@descix/cli 1.0.9 → 1.0.11

package/agent-assets/instructions/onboarding-flow.md

@@ -294,6 +294,12 @@ User is trying to configure an app they don't own. Suggest purchase.
 ### "Folder already configured"
 Check if configuration matches intended app. Offer to update or skip.
 
+### Powch sign-in under `descix serve` fails ("User cancelled", or "WebAuthn is not supported on sites with TLS certificate errors")
+The gateway's dev certificate is not trusted. Run `descix dev-certs check`. If it is not trusted,
+give the user exactly `descix dev-certs trust` to run (macOS password prompt), tell them to quit
+and reopen Chrome, then run `descix dev-certs check` again. Details and Linux/Windows notes:
+`reference/local-dev.md` §3.3.
+
 ## Next Steps After Onboarding
 
 Once configured, guide user to:

package/agent-assets/reference/local-dev.md

@@ -187,11 +187,51 @@ a different case and FAILS LOUD naming the remedy, rather than quietly becoming
 
 The derived Shell rule is the important one: point the API at a cloud environment and the shell comes from that same origin, so one origin carries shell + app + `/apifront` with nothing else configured. A stale `env.platform.site` block does not hijack the root once the API is remote. Platform developers opt IN to a local shell by naming it (`--site-url https://localhost:5174` or `env.siteUrl`).
 
+### 3.3. The dev certificate and passkey sign-in
+
+The gateway's HTTPS uses one certificate: the pair named by `env.devCerts`
+(`descix config set-dev-certs`), else the SDK's shipped self-signed pair (SAN `localhost`,
+`127.0.0.1`, `::1`). The same pair serves every app dev server behind the gateway.
+
+**Chrome refuses WebAuthn on a tab whose certificate is untrusted, and it judges the whole tab.**
+The Powch sign-in iframe on `powch.descix.net` therefore fails inside an untrusted
+`https://localhost:<port>` tab, even though Powch's own origin is trusted.
+
+| Symptom | Meaning |
+|---|---|
+| Powch/ChatWidget log shows only `User cancelled` on sign-in | Powch's panel-close rejection; the real error is in the browser console |
+| Browser console: `NotAllowedError: WebAuthn is not supported on sites with TLS certificate errors` | the gateway certificate is not trusted (or has expired) |
+| `descix serve` banner: `Passkey sign-in will fail: <reason>. Run: descix dev-certs trust` | same, detected before the browser opens |
+| `descix doctor` row "Dev certificate" not passing | same |
+
+| Command | Who runs it | Effect |
+|---|---|---|
+| `descix dev-certs check [--json]` | the assisting agent | resolves the cert the gateway uses; exits 0 only when it is trusted; reports `unverifiable` with the reason on Linux and Windows |
+| `descix doctor` | the assisting agent | includes the same result as the "Dev certificate" check |
+| `descix dev-certs trust` | **the human** | macOS: trusts that cert in the login keychain (password prompt), re-checks, and says to quit and reopen Chrome |
+
+**Assisting-agent flow:**
+
+1. Run `descix doctor` (or `descix dev-certs check`).
+2. If the dev certificate is trusted, continue. Passkey failures then have another cause.
+3. If it is not trusted, hand the human exactly one command to run in their own terminal:
+   `descix dev-certs trust`. It needs their macOS login password, so the agent does not run it.
+4. When they report it succeeded, tell them to **quit Chrome completely and reopen it** (a running
+   Chrome keeps its old verdict), reload the `descix serve` URL, and sign in again.
+5. Run `descix dev-certs check` again and confirm exit 0 before calling sign-in fixed.
+
+**Linux and Windows:** the check reports `unverifiable`, and `descix dev-certs trust` is macOS-only.
+Trusting the cert is a manual step that has not been verified on DeSciX: import the gateway's
+`cert.pem` into the store Chrome reads (Linux: the NSS database,
+`certutil -d sql:$HOME/.pki/nssdb -A -t "P,," -n descix-localhost -i <cert.pem>`; Windows: the
+Current User "Trusted Root Certification Authorities" store), then restart the browser. Tell the
+human it is unverified.
+
 ## 4. Scenarios
 
 ### Scenario 1: App Dev (the common case)
 *   **User**: App developer. No platform checkout.
-*   **Setup**: `descix config set-env dev`, `descix app init`, `descix app set-site`, then `cd my-app && descix serve`.
+*   **Setup**: `descix config set-env dev`, `descix app init`, `descix app set-site`, `descix dev-certs check` (if not trusted, the human runs `descix dev-certs trust` once — §3.3), then `cd my-app && descix serve`.
 *   **Result**: one HTTPS origin. `/` → the cloud App Shell, `/apifront` → the cloud API, `/p/my-app` → your app, and the shell boots **standalone into your app** — its INITIAL VIEW is your app rather than the store, with the platform views still reachable from there.
 
 ### Scenario 2: Platform Dev ("Dogfooding")

package/agent-assets/reference/workspace-config.md

@@ -33,7 +33,7 @@ is the schema and the verb that owns each key.
 | `env.environment` | `DEV`, `DEMO` or `PROD` | `descix config init --env dev\|demo\|prod` |
 | `env.apiUrl` | the API origin this workspace talks to | `descix config init --env …` (known envs) / `descix config set-env <name> --url <origin>` (custom) |
 | `env.gateway.port` | the port `descix serve` listens on | `descix config set-gateway-port <port>` |
-| `env.devCerts` | `dir`, or `cert` + `key` — the TLS pair for the gateway and every app behind it | `descix config set-dev-certs --dir\|--cert\|--key\|--clear` |
+| `env.devCerts` | `dir`, or `cert` + `key` — the TLS pair for the gateway and every app behind it; unset means the SDK's shipped pair. Passkey sign-in needs it trusted: `descix dev-certs check`, then the human runs `descix dev-certs trust` (see `reference/local-dev.md` §3.3) | `descix config set-dev-certs --dir\|--cert\|--key\|--clear` |
 | `env.powchUrl` | Powch's own origin (optional; Powch is cross-origin from the shell by design) | `descix config set-powch-url <url>` |
 | `env.siteUrl` | the App Shell origin the gateway proxies `/` to (optional; default is the API origin) | `descix config set-site-url <url>` |
 | `env.products[]` | one entry per app in this workspace, see below | `descix app init`, `app set-site`, `app set-port`, `app set-localpath`, `app unmap` |

