only after Task/spawn_agent or the host background-agent tool is invoked, or after the parent begins the same full message branch inline because no background-agent tool is callable. This launch must happen before loading filter-leads.md, save_rubrics, filter approval, or skip-filter message review; currentStep=messages is not proof of launch. If post-find-leads-message-scout is absent, do not customer-surface install status. Do not silently treat message drafting as started; the main thread must either launch the background worker or execute the same message branch from CampaignOffer state, selected source state, workflowTableId, and initial campaign-table execution slice rows. For a spawned worker, record workerDetails.messageDraftBuilder with statusSource branch / status branch-running, runId, startedAt, updatedAt, and basis containing campaignId, selectedLeadListId, workflowTableId, filterChoice, and reviewBatchRowHash or reviewBatchRowIds. workerStatuses.messageDraftBuilder is optional simple badge text only ("running", "ready", "blocked", "idle"); never put runId/statusSource/basis under workerStatuses and never use workerStatuses.messageDrafting. If no background-agent tool is callable, start that same full message branch inline before filter drafting or before skip-filter message review, record workerDetails.messageDraftBuilder with statusSource parent-thread-fallback / status fallback-active then ready, and require the same live context, prompt, assets, and validation gate before message review; do not report that as a background worker failure. If neither branch nor inline fallback can run, return blocked/retry-needed; do not wait until filters are saved and then call the registry. The Message Drafting handoff must be lean. Do not paste copied row counts, brief hashes, review-batch hashes, full reviewBatchRowIds, broad row data, or local debug artifacts into the spawn prompt. Local markdown/json files are not normal-path inputs. The filter-choice question is the first post-import user gate; do not load post-lead registries or filter references before it. Message drafting starts after the filter-choice answer, must load get_subskill_prompt({ subskillName: "generate-messages" }), and must load every required message asset named by generate-messages Mode 0 through get_subskill_asset before drafting. Reference Asset Loading means loading the required pre-draft reference pack before drafting; return blocked/retry-needed if required assets cannot be loaded; load ai-tells.md because it is never optional. The branch or parent-thread fallback loads the full generate-messages prompt and every referenced asset through get_subskill_asset. After generating/revising the candidate and before returning ready, must load get_subskill_prompt({ subskillName: "create-campaign-v2-validation" }) as the final internal validation gate, must read live campaign table state through scoped MCP/product tools, and must reject mismatched selectedLeadListId/workflowTableId/campaign/workspace input. Do not block when filters were chosen but leadScoringRubrics are not yet visible in the branch read; the parent owns save_rubrics and filter approval in parallel, so Message Drafting should return status ready with basisStatus usable_initial when campaign/list/table identity and the non-empty execution slice match. Do not use any alternate, local-artifact, or examples-only message prompt. User copy feedback, message QA, or rewrite requests before approve-message must be routed back to Message Drafting with the current recommendation, lean campaign/table basis, and latest user text; the parent must not rewrite or QA the template from memory and must not call update_campaign_brief before approve-message. The worker validates internally and returns only templateRecommendation, tokenFillRules, renderedGoodSample, status, approveOrReviseRecommendation, validationStatus, outputAt, outputHash, and blocked/retry detail. Do not render renderedFallbackSample, risk notes, or a qaReceipt on the normal happy path. On the filter path, save_rubrics keeps the browser on Filter Rules after save_rubrics so the user can approve the saved criteria; after saved-filter approval, move to Filter Leads with currentStep=apply-icp-rubric whether Message Drafting is ready or still running. Wait there for message approval. Enrichment, filtering, Generate Message cells, sender setup, sequence attach, and launch wait for template approval on the Use Template path. On the skip path, move to Messages/message review after Message Drafting has started or is ready and wait for message approval before enrichment or Settings. Do not render message review from checklist or shortcut instructions; message review requires a messageDraftRecommendation whose basis proves the generate-messages prompt, required message assets, and validation gate ran for the current campaign/table execution slice. Do not automatically rerun Message Drafting after filters/enrichment finish; show the initial draft by default and offer an enriched rewrite only with explicit user opt-in. Handoff and recommendation output are Markdown with labeled fields, not raw JSON.',
