@skyphusion/search-mcp 0.1.3 → 0.2.0

package/README.md

@@ -108,7 +108,31 @@ npm run sync:run     # isolated clone root, sync all targets, optional reindex
 
 The sync remaps non-native extensions (`.ts`, `.tsx`, extensionless `Dockerfile`, `.service`, etc.) to `.txt` keys so AI Search indexes them. See `scripts/sync-ingest.mjs`.
 
-If you name a target `public`, optional boundary checks refuse overlap with `restrictedRepos` and can verify GitHub repo visibility before upload.
+### Reindex dispatch
+
+AI Search rejects a new reindex job for two distinct reasons, and `sync-runner` clears both
+before dispatching:
+
+1. **A job is in flight.** Firing anyway does not queue behind it; Cloudflare ends the running
+   job with `end_reason: "new_job_has_started"` and restarts. So we wait for `ended_at`.
+2. **The post-job cooldown.** Even once a job ends, a new one is refused for a cooldown window
+   with `sync_in_cooldown [code: 7020]`. Waiting for the job to end is necessary but not
+   sufficient, so we retry until it clears.
+
+Waiting (rather than skipping) means the job we start always lands strictly after our own
+upload, so it sees every object this run wrote. Merge bursts still coalesce: a waiting run holds
+the workflow concurrency group, and GitHub keeps only the newest queued run, so the runs behind
+it collapse instead of each firing their own reindex.
+
+Each wait has its own budget (10 min in-flight, 10 min cooldown) rather than one shared
+deadline, since the two are additive on a perfectly healthy path: a run can wait minutes for an
+in-flight reindex and then still owe a cooldown wait.
+
+The measured cooldown is short (rejected at 10s after a job ends, accepted at 32s), so the
+budgets are far larger than they need to be today. That is deliberate. The measurement is an
+observation, not a contract, and a budget sized to it would turn ordinary upstream variance into
+red builds. If a budget is exhausted the run fails loudly and says what it means: the R2 corpus
+uploaded fine, nothing is lost, the index lags until the next sync or the daily backstop.
 
 ## Ask widget
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@skyphusion/search-mcp",
-  "version": "0.1.3",
+  "version": "0.2.0",
   "description": "Bearer-gated MCP server and streaming query Worker over Cloudflare AI Search, plus a git-to-R2 corpus sync toolkit.",
   "license": "AGPL-3.0-only",
   "author": "Conrad Rockenhaus <conrad@skyphusion.org>",

package/scripts/sync-runner.mjs

@@ -6,6 +6,7 @@
 //   node scripts/sync-runner.mjs corpus          # one target
 //   node scripts/sync-runner.mjs --dry-run
 //   node scripts/sync-runner.mjs --no-reindex
+//   node scripts/sync-runner.mjs --reindex-only     # skip clone+sync, dispatch reindex only
 //   node scripts/sync-runner.mjs --no-github-verify
 //
 // Env:
@@ -31,9 +32,11 @@ export function parseArgs(argv, allTargetNames = []) {
   const flags = new Set(argv.filter((a) => a.startsWith("--")));
   const positional = argv.filter((a) => !a.startsWith("--"));
   const requested = positional[0];
+  const reindexOnly = flags.has("--reindex-only");
   return {
     targets: requested ? [requested] : [...allTargetNames],
     dryRun: flags.has("--dry-run"),
+    sync: !reindexOnly,
     reindex: !flags.has("--no-reindex") && !flags.has("--dry-run"),
     passThrough: [...flags].filter((f) => f === "--dry-run" || f === "--no-github-verify"),
   };
@@ -109,40 +112,209 @@ function syncRepoTree(org, repo, root, authArgs) {
   return mode;
 }
 