package/bin/descix.js

@@ -2187,6 +2187,7 @@ corpusCommand
   .option('--dry-run', 'Enumerate would-be-purged file_ids and would-be-upserted chunks without ANY Pinecone writes. Exit 0 if no drift, 1 if drift. Read-only.')
   .option('--show-walk', 'Print the resolved ref + the first 50 walked files BEFORE any Pinecone operations. Useful for verifying --ref / manifest source resolution.')
   .option('--yes', 'Skip the interactive purge confirmation in --rebuild mode. Use in scripting/CI.')
+  .option('--skip-retrieval-canary', 'Skip the post-sync retrieval canary that confirms a synced chunk actually retrieves before reporting success (measured 2026-09-17: presence in Pinecone is not the same as searchable — the index can lag ~20 minutes after a large upsert/purge). Skipping costs nothing but SAVES one credit-metered query_knowledge_base call and up to ~25s; the sync then reports success on presence only, same as before this check existed. Prefer leaving this ON for anything user-facing; use it for CI/cost-sensitive automation that will verify separately.')
   .action(async (options) => {
     try {
       const apiClient = new DeSciXApiClient();
@@ -4493,6 +4494,37 @@ program
     }
   });
 
+// ============ Dev Certs Commands ============
+
+const devCertsCommand = program
+  .command('dev-certs')
+  .description('TLS dev-cert trust status for `descix serve` — passkey sign-in needs this cert trusted once per machine');
+
+devCertsCommand
+  .command('check')
+  .description('Check whether the dev-server cert is trusted for https://localhost (exit 0 only when trusted)')
+  .option('--json', 'Output raw JSON')
+  .action(async (options) => {
+    try {
+      const { runDevCertsCheck } = await import('../lib/commands/dev-certs.js');
+      await runDevCertsCheck(options);
+    } catch (error) {
+      fail(error);
+    }
+  });
+
+devCertsCommand
+  .command('trust')
+  .description('Trust the dev-server cert in the macOS keychain (prompts for your password)')
+  .action(async () => {
+    try {
+      const { runDevCertsTrust } = await import('../lib/commands/dev-certs.js');
+      await runDevCertsTrust();
+    } catch (error) {
+      fail(error);
+    }
+  });
+
 // ============ Quickstart Command ============
 
 program

package/bin/mcp-server.js

@@ -21,6 +21,9 @@ import { WorkspaceConfig } from '../lib/workspace-config.js';
 import { DeSciXApiClient } from '../lib/api-client.js';
 import { WalletFileManager } from '../lib/wallet-file.js';
 import { CLI_VERSION } from '../lib/cli-version.js';
+import { checkDevCert } from '@descix/app-sdk/dev';
+import { resolveGatewayCertContext } from '../lib/dev-cert-resolver.js';
+import { originStampFor, stampToolResult } from '../lib/mcp-origin-stamp.js';
 import * as fs from 'fs/promises';
 import * as path from 'path';
 // WS-MCP-SSOT-TIER2 (audit §5.B-1): the curated HTTP-valid core tools are the SINGLE SOURCE
@@ -151,7 +154,7 @@ try {
   // descix_doctor — local diagnostic (no backend call)
   // -------------------------------------------------------------------------
 
-  async function runDoctor(args) {
+  async function runDoctor(args, originStamp) {
     const verifyRemote = args?.verify_remote !== false;
     const report = {
       auth: { connected: false },
@@ -197,7 +200,12 @@ try {
           app_id: freshContext.appId,
           community_id: freshContext.communityId,
           kb_id: freshContext.kbId || 'General',
-          api_url: freshConfig.apiUrl || apiClient.baseUrl,
+          // ONE OWNER: `freshConfig.apiUrl` is a legacy top-level key WorkspaceConfig never
+          // populates (the real value lives at `env.apiUrl`, read via `resolveOrigin()`) — the
+          // `||` here was always falling through to `apiClient.baseUrl`. Read the resolved
+          // origin stamp directly rather than keeping a dead alternative that looked like a
+          // second source.
+          api_url: originStamp?.origin ?? apiClient.baseUrl,
           local_apps: [
             platform.appId,
             ...products.map(p => p.appId),
@@ -322,6 +330,28 @@ try {
       }
     }
 
+    // Dev certificate — the ONE owner (checkDevCert, shared with `descix
+    // dev-certs`, `descix doctor` and the `descix serve` banner). An untrusted
+    // cert is invisible from inside a WebAuthn ceremony: Chrome judges the
+    // whole tab's TLS state and Powch only ever sees "User cancelled", so an
+    // agent assisting a stuck developer needs this named explicitly, not
+    // inferred from a login failure.
+    try {
+      const { certPath } = resolveGatewayCertContext(checkRoot);
+      const certResult = checkDevCert({ certPath });
+      report.dev_cert = { cert_path: certPath, ...certResult };
+      if (certResult.status !== 'trusted') {
+        report.warnings.push(
+          'Dev certificate not trusted: passkey sign-in under descix serve will fail ' +
+          '(Powch logs "User cancelled"; the console says "WebAuthn is not supported on sites ' +
+          'with TLS certificate errors"). Ask the user to run exactly: descix dev-certs trust ' +
+          '— then quit and reopen Chrome, and re-run descix dev-certs check.'
+        );
+      }
+    } catch (err) {
+      report.dev_cert = { status: 'unverifiable', detail: `Could not resolve dev cert: ${err.message}`, next: null };
+    }
+
     // Suggest cloud brain alignment when workspace is configured but alignment is missing
     if (!report.alignment && report.workspace?.configured) {
       report.suggestions = report.suggestions || [];
@@ -337,6 +367,13 @@ try {
       report.status = 'warnings';
     }
 
+    // The same origin stamp every other tool result carries (see mcp-origin-stamp.js) — doctor
+    // is the diagnostic surface, so it states the fact explicitly rather than only via the
+    // appended content block every tool gets.
+    if (originStamp) {
+      report.origin_stamp = originStamp;
+    }
+
     return report;
   }
 
