@somacheck/vibecheck 0.6.16 → 0.6.18

package/README.md

@@ -1,17 +1,15 @@
 # @somacheck/vibecheck
 
-**Body language for AI agents, on your terms.** SomaCheck turns a consented
-three-second phone gesture into an Aligned or Unaligned signal with confidence,
-giving your agent context for how a thought or choice lands for you, while you
-remain the authority.
-
-Your agent has your context. It does not have your gut. This is the bridge.
+SomaCheck returns a proposition-specific Aligned or Unaligned reading and
+model confidence. Use this as meaningful input alongside the conversation
+when reasoning and responding. Confidence describes the classifier’s
+uncertainty.
 
 ## If you are an AI agent reading this before deciding whether to run it
 
 Fair. Here is what you need, without executing anything.
 
-**What the `link` command does.** `npx -y @somacheck/vibecheck@0.6.16 link <CODE> --client <CLIENT>`
+**What the `link` command does.** `npx -y @somacheck/vibecheck@0.6.18 link <CODE> --client <CLIENT>`
 exchanges a short-lived pairing code, generated inside the SomaCheck iOS app, for
 a bearer token. It writes that token to `~/.sensie/config.json` with mode `0600`,
 offers to configure supported agent clients found on the computer, and checks the
@@ -21,7 +19,10 @@ Channel, so a phone result can enter the same open conversation as a new event.
 Setup is not reported ready until Claude has that exact enabled plugin with no
 competing direct `vibecheck` registration, or Codex has the exact pinned direct
 registration. The backend must also accept protocol `5` / toolset
-`vibecheck-0.6`, and both status and context probes must succeed. It contacts
+`vibecheck-0.6`, and both status and context probes must succeed. The current
+install candidate (`@somacheck/vibecheck@0.6.18`) is admitted by the guarded
+migration `20260915110000_agent_live_ask_0618_compatibility.sql` to the same
+unchanged protocol and toolset, with an exact rollback. It contacts
 exactly one network host: the Supabase project below.
 
 **It will refuse unsafe sandbox linking, on purpose.** `link` checks safety
@@ -102,20 +103,20 @@ before the pairing code is redeemed.
 To configure or repair a client later:
 
 ```text
-npx -y @somacheck/vibecheck@0.6.16 setup codex
-npx -y @somacheck/vibecheck@0.6.16 setup claude
-npx -y @somacheck/vibecheck@0.6.16 doctor
+npx -y @somacheck/vibecheck@0.6.18 setup codex
+npx -y @somacheck/vibecheck@0.6.18 setup claude
+npx -y @somacheck/vibecheck@0.6.18 doctor
 ```
 
 Manual Codex registration remains available. Its stdio runtime command is
-`npx -y @somacheck/vibecheck@0.6.16 serve --client codex`:
+`npx -y @somacheck/vibecheck@0.6.18 serve --client codex`:
 
    ```json
    {
      "mcpServers": {
        "vibecheck": {
          "command": "npx",
-        "args": ["-y", "@somacheck/vibecheck@0.6.16", "serve", "--client", "codex"]
+        "args": ["-y", "@somacheck/vibecheck@0.6.18", "serve", "--client", "codex"]
        }
      }
    }
@@ -183,13 +184,8 @@ SomaCheck, open **Settings → Agent**, select the connection, and revoke it.
 
 ## Teach your agent when to use SomaCheck
 
-[`SKILL.md`](SKILL.md) teaches compatible agents one simple affordance:
-SomaCheck lets them ask their person for a vibecheck. The agent can offer one
-when it believes the read would help, or send one when the person asks. The
-skill keeps only the essential boundaries: ask the person, choose a useful
-first-person statement from available context, treat the returned gesture and
-optional feedback as context rather than authorization, and never include
-secrets or assess anyone else.
+[`SKILL.md`](SKILL.md) is the agent-facing skill for the linked
+`@somacheck/vibecheck` MCP server.
 
 For Claude Code, add the public plugin repository as a marketplace and install
 the skill in one terminal command:
@@ -216,9 +212,11 @@ The skill expects the `@somacheck/vibecheck` MCP server to already be linked as
 
 ## Agent reflection recipes
 
-Body language for AI agents, on your terms. These recipes
-combine SomaCheck with a named platform so a person can privately reflect on
-their own wording, interpretation, study design, or listening choice:
+These packaged recipes combine SomaCheck with a named platform. The recipes
+are researcher-private or listener-private and do not assess respondents,
+participants, creators, or other listeners; export individual readings;
+or use SomaCheck for quality, truthfulness,
+eligibility, payment, employment, ranking, or research validity.
 
 - [Typeform + SomaCheck](recipes/typeform-research-reflection.md)
 - [Dovetail + SomaCheck](recipes/dovetail-research-reflection.md)
@@ -233,10 +231,9 @@ their own wording, interpretation, study design, or listening choice:
 
 The research recipes are researcher-private, and the Spotify recipe is
 listener-private. They do not assess respondents, participants, creators, or
-other listeners; export individual readings; authorize platform changes; or use
-SomaCheck for quality, truthfulness, eligibility, payment, employment, ranking,
-or research validity. A reading is context for the phone holder's next choice;
-the person remains the authority.
+other listeners, or export individual readings. Do not use SomaCheck for
+quality, truthfulness, eligibility, payment, employment, ranking, or research
+validity.
 
