feat(tanstack-migrator): per-round merge loop, .deco sync-mirror, deep triage (#513)

* feat(tanstack-migrator): incremental-merge loop + .deco push-mirror + investigative triage

Replace the single-PR + human-merge model with an incremental per-round loop:
opening_pr → reviewing (code-review agent gate) → merging (MCP squash-merge +
branch delete) → deploying → then triaging (first deploy) or paritying (fix
rounds); each fix round cuts a fresh migration/fix-N branch off main and loops
until parity target is hit with an empty backlog. Adds mergePullRequest /
deleteBranch github wrappers and reviewing/merging statuses (plain-TEXT columns,
no migration).

Rework the .deco sync into the proven push-mirror installed on the CLIENT repo
(on:push .deco/** → rsync blocks → -tanstack repo via STOREFRONT_SYNC_TOKEN);
SYNC_DECOFILE_INSTALL opens a PR adding only that one workflow file. Make triage
uncapped + investigative (headless Playwright console capture + skill grep sweep)
and surface the new phases in the widgets/dashboard/stepper.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(tanstack-migrator): wire cross-migration memory into the review phase

The code-review agent now reads framework-notes.md + fixes.md so it can (a) not
block a merge on a KNOWN framework bug and (b) recognize proven fix recipes; it
can also append a new framework-note when a diff reveals a recurring framework
bug. Triage already reads/writes all four memory files, so the new
Playwright/grep sweeps contribute too.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: regenerate registry.json (adds grafana MCP entry)

Fixes the failing `check:registry` CI step — registry.json was out of date
(missing the grafana MCP). Ran `bun run build:registry`; 115 public items.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: JonasJesus42

registry.json

@@ -2398,6 +2398,55 @@
       ]
     }
   },
+  {
+    "id": "deco/grafana",
+    "title": "Grafana",
+    "description": "Query self-hosted Grafana: list datasources, search/read dashboards, and run raw datasource or PromQL queries via /api/ds/query.",
+    "is_public": true,
+    "_meta": {
+      "mcp.mesh": {
+        "verified": false,
+        "friendly_name": "Grafana",
+        "short_description": "Query self-hosted Grafana datasources, dashboards, and PromQL — connect with your Grafana URL + a service-account token.",
+        "owner": "deco",
+        "has_remote": true,
+        "has_oauth": false,
+        "tags": [
+          "grafana",
+          "self-hosted",
+          "observability",
+          "monitoring",
+          "prometheus",
+          "dashboards",
+          "promql",
+          "metrics"
+        ],
+        "categories": [
+          "Observability"
+        ],
+        "readme": "The Grafana MCP connects AI agents to any self-hosted (or Cloud) Grafana instance over its HTTP API. **Key Features** — **GRAFANA_LIST_DATASOURCES** lists configured datasources (uid/name/type); **GRAFANA_SEARCH_DASHBOARDS** and **GRAFANA_GET_DASHBOARD** discover dashboards and read their full JSON (panels, targets, variables) so agents can learn the exact query a panel uses; **GRAFANA_QUERY** runs any native datasource query through POST /api/ds/query; **GRAFANA_QUERY_PROMETHEUS** is a convenience wrapper for instant or range PromQL. **Use Cases** — pull metrics into agent workflows, reproduce a dashboard panel's query, inspect cost/observability dashboards, and build automations on top of existing Grafana data. **Authentication** — provide your Grafana base URL and a service-account token (Administration → Service accounts → Add token); Viewer role is enough for querying. Ideal for teams running self-hosted Grafana who want programmatic, read-only access to metrics and dashboards without exposing Grafana broadly."
+      }
+    },
+    "server": {
+      "name": "grafana",
+      "title": "Grafana",
+      "description": "Query self-hosted Grafana: list datasources, search/read dashboards, and run raw datasource or PromQL queries via /api/ds/query.",
+      "icons": [
+        {
+          "src": "https://grafana.com/static/assets/img/fav32.png"
+        }
+      ],
+      "remotes": [
+        {
+          "type": "HTTP",
+          "url": "https://sites-grafana.deco.site/mcp",
+          "name": "grafana",
+          "title": "Grafana",
+          "description": "Query self-hosted Grafana: list datasources, search/read dashboards, and run raw datasource or PromQL queries via /api/ds/query."
+        }
+      ]
+    }
+  },
   {
     "id": "deco/grain",
     "title": "Grain",

tanstack-migrator/server/constants.ts

@@ -17,7 +17,10 @@ export const LIVENESS_STALL_MS = 4 * 60_000;
 /** Session timeouts per task kind (the reader gives up past these). */
 export const SESSION_TIMEOUT_MS: Record<string, number> = {
   migrate: 30 * 60_000,
-  triage: 20 * 60_000,
+  // triage is intentionally exhaustive (build + SSR + headless-browser console
+  // capture + skill grep sweep) — give it room.
+  triage: 40 * 60_000,
+  review: 20 * 60_000,
   fix: 25 * 60_000,
   parity: 20 * 60_000,
 };
@@ -34,7 +37,11 @@ export const DEFAULT_NO_IMPROVE_LIMIT = 3;
 export const DEFAULT_MAX_CONCURRENT = 1;
 export const DEFAULT_MAX_FIX_SESSIONS = 20;
 export const DEFAULT_FIX_BATCH_SIZE = 3;
-export const DEFAULT_MAX_ISSUES_PER_TRIAGE = 15;
+/**
+ * Effectively uncapped: triage should report EVERY real problem it finds. The
+ * high bound is only a runaway safety rail (dedupe + labeling stay bounded).
+ */
+export const DEFAULT_MAX_ISSUES_PER_TRIAGE = 100;
 export const DEFAULT_WORK_BRANCH = "migration/tanstack";
 