@@ -383,13 +420,35 @@ try {
     const { name, arguments: args } = request.params;
     console.error(`[MCP] tools/call — ${name}`);
 
+    // Resolve the origin stamp ONCE per call, from the one owner (api-client.js's own
+    // resolution — the same baseUrl/originSource the CLI's `env:` line reports), and carry it
+    // onto EVERY result this call returns, success or error. MEASURED 2026-09-17: two sessions
+    // spent hours reading DEV through this connector while believing they were reading PROD
+    // because no tool result said where it came from — see mcp-origin-stamp.js.
+    let originStamp = null;
+    let originStampError = null;
+    try {
+      await apiClient.ensureInitialized();
+      originStamp = originStampFor(apiClient);
+    } catch (err) {
+      // The origin itself failed to resolve (e.g. a configured-but-invalid origin). There is
+      // nothing valid to stamp — the error's own message already names the fix — so this call
+      // fails loud below rather than stamping a guess.
+      originStampError = err;
+    }
+    const withStamp = (result) => (originStamp ? stampToolResult(result, originStamp) : result);
+
     try {
+      if (originStampError) {
+        throw originStampError;
+      }
+
       // descix_doctor is handled locally
       if (name === 'descix_doctor') {
-        const report = await runDoctor(args);
-        return {
+        const report = await runDoctor(args, originStamp);
+        return withStamp({
           content: [{ type: 'text', text: JSON.stringify(report, null, 2) }],
-        };
+        });
       }
 
       // Merge default context for convenience (caller can override). Only auto-fill when the
@@ -431,12 +490,12 @@ try {
       });
 
       const result = await apiClient.invoke(name, params);
-      return {
+      return withStamp({
         content: [{
           type: 'text',
           text: typeof result === 'string' ? result : JSON.stringify(result, null, 2),
         }],
-      };
+      });
     } catch (error) {
       console.error(`[MCP] Tool error: ${error.message}`);
       // WS-HEADLESS-MVP-A3: when the backend attached structured error fields (code +
@@ -447,10 +506,10 @@ try {
       const structured = (error.code || error.data)
         ? `\n${JSON.stringify({ code: error.code, data: error.data }, null, 2)}`
         : '';
-      return {
+      return withStamp({
         content: [{ type: 'text', text: `Error: ${error.message}${structured}` }],
         isError: true,
-      };
+      });
     }
   });
 

package/lib/commands/corpus.js

@@ -32,6 +32,7 @@ import {
   isReportedCount, accumulateReportedCount, UNREPORTED_COUNT
 } from '../core/Syncer.js';
 import { lintManifestForDenyClasses, assertNoViolations } from '../core/CorpusDenyLint.js';
+import { pickCanaryChunk, runRetrievalCanary, DEFAULT_CANARY_DELAYS_MS } from '../core/RetrievalCanary.js';
 
 /**
  * Resolve community_id for corpus sync — same precedence as site upload / update kb:
@@ -467,6 +468,11 @@ export async function runCorpusSync(apiClient, options) {
     let totalChunksCreated = 0;
     let totalFilesSkipped = 0;
     let totalFilesDeleted = 0;
+    // The retrieval canary's verdict per KB — a sync/rebuild does not get to report a bare
+    // success while any of these are unverified (see RetrievalCanary.js). Collected across
+    // manifests so ONE loud failure at the end names every KB still unconfirmed, rather than
+    // failing fast on the first and leaving later manifests unsynced.
+    const canaryResults = [];
 
     for (const rawManifest of manifests) {
       // ── Deliverable B: resolve --ref override (CLI > manifest > default) ──
@@ -877,6 +883,49 @@ export async function runCorpusSync(apiClient, options) {
         }
       }
 
+      // 3f-2. RETRIEVAL CANARY (measured 2026-09-17): presence in Pinecone (reconciled above)
+      // is NOT searchability — the PROD index can take ~20 minutes to catch up after a
+      // purge+re-upsert, during which every status surface still says healthy. A sync does not
+      // get to report bare success until a chunk it just wrote actually RETRIEVES. Bounded
+      // (RetrievalCanary.js) rather than a loop: query_knowledge_base is credit-metered.
+      let retrievalCanary = null;
+      if (allChunks.length > 0 && options.skipRetrievalCanary) {
+        console.log(chalk.gray('  [canary] Skipped (--skip-retrieval-canary): searchability NOT confirmed.'));
+      } else if (allChunks.length > 0) {
+        const canaryChunk = pickCanaryChunk(allChunks, syncFailures);
+        if (canaryChunk) {
+          const canarySpinner = ora('  [canary] Confirming a synced chunk actually retrieves...').start();
+          // TEST SEAM ONLY: options._canaryDelaysMs / options._canarySleep let tests exercise
+          // this path without real wall-clock waits or a live query_knowledge_base backend.
+          // Production callers never set these — runRetrievalCanary defaults to the real
+          // bounded backoff and real timers when they are undefined.
+          const result = await runRetrievalCanary(apiClient, {
+            appId, kbName, chunk: canaryChunk,
+            ...(options._canaryDelaysMs ? { delaysMs: options._canaryDelaysMs } : {}),
+            ...(options._canarySleep ? { sleep: options._canarySleep } : {}),
+          });
+          retrievalCanary = { ...result, checked_at: new Date().toISOString() };
+          if (result.ok) {
+            canarySpinner.succeed(
+              `  [canary] Confirmed searchable after ${(result.elapsedMs / 1000).toFixed(1)}s ` +
+              `(attempt ${result.attempts}/${DEFAULT_CANARY_DELAYS_MS.length})`
+            );
+          } else {
+            canarySpinner.warn(
+              `  [canary] NOT YET SEARCHABLE after ${(result.elapsedMs / 1000).toFixed(1)}s and ` +
+              `${result.attempts} attempt(s)${result.lastError ? ` (last error: ${result.lastError})` : ''}. ` +
+              `The vectors ARE present in Pinecone (reconciled above) but the index has not caught up yet — ` +
+              `this is a known ~20-minute propagation delay after a large upsert/purge. ` +
+              `Re-run 'descix kb corpus sync -a ${appId} -k ${kbName}' later to re-check ` +
+              `(unchanged files are cheap; this canary re-runs).`
+            );
+          }
+          canaryResults.push({ kbName, ...retrievalCanary });
+        } else {
+          console.log(chalk.yellow('  [canary] Skipped: every upsert batch failed, nothing new to confirm.'));
+        }
+      }
+
       // 3g. Save sync state with blob SHAs for future delta computation
       const allSyncedBlobShas = [...localBlobShas]; // All current files
       await saveSyncState(appRoot, kbName, {
@@ -894,7 +943,10 @@ export async function runCorpusSync(apiClient, options) {
         files_unchanged: unchangedCount,
         chunks_upserted: isReportedCount(upserted) ? upserted : null,
         chunks_deleted: isReportedCount(deleted) ? deleted : null,
-        synced_blob_shas: allSyncedBlobShas
+        synced_blob_shas: allSyncedBlobShas,
+        // The retrieval canary's verdict — read this, not `total_chunks`, to know whether the
+        // last sync's content was CONFIRMED searchable or only confirmed present.
+        retrieval_canary: retrievalCanary
       });
 