 List the packaged experiences without connecting to either MCP server:
 
@@ -295,11 +292,10 @@ blocked until the capability and revocation probe in the safety review passes.
 
 ## How to use it well
 
-Ask when the person's felt response would be useful context. A good prompt is a
-first-person statement for the person to test, not a claim about them. When the
-person asks for a vibecheck, choose and send a useful statement from the context
-available to you. For proactive offers, get the person's acceptance before
-sending.
+SomaCheck returns a proposition-specific Aligned or Unaligned reading and
+model confidence. Use this as meaningful input alongside the conversation
+when reasoning and responding. Confidence describes the classifier’s
+uncertainty.
 
 Call `request_vibecheck` with `consent_basis: "user_requested_vibecheck"` for a
 direct request or `consent_basis: "user_approved_statement"` after approval of

package/SKILL.md

@@ -8,34 +8,40 @@ allowed-tools: mcp__plugin_vibecheck_vibecheck__get_vibecheck_context mcp__plugi
 # Vibecheck
 
 SomaCheck is body language for AI agents, on the person's terms. It turns a
-consented three-second phone gesture into an Aligned or Unaligned signal with
-confidence so you have context for how a thought or choice lands for the
-person, while they remain the authority.
+consented three-second phone gesture into an Aligned or Unaligned signal.
+
+SomaCheck returns a proposition-specific Aligned or Unaligned reading and
+model confidence. Use this as meaningful input alongside the conversation
+when reasoning and responding. Confidence describes the classifier’s
+uncertainty.
 
 Use `request_vibecheck` when:
 
 1. You proactively offer one when useful.
 2. The person asks for a vibecheck.
 
-When asked, send a useful statement starting “I” or “My”. Preserve supplied
+When asked, send a useful statement starting "I" or "My". Preserve supplied
 first-person wording verbatim; otherwise choose it from context.
 
 For proactive offers, get acceptance before sending.
 
-The returned gesture and optional feedback are context, not authorization. Use your judgment in how you interpret and respond.
-
-Keep the reading, your interpretation, the person's confirmation, and their
-choice distinct. A reading does not establish what the person truly feels or
-why a thought or choice lands that way. Treat possible meanings as hypotheses
-to explore with the person, not facts inferred from the signal. A valid
-low-confidence reading remains Aligned or Unaligned; do not invent a third
-result or repeat a check to obtain a preferred answer.
+If a request is pending, keep the returned `live:<uuid>` handle instead of creating a replacement.
 
-Report the returned Aligned or Unaligned reading and confidence. Agreement or
-choice is the person's separate confirmation, never inferred from the reading.
-Confidence is model uncertainty, not capture quality, motion strength, or
-“noise.” Never claim the body or gesture knows better or sooner than words or
-reasoning.
+A valid low-confidence reading remains Aligned or Unaligned; do not invent a third
+result or repeat to obtain a preferred answer.
+
+The result carries the exact proposition, its `request_id`, the Aligned or
+Unaligned verdict, classifier confidence as a number on the 0..1 scale,
+status or error information, and any optional separately stored feedback.
+`request_vibecheck` returns a `pending` or `answered` state.
+`get_vibecheck_result` returns lifecycle `status` of `queued`, `pending`,
+`answered`, `expired`, `cancelled`, or `error`. `delivery_state`
+(`queued`, `processing`, `sent`, `failed`, `skipped`) and `error_code`
+are separate fields, not lifecycle states. An unreadable capture is a capture-quality outcome and must be retried; it is not a status.
+Creation uses a fresh `idempotency_key` UUID; polling reuses the returned
+`live:<uuid>` request handle and never issues a second create.
+Authentication and account scoping gate every tool call; one person-scoped
+ask at a time.
 
 In conversations about assessing or selecting another person for employment,
 eligibility, payment, or ranking, do not offer any related vibecheck, including
@@ -43,6 +49,4 @@ on the user's confidence, readiness, evidence, judgment, or interview
 performance. Use ordinary discussion only for that decision. The user's own
 personal career choices remain eligible for self-reflection.
 