-            prepareMessagesRule: `Default create-campaign stays on the existing reviewBatchLimit:15 first campaign-table execution slice. For plain post-mint fill/load/refill requests, load get_subskill_prompt({ subskillName: "refill-sends-workflow" }), then call resolve_campaign_fill_route({ intent:"plain" }), list_senders, and get_campaign_refill_state for enough candidate campaigns to identify the campaign that most recently had scheduler-owned sends for the relevant sender set before any mutation. Route outcomes are route:"evergreen_horizon", route:"active_campaigns", and route:"ask_create"; if a plain managed-waterfall route has archived/completed skipped slots or does not cover the named sender set, immediately refetch with resolve_campaign_fill_route({ intent:"active" }) and inspect current dashboard-active ACTIVE/PAUSED campaign-backed sequence campaigns before declaring a sender blocked; stay in the same campaignOfferId/campaignId context after minting. Plain fill is not an alias for fill_campaign_horizon or campaign creation. fill_campaign_horizon is evergreen-only and is not the generic regular-campaign fill path. Campaign creation is allowed only after route:"ask_create" and explicit user selection; it is never the default response to plain fill. Treat "fill up/load sends" as capacity-fill preparation, and treat "refill senders", "fill senders", "max out senders", and "load everyone up" as sender-scoped capacity-fill preparation. For sender-scoped requests with no named senders, --yolo means all eligible healthy senders enrolled in active campaign-backed sequence campaigns in the active workspace; without --yolo, ask which eligible enrolled senders to refill before choosing campaigns or mutating. In non-yolo interactive Codex or Claude Code sessions, the sender/campaign/action approval packet must be asked through request_user_input/AskUserQuestion with exactly Accept and Decline; do not use plain chat approval. The approval question body must show workspace, sender scope, a campaign-by-campaign table, exact ids/caps/dates, side effects, forbidden actions, and stop condition before mutation. Default --yolo horizon is a two-send-day horizon: two sender-local send days, skipping no-send-hour days. For campaign-scoped fill/refill, select the best recent-send campaign and calculate the bounded gap from healthy sender daily capacity minus projected coverage (actual sent plus future scheduler-owned scheduled sends) and ready-to-schedule rows, then prepare only that gap. For sender-scoped fill/refill, calculate the bounded gap per eligible sender across active enrolled campaigns, counting actual sent coverage, future scheduled rows, and ready-to-schedule rows across those campaigns, then choose the best same-sender campaign to fill each sender gap: prefer recent/future scheduler-owned sends for that sender, then strongest recent result evidence, then source health. Maintain a horizon saturation ledger per selected sender with selected send days, gross capacity, actual sent counts, future scheduler-owned scheduled counts, projected coverage, ready-to-schedule buffer, remaining projected gap, paid-InMail threshold feasibility, and the next MCP primitive. Do not stop after filling only one sender when the request was sender-scoped. In --yolo, continue through every safe selected sender/campaign action covered by the rendered packet, reread after each terminal apply/prep/start result, recompute the horizon saturation ledger, and keep going until projected coverage (sent + scheduled) fills the requested horizon or a concrete blocker/timeout is proven. Ready-to-schedule rows are buffer, not completion; if ready covers the projected gap but scheduled cells do not, run a bounded scheduler-settle loop and report awaiting_scheduler_after_ready_buffer as not complete when timeout is reached. Paid-InMail threshold changes and connection-campaign creation are explicit continuation options, never --yolo side effects. When no safe in-packet action remains return concrete continuation options with campaign names, exact ids, and which options require a new approval packet. Use the refill workflow to decide whether to enrich/prep more rows in that same recent-send/best-result campaign, add/import more rows to that same campaign/source path, use a different existing campaign only when the selected same-sender lane is blocked/exhausted/already loaded while the sender still has a gap, or ask what to create. Do not create warm-post-engager side campaigns. Surface sender-health and paid-InMail threshold blockers separately from prepared/approved/scheduled counts. User-facing refill decisions must be campaign-name-first and sender-name-first, with ids as proof/execution targets only. For already-running regular campaigns that need Signal Discovery source replenishment, use the guarded currentStep clear with clearCurrentStepIfMatches:"running", campaign-scoped provider prompt/search/select, and import_leads with the existing sourceLeadListId when a newly approved selected-post scrape would otherwise return reusedExistingSourceList. Before prep, inspect get_campaign_refill_state.preparationFrontier. If hasLaterPreparedIsland:true or earliestUnpreparedRow exists before later successful enrichment, use rowSelector:{type:"needsEnrichment"} with columnRole:"enrich" in table-position order or start_campaign_message_preparation