       totalFilesProcessed += newOrChangedFiles.length;
@@ -935,6 +987,26 @@ export async function runCorpusSync(apiClient, options) {
     const deletedLine = `  Chunks deleted:  ${renderCount(totalFilesDeleted)}`;
     console.log(totalFilesDeleted === 0 ? chalk.gray(deletedLine) : chalk.white(deletedLine));
     console.log('');
+
+    // NEVER A BARE SUCCESS: if any KB's retrieval canary did not confirm searchability, this
+    // command fails loud here rather than letting the green summary above stand as the last
+    // word. The summary itself is left intact (it is a true report of what was written) —
+    // this is a SEPARATE fact ("can it be found yet") that gets its own loud verdict.
+    const unverified = canaryResults.filter((r) => !r.ok);
+    if (unverified.length > 0) {
+      throw new Error(
+        `Sync wrote vectors successfully, but ${unverified.length} KB(s) did NOT confirm ` +
+        `searchable within the canary window:\n` +
+        unverified.map((r) =>
+          `  - ${r.kbName}: not found after ${(r.elapsedMs / 1000).toFixed(1)}s / ${r.attempts} attempt(s)` +
+          (r.lastError ? ` (last error: ${r.lastError})` : '')
+        ).join('\n') +
+        `\nThe vectors ARE present in Pinecone (reconciled above) — this is the store's own ` +
+        `~20-minute post-upsert propagation delay, not data loss. Re-run the same sync command ` +
+        `later to re-check; unchanged files make the re-run cheap and it re-runs this canary.`
+      );
+    }
+
     return { dryRun: false };
 
   } catch (error) {
@@ -1017,11 +1089,32 @@ export async function runCorpusStatus(apiClient, options) {
       console.log(chalk.white(`  Last sync:       ${syncState.timestamp}`));
       console.log(chalk.white(`  Commit:          ${syncState.last_sync_commit?.substring(0, 8) || 'unknown'}`));
       console.log(chalk.white(`  Files tracked:   ${syncState.synced_files_count}`));
+      // NAME THE SOURCE OF THIS NUMBER: this is a LOCAL RECORD of what the store reported
+      // AT THE TIME OF THE LAST SYNC's reconcile — not a fresh read of the store now. It can be
+      // stale (the store's contents can change from other syncs/purges) and, even when fresh,
+      // "present" is not "searchable" (see the Retrieval line below). `descix kb doctor -a
+      // <app> -k <kb> --live` reads Pinecone directly for the current truth.
       console.log(chalk.white(
-        `  Chunks in store: ${Number.isFinite(syncState.total_chunks)
+        `  Chunks (local record, as of last sync): ${Number.isFinite(syncState.total_chunks)
           ? syncState.total_chunks
           : 'unknown (the last sync reported no live count)'}`
       ));
+      // Retrieval canary verdict from the last sync (see RetrievalCanary.js) — the only field
+      // in this file that speaks to SEARCHABILITY rather than presence.
+      const canary = syncState.retrieval_canary;
+      if (canary === undefined) {
+        console.log(chalk.gray('  Retrieval:       unknown (synced before the retrieval-canary check existed)'));
+      } else if (canary === null) {
+        console.log(chalk.gray('  Retrieval:       not checked (last sync upserted nothing new)'));
+      } else if (canary.ok) {
+        console.log(chalk.green(`  Retrieval:       CONFIRMED searchable (as of last sync, ${canary.checked_at})`));
+      } else {
+        console.log(chalk.yellow(
+          `  Retrieval:       NOT CONFIRMED as of last sync (${canary.checked_at}) — vectors were ` +
+          `present but did not retrieve within the canary window. Re-run 'descix kb corpus sync ` +
+          `-a ${appId} -k ${kbName}' to re-check.`
+        ));
+      }
 
       // Quick change detection: walk and compare blob SHAs
       if (options.verbose) {