-function reindexInstance(instance) {
-  execFileSync("npx", ["wrangler", "ai-search", "jobs", "create", instance, "--json"], {
-    stdio: "inherit",
-    cwd: REPO,
-  });
+/**
+ * One dispatch attempt. stderr is PIPED rather than inherited: wrangler reports
+ * `sync_in_cooldown [code: 7020]` on stderr, and with stdio "inherit" that text never reaches
+ * the thrown error, so the retry logic could not tell a cooldown from a real failure. We echo
+ * the captured output so the run log is unchanged.
+ */
+function reindexOnce(instance) {
+  try {
+    const out = execFileSync(
+      "npx",
+      ["wrangler", "ai-search", "jobs", "create", instance, "--json"],
+      { cwd: REPO, encoding: "utf8", stdio: ["ignore", "pipe", "pipe"] },
+    );
+    if (out) process.stdout.write(out);
+  } catch (err) {
+    if (err.stdout) process.stdout.write(err.stdout);
+    if (err.stderr) process.stderr.write(err.stderr);
+    throw err;
+  }
+}
+
+/**
+ * True when any reindex job for the instance is still running. AI Search reports an
+ * unfinished job as `ended_at: null`.
+ *
+ * Firing a job while one is in flight does not queue it: Cloudflare ends the running job
+ * with `end_reason: "new_job_has_started"` and starts over. On the biggest corpus the
+ * reindex outlives the sync that triggers it, so an unguarded dispatch during a merge burst
+ * restarts the index pass repeatedly and it never settles (#12).
+ */
+export function jobInFlight(jobs) {
+  return (jobs || []).some((job) => !job.ended_at);
 }
 
-export function run(plan, deps) {
-  const { repos, targets, instances, reindex } = plan;
-  const { cloneTree, runSync, runReindex, log = console.log, error = console.error } = deps;
+function listJobs(instance) {
+  const out = execFileSync(
+    "npx",
+    ["wrangler", "ai-search", "jobs", "list", instance, "--json"],
+    { cwd: REPO, encoding: "utf8" },
+  );
+  return JSON.parse(out);
+}
+
+// Two independent budgets, deliberately NOT one shared deadline: the phases are additive on a
+// healthy path (wait out an in-flight reindex, ~6 min on the biggest corpus, and THEN still owe
+// a cooldown wait), so one shared bound would fail a run that did nothing wrong.
+//
+// Both budgets are far larger than the measured cooldown (rejected 10s after a job ends,
+// accepted at 32s; 2026-07-16). That is intentional. The measurement is an observation, not a
+// contract Cloudflare owes us, and it may move with corpus size or load. Sizing to today's
+// number would convert ordinary upstream variance into red builds; sizing in minutes costs
+// nothing when the cooldown is short, because the retry clears and the run moves on.
+export const REINDEX_INFLIGHT_TIMEOUT_MS = 10 * 60 * 1000;
+export const REINDEX_COOLDOWN_TIMEOUT_MS = 10 * 60 * 1000;
+export const REINDEX_POLL_MS = 15 * 1000;
 