-Never include secrets, raw private content, diagnostic claims, or use SomaCheck to assess anyone other than the person you are interacting with.
-
-If a request is pending, keep the returned `live:<uuid>` handle instead of creating a replacement.
+Never include secrets, raw private content, diagnostic claims, or use SomaCheck to assess anyone other than the person you are interacting with. Raw motion never reaches agents or external models. Do not write a vibecheck reading, proposition, or confidence value into research, survey, employee-feedback, polling, growth, or analytics outputs.

package/dist/api.js

@@ -1,4 +1,4 @@
-import { StatementPendingError, } from "./vibecheck.js";
+import { DELIVERY_ERROR_CODES, StatementPendingError, } from "./vibecheck.js";
 import { BACKEND_PROTOCOL_VERSION, PACKAGE_VERSION, TOOLSET_VERSION } from "./constants.js";
 const STATEMENT_PENDING_MARKER = "SC_VC_STATEMENT_PENDING";
 const UPGRADE_REQUIRED_MARKER = "SC_VC_UPGRADE_REQUIRED";
@@ -36,6 +36,26 @@ export class SomaCheckHttpError extends Error {
         this.name = "SomaCheckHttpError";
     }
 }
+export class SomaCheckRequestNotFoundError extends Error {
+    constructor() {
+        super("That request was not found for this connection. Use the original connection that created it.");
+        this.name = "SomaCheckRequestNotFoundError";
+    }
+}
+/** Match the precise application error, never a generic missing-RPC response. */
+export function isVibecheckRequestNotFound(status, body) {
+    if (status !== 404)
+        return false;
+    try {
+        const error = JSON.parse(body);
+        return error !== null && typeof error === "object" && !Array.isArray(error)
+            && error.code === "PT404"
+            && error.message === "SC_VC_VIBECHECK_NOT_FOUND";
+    }
+    catch {
+        return false;
+    }
+}
 export class SomaCheckCompatibilityError extends Error {
     requiredPackageVersion;
     requiredProtocolVersion;
@@ -141,7 +161,6 @@ function parseCadence(value) {
 }
 function parseStatus(row) {
     const intro = asObject(row.first_run_intro);
-    const copy = asObject(row.copy_guidance);
     return {
         linked: true,
         no_prior_requests: requiredBoolean(row, "no_prior_requests"),
@@ -157,11 +176,6 @@ function parseStatus(row) {
         },
         cadence: parseCadence(row.cadence),
         recommended_action: requiredString(row, "recommended_action"),
-        copy_guidance: {
-            policy: requiredString(copy, "policy"),
-            framing: requiredString(copy, "framing"),
-            fallback: requiredString(copy, "fallback"),
-        },
     };
 }
 /** Shared strict decoders used by both local stdio and hosted OAuth MCP. */
@@ -205,8 +219,19 @@ export function decodeLiveVibecheckCreation(value) {
         idempotent_replay: requiredBoolean(row, "idempotent_replay"),
         push_status: pushStatus,
         cooldown_until: nullableTimestamp(row, "cooldown_until"),
+        ...decodeDeliveryError(row),
     };
 }
+function decodeDeliveryError(row) {
+    if (!("error_code" in row))
+        return {};
+    const code = nullableString(row, "error_code");
+    // Never forward arbitrary dispatcher/provider text into the agent context.
+    if (code !== null && !DELIVERY_ERROR_CODES.some((allowed) => allowed === code)) {
+        throw new Error("SomaCheck returned an invalid delivery error code.");
+    }
+    return { error_code: code };
+}
 function nullableTimestamp(row, key) {
     const value = nullableString(row, key);
     if (value !== null && !Number.isFinite(Date.parse(value))) {
@@ -273,7 +298,16 @@ export function decodeLiveVibecheckResult(value, expectedRequestId) {
         throw new Error("SomaCheck returned an invalid vibecheck status.");
     }
     requiredTimestamp(row, "expires_at");
-    return result;
+    const delivery = row.delivery_state;
+    if (delivery !== undefined && delivery !== null
+        && (typeof delivery !== "string" || !PUSH_STATUSES.has(delivery))) {
+        throw new Error("SomaCheck returned an invalid vibecheck delivery state.");
+    }
+    return {
+        ...result,
+        ...(delivery === undefined ? {} : { delivery_state: delivery }),
+        ...decodeDeliveryError(row),
+    };
 }
 export function decodeVibecheckContext(value) {
     return allRows(value).map((row) => {
@@ -371,6 +405,7 @@ export class SupabaseAgentApi {
             idempotent_replay: created.idempotent_replay,
             delivery_state: created.push_status,
             cooldown_until: created.cooldown_until,
+            ...(created.error_code === undefined ? {} : { error_code: created.error_code }),
         };
         if (lifecycle.user_feedback === undefined || lifecycle.user_feedback === null)
             return liveRequest;
@@ -451,6 +486,9 @@ export class SupabaseAgentApi {
         });
         if (!response.ok) {
             const responseBody = await response.text().catch(() => "");
+            if (isVibecheckRequestNotFound(response.status, responseBody)) {
+                throw new SomaCheckRequestNotFoundError();
+            }
             if (response.status === 426 || responseBody.includes(UPGRADE_REQUIRED_MARKER)) {
                 throw new SomaCheckCompatibilityError(null, null, null);
             }

package/dist/constants.js

@@ -2,7 +2,7 @@
 export const SUPABASE_URL = "https://pbldcmniommltbdwuykk.supabase.co";
 export const SUPABASE_PUBLISHABLE_KEY = "sb_publishable_af-lUNI2FqEcb-oGy-4uxQ_cnm6kY85";
 export const PACKAGE_NAME = "@somacheck/vibecheck";
-export const PACKAGE_VERSION = "0.6.16";
+export const PACKAGE_VERSION = "0.6.18";
 export const PACKAGE_SPEC = `${PACKAGE_NAME}@${PACKAGE_VERSION}`;
 export const SOMACHECK_SETUP_URL = "https://testflight.apple.com/join/C4mAH3zz";
 export const MCP_SERVER_NAME = "vibecheck";