package/lib/commands/dev-certs.js

@@ -0,0 +1,93 @@
+/**
+ * dev-certs — report and (on trust) apply this machine's TLS trust status for
+ * the cert `descix serve` uses for its localhost HTTPS origin.
+ *
+ * Cert resolution is the ONE owner shared with `descix doctor` and the
+ * `descix serve` banner: lib/dev-cert-resolver.js -> @descix/app-sdk/dev's
+ * resolveDevCertOptions + resolveCertPaths. This file never re-derives which
+ * cert is "the" dev cert, and never mints or replaces one — the shipped
+ * SDK-tracked pair (descix-app-sdk/src/dev/certs/) is the only cert this
+ * command ever points at, same as every other consumer.
+ */
+import { execSync } from 'child_process';
+import chalk from 'chalk';
+import { checkDevCert, trustCertCommand } from '@descix/app-sdk/dev';
+import { resolveGatewayCertContext } from '../dev-cert-resolver.js';
+
+/**
+ * `descix dev-certs check [--json]` — exits 0 only when the cert is trusted.
+ */
+export async function runDevCertsCheck(options = {}) {
+  const { certPath, port } = resolveGatewayCertContext(process.cwd());
+  const result = checkDevCert({ certPath });
+
+  if (options.json) {
+    console.log(JSON.stringify({ certPath, port, ...result }, null, 2));
+  } else {
+    console.log(chalk.cyan('\nDev certificate\n'));
+    console.log(`  Cert:   ${certPath}`);
+    console.log(`  Status: ${result.status === 'trusted' ? chalk.green(result.status) : chalk.red(result.status)}`);
+    console.log(`  Detail: ${result.detail}`);
+    if (result.next) {
+      console.log(`\n  Next:\n    ${String(result.next).replace(/\n/g, '\n    ')}`);
+    }
+    console.log('');
+  }
+
+  process.exitCode = result.status === 'trusted' ? 0 : 1;
+  return result;
+}
+
+/**
+ * `descix dev-certs trust` — the ONE final human command. Runs the exact
+ * trustCertCommand() string (macOS prompts for the keychain password), then
+ * re-checks and tells the human to reload Chrome. Refuses to run against a
+ * cert that is structurally broken (missing / no SAN / expired) or on a
+ * platform this cannot verify — those are not this command's job to paper
+ * over, and running `security add-trusted-cert` on a bad cert would silently
+ * "succeed" while passkeys still fail.
+ */
+export async function runDevCertsTrust() {
+  const { certPath, port } = resolveGatewayCertContext(process.cwd());
+  const before = checkDevCert({ certPath });
+
+  if (before.status === 'trusted') {
+    console.log(chalk.green(`\n${certPath} is already trusted.\n`));
+    return before;
+  }
+
+  if (before.status === 'unverifiable') {
+    // `security add-trusted-cert` is macOS-only — there is no automated trust
+    // step on this platform. Refuse loud rather than silently no-op, and hand
+    // the human the exact file so they can import it into their own OS/browser
+    // trust store by hand.
+    console.error(chalk.red(`\nCannot trust automatically on ${process.platform}: ${before.detail}`));
+    console.error(`\nImport this certificate into your OS/browser trust store manually:\n  ${certPath}\n`);
+    process.exitCode = 1;
+    return before;
+  }
+
+  if (before.status !== 'untrusted') {
+    console.error(chalk.red(`\nCannot trust this cert — status: ${before.status}`));
+    console.error(`  ${before.detail}`);
+    if (before.next) {
+      console.error(`\n  Next:\n    ${String(before.next).replace(/\n/g, '\n    ')}\n`);
+    }
+    process.exitCode = 1;
+    return before;
+  }
+
+  console.log(chalk.cyan('\nTrusting the dev certificate — macOS will prompt for your password.\n'));
+  execSync(trustCertCommand(certPath), { stdio: 'inherit' });
+
+  const after = checkDevCert({ certPath });
+  if (after.status === 'trusted') {
+    console.log(chalk.green('\nTrusted.'));
+    console.log(`Quit and reopen Chrome, then reload https://localhost:${port}/\n`);
+  } else {
+    console.error(chalk.red(`\nStill not trusted after running the trust command (${after.status}).`));
+    console.error(`  ${after.detail}\n`);
+    process.exitCode = 1;
+  }
+  return after;
+}

package/lib/commands/doctor.js

@@ -14,6 +14,8 @@ import { exec } from 'child_process';
 import { promisify } from 'util';
 import { DeSciXApiClient } from '../api-client.js';
 import { isAuthenticated } from '../auth-guard.js';
+import { checkDevCert } from '@descix/app-sdk/dev';
+import { resolveGatewayCertContext } from '../dev-cert-resolver.js';
 
 const execAsync = promisify(exec);
 
@@ -223,6 +225,26 @@ export async function checkAdc() {
   }
 }
 
+/**
+ * Check the TLS cert `descix serve` would load for the cwd. An untrusted or
+ * structurally broken cert breaks passkey (WebAuthn) sign-in — the only
+ * DeSciX auth — so this is a real failure, not an admin-only advisory.
+ * Resolution mirrors the gateway exactly (dev-cert-resolver.js, the ONE
+ * owner shared with `descix dev-certs` and the `descix serve` banner).
+ */
+async function checkDevCertificate() {
+  const { certPath } = resolveGatewayCertContext(process.cwd());
+  const result = checkDevCert({ certPath });
+  if (result.status === 'trusted') {
+    return { ok: true, message: `${certPath} (trusted)` };
+  }
+  return {
+    ok: false,
+    message: `${result.status} — ${result.detail}`,
+    remediation: result.next || 'Run: descix dev-certs check',
+  };
+}
+
 /**
  * Run the doctor command
  */
@@ -282,7 +304,12 @@ export async function runDoctor(options = {}) {
     spinner.text = 'Checking ADC Credentials...';
     const adcRes = await checkAdc();
     results.push({ label: 'ADC Credentials', ...adcRes });
-    
+
+    // 7. Dev certificate (passkey sign-in under `descix serve`)
+    spinner.text = 'Checking dev certificate...';
+    const devCertRes = await checkDevCertificate();
+    results.push({ label: 'Dev certificate', ...devCertRes });
+
     spinner.stop();
     
     // Print Results
@@ -317,8 +344,9 @@ export async function runDoctor(options = {}) {
     // the pass stops.
     console.log(chalk.gray(
       'Checked: Node version, API origin resolution + reachability, session, local write\n' +
-      'permissions. [-] rows are admin-only dependencies and never fail this command. NOT\n' +
-      'checked: your app\'s own build, chain state, or any credential\'s permissions at the API.'
+      'permissions, dev-cert trust for `descix serve`. [-] rows are admin-only dependencies and\n' +
+      'never fail this command. NOT checked: your app\'s own build, chain state, or any\n' +
+      'credential\'s permissions at the API.'
     ));
     console.log();
 