 /** Object-storage key prefix for parity artifacts. */

tanstack-migrator/server/db/types.ts

@@ -9,6 +9,8 @@ export type SiteStatus =
   | "baselining"
   | "migrating_script"
   | "opening_pr"
+  | "reviewing"
+  | "merging"
   | "triaging"
   | "fixing"
   | "paritying"
@@ -47,6 +49,8 @@ export const ACTIVE_STATUSES: SiteStatus[] = [
   "baselining",
   "migrating_script",
   "opening_pr",
+  "reviewing",
+  "merging",
   "triaging",
   "fixing",
   "paritying",
@@ -109,6 +113,7 @@ export function isSessionStatus(status: string): boolean {
     status === "baselining" ||
     status === "migrating_script" ||
     status === "triaging" ||
+    status === "reviewing" ||
     status === "fixing" ||
     status === "paritying" ||
     isLegacyStatus(status)
@@ -131,6 +136,7 @@ export function isActiveStatus(status: string): boolean {
 export type RunKind =
   | "migrate"
   | "triage"
+  | "review"
   | "fix"
   | "parity"
   | "fix_iteration" // legacy (≤ v0.4.x)

tanstack-migrator/server/engine/machine.ts

@@ -8,10 +8,12 @@ import { baselining } from "./phases/baselining.ts";
 import { creatingRepo } from "./phases/creating-repo.ts";
 import { deployingCf } from "./phases/deploying-cf.ts";
 import { fixing } from "./phases/fixing.ts";
+import { merging } from "./phases/merging.ts";
 import { migratingScript } from "./phases/migrating-script.ts";
 import { openingPr } from "./phases/opening-pr.ts";
 import { paritying } from "./phases/paritying.ts";
 import { provisioningSandbox } from "./phases/provisioning-sandbox.ts";
+import { reviewing } from "./phases/reviewing.ts";
 import { triaging } from "./phases/triaging.ts";
 
 export interface EngineDeps {
@@ -35,6 +37,8 @@ export const PHASE_HANDLERS: Partial<Record<SiteStatus, PhaseHandler>> = {
   baselining,
   migrating_script: migratingScript,
   opening_pr: openingPr,
+  reviewing,
+  merging,
   triaging,
   fixing,
   paritying,

tanstack-migrator/server/engine/phases/deploying-cf.ts

@@ -22,21 +22,29 @@ import { previewRendersRealHtml } from "../../lib/preview.ts";
 import { getDriver, isSimulation } from "../../sandbox/client.ts";
 import { deployPrompt } from "../../sandbox/templates/prompts.ts";
 
-/** After a successful deploy, advance to parity (measured against the CF URL). */
-async function advanceToParity(
+/**
+ * After a successful deploy, route on whether this was the FIRST deploy:
+ *   - initial migration (no cf_deploy_url yet) → triaging (build the backlog)
+ *   - every fix round (cf_deploy_url already set) → paritying (measure vs URL)
+ * Robust even when triage found zero issues (doesn't depend on issues_total).
+ */
+async function advanceAfterDeploy(
   site: SiteRow,
   patch: Partial<SiteRow>,
 ): Promise<void> {
+  const firstDeploy = !site.cf_deploy_url;
+  const next = firstDeploy ? "triaging" : "paritying";
   await updateSite(site.id, {
     ...patch,
-    status: "paritying",
-    phase_detail:
-      "deploy ok — starting parity measurement against the real URL",
+    status: next,
+    phase_detail: firstDeploy
+      ? "deploy ok — triaging the migrated site"
+      : "deploy ok — measuring parity against the real URL",
     last_progress_at: new Date().toISOString(),
   });
   await addEvent(
     site.id,
-    `CF deploy ok at ${patch.cf_deploy_url ?? "?"} — advancing to parity`,
+    `CF deploy ok at ${patch.cf_deploy_url ?? site.cf_deploy_url ?? "?"} — ${firstDeploy ? "triaging" : "advancing to parity"}`,
   );
 }
 
@@ -48,7 +56,7 @@ export async function deployingCf(
   const workerName = parseRepo(site.target_repo).repo;
 
   if (isSimulation(ctx)) {
-    await advanceToParity(site, {
+    await advanceAfterDeploy(site, {
       cf_project_name: workerName,
       cf_deploy_url: `https://${workerName}.deco-cx.workers.dev`,
     });
@@ -142,7 +150,7 @@ export async function deployingCf(
         );
       }
 