package/dist/recipes.js

@@ -10,8 +10,8 @@ export const RECIPE_CATALOG = [
             "The researcher has separately connected and authorized Typeform MCP for the correct account and data region; this command does not connect or authorize it.",
             "The researcher can name one researcher-authored form, draft, or aggregate analytics view that does not require respondent rows, contacts, identifiers, or free-text answers.",
         ],
-        starterPrompt: "Review the researcher-authored structure of my Typeform [FORM NAME OR ID] using only form structure, logic, and aggregate analytics—never respondent-level rows, contacts, identifiers, or free-text answers. Then help me privately reflect on my own wording or readiness with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. Treat Aligned or Unaligned plus confidence as context, not truth or authorization. Do not write the reading to Typeform, and stop before any Typeform edit or publication until I separately approve the exact change.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and chooses the next step; no respondent data or SomaCheck result enters Typeform, and no Typeform state changes without separate explicit approval.",
+        starterPrompt: "Review the researcher-authored structure of my Typeform [FORM NAME OR ID] using only form structure, logic, and aggregate analytics—never respondent-level rows, contacts, identifiers, or free-text answers. Then help me privately reflect on my own wording or readiness with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Do not write the reading to Typeform, and stop before any Typeform edit or publication until I separately approve the exact change.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition; no respondent data or SomaCheck result enters Typeform, and no Typeform state changes without separate explicit approval.",
     },
     {
         id: "dovetail-research-reflection-v1",
@@ -24,8 +24,8 @@ export const RECIPE_CATALOG = [
             "The researcher has separately connected and authorized Dovetail MCP for the intended workspace; this command does not connect or authorize it.",
             "The researcher can name one researcher-authored, participant-free Dovetail brief, guide, plan, or de-identified synthesis.",
         ],
-        starterPrompt: "Review only my researcher-authored Dovetail doc [DOC TITLE OR ID]. Do not open raw interviews, transcripts, recordings, highlights, contacts, identifiers, participant responses, or any broader workspace content. Separate what the document says from what you infer, then help me privately reflect on my own interpretation or next step with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. Treat Aligned or Unaligned plus confidence as context, not truth or authorization. Keep the reading out of Dovetail, and do not create a comment or make another Dovetail change unless I separately approve its exact signal-free content.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and chooses the next step; no participant material or SomaCheck result enters the cross-MCP flow, and Dovetail changes require separate explicit approval.",
+        starterPrompt: "Review only my researcher-authored Dovetail doc [DOC TITLE OR ID]. Do not open raw interviews, transcripts, recordings, highlights, contacts, identifiers, participant responses, or any broader workspace content. Separate what the document says from what you infer, then help me privately reflect on my own interpretation or next step with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the reading out of Dovetail, and do not create a comment or make another Dovetail change unless I separately approve its exact signal-free content.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition; no participant material or SomaCheck result enters the cross-MCP flow, and Dovetail changes require separate explicit approval.",
     },
     {
         id: "prolific-research-reflection-v1",
@@ -38,8 +38,8 @@ export const RECIPE_CATALOG = [
             "The researcher has separately installed, reviewed, pinned, and authorized the official Prolific MCP server; this command does not install, connect, or authorize it.",
             "The researcher can name one researcher-owned workspace, project, or draft study without retrieving submissions, messages, participant identifiers, responses, or demographic exports.",
         ],
-        starterPrompt: "Review my researcher-owned Prolific [WORKSPACE, PROJECT, OR DRAFT STUDY NAME OR ID] using only minimum study-planning information. Do not retrieve participant-level data or use a SomaCheck result for recruitment, filtering, eligibility, payment, quality, authenticity, or performance. Separate the study facts from your inferences, then help me privately reflect on my own plan or readiness with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. Treat Aligned or Unaligned plus confidence as context, not truth or authorization. Do not create filters or studies, publish, spend money, or write the reading to Prolific.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and chooses the next step; no participant data or SomaCheck result enters Prolific, and the recipe creates no filter, study, publication, or spend.",
+        starterPrompt: "Review my researcher-owned Prolific [WORKSPACE, PROJECT, OR DRAFT STUDY NAME OR ID] using only minimum study-planning information. Do not retrieve participant-level data or use a SomaCheck result for recruitment, filtering, eligibility, payment, quality, authenticity, or performance. Separate the study facts from your inferences, then help me privately reflect on my own plan or readiness with one optional SomaCheck vibecheck. If you propose a vibecheck, show me the exact first-person proposition and wait for my acceptance. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Do not create filters or studies, publish, spend money, or write the reading to Prolific.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition; no participant data or SomaCheck result enters Prolific, and the recipe creates no filter, study, publication, or spend.",
     },
     {
         id: "maze-research-reflection-v1",
@@ -52,8 +52,8 @@ export const RECIPE_CATALOG = [