package/lib/core/RetrievalCanary.js

@@ -0,0 +1,128 @@
+/**
+ * RetrievalCanary — a corpus sync does not report success until a KNOWN chunk it just
+ * wrote actually RETRIEVES from the store.
+ *
+ * MEASURED 2026-09-17: `descix kb corpus sync --rebuild` purges and re-upserts, then
+ * reports success and a reconciled live count, while the PROD search index takes ~20
+ * minutes to catch up. Every status surface said healthy over a KB that returned zero
+ * results for hours. The reconcile step (`get_kb_rag_status {reconcile:true}`, driven from
+ * `corpus.js::runCorpusSync`) counts by Pinecone ID PREFIX — presence, not searchability —
+ * so a clean reconcile is not evidence the index can answer a query yet.
+ *
+ * THE CANARY: after upserting, pick ONE chunk that was actually upserted (never one from a
+ * failed batch), run `query_knowledge_base` scoped to that chunk's own document
+ * (`file_filter`) using the chunk's own text as the query, and check the result set for a
+ * row whose metadata identifies THIS chunk (blob_sha + chunk_idx — the two fields
+ * `corpus.js::chunkCorpusFile` always writes, so this needs no second identity scheme).
+ *
+ * BOUNDED, NOT A LOOP: `query_knowledge_base` is a real, credit-metered read (WS-HEADLESS-MVP-A3
+ * flat debit per call). A canary that polled every few seconds for the full ~20-minute
+ * propagation window would multiply that cost by hundreds of calls for a single sync. This
+ * canary makes a SMALL, FIXED number of attempts (`DEFAULT_CANARY_DELAYS_MS`, 3 calls across
+ * ~25s) — enough to catch the common case where the index catches up quickly — and REFUSES to
+ * claim success if it does not, rather than waiting out the full window synchronously inside a
+ * CLI command. See corpus.js's hand-back for the case this does not close and the remedy it
+ * names when it doesn't (re-run the sync; unchanged blob SHAs make the re-run cheap and it
+ * re-runs this same canary).
+ */
+
+/** Fixed, small backoff schedule — see the module doc for why this is bounded, not a loop. */
+export const DEFAULT_CANARY_DELAYS_MS = [3000, 7000, 15000];
+
+/** Pinecone metadata budget note in chunkCorpusFile caps chunk text at 35000 bytes; a canary
+ * query only needs enough text to be a near-exact self-match, so it is truncated far below that
+ * — a long query costs more to embed for no better discrimination.
+ */
+const MAX_CANARY_QUERY_CHARS = 300;
+
+/**
+ * The query text for a canary check: the chunk's own text is guaranteed to retrieve itself
+ * with the highest similarity of anything in the index, so no second corpus of "known good"
+ * queries is needed.
+ *
+ * @param {{text: string}} chunk
+ * @returns {string}
+ */
+export function canaryQueryText(chunk) {
+    return String(chunk.text || '').trim().slice(0, MAX_CANARY_QUERY_CHARS);
+}
+
+/**
+ * Does this `query_knowledge_base` result row identify the given chunk?
+ *
+ * Matches on `blob_sha` + `chunk_idx` — the two fields `chunkCorpusFile()` always writes as
+ * Pinecone metadata, content-addressed and stable regardless of any server-side text
+ * normalization. This is the SAME identity the sync's own delta computation uses, not a
+ * second one invented for this check.
+ *
+ * @param {{metadata?: Record<string, any>}} row - one row of `query_knowledge_base`'s `results`
+ * @param {{blob_sha: string, chunk_idx: number}} chunk
+ * @returns {boolean}
+ */
+export function chunkMatchesCanary(row, chunk) {
+    const md = row?.metadata || {};
+    return md.blob_sha === chunk.blob_sha && md.chunk_idx === chunk.chunk_idx;
+}
+
+/**
+ * Pick one upserted chunk to canary-check: the first chunk that was NOT part of a batch the
+ * store reported as failed. Returns null when there is nothing safe to check (every batch
+ * failed) — the caller's existing batch-failure reporting already covers that case loudly;
+ * this module does not re-report it.
+ *
+ * @param {Array<{id: string}>} allChunks - every chunk this sync attempted to upsert
+ * @param {Array<{chunks: string[]}>} syncFailures - failed batches, each naming its chunk ids
+ * @returns {object|null}
+ */
+export function pickCanaryChunk(allChunks, syncFailures) {
+    const failedIds = new Set((syncFailures || []).flatMap((f) => f.chunks || []));
+    return allChunks.find((c) => !failedIds.has(c.id)) || null;
+}
+
+/**
+ * Run the bounded-backoff retrieval canary against a freshly-upserted chunk.
+ *
+ * @param {object} apiClient - an initialized DeSciXApiClient
+ * @param {object} args
+ * @param {string} args.appId
+ * @param {string} args.kbName
+ * @param {object} args.chunk - one entry from `allChunks` (must carry text/file_id/blob_sha/chunk_idx)
+ * @param {number[]} [args.delaysMs] - injectable for tests; production default is
+ *        DEFAULT_CANARY_DELAYS_MS
+ * @param {(ms: number) => Promise<void>} [args.sleep] - injectable for tests (real timers by default)
+ * @returns {Promise<{ok: boolean, attempts: number, elapsedMs: number, lastError: string|null}>}
+ */
+export async function runRetrievalCanary(apiClient, {
+    appId, kbName, chunk,
+    delaysMs = DEFAULT_CANARY_DELAYS_MS,
+    sleep = (ms) => new Promise((r) => setTimeout(r, ms)),
+}) {
+    const startedAt = Date.now();
+    let attempts = 0;
+    let lastError = null;
+
+    for (const delay of delaysMs) {
+        attempts++;
+        await sleep(delay);
+        try {
+            const resp = await apiClient.invoke('query_knowledge_base', {
+                app_id: appId,
+                kb_id: kbName,
+                query: canaryQueryText(chunk),
+                limit: 5,
+                file_filter: chunk.file_id,
+            }, { allowGuest: false });
+            const data = resp?.message || resp;
+            const results = data?.results || [];
+            if (results.some((row) => chunkMatchesCanary(row, chunk))) {
+                return { ok: true, attempts, elapsedMs: Date.now() - startedAt, lastError: null };
+            }
+        } catch (err) {
+            // A transient failure on ONE attempt does not end the canary early — the schedule is
+            // already bounded by delaysMs.length, so this is not an unbounded retry-on-error loop.
+            lastError = err.message;
+        }
+    }
+
+    return { ok: false, attempts, elapsedMs: Date.now() - startedAt, lastError };
+}