-  log(`Refreshing ${repos.length} repo tree(s) ...`);
-  const failedClones = [];
-  for (const repo of repos) {
+/**
+ * AI Search refuses a new job for a cooldown window after the previous one ENDS, distinct from
+ * the job being in flight. Measured 2026-07-16: rejected 10s after a job ended, accepted at
+ * 32s. Waiting for `ended_at` alone is necessary but not sufficient.
+ */
+export function isCooldownError(err) {
+  const text = `${err?.stdout || ""}${err?.stderr || ""}${err?.message || ""}`;
+  return /sync_in_cooldown|code:\s*7020/.test(text);
+}
+
+/**
+ * Block until no reindex job is in flight for the instance, so our dispatch lands strictly
+ * after the running job finishes instead of superseding it.
+ *
+ * On timeout we dispatch ANYWAY and say so loudly. A job still running past the bound (well
+ * over the ~6 min a full internal reindex takes) is itself anomalous, and superseding it is
+ * the correct recovery: the replacement job sees every object currently in R2. The failure
+ * mode we want is "one supersession in a pathological case", never "the corpus quietly goes
+ * unindexed", which is the class of bug that a trailing-pass design introduced.
+ */
+export async function awaitReindexSlot(instance, deps) {
+  const {
+    listJobs,
+    sleep,
+    now = () => Date.now(),
+    timeoutMs = REINDEX_INFLIGHT_TIMEOUT_MS,
+    pollMs = REINDEX_POLL_MS,
+    log = console.log,
+  } = deps;
+  const deadline = now() + timeoutMs;
+  let waited = false;
+  while (jobInFlight(listJobs(instance))) {
+    if (now() >= deadline) {
+      log(
+        `  !! reindex for ${instance} still in flight after ${Math.round(timeoutMs / 1000)}s; ` +
+          "dispatching anyway and superseding it. A job running this long is anomalous; " +
+          "the replacement job reindexes the full corpus, so this is safe but worth a look.",
+      );
+      return { waited, timedOut: true };
+    }
+    if (!waited) {
+      log(`  reindex in flight for ${instance}; waiting for it to finish before dispatching.`);
+    }
+    waited = true;
+    await sleep(pollMs);
+  }
+  return { waited, timedOut: false };
+}
+
+/**
+ * Dispatch a reindex, retrying while AI Search reports its post-job cooldown.
+ *
+ * Cooldown is transient and short (measured well under the budget), so the routine burst case
+ * clears here and never goes red. Exhausting the budget therefore means something genuinely
+ * anomalous upstream, and we fail loud rather than exit green: a green run while indexing is
+ * actually stalled is the work-blind failure mode this whole issue exists to remove. The
+ * message carries the honest blast radius so the red is actionable, not alarming.
+ */
+export async function dispatchWithCooldownRetry(instance, deps) {
+  const {
+    runReindexOnce,
+    sleep,
+    now = () => Date.now(),
+    timeoutMs = REINDEX_COOLDOWN_TIMEOUT_MS,
+    pollMs = REINDEX_POLL_MS,
+    log = console.log,
+  } = deps;
+  const deadline = now() + timeoutMs;
+  let announced = false;
+  for (;;) {
     try {
-      const mode = cloneTree(repo);
-      log(`  ok  ${repo} (${mode})`);
+      runReindexOnce(instance);
+      return { retried: announced };
     } catch (err) {
-      failedClones.push(repo);
-      error(`  !!  ${repo}: git failed (${describeExit(err)})`);
+      if (!isCooldownError(err)) throw err;
+      if (now() >= deadline) {
+        throw new Error(
+          `reindex for ${instance} blocked by AI Search cooldown for longer than ` +
+            `${Math.round(timeoutMs / 1000)}s. The R2 corpus uploaded OK and no data is lost; ` +
+            "the index lags until the next sync or the daily backstop. A cooldown this " +
+            "persistent is upstream and anomalous, so this run fails loudly on purpose.",
+        );
+      }
+      if (!announced) {
+        log(`  ${instance} is in AI Search cooldown; retrying until it clears.`);
+        announced = true;
+      }
+      await sleep(pollMs);
     }
   }
+}
 
-  if (failedClones.length) {
-    error(
-      `BLOCKER: ${failedClones.length} clone/fetch failure(s) [${failedClones.join(", ")}]; ` +
-        "aborting before sync so no repo corpus is pruned or synced stale.",
-    );
-    return { exitCode: 1, synced: false };
-  }
+/**
+ * The real dependencies main() hands to awaitReindexSlot.
+ *
+ * Exported and built here on purpose. Every other seam in this file is injected, so a unit
+ * suite full of stubs proves nothing about the production wiring: an undefined symbol here
+ * rides a green suite and a clean `node --check` all the way to prod, then throws
+ * ReferenceError on the first real dispatch. Constructing this object evaluates each binding,
+ * so the wiring test fails loudly if one of them stops existing.
+ */
+export function productionReindexDeps() {
+  return {
+    listJobs,
+    runReindexOnce: reindexOnce,
+    sleep: (ms) => new Promise((resolve) => setTimeout(resolve, ms)),
+  };
+}
 