adaptive defaults; do not use the UI Jump anchor or needsGeneratedMessage as the refill cursor. After confirm_lead_list copies rows into an existing table, avoid fixed maxRowsToCheck:100; if confirm_lead_list returns USER_ADDED_ROWS_LIMIT_EXCEEDED, create a bounded same-source split from selectedLeadListId with get_rows_minimal/load_csv_linkedin_leads, confirm that smaller source list into the same campaign, and inspect reviewBatch only as diagnostics unless the approved packet explicitly prioritizes the just-copied split and earlier needsEnrichment rows are exhausted, dependency-blocked, or excluded. For approval diagnostics use rowSelector:{type:"needsApproval"} after current generated messages exist; mutating Approved cells still requires approvalMode:approve and exact bounded approval. Do not interpret checkedRows as enriched rows; it is only the table cursor. Prepared, approved, and ready_to_schedule rows are intermediate states; never call them scheduled unless a re-read proves scheduler-owned scheduled cells with non-null scheduledFor contribute to projected coverage, and never call them complete unless a final re-read proves projected coverage fills the requested horizon. Before source import, prep, approval, or selected paused campaign start, require exact visible approval or --yolo packet auto-accept and a fresh get_campaign_refill_state reread; stop if freshness.stateHash or exact ids changed. For "approve X messages", use approvalMode:approve only when explicitly requested. For "schedule X sends" or "fill sender sends", approve only when explicitly requested, then re-read campaign/table scheduled counts; if projected coverage does not fill the horizon, report prepared/approved/ready - awaiting scheduler instead of success. Do not call start_campaign as part of fill/schedule horizon. start_campaign is allowed in refill only for exact selected PAUSED, dashboard-active, campaign-backed sequence targets named in the bounded packet, and the packet must state that starting can let the product scheduler schedule/send approved eligible sequence actions. Never start unrelated, archived, completed, draft, direct, or non-selected campaigns, never broad approve-all, and never use direct scheduler writes. campaignId is CampaignOffer.id. If the user asks to stop preparation, the target is wrong, or status shows the wrong campaign/table, use cancel_campaign_message_preparation only for the exact active job. Low-level selectors are diagnostics and recovery only for this lane.`,
+            prepareMessagesRule: `Default create-campaign stays on the existing reviewBatchLimit:15 first campaign-table execution slice. For plain post-mint fill/load/refill requests, load get_subskill_prompt({ subskillName: "refill-sends-workflow" }), then call resolve_campaign_fill_route({ intent:"plain" }), list_senders, and get_campaign_refill_state for enough candidate campaigns to identify the campaign that most recently had scheduler-owned sends for the relevant sender set before any mutation. Route outcomes are route:"evergreen_horizon", route:"active_campaigns", and route:"ask_create"; if a plain managed-waterfall route has archived/completed skipped slots or does not cover the named sender set, immediately refetch with resolve_campaign_fill_route({ intent:"active" }) and inspect current dashboard-active ACTIVE/PAUSED campaign-backed sequence campaigns before declaring a sender blocked; stay in the same campaignOfferId/campaignId context after minting. Plain fill is not an alias for fill_campaign_horizon or campaign creation. fill_campaign_horizon is evergreen-only and is not the generic regular-campaign fill path. Campaign creation is allowed only after route:"ask_create" and explicit user selection; it is never the default response to plain fill. Treat "fill up/load sends" as capacity-fill preparation, and treat "refill senders", "fill senders", "max out senders", and "load everyone up" as sender-scoped capacity-fill preparation. For sender-scoped requests with no named senders, --yolo means all eligible healthy senders enrolled in active campaign-backed sequence campaigns in the active workspace; without --yolo, ask which eligible enrolled senders to refill before choosing campaigns or mutating. In non-yolo interactive Codex or Claude Code sessions, render the sender/campaign/action approval packet in normal chat using Markdown first, then ask a compact request_user_input/AskUserQuestion with exactly Accept and Decline; do not use plain chat approval, and do not duplicate the full packet inside the approval panel. The chat-rendered approval packet must show workspace, sender scope, a campaign-by-campaign table, exact ids/caps/dates, side effects, forbidden actions, and stop condition before mutation. Default --yolo horizon is a two-send-day horizon: two sender-local send days, skipping no-send-hour days. For campaign-scoped fill/refill, select the best recent-send campaign and calculate the bounded gap from healthy sender daily capacity minus projected coverage (actual sent plus future scheduler-owned scheduled sends) and ready-to-schedule rows, then prepare only that gap. For sender-scoped fill/refill, calculate the bounded gap per eligible sender across active enrolled campaigns, counting actual sent