package/lib/dev-cert-resolver.js

@@ -0,0 +1,50 @@
+/**
+ * resolveGatewayCertContext — the CLI-side mirror of which cert `descix serve`
+ * will load for a given cwd. ONE owner within the CLI: `doctor.js` and
+ * `dev-certs.js` both call this instead of re-deriving workspace lookup +
+ * resolveDevCertOptions + resolveCertPaths by hand — the same two-derivations
+ * risk `getViteHttpsConfig.js` names for the gateway itself.
+ *
+ * Mirrors gateway.js exactly: resolveDevCertOptions(workspaceRoot, {}, config)
+ * then resolveCertPaths(certOpts) — both from @descix/app-sdk/dev, the actual
+ * owner of "which TLS cert a local dev server uses". This module only supplies
+ * the workspaceRoot/config a CLI command has to hand; it never re-derives the
+ * precedence itself.
+ */
+import {
+  findWorkspaceRoot,
+  readWorkspaceConfig,
+  resolveDevCertOptions,
+  resolveCertPaths,
+  resolveGatewayPort,
+} from '@descix/app-sdk/dev';
+
+/**
+ * @param {string} cwd
+ * @returns {{
+ *   certPath: string,
+ *   keyPath: string,
+ *   workspaceRoot: string|null,
+ *   config: Object|null,
+ *   port: number,
+ *   portSource: string,
+ * }}
+ */
+export function resolveGatewayCertContext(cwd) {
+  const workspaceRoot = findWorkspaceRoot(cwd);
+  let config = null;
+  if (workspaceRoot) {
+    try {
+      config = readWorkspaceConfig(workspaceRoot);
+    } catch {
+      // A malformed workspace is the workspace resolver's problem to report
+      // elsewhere; this resolver just falls back to "no workspace opinion",
+      // same as resolveDevCertOptions does for a missing/malformed file.
+      config = null;
+    }
+  }
+  const certOpts = resolveDevCertOptions(workspaceRoot || cwd, {}, config);
+  const { certPath, keyPath } = resolveCertPaths(certOpts);
+  const { port, portSource } = resolveGatewayPort(config || {}, {});
+  return { certPath, keyPath, workspaceRoot, config, port, portSource };
+}

package/lib/mcp-origin-stamp.js

@@ -0,0 +1,84 @@
+/**
+ * THE ORIGIN STAMP EVERY MCP TOOL RESULT CARRIES.
+ *
+ * MEASURED 2026-09-17: two sessions spent hours investigating a knowledge base as broken.
+ * Both were reading DEV through the MCP connector while believing they were reading PROD,
+ * because tool results said nothing about where they came from. `bin/mcp-server.js` returned
+ * every result as a bare `JSON.stringify(result)` — no origin, no environment. The envelope
+ * carries `community_id`/`app_id`, which READ like provenance and are not: they say which app,
+ * never which deployment of the platform answered.
+ *
+ * By contrast the CLI prints `env: <name> (<source>) <origin>` on EVERY network-bound command,
+ * unprompted (`lib/environment-report.js` `reportEnvironment()`), sourced from the ONE owner of
+ * "which origin did this invocation resolve to" (`lib/origin.js` `resolveOrigin()`, consumed by
+ * `DeSciXApiClient.initialize()` and kept on the client as `baseUrl` + `originSource`).
+ *
+ * This module adds NO second derivation. It reads the SAME two facts the CLI already keeps on
+ * an initialized `DeSciXApiClient` and combines them with the SAME env-name owner
+ * (`environmentNameFor`) the CLI's own status line uses. There is exactly one answer to "what
+ * origin, what environment" for a given client, and every surface — the CLI's stderr line, the
+ * MCP stamp — reads it from here.
+ */
+import { environmentNameFor } from './environment-report.js';
+
+/**
+ * Build the origin stamp for an MCP tool result FROM the one owner: an already-initialized
+ * `DeSciXApiClient`'s `baseUrl` + `originSource` (set once, in `initialize()`, by
+ * `resolveOrigin()` — see `api-client.js`).
+ *
+ * @param {import('./api-client.js').DeSciXApiClient} apiClient - must already be initialized
+ *   (`ensureInitialized()` resolved) so `baseUrl`/`originSource` are populated. Calling this
+ *   before initialization is a caller defect, not a state to paper over with a guess.
+ * @returns {{origin: string, source: string, environment: string}}
+ */
+export function originStampFor(apiClient) {
+    const origin = apiClient?.baseUrl;
+    const source = apiClient?.originSource;
+    if (!origin || !source) {
+        throw new Error(
+            'originStampFor: apiClient has no resolved baseUrl/originSource — ' +
+            'call ensureInitialized() before stamping a result.',
+        );
+    }
+    return { origin, source, environment: environmentNameFor(origin) };
+}
+
+/**
+ * The ONE spelling of the human-visible stamp line. Mirrors `formatEnvLine()`
+ * (`environment-report.js`) in shape so a reader who has seen the CLI's own `env:` line
+ * recognizes this as the same fact, not a second convention.
+ *
+ * @param {{origin: string, source: string, environment: string}} stamp
+ * @returns {string}
+ */
+export function formatOriginStampLine({ origin, source, environment }) {
+    return `[DeSciX origin] env: ${environment} (${source}) ${origin}`;
+}
+
+/**
+ * Attach the origin stamp to an MCP `CallToolResult` — success or error alike.
+ *
+ * TWO PLACES, DELIBERATELY, per the same fact:
+ *   1. An appended `content` block (visible text). This is the fix for the actual measured
+ *      failure: the reader is an LLM/agent consuming the tool's TEXT output, not a program
+ *      inspecting `_meta` — a stamp only a machine can see would not have caught the two
+ *      sessions this closes. The ORIGINAL content entries are left untouched and in place, so
+ *      an existing caller's `JSON.parse(content[0].text)` keeps working exactly as before —
+ *      this only ever APPENDS a new block, never rewrites or reorders the old one.
+ *   2. `_meta.descix_origin` — the MCP-spec extension point (`CallToolResult._meta`, a sibling
+ *      of `content`, not nested inside it) for a programmatic caller that wants the fact as
+ *      structured data without re-parsing the human-readable line.
+ *
+ * @param {{content?: Array<object>, isError?: boolean, [k: string]: any}} result
+ * @param {{origin: string, source: string, environment: string}} stamp
+ * @returns {object} a NEW result object; `result` is not mutated.
+ */
+export function stampToolResult(result, stamp) {
+    const content = Array.isArray(result.content) ? [...result.content] : [];
+    content.push({ type: 'text', text: formatOriginStampLine(stamp) });
+    return {
+        ...result,
+        content,
+        _meta: { ...(result._meta || {}), descix_origin: stamp },
+    };
+}