-  for (const target of targets) {
-    log(`\n=== sync ${target} ===`);
-    runSync(target);
+export async function run(plan, deps) {
+  const { repos, targets, instances, reindex, sync = true } = plan;
+  const {
+    cloneTree,
+    runSync,
+    runReindex,
+    awaitSlot,
+    log = console.log,
+    error = console.error,
+  } = deps;
+
+  if (sync) {
+    log(`Refreshing ${repos.length} repo tree(s) ...`);
+    const failedClones = [];
+    for (const repo of repos) {
+      try {
+        const mode = cloneTree(repo);
+        log(`  ok  ${repo} (${mode})`);
+      } catch (err) {
+        failedClones.push(repo);
+        error(`  !!  ${repo}: git failed (${describeExit(err)})`);
+      }
+    }
+
+    if (failedClones.length) {
+      error(
+        `BLOCKER: ${failedClones.length} clone/fetch failure(s) [${failedClones.join(", ")}]; ` +
+          "aborting before sync so no repo corpus is pruned or synced stale.",
+      );
+      return { exitCode: 1, synced: false };
+    }
+
+    for (const target of targets) {
+      log(`\n=== sync ${target} ===`);
+      runSync(target);
+    }
+  } else {
+    log("Reindex-only: skipping clone and sync.");
   }
 
   const failedReindex = [];
@@ -150,7 +322,8 @@ export function run(plan, deps) {
     for (const instance of instances) {
       log(`\n=== reindex ${instance} ===`);
       try {
-        runReindex(instance);
+        if (awaitSlot) await awaitSlot(instance);
+        await runReindex(instance);
       } catch (err) {
         failedReindex.push(instance);
         error(`  !! reindex failed for ${instance} (${describeExit(err)})`);
@@ -162,10 +335,10 @@ export function run(plan, deps) {
 
   if (failedReindex.length) {
     error(`BLOCKER: ${failedReindex.length} reindex failure(s) [${failedReindex.join(", ")}].`);
-    return { exitCode: 1, synced: true };
+    return { exitCode: 1, synced: sync };
   }
   log("\nCorpus sync complete.");
-  return { exitCode: 0, synced: true };
+  return { exitCode: 0, synced: sync };
 }
 
 function loadConfig() {
@@ -177,7 +350,7 @@ function loadConfig() {
   return JSON.parse(readFileSync(path, "utf8"));
 }
 
-function main() {
+async function main() {
   const cfg = loadConfig();
   const allTargets = Object.keys(cfg.targets || {});
   const opts = parseArgs(process.argv.slice(2), allTargets);
@@ -196,8 +369,8 @@ function main() {
   console.log(`Corpus root: ${root}`);
   console.log(`Org: ${org}`);
 
-  const result = run(
-    { repos, targets: opts.targets, instances, reindex: opts.reindex },
+  const result = await run(
+    { repos, targets: opts.targets, instances, reindex: opts.reindex, sync: opts.sync },
     {
       cloneTree: (repo) => syncRepoTree(org, repo, root, authArgs),
       runSync: (target) =>
@@ -206,11 +379,12 @@ function main() {
           cwd: REPO,
           env: { ...process.env, SYNC_REPO_ROOT: root },
         }),
-      runReindex: reindexInstance,
+      awaitSlot: (instance) => awaitReindexSlot(instance, productionReindexDeps()),
+      runReindex: (instance) => dispatchWithCooldownRetry(instance, productionReindexDeps()),
     },
   );
   process.exit(result.exitCode);
 }
 
 const invokedDirectly = process.argv[1] && process.argv[1] === fileURLToPath(import.meta.url);
-if (invokedDirectly) main();
+if (invokedDirectly) await main();