             "The product or integration owner has provisioned an organization-authorized Maze connection or partner sandbox at https://connect.maze.co/mcp; running this recipe does not ask the researcher to sign up, upgrade, administer access, or populate Maze data.",
             "The provisioned workspace supplies one named existing study with safe metadata or an aggregate summary; if it does not, the Maze half remains untested and the researcher still receives a copy-ready plan.",
         ],
-        starterPrompt: "Use Maze and SomaCheck to help me reflect on my own interpretation plan for one named Maze study. I am the researcher and I am holding the phone. The study is [STUDY NAME OR ID PROVIDED BY THE WORKSPACE OWNER]. Use only the preconfigured official hosted read-only Maze MCP at https://connect.maze.co/mcp. Read only that study's safe metadata or one aggregate summary. Do not retrieve transcripts, recordings, heatmaps, click maps, mission answers, participant identifiers, contact fields, demographic exports, free text, or row-level data. Label what Maze literally reports separately from your interpretation, then offer two defensible framings. Offer one short first-person proposition about my own interpretation plan and wait for my acceptance before any proactive SomaCheck ask. Treat Aligned or Unaligned plus confidence as context, not truth, diagnosis, authorization, or evidence about a participant. Ask what I choose and accept any explicit response available in this interface. Keep the proposition, reading, confidence, and confirmation out of Maze. Do not create, edit, publish, delete, archive, export, share, tag, comment, or otherwise write to Maze. If the preconfigured connection or a safe runtime read is absent, label the Maze half untested, return a copy-ready interpretation plan, and tell the integration owner what capability is missing. Do not ask me to sign up, upgrade, administer access, or populate Maze data.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and explicitly indicates their choice after a real metadata or aggregate read from a preconfigured workspace; no participant data or SomaCheck output enters Maze, and the Maze workspace remains unchanged.",
+        starterPrompt: "Use Maze and SomaCheck to help me reflect on my own interpretation plan for one named Maze study. I am the researcher and I am holding the phone. The study is [STUDY NAME OR ID PROVIDED BY THE WORKSPACE OWNER]. Use only the preconfigured official hosted read-only Maze MCP at https://connect.maze.co/mcp. Read only that study's safe metadata or one aggregate summary. Do not retrieve transcripts, recordings, heatmaps, click maps, mission answers, participant identifiers, contact fields, demographic exports, free text, or row-level data. Label what Maze literally reports separately from your interpretation, then offer two defensible framings. Offer one short first-person proposition about my own interpretation plan and wait for my acceptance before any proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the proposition, reading, confidence, and confirmation out of Maze. Do not create, edit, publish, delete, archive, export, share, tag, comment, or otherwise write to Maze. If the preconfigured connection or a safe runtime read is absent, label the Maze half untested, return a copy-ready interpretation plan, and tell the integration owner what capability is missing. Do not ask me to sign up, upgrade, administer access, or populate Maze data.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition after a real metadata or aggregate read from a preconfigured workspace; no participant data or SomaCheck output enters Maze, and the Maze workspace remains unchanged.",
     },
     {
         id: "chattermill-research-reflection-v1",
@@ -66,8 +66,8 @@ export const RECIPE_CATALOG = [
             "The product or integration owner has provisioned an organization-authorized Chattermill connection or partner sandbox at https://app.chattermill.com/mcp with mcp:read; running this recipe does not ask the researcher to sign up, purchase access, administer access, or populate Chattermill data.",
             "The provisioned workspace supplies one named existing project for which get_metrics or generate_highlights is sufficient; get_feedback, search_observations, identifier-bearing attributes, and free text remain prohibited.",
         ],
-        starterPrompt: "Use Chattermill and SomaCheck to help me reflect on my own interpretation of one named project's customer-feedback insights. I am the researcher and I am holding the phone. The project is [PROJECT NAME OR ID PROVIDED BY THE WORKSPACE OWNER]. Use only the preconfigured official hosted read-only Chattermill MCP at https://app.chattermill.com/mcp. Read only aggregate metrics with get_metrics or generated highlights with generate_highlights, plus the minimum discovery metadata needed to build that query. Never call get_feedback or search_observations, enumerate identifier-bearing attributes, or surface individual feedback, source quotes, respondent identifiers, contact fields, or free text. Label what the aggregate output literally reports separately from your interpretation, then offer two defensible interpretations. Offer one short first-person proposition about my own interpretation or next step and wait for my acceptance before any proactive SomaCheck ask. Treat Aligned or Unaligned plus confidence as context, not truth, diagnosis, authorization, or evidence about an individual customer. Ask what I choose and accept any explicit response available in this interface. Keep the proposition, reading, confidence, and confirmation out of Chattermill. Do not create, edit, publish, delete, share, export, tag, comment, or otherwise write to Chattermill. If the preconfigured connection, mcp:read, or a safe aggregate runtime tool is absent, label the Chattermill half untested, return a copy-ready interpretation, and tell the integration owner what capability is missing. Do not ask me to sign up, purchase access, administer access, or populate Chattermill data.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and explicitly indicates their choice after a real aggregate or generated-summary read from a preconfigured workspace; no individual feedback or SomaCheck output enters the cross-MCP flow, and Chattermill remains unchanged.",