package/lib/verb-invokes.generated.json

@@ -116,6 +116,8 @@
   "credits refund": [
     "refund_credits"
   ],
+  "dev-certs check": [],
+  "dev-certs trust": [],
   "doctor": [],
   "drive pull": [],
   "drive push": [],
@@ -134,7 +136,8 @@
     "kb_delete_chunks",
     "kb_list_file_ids",
     "kb_sync_chunks",
-    "list_knowledge_bases"
+    "list_knowledge_bases",
+    "query_knowledge_base"
   ],
   "kb delete": [
     "delete_knowledge_base"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@descix/cli",
-  "version": "1.0.9",
+  "version": "1.0.11",
   "description": "DeSciX CLI and MCP Server - Unified command-line interface for DeSciX platform",
   "repository": {
     "type": "git",
@@ -14,7 +14,7 @@
     "descix-mcp": "bin/mcp-server.js"
   },
   "scripts": {
-    "test": "node --test \"tests/*.test.js\"",
+    "test": "node --experimental-test-module-mocks --test \"tests/*.test.js\"",
     "pack": "npm pack",
     "check:doc-paths": "node scripts/check-doc-paths.js",
     "briefer": "node scripts/briefer/run.mjs"
@@ -30,7 +30,7 @@
   "author": "DeSciX",
   "license": "SEE LICENSE IN LICENSE.md",
   "dependencies": {
-    "@descix/app-sdk": "^0.1.5",
+    "@descix/app-sdk": "^0.1.8",
     "@descix/sdk": "^2.0.0",
     "@modelcontextprotocol/sdk": "^1.25.1",
     "axios": "^1.7.9",

package/templates/agent-claude.md

@@ -174,6 +174,16 @@ resolves `--port` > `env.gateway.port` > 5173 and is printed with its source, so
 never hardcode `:5173`. Set your framework's base path to `/p/<appId>` — the
 gateway does not rewrite paths.
 
+**Passkey sign-in needs a trusted dev certificate.** If Powch sign-in under
+`descix serve` fails with `User cancelled`, or the browser console says
+`WebAuthn is not supported on sites with TLS certificate errors`, the gateway's
+certificate is not trusted. Run `descix dev-certs check` (or `descix doctor`, row
+"Dev certificate"). If it is not trusted, do not try to trust it yourself: give
+the user exactly `descix dev-certs trust` to run (macOS asks for their password),
+tell them to quit and reopen Chrome, then run `descix dev-certs check` again. On
+Linux and Windows the check reports `unverifiable` and trusting is a manual,
+unverified step — see `.descix/sdk-assets/reference/local-dev.md` §3.3.
+
 ## Integrating Existing Content
 
 If the repo already has content (an HTML page, a React app, an SDK with docs),

package/templates/agent-cline.md

@@ -136,6 +136,16 @@ resolves `--port` > `env.gateway.port` > 5173 and is printed with its source, so
 never hardcode `:5173`. Set your framework's base path to `/p/<appId>` — the
 gateway does not rewrite paths.
 
+**Passkey sign-in needs a trusted dev certificate.** If Powch sign-in under
+`descix serve` fails with `User cancelled`, or the browser console says
+`WebAuthn is not supported on sites with TLS certificate errors`, the gateway's
+certificate is not trusted. Run `descix dev-certs check` (or `descix doctor`, row
+"Dev certificate"). If it is not trusted, do not try to trust it yourself: give
+the user exactly `descix dev-certs trust` to run (macOS asks for their password),
+tell them to quit and reopen Chrome, then run `descix dev-certs check` again. On
+Linux and Windows the check reports `unverifiable` and trusting is a manual,
+unverified step — see `.descix/sdk-assets/reference/local-dev.md` §3.3.
+
 ## MCP Tools
 
 - `descix_doctor` — Startup diagnostic (call first)

package/templates/agent-copilot.md

@@ -136,6 +136,16 @@ resolves `--port` > `env.gateway.port` > 5173 and is printed with its source, so
 never hardcode `:5173`. Set your framework's base path to `/p/<appId>` — the
 gateway does not rewrite paths.
 
+**Passkey sign-in needs a trusted dev certificate.** If Powch sign-in under
+`descix serve` fails with `User cancelled`, or the browser console says
+`WebAuthn is not supported on sites with TLS certificate errors`, the gateway's
+certificate is not trusted. Run `descix dev-certs check` (or `descix doctor`, row
+"Dev certificate"). If it is not trusted, do not try to trust it yourself: give
+the user exactly `descix dev-certs trust` to run (macOS asks for their password),
+tell them to quit and reopen Chrome, then run `descix dev-certs check` again. On
+Linux and Windows the check reports `unverifiable` and trusting is a manual,
+unverified step — see `.descix/sdk-assets/reference/local-dev.md` §3.3.
+
 ## MCP Tools
 
 - `descix_doctor` — Startup diagnostic (call first)

package/templates/agent-cursor.md

@@ -136,6 +136,16 @@ resolves `--port` > `env.gateway.port` > 5173 and is printed with its source, so
 never hardcode `:5173`. Set your framework's base path to `/p/<appId>` — the
 gateway does not rewrite paths.
 
+**Passkey sign-in needs a trusted dev certificate.** If Powch sign-in under
+`descix serve` fails with `User cancelled`, or the browser console says
+`WebAuthn is not supported on sites with TLS certificate errors`, the gateway's
+certificate is not trusted. Run `descix dev-certs check` (or `descix doctor`, row
+"Dev certificate"). If it is not trusted, do not try to trust it yourself: give
+the user exactly `descix dev-certs trust` to run (macOS asks for their password),
+tell them to quit and reopen Chrome, then run `descix dev-certs check` again. On
+Linux and Windows the check reports `unverifiable` and trusting is a manual,
+unverified step — see `.descix/sdk-assets/reference/local-dev.md` §3.3.
+
 ## MCP Tools
 
 - `descix_doctor` — Startup diagnostic (call first)