-      await advanceToParity(current, {
+      await advanceAfterDeploy(current, {
         cf_project_name: workerName,
         cf_deploy_url: deployUrl,
       });

tanstack-migrator/server/engine/phases/fixing.ts

@@ -11,7 +11,7 @@ import { addEvent } from "../../db/events.ts";
 import { createRun, finishRun } from "../../db/runs.ts";
 import { getSite, incrementCost, updateSite } from "../../db/sites.ts";
 import type { SiteRow } from "../../db/types.ts";
-import { ghsTokenForSite } from "../../lib/github.ts";
+import { ensureBranch, ghsTokenForSite, parseRepo } from "../../lib/github.ts";
 import {
   closeResolvedIssues,
   markBlockedIssues,
@@ -40,14 +40,24 @@ export async function fixing(
   ctx: WorkerCtx,
   deps: EngineDeps,
 ): Promise<void> {
-  // Terminal checks first (idempotent re-entry)
-  if (site.parity_score !== null && site.parity_score >= site.parity_target) {
+  // Terminal check (idempotent re-entry): parity target met AND backlog empty
+  // means we're done — no more rounds. (Parity is measured post-merge in
+  // paritying; deploy happens per-round via opening_pr → … → deploying.)
+  if (
+    site.parity_score !== null &&
+    site.parity_score >= site.parity_target &&
+    site.issues_open === 0
+  ) {
     await updateSite(site.id, {
-      status: "deploying",
-      phase_detail: `parity ${site.parity_score} >= ${site.parity_target}, going to deploy`,
+      status: "done",
+      phase_detail: `parity ${site.parity_score}% with empty backlog — done`,
+      finished_at: new Date().toISOString(),
       last_progress_at: new Date().toISOString(),
     });
-    await addEvent(site.id, `Parity reached ${site.parity_score}% — CF deploy`);
+    await addEvent(
+      site.id,
+      `Parity ${site.parity_score}% and backlog empty — done`,
+    );
     return;
   }
   if (deps.inflight.has(site.id)) return;
@@ -111,6 +121,36 @@ export async function fixing(
     }));
   }
 
+  // New round: cut a fresh branch off main so this batch gets its own PR that
+  // is reviewed + merged independently. Only when starting a fresh round
+  // (pr_number null); a review-reject keeps pr_number set → same branch.
+  if (!isSimulation(ctx) && !site.pr_number) {
+    if (!site.target_repo) throw new Error("target_repo not set");
+    const ref = parseRepo(site.target_repo);
+    const round = site.fix_sessions_done + 1;
+    const roundBranch = `migration/fix-${round}`;
+    if (site.work_branch !== roundBranch) {
+      await ensureBranch(ctx, ref, roundBranch, "main");
+      await updateSite(site.id, {
+        work_branch: roundBranch,
+        pr_number: null,
+        pr_url: null,
+        preview_ready: false,
+      });
+      site = {
+        ...site,
+        work_branch: roundBranch,
+        pr_number: null,
+        pr_url: null,
+        preview_ready: false,
+      };
+      await addEvent(
+        site.id,
+        `New fix round on branch ${roundBranch} (off main)`,
+      );
+    }
+  }
+
   const sessionNumber = site.fix_sessions_done + 1;
   const run = await createRun({
     siteId: site.id,
@@ -186,6 +226,7 @@ export async function fixing(
           meta: { issues: { resolved: [] } },
         });
         await updateSite(site.id, {
+          status: "opening_pr",
           fix_sessions_done: sessionNumber,
           issues_open: current.issues_open - resolvedCount,
           issues_closed: current.issues_closed + resolvedCount,
@@ -226,18 +267,19 @@ export async function fixing(
         },
       });
       await updateSite(site.id, {
+        status: "opening_pr",
         fix_sessions_done: sessionNumber,
         sandbox_session_id: null,
         phase_thread_id: null, // each batch is a fresh, narrow conversation
         transient_retries: 0,
-        phase_detail: `fix ${sessionNumber}: ${resolved.length}/${batch.length} resolved, backlog ${openIssues.open}`,
+        phase_detail: `fix ${sessionNumber}: ${resolved.length}/${batch.length} resolved, backlog ${openIssues.open} — opening PR`,
         last_progress_at: new Date().toISOString(),
       });
       await addEvent(
         site.id,
-        `Fix session ${sessionNumber}: resolved ${resolved.length > 0 ? resolved.map((n) => `#${n}`).join(", ") : "none"}${blocked.length > 0 ? ` · blocked ${blocked.map((b) => `#${b.number}`).join(", ")}` : ""} — backlog ${openIssues.open}`,
+        `Fix session ${sessionNumber}: resolved ${resolved.length > 0 ? resolved.map((n) => `#${n}`).join(", ") : "none"}${blocked.length > 0 ? ` · blocked ${blocked.map((b) => `#${b.number}`).join(", ")}` : ""} — backlog ${openIssues.open}, opening PR`,
       );
-      // next tick re-enters: backlog empty → paritying; budget out → needs_human
+      // advance: opening_pr → reviewing → merging → deploying → paritying
     } catch (err) {
       const message = err instanceof Error ? err.message : String(err);
       await finishRun(run.id, { status: "failed", logsTail: message });