coverage, future scheduled rows, and ready-to-schedule rows across those campaigns, then choose the best same-sender campaign to fill each sender gap: prefer recent/future scheduler-owned sends for that sender, then strongest recent result evidence, then source health. Maintain a horizon saturation ledger per selected sender with selected send days, gross capacity, actual sent counts, future scheduler-owned scheduled counts, projected coverage, ready-to-schedule buffer, remaining projected gap, paid-InMail threshold feasibility, and the next MCP primitive. Do not stop after filling only one sender when the request was sender-scoped. In --yolo, continue through every safe selected sender/campaign action covered by the rendered packet, reread after each terminal apply/prep/start result, recompute the horizon saturation ledger, and keep going until projected coverage (sent + scheduled) fills the requested horizon or a concrete blocker/timeout is proven. Ready-to-schedule rows are buffer, not completion; if ready covers the projected gap but scheduled cells do not, run a bounded scheduler-settle loop and report awaiting_scheduler_after_ready_buffer as not complete when timeout is reached. Paid-InMail threshold changes and connection-campaign creation are explicit continuation options, never --yolo side effects. When no safe in-packet action remains return concrete continuation options with campaign names, exact ids, and which options require a new approval packet. Use the refill workflow to decide whether to enrich/prep more rows in that same recent-send/best-result campaign, add/import more rows to that same campaign/source path, use a different existing campaign only when the selected same-sender lane is blocked/exhausted/already loaded while the sender still has a gap, or ask what to create. Do not create warm-post-engager side campaigns. Surface sender-health and paid-InMail threshold blockers separately from prepared/approved/scheduled counts. User-facing refill decisions must be campaign-name-first and sender-name-first, with ids as proof/execution targets only. For already-running regular campaigns that need Signal Discovery source replenishment, use the guarded currentStep clear with clearCurrentStepIfMatches:"running", campaign-scoped provider prompt/search/select, and import_leads with the existing sourceLeadListId when a newly approved selected-post scrape would otherwise return reusedExistingSourceList. Before prep, inspect get_campaign_refill_state.preparationFrontier. If hasLaterPreparedIsland:true or earliestUnpreparedRow exists before later successful enrichment, use rowSelector:{type:"needsEnrichment"} with columnRole:"enrich" in table-position order or start_campaign_message_preparation adaptive defaults; do not use the UI Jump anchor or needsGeneratedMessage as the refill cursor. After confirm_lead_list copies rows into an existing table, avoid fixed maxRowsToCheck:100; if confirm_lead_list returns USER_ADDED_ROWS_LIMIT_EXCEEDED, create a bounded same-source split from selectedLeadListId with get_rows_minimal/load_csv_linkedin_leads, confirm that smaller source list into the same campaign, and inspect reviewBatch only as diagnostics unless the approved packet explicitly prioritizes the just-copied split and earlier needsEnrichment rows are exhausted, dependency-blocked, or excluded. For approval diagnostics use rowSelector:{type:"needsApproval"} after current generated messages exist; mutating Approved cells still requires approvalMode:approve and exact bounded approval. Do not interpret checkedRows as enriched rows; it is only the table cursor. Prepared, approved, and ready_to_schedule rows are intermediate states; never call them scheduled unless a re-read proves scheduler-owned scheduled cells with non-null scheduledFor contribute to projected coverage, and never call them complete unless a final re-read proves projected coverage fills the requested horizon. Before source import, prep, approval, or selected paused campaign start, require exact visible approval or --yolo packet auto-accept and a fresh get_campaign_refill_state reread; stop if freshness.stateHash or exact ids changed. For "approve X messages", use approvalMode:approve only when explicitly requested. For "schedule X sends" or "fill sender sends", approve only when explicitly requested, then re-read campaign/table scheduled counts; if projected coverage does not fill the horizon, report prepared/approved/ready - awaiting scheduler instead of success. Do not call start_campaign as part of fill/schedule horizon. start_campaign is allowed in refill only for exact selected PAUSED, dashboard-active, campaign-backed sequence targets named in the bounded packet, and the packet must state that starting can let the product scheduler schedule/send approved eligible sequence actions. Never start unrelated, archived, completed, draft, direct, or non-selected campaigns, never broad approve-all, and never use direct scheduler writes. campaignId is CampaignOffer.id. If the user asks to stop preparation, the target is wrong, or status shows the wrong campaign/table, use cancel_campaign_message_preparation only for the exact active job. Low-level selectors are diagnostics and recovery only for this lane.`,
         },
     };
 }

