diff --git a/AGENTS.md b/AGENTS.md
index ebf8fd7..7f4cc91 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -26,6 +26,21 @@ Please bear in mind what parts of the system are secrets and look after the secu
 In particular, all web pieces are made available on the internet with a strict CSP.  That means
 no inline styles or scripts.  You can find the web pieces (JS, HTML, css) in ./assets/
 
+## Churn management
+
+When you produced fixes, if possible (main branch, target is within 8 patches back,
+no intervening non-sai- tag) it's preferable to --amend apply the fixes directly to
+the patch that originated the problem, essentially editing the history, even if
+it means just doing that and not adding any fix patch or explanation.  Similarly, unless
+asked to produce a new patch or the goal is an explicit phased series, if it's on the
+main branch and we are iterating on the same work, and HEAD patch is yours from the
+last iteration, it's preferable to directly use --amend on it to commit.
+
+If the changes are for mixed purposes, if you initiate new core library changes or fixes,
+these should be broken out into their own patch, even if the rest relates to recent
+changes and is squashed in with those.
+
+
 ## Build testing
 
 Please don't worry about build-testing, just push patches when you are confident they are complete and have considered all affected code (ie, not half-assed) and ready and I will try them and report back with grounded information.
diff --git a/assets/sai.js b/assets/sai.js
index a76f7c9..887ac9f 100644
--- a/assets/sai.js
+++ b/assets/sai.js
@@ -4703,7 +4703,7 @@ function ws_open_sai()
 			case "com-warmcat-sai-artifact":
 				console.log(jso);
 
-				sai_arts += "<div class=\"sai_arts\"><img src=\"artifact.svg\">&nbsp;<a href=\"artifacts/" +
+				sai_arts += "<div class=\"sai_arts\"><img src=\"artifact.svg\">&nbsp;<a href=\"/sai/artifacts/" +
 					san(jso.task_uuid) + "/" +
 					san(jso.artifact_down_nonce) + "/" +
 					san(jso.blob_filename) + "\">" +
diff --git a/src/builder/b-artifacts.c b/src/builder/b-artifacts.c
index 7636431..483b9d5 100644
--- a/src/builder/b-artifacts.c
+++ b/src/builder/b-artifacts.c
@@ -176,8 +176,14 @@ saib_artifact_state(void *userobj, void *sh, lws_ss_constate_t state,
 		}
 		unlink(ap->path);
 
-		ap->ns->count_artifacts--;
-		if (!ap->ns->count_artifacts) {
+		/*
+		 * Account for us on the ns, and if we were what it was waiting
+		 * for, destroy it.  If the ns is already destroying us via
+		 * saib_task_destroy(), it will complete itself.
+		 */
+
+		if (ap->ns && ap->ns->count_artifacts &&
+		    !--ap->ns->count_artifacts && !ap->ns->destroying) {
 			lwsl_notice("%s: last artifact completed, destroying ns now\n", __func__);
 			saib_task_destroy(ap->ns);
 		}
diff --git a/src/builder/b-sai.c b/src/builder/b-sai.c
index 53e8039..88320a8 100644
--- a/src/builder/b-sai.c
+++ b/src/builder/b-sai.c
@@ -504,6 +504,24 @@ void sigint_handler(int sig)
 void
 sai_ns_destroy(struct sai_nspawn *ns)
 {
+	/*
+	 * In-flight artifact uploads still reference the ns; their SS handles
+	 * outlive it until lws_context_destroy() tears them down, which would
+	 * then touch freed ns.  Destroy them first, and mark the ns as already
+	 * destroying so their DESTROYING doesn't try to run the full task
+	 * teardown from inside builder shutdown.
+	 */
+
+	ns->destroying = 1;
+
+	lws_start_foreach_dll_safe(struct lws_dll2 *, d, d1,
+				   ns->artifact_owner.head) {
+		sai_artifact_t *ap = lws_container_of(d, sai_artifact_t, list);
+		struct lws_ss_handle *h = ap->ss;
+
+		lws_ss_destroy(&h);
+	} lws_end_foreach_dll_safe(d, d1);
+
 	lws_dll2_remove(&ns->list);
 	free(ns);
 }
diff --git a/src/builder/b-task.c b/src/builder/b-task.c
index 749cd84..c553241 100644
--- a/src/builder/b-task.c
+++ b/src/builder/b-task.c
@@ -351,7 +351,18 @@ saib_task_destroy(struct sai_nspawn *ns)
 {
 	int n;
 
-	lwsl_notice("====== saib_task_destroy START (ns=%p, uuid=%s, spm=%p) ======\n", 
+	/*
+	 * The last in-flight artifact upload calls this from its own
+	 * DESTROYING; if we got here first via the cleaner or cancel paths, the
+	 * outstanding artifact destroys below would otherwise call us back
+	 * reentrantly.
+	 */
+
+	if (ns->destroying)
+		return;
+	ns->destroying = 1;
+
+	lwsl_notice("====== saib_task_destroy START (ns=%p, uuid=%s, spm=%p) ======\n",
 		(void*)ns, ns->task ? ns->task->uuid : "null", (void*)ns->spm);
 
 	lwsl_notice("%s: destroying task %s\n", __func__,
@@ -361,6 +372,24 @@ saib_task_destroy(struct sai_nspawn *ns)
 	lws_sul_cancel(&ns->sul_task_cancel);
 