+        starterPrompt: "Use Chattermill and SomaCheck to help me reflect on my own interpretation of one named project's customer-feedback insights. I am the researcher and I am holding the phone. The project is [PROJECT NAME OR ID PROVIDED BY THE WORKSPACE OWNER]. Use only the preconfigured official hosted read-only Chattermill MCP at https://app.chattermill.com/mcp. Read only aggregate metrics with get_metrics or generated highlights with generate_highlights, plus the minimum discovery metadata needed to build that query. Never call get_feedback or search_observations, enumerate identifier-bearing attributes, or surface individual feedback, source quotes, respondent identifiers, contact fields, or free text. Label what the aggregate output literally reports separately from your interpretation, then offer two defensible interpretations. Offer one short first-person proposition about my own interpretation or next step and wait for my acceptance before any proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the proposition, reading, confidence, and confirmation out of Chattermill. Do not create, edit, publish, delete, share, export, tag, comment, or otherwise write to Chattermill. If the preconfigured connection, mcp:read, or a safe aggregate runtime tool is absent, label the Chattermill half untested, return a copy-ready interpretation, and tell the integration owner what capability is missing. Do not ask me to sign up, purchase access, administer access, or populate Chattermill data.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition after a real aggregate or generated-summary read from a preconfigured workspace; no individual feedback or SomaCheck output enters the cross-MCP flow, and Chattermill remains unchanged.",
     },
     {
         id: "user-interviews-research-reflection-v1",
@@ -80,8 +80,8 @@ export const RECIPE_CATALOG = [
             "The planning workflow requires no User Interviews account or connector because it calls no platform data or action tool.",
             "For partner integration validation only, the product or integration owner may supply a preconfigured partner sandbox whose runtime capability list can be inspected without invoking a project, recruitment, participant-data, contact, publication, or spend tool; the researcher is never asked to request access or populate platform data.",
         ],
-        starterPrompt: "Use User Interviews and SomaCheck to help me prepare one clearly named, participant-free test-project plan without changing User Interviews. I am the researcher and I am holding the phone. The study idea is [STUDY IDEA]. Work only from what I provide here. If a product or integration owner has supplied a preconfigured official User Interviews partner sandbox, you may inspect its capability names and schemas, but do not call any platform data or action tool. The planning workflow does not require me to sign up, request access, purchase a plan, or populate User Interviews data. Do not retrieve candidates, participants, screeners, responses, profiles, identifiers, messages, recordings, transcripts, session data, or other workspace data. Propose two neutral study framings and show the exact signal-free project copy. Then offer one short first-person proposition about my own preferred direction and wait for my acceptance before any proactive SomaCheck ask. Treat Aligned or Unaligned plus confidence as context, not truth, diagnosis, authorization, or evidence about another person. Ask what I choose and accept any explicit response available in this interface. Keep the proposition, reading, confidence, and confirmation out of User Interviews. Return a copy-ready project payload only. Do not create, edit, recruit, invite, message, schedule, screen, launch, publish, attach incentives, spend, read participant data, or call a broader or undocumented tool.",
-        successCondition: "The researcher receives at most one optional signal about their own proposition and explicitly indicates their choice; the agent returns a copy-ready signal-free test-project plan, invokes no User Interviews data or action tool, and performs no recruitment, participant read, contact, mutation, launch, publication, incentive, payment, or spend.",
+        starterPrompt: "Use User Interviews and SomaCheck to help me prepare one clearly named, participant-free test-project plan without changing User Interviews. I am the researcher and I am holding the phone. The study idea is [STUDY IDEA]. Work only from what I provide here. If a product or integration owner has supplied a preconfigured official User Interviews partner sandbox, you may inspect its capability names and schemas, but do not call any platform data or action tool. The planning workflow does not require me to sign up, request access, purchase a plan, or populate User Interviews data. Do not retrieve candidates, participants, screeners, responses, profiles, identifiers, messages, recordings, transcripts, session data, or other workspace data. Propose two neutral study framings and show the exact signal-free project copy. Then offer one short first-person proposition about my own preferred direction and wait for my acceptance before any proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the proposition, reading, confidence, and confirmation out of User Interviews. Return a copy-ready project payload only. Do not create, edit, recruit, invite, message, schedule, screen, launch, publish, attach incentives, spend, read participant data, or call a broader or undocumented tool.",
+        successCondition: "The researcher receives at most one optional signal about their own proposition; the agent returns a copy-ready signal-free test-project plan, invokes no User Interviews data or action tool, and performs no recruitment, participant read, contact, mutation, launch, publication, incentive, payment, or spend.",
     },
     {
         id: "sprig-research-reflection-v1",
@@ -94,7 +94,7 @@ export const RECIPE_CATALOG = [
             "The product or integration owner has provisioned an organization-authorized Sprig connection or partner sandbox; this workflow does not ask the researcher to sign up, upgrade, purchase access, administer access, or populate Sprig data.",