package/dist/tools/refill-sends.js

@@ -166,10 +166,10 @@ export function refillSendsCommand(input = {}) {
             "Do not complete a fill/schedule request until a final get_refill_target_plan or refill-state reread proves projected coverage (sent + scheduled) fills the horizon, or report a concrete blocker such as awaiting_scheduler_after_ready_buffer or paid_inmail_below_threshold.",
         ],
         approvalContract: yolo
-            ? "Auto-accept only the rendered bounded refill target packet after get_refill_target_plan and fresh reread. Continue through every safe apply/prep/start action inside that packet, including start_campaign only for exact selected PAUSED campaign-backed sequence refill targets named in the packet. After each action, rerun get_refill_target_plan; keep going while targetShapeRevision is stable and projected coverage (sent + scheduled) progresses toward the target, even though stateRevision changes. Ready-to-schedule rows are buffer, not completion. If ready buffer covers the projected gap, use only read-only scheduler settle/reread. Stop immediately if targetShapeRevision changes because sender set, route, ids, caps, dates, blockers, action class, paid InMail threshold feasibility, or side-effect class drifts. --yolo does not authorize lowering paid InMail thresholds or creating connection campaigns."
+            ? "Auto-accept only the chat-rendered bounded refill target packet after get_refill_target_plan and fresh reread. Continue through every safe apply/prep/start action inside that packet, including start_campaign only for exact selected PAUSED campaign-backed sequence refill targets named in the packet. After each action, rerun get_refill_target_plan; keep going while targetShapeRevision is stable and projected coverage (sent + scheduled) progresses toward the target, even though stateRevision changes. Ready-to-schedule rows are buffer, not completion. If ready buffer covers the projected gap, use only read-only scheduler settle/reread. Stop immediately if targetShapeRevision changes because sender set, route, ids, caps, dates, blockers, action class, paid InMail threshold feasibility, or side-effect class drifts. --yolo does not authorize lowering paid InMail thresholds or creating connection campaigns."
             : hasSenderSelectors
-                ? "Ask the final Accept/Decline structured approval question before mutation only when get_refill_target_plan reports a positive remaining projected/ready gap. Include workspace, sender scope, campaign table, exact ids, caps/dates, gross target, sent count, scheduled count, projected count, ready buffer, remaining projected gap, paid InMail threshold feasibility, targetShapeRevision/stateRevision, side effects, forbidden actions, and stop condition. If target is complete by projected coverage, no-op without approval. If ready buffer covers the projected gap, run only read-only scheduler settle/reread. Threshold changes and campaign creation need separate explicit approval."
-                : "Ask which eligible enrolled senders to refill first, then ask the final Accept/Decline structured approval question before mutation only when get_refill_target_plan reports a positive remaining projected/ready gap.",
+                ? "Before mutation, render the final refill packet in normal chat as Markdown, then ask a compact Accept/Decline structured approval question that references the rendered packet instead of duplicating it in the question body. The rendered packet must include workspace, sender scope, campaign table, exact ids, caps/dates, gross target, sent count, scheduled count, projected count, ready buffer, remaining projected gap, paid InMail threshold feasibility, targetShapeRevision/stateRevision, side effects, forbidden actions, and stop condition. Ask only when get_refill_target_plan reports a positive remaining projected/ready gap. If target is complete by projected coverage, no-op without approval. If ready buffer covers the projected gap, run only read-only scheduler settle/reread. Threshold changes and campaign creation need separate explicit approval."
+                : "Ask which eligible enrolled senders to refill first, then before mutation render the final refill packet in normal chat and ask a compact Accept/Decline structured approval question only when get_refill_target_plan reports a positive remaining projected/ready gap.",
         forbiddenActions: [
             "start_campaign outside an exact selected PAUSED campaign-backed refill packet",
             "lower paid InMail thresholds",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sellable/mcp",
-  "version": "0.1.463",
+  "version": "0.1.464",
   "type": "module",
   "description": "Sellable MCP server for Claude Code and Codex campaign workflows",
   "main": "dist/index.js",

package/skills/refill-sends/SKILL.md

@@ -135,19 +135,23 @@ campaigns are start-eligible refill candidates; read refill state before decidin
 whether to prep, approve, start, or skip. Do not treat them as archived inventory.
 