 	/*
+	 * Any artifact uploads still referencing us must go first... their
+	 * DESTROYING unlinks their temp file and accounts against
+	 * ns->count_artifacts, but skips the recursive destroy since we are
+	 * already destroying.
+	 */
+
+	lws_start_foreach_dll_safe(struct lws_dll2 *, d, d1,
+				   ns->artifact_owner.head) {
+		sai_artifact_t *ap = lws_container_of(d, sai_artifact_t, list);
+		struct lws_ss_handle *h = ap->ss;
+
+		lwsl_notice("%s: destroying in-flight artifact %s\n", __func__,
+			    ap->path);
+
+		lws_ss_destroy(&h);
+	} lws_end_foreach_dll_safe(d, d1);
+
+	/*
 	 * If able, builder should reintroduce himself to get
 	 * another task
 	 */
@@ -617,6 +646,17 @@ artifact_glob_cb(void *data, const char *path)
 	/* take a copy so we can unlink the path later */
 	lws_strncpy(ap->path, upp, sizeof(ap->path));
 
+	/*
+	 * The upload outlives the task's own steps and must know its ns, so
+	 * DESTROYING can account for it and destroy the ns when the last one
+	 * finishes; and saib_task_destroy() can find and kill in-flight uploads
+	 * if it goes first.  The SS user object is zalloc'd, so without this
+	 * ap->ns is NULL.
+	 */
+
+	ap->ns = ns;
+	lws_dll2_add_tail(&ap->list, &ns->artifact_owner);
+
 	lwsl_notice("%s: artifact ss created '%s'\n", __func__, ap->path);
 	ns->count_artifacts++;
 
@@ -646,6 +686,9 @@ artifact_glob_cb(void *data, const char *path)
  * of the task and reset the nspawn.
  */
 
+/* cap on how long we let artifact uploads hold the ns alive */
+#define SAIB_ARTIFACT_UPLOAD_MAX_US (10 * 60 * LWS_USEC_PER_SEC)
+
 static void
 saib_start_artifact_upload(struct sai_nspawn *ns)
 {
@@ -776,9 +819,21 @@ scan:
 		/* no artifacts to hang around for... nuke the ns now */
 		lws_sul_cancel(&ns->sul_cleaner);
 		saib_task_destroy(ns);
-	} else
+	} else {
 		lwsl_notice("%s: created / waiting on %d artifact uploads\n",
 				__func__, ns->count_artifacts);
+
+		/*
+		 * The ns now lives until the last upload destroys it, not the
+		 * 20s task grace timer... but keep a much longer hard cap so
+		 * a stream stuck retrying can't pin the ns (and the builder's
+		 * idle / power state) forever.
+		 */
+
+		lws_sul_schedule(builder.context, 0, &ns->sul_cleaner,
+				 saib_sub_cleaner_cb,
+				 SAIB_ARTIFACT_UPLOAD_MAX_US);
+	}
 }
 
 void
diff --git a/src/web/w-comms.c b/src/web/w-comms.c
index 0280d50..55a3ef1 100644
--- a/src/web/w-comms.c
+++ b/src/web/w-comms.c
@@ -52,6 +52,7 @@ typedef enum {
 	SHMUT_BROWSE,
 	SHMUT_STATUS,
 	SHMUT_ARTIFACTS,
+	SHMUT_ARTIFACTS_SAI,
 	SHMUT_LOGIN
 } sai_http_murl_t;
 
@@ -60,6 +61,7 @@ static const char * const well_known[] = {
 	"/sai/browse",
 	"/status",
 	"/artifacts/", /* HTTP api for accessing build artifacts */
+	"/sai/artifacts/", /* same, via the /sai mount the pages live under */
 	"/login"
 };
 
@@ -451,14 +453,17 @@ w_callback_ws(struct lws *wsi, enum lws_callback_reasons reason, void *user,
 			goto passthru;
 
 		case SHMUT_ARTIFACTS:
+		case SHMUT_ARTIFACTS_SAI:
 			/*
 			 * HTTP Bulk GET interface for artifact download
 			 *
-			 * /artifacts/<taskhash>/<down_nonce>/filename
+			 * [/sai]/artifacts/<taskhash>/<down_nonce>/filename
 			 */
 			lwsl_notice("%s: SHMUT_ARTIFACTS\n", __func__);
 			pss->artifact_offset = 0;
-			if (saiw_get_blob(vhd, (const char *)in + 11,
+			if (saiw_get_blob(vhd,
+					  (const char *)in +
+					  (mu == SHMUT_ARTIFACTS ? 11 : 15),
 					  &pss->pdb_artifact,
 					  &pss->blob_artifact,
 					  &pss->artifact_length)) {