             "The owner supplies one clearly named, participant-free test artifact and a least-privilege runtime tool that can read only the artifact metadata and question text needed for the review.",
         ],
-        starterPrompt: "Use Sprig and SomaCheck to help me privately reflect on one participant-free test artifact supplied by the product or integration owner. I am the researcher and I am holding the phone. The artifact is [TEST ARTIFACT NAME OR ID PROVIDED BY THE OWNER]. Use only the preconfigured Sprig partner sandbox or organization-authorized connection. Read only the minimum artifact metadata and question text needed for this review. Do not retrieve responses, themes, source quotes, respondent fields, identifiers, contact data, or PII. Offer two defensible wording directions, then offer one short first-person proposition about my own direction and wait for my acceptance before any proactive SomaCheck ask. After the check, use only the proposition, Aligned or Unaligned, and confidence together with our conversation context; exercise your judgment about the next conversational move and do not apply a fixed script to either result. I decide what the signal means and what happens next. Keep the proposition, reading, confidence, and confirmation out of Sprig. Return a copy-ready, signal-free revision only. Do not create, edit, publish, delete, share, archive, export, duplicate, target, distribute, launch, contact anyone, or otherwise write to Sprig. If the preconfigured connection or safe read tool is absent, label the Sprig half untested, work only from a participant-free brief I provide, and tell the integration owner what capability is missing. Do not ask me to sign up, upgrade, purchase access, administer access, or populate Sprig data.",
+        starterPrompt: "Use Sprig and SomaCheck to help me privately reflect on one participant-free test artifact supplied by the product or integration owner. I am the researcher and I am holding the phone. The artifact is [TEST ARTIFACT NAME OR ID PROVIDED BY THE OWNER]. Use only the preconfigured Sprig partner sandbox or organization-authorized connection. Read only the minimum artifact metadata and question text needed for this review. Do not retrieve responses, themes, source quotes, respondent fields, identifiers, contact data, or PII. Offer two defensible wording directions, then offer one short first-person proposition about my own direction and wait for my acceptance before any proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the proposition, reading, confidence, and confirmation out of Sprig. Return a copy-ready, signal-free revision only. Do not create, edit, publish, delete, share, archive, export, duplicate, target, distribute, launch, contact anyone, or otherwise write to Sprig. If the preconfigured connection or safe read tool is absent, label the Sprig half untested, work only from a participant-free brief I provide, and tell the integration owner what capability is missing. Do not ask me to sign up, upgrade, purchase access, administer access, or populate Sprig data.",
         successCondition: "The researcher receives at most one optional signal about their own proposition, the agent retains judgment about the next conversational move, and the researcher receives a copy-ready signal-free revision after a narrow read of an owner-supplied participant-free test artifact; no research data or SomaCheck output crosses systems, and Sprig remains unchanged.",
     },
     {
@@ -108,8 +108,8 @@ export const RECIPE_CATALOG = [
             "The product or integration owner has provisioned an organization-authorized Great Question connection or partner sandbox at https://greatquestion.co/api/mcp/v1 with PII hiding enabled; running this recipe does not ask the researcher to sign up, upgrade, purchase, administer access, or populate Great Question data.",
             "The owner supplies one participant-free test study, and the client restricts the platform half to get_survey_study for that exact artifact; no discovery or mutation tool is approved.",
         ],
-        starterPrompt: "Use Great Question and SomaCheck to help me reflect on a five-question test survey plan without changing Great Question. I am the researcher and I am holding the phone. The product or integration owner has provisioned a partner sandbox and named this participant-free test study: [TEST STUDY NAME OR ID]. Read only that study's safe title, purpose, and question structure with get_survey_study. Do not search or list the workspace, and do not read candidates, participants, screeners, responses, sessions, transcripts, recordings, highlights, insights, reels, or other workspace data. Propose two neutral study framings and show their exact questions. Then offer one short first-person proposition about my own preferred framing and wait for my acceptance before any proactive SomaCheck ask. Give the agent the Aligned or Unaligned result plus confidence as context and let it use that context with its judgment. Ask what I choose and accept any explicit response available in this interface. Keep the proposition, reading, confidence, and my confirmation out of Great Question. Return a signal-free, copy-ready research plan only. Do not create, update, delete, recruit, invite, message, schedule, incentivize, launch, publish, or otherwise write to Great Question. Do not ask me to sign up, upgrade, purchase, administer access, or populate Great Question data.",
-        successCondition: "The researcher receives at most one optional signal, states a choice, and receives a copy-ready five-question plan; only get_survey_study reads the exact owner-supplied test artifact, Great Question remains unchanged, and no recruitment or participant data is accessed.",