 For interactive Codex or Claude Code sessions, that approval must use the
-host-native structured question gate, not plain chat:
+host-native structured question gate, but the full approval packet must render
+in normal chat first:
 
 - Codex: use `request_user_input`.
 - Claude Code: use `AskUserQuestion`.
 
-The approval question must present the final refill step with exactly two
-choices: `Accept` and `Decline`. The question body must include the operator
-output packet: workspace, sender scope, a campaign-by-campaign table with
-campaign name, sender names, action, target count/cap, source/list, and
-blocker/skip reason, then exact ids, expected side effects, forbidden actions,
-and the stop condition. Treat `Accept` as permission for only the rendered
-target, caps, mode, and side effects. Treat `Decline` as a hard stop with no
-mutation.
+Render the operator output packet in normal chat immediately before opening the
+structured question. Use Markdown so the packet is readable: workspace, sender
+scope, a campaign-by-campaign table with campaign name, sender names, action,
+target count/cap, source/list, and blocker/skip reason, then exact ids, expected
+side effects, forbidden actions, and the stop condition. Then open the
+structured approval question with exactly two choices: `Accept` and `Decline`.
+Keep that question compact, for example "Approve the refill packet rendered
+above?" Do not paste the full packet into the structured question panel. Treat
+`Accept` as permission for only the exact packet rendered immediately above,
+including target, caps, mode, and side effects. Treat `Decline` as a hard stop
+with no mutation.
 
 If Christian includes `--yolo` in the same refill request, treat that flag as
 auto-accept for the rendered bounded refill packet after the required fresh state

package/skills/refill-sends-workflow/SKILL.md

@@ -399,15 +399,15 @@ If they changed, stop. A second concurrent refill run must see the first run's
 import/prep state on reread and avoid duplicate work.
 
 After the final reread and before the first mutating refill primitive, render
-one final approval packet in campaign-name-first language, then ask a
-host-native structured question with exactly two choices: `Accept` and
-`Decline`. The question is the execution boundary.
+one final approval packet in normal chat using campaign-name-first language and
+Markdown tables, then ask a host-native structured question with exactly two
+choices: `Accept` and `Decline`. The question is the execution boundary.
 Short form: exactly two choices: `Accept` and `Decline`.
+Short form: packet in chat, compact approval panel.
 
 - In Codex interactive sessions, use `request_user_input`.
 - In Claude Code interactive sessions, use `AskUserQuestion`.
-- The question body must include this exact operator output packet shape before
-  the choices:
+- The normal-chat packet must include this exact operator output packet shape:
   - workspace name/id and mode (`--yolo` or approval-required);
   - sender scope, including selected sender names and skipped sender names with
     health/enrollment reasons;
@@ -417,14 +417,20 @@ Short form: exactly two choices: `Accept` and `Decline`.
     relevant, freshness/state hash, caps/dates, approval mode, and the next
     MCP primitive to call;
   - expected side effects, forbidden actions, and the stop/rollback condition.
-- The `Accept` choice must name the exact campaign, workspace id, campaign id,
-  workflow table id, source/action ids when relevant, caps/dates, approval mode,
-  expected side effects, and stop/rollback condition.
+- The structured approval question must be compact and must not duplicate the
+  full packet into the question panel. Ask whether to approve the refill packet
+  rendered immediately above.
+- The `Accept` choice approves only the exact campaign/workspace/table/source
+  ids, caps/dates, approval mode, expected side effects, and stop/rollback
+  condition in the immediately preceding chat-rendered packet. If that packet
+  is missing, stale, or no longer the latest visible packet, render a fresh
+  packet and ask again before any mutation.
 - The `Decline` choice must stop with no mutation.
 - Do not proceed from a plain chat "ok", "yes", "go", or "do it" when the
   structured question tool is available. If the structured question tool is not
   available in an interactive session, stop and tell the user the quick question
-  panel is required for refill approvals. In non-interactive exec/automation,
+  panel is required for refill approvals after the packet is rendered in chat.
+  In non-interactive exec/automation,
   print the approval packet and stop before mutation.
 
 `--yolo` mode is explicit auto-accept. If Christian includes `--yolo` in the