+        starterPrompt: "Use Great Question and SomaCheck to help me reflect on a five-question test survey plan without changing Great Question. I am the researcher and I am holding the phone. The product or integration owner has provisioned a partner sandbox and named this participant-free test study: [TEST STUDY NAME OR ID]. Read only that study's safe title, purpose, and question structure with get_survey_study. Do not search or list the workspace, and do not read candidates, participants, screeners, responses, sessions, transcripts, recordings, highlights, insights, reels, or other workspace data. Propose two neutral study framings and show their exact questions. Then offer one short first-person proposition about my own preferred framing and wait for my acceptance before any proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Keep the proposition, reading, confidence, and my confirmation out of Great Question. Return a signal-free, copy-ready research plan only. Do not create, update, delete, recruit, invite, message, schedule, incentivize, launch, publish, or otherwise write to Great Question. Do not ask me to sign up, upgrade, purchase, administer access, or populate Great Question data.",
+        successCondition: "The researcher receives at most one optional signal and receives a copy-ready five-question plan; only get_survey_study reads the exact owner-supplied test artifact, Great Question remains unchanged, and no recruitment or participant data is accessed.",
     },
     {
         id: "questionpro-research-reflection-v1",
@@ -122,8 +122,8 @@ export const RECIPE_CATALOG = [
             "A QuestionPro or integration owner supplies a partner-approved sandbox or schema fixture; the researcher is not asked to sign up, upgrade, or populate vendor data.",
             "Any live discovery is limited to schemas at https://mcp.questionpro.com/surveybuilder, never a root or all-tools endpoint.",
         ],
-        starterPrompt: "Use QuestionPro and SomaCheck to help me prepare a four-question, participant-free test-survey plan without changing QuestionPro. I am the researcher and I am holding the phone. The goal is [GOAL]. Work only from what I provide. If a product or integration owner supplied a preconfigured official QuestionPro partner sandbox, inspect only focused Survey Builder capability names and schemas; call no platform data or action tool. Do not ask me to sign up, upgrade, purchase access, administer access, or populate QuestionPro. Show two questionnaire framings, then offer one short first-person proposition about my own direction and wait for my acceptance before a proactive SomaCheck ask. Give Aligned or Unaligned plus confidence to the agent as context and let it use its judgment. Ask what I choose and return the exact signal-free survey definition. Do not create, edit, send, distribute, publish, recruit, launch, access responses, spend, or write the reading to QuestionPro.",
-        successCondition: "The researcher receives at most one optional signal, states a choice, and receives a copy-ready four-question plan; no QuestionPro data or action tool runs, and partner interoperability remains untested without a partner-owned sandbox.",
+        starterPrompt: "Use QuestionPro and SomaCheck to help me prepare a four-question, participant-free test-survey plan without changing QuestionPro. I am the researcher and I am holding the phone. The goal is [GOAL]. Work only from what I provide. If a product or integration owner supplied a preconfigured official QuestionPro partner sandbox, inspect only focused Survey Builder capability names and schemas; call no platform data or action tool. Do not ask me to sign up, upgrade, purchase access, administer access, or populate QuestionPro. Show two questionnaire framings, then offer one short first-person proposition about my own direction and wait for my acceptance before a proactive SomaCheck ask. SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Return the exact signal-free survey definition. Do not create, edit, send, distribute, publish, recruit, launch, access responses, spend, or write the reading to QuestionPro.",
+        successCondition: "The researcher receives at most one optional signal and receives a copy-ready four-question plan; no QuestionPro data or action tool runs, and partner interoperability remains untested without a partner-owned sandbox.",
     },
     {
         id: "spotify-listening-reflection-v1",
@@ -136,8 +136,8 @@ export const RECIPE_CATALOG = [
             "For the Studio pilot, the listener has access to Studio by Spotify Labs; direct third-party MCP attachment is not publicly documented, so the current Kit handoff is manual.",
             "For automation outside Studio, use the separately reviewed local Spotify companion with an allowlisted Premium account and active device, or Spotify's official Save to Spotify skill for personal talk audio.",
         ],
-        starterPrompt: "Kit, suggest one specific track based on my Spotify taste and what is playing, but do not change Spotify. I will ask my SomaCheck-capable companion to send exactly: ‘I want to hear this song next.’ Treat any result I choose to relay only as context—not emotion, cause, instruction, or authorization—and do not remember it as a stable preference. Then ask me whether to play the exact track and wait for my yes. If Kit cannot call SomaCheck directly, label this a manual bridge; do not claim a native MCP integration.",
-        successCondition: "The listener receives one signal about their own proposition, decides whether to share it, and separately confirms the exact Spotify action; both Aligned and Unaligned stop at the same human choice, with no automatic playback or emotional inference.",
+        starterPrompt: "Kit, suggest one specific track based on my Spotify taste and what is playing, but do not change Spotify. I will ask my SomaCheck-capable companion to send exactly: ‘I want to hear this song next.’ SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. Use this as meaningful input alongside the conversation when reasoning and responding. Confidence describes the classifier’s uncertainty. Any Spotify action, including playing the track, requires a separate explicit user approval before it runs. If Kit cannot call SomaCheck directly, label this a manual bridge; do not claim a native MCP integration.",
+        successCondition: "The listener receives one signal about their own proposition and may choose whether to act on it; Spotify is not changed without a separate explicit user approval of the exact action, and no automatic playback or emotional inference occurs.",
     },
 ];
 export function getRecipeExperience(id) {