@gamaze/hicortex 0.23.1 → 0.24.0

package/assets/dashboard.html

@@ -76,7 +76,7 @@
       badged "best match" (server composite order kept — ranking is #425).
       All-below-floor renders "no matches for '<q>'" — NN junk never shows.
     - ONE-SCREEN LAYOUT: header + band + field fit the first viewport at
-      ~1440×900 (MBP5 reference; 1280×800 verified). Band compacted; the
+      ~1440×900 (reference resolution; 1280×800 verified). Band compacted; the
       field takes the viewport remainder (clamp with a measured offset —
       see the #brainwrap CSS note). Scroll stays available as overflow.
 
@@ -574,7 +574,7 @@
 
   /* ---- the brain-field ---- */
   /* ONE-SCREEN AT-REST LAYOUT (#409 polish, owner ruling 2026-09-13): header
-     + stats band + field fit the first viewport at ~1440×900 (MBP5 is the
+     + stats band + field fit the first viewport at ~1440×900 (that is the
      reference; 1280×800 verified too). The field takes the viewport REMAINDER
      — the subtracted constant is the MEASURED header+band+paddings total on
      the reference store (Chrome 152, headless): 412px at the 4-column width,
@@ -865,6 +865,15 @@
   /* #423 phase 3: the capture-pause badge — the same amber family as the
      held badge (an operator state, not an error) */
   .cap-paused{ color:var(--amber); margin-left:5px; font-size:9px; letter-spacing:.08em; }
+  /* #530 follow-up: the distill-inbox row — stale badge + bar fill in the
+     amber family once the oldest queued item passes the payload's ECHOED warn
+     age. [hidden] must win over .cap-row2's flex (author display beats the
+     UA's hidden — the #billing-banner idiom) so a pre-#529 server shows
+     nothing, theme tokens only */
+  .cap-qrow[hidden]{ display:none; }
+  .cap-stale{ color:var(--amber); margin-left:5px; }
+  .cap-qrow .fill.warn{ background:var(--amber); }
+  .cap-qrow.stale .cap-dot{ background:var(--amber); }
 
   /* ================= the memory dropdown — opens UNDER the left search ================= */
   #panel{
@@ -1277,6 +1286,11 @@
     <div class="brow">
       <div class="bseg" id="g-capture">
         <div class="gh"><span class="t">Capture health</span><span class="sub" id="cap-sub">memory counts</span></div>
+        <!-- #530 follow-up: the distill inbox — the drain-side counterpart of
+             this card's delivery-side health. Pinned above the scrollable
+             #cap-rows; hidden until buildCapture's payload guard unhides it
+             (pre-#529 servers omit the block). -->
+        <div class="cap-row2 cap-qrow" id="cap-queue" hidden></div>
         <div id="cap-rows"></div>
       </div>
       <div class="bseg" id="g-roll">
@@ -1395,7 +1409,7 @@
       </label>
       <label class="mm-check"><input type="checkbox" id="mm-think"> enable thinking <small>(local endpoints only)</small></label>
     </div>
-    <div class="mm-note">applies on server restart</div>
+    <div class="mm-note" id="mm-note">applies on server restart</div>
     <div class="mm-foot">
       <button class="btn" id="mm-cancel">Cancel</button>
       <button class="btn save" id="mm-save">Save</button>
@@ -2414,7 +2428,7 @@ function draw(now, dt){
     if(L.rk>=lodK) continue;                             // zoom LOD (lines only)
     const a=mems.get(L.a), b=mems.get(L.b);
     if(!a||!b||!aliveAt(a,t)||!aliveAt(b,t)) continue;
-    /* #474 final polish: the both-endpoints-in-window gate killed the mesh — links draw between ANY placed memories (bedrock parity: the full LOD-tamed mesh), endpoints carry positions regardless of window membership */
+    /* #474 final polish: the both-endpoints-in-window gate killed the mesh — links draw between ANY placed memories (reference parity: the full LOD-tamed mesh), endpoints carry positions regardless of window membership */
     if(a._emerged===false||b._emerged===false) continue; // thread waits for both endpoints
     const [ax,ay]=bpos(a,now,t), [bx,by]=bpos(b,now,t);
     const ctrl=segCtrlN(ax,ay,bx,by,(L.use>18?1:-1));
@@ -3398,8 +3412,8 @@ function renderRecBand(ups){
   /* No trend arrow (owner 2026-09-20: removed — unclear). Its renderer and
      the element writes are removed WITH it: a leftover write to a removed
      element is a boot-killing TypeError whenever the recall band carries
-     real data (bedrock 0.22.2: "0 memories" — boot died before the counter
-     was ever seeded). */
+     real data (a production instance on 0.22.2: "0 memories" — boot died
+     before the counter was ever seeded). */
 }
 
 /* ---------- Memory Precision — Level 1, the pushed index (#476). A compact
@@ -3509,32 +3523,44 @@ function renderModelCard(){
    only as api_key_set), writes PUT with the subset the user touched. The
    api-key input is WRITE-ONLY: prefilled with nothing, sent only when
    non-empty ("leave blank to keep"). enable_thinking is sent only when the
-   user toggles the checkbox (tri-state not needed). ---------- */
+   user toggles the checkbox (tri-state not needed).
+   Hosted mode (#514): the GET carries managed:true — the hosting service
+   owns the endpoint + credentials, so their inputs are disabled and never
+   ride the PUT body (a save cannot hit the server's managed-field 403). */
 let mmThinkTouched=false;
 function openModelModal(){
   if(!M) return;               // no settings payload → no dialog (degrade)
   const mdl=el('model-modal');
+  const managed=!!M.managed;
   el('mm-backend').value=M.backend||'';   // '' = auto / OpenAI-compatible
   el('mm-baseurl').value=M.base_url||'';
   el('mm-model').value=M.model||'';
   const ak=el('mm-apikey'); ak.value='';
-  ak.placeholder=M.api_key_set?'configured — leave blank to keep':'api key (optional for local endpoints)';
+  ak.placeholder=managed?'managed by the hosting service'
+    :(M.api_key_set?'configured — leave blank to keep':'api key (optional for local endpoints)');
   el('mm-maxtok').value=M.max_tokens!=null?String(M.max_tokens):'';
   el('mm-think').checked=!!M.enable_thinking;
   mmThinkTouched=false;
+  for(const id of ['mm-backend','mm-baseurl','mm-apikey']) el(id).disabled=managed;
+  el('mm-note').textContent=managed
+    ?'endpoint and credentials are managed by the hosting service — applies on server restart'
+    :'applies on server restart';
   mdl.hidden=false;
-  el('mm-baseurl').focus();
+  (managed?el('mm-model'):el('mm-baseurl')).focus();
 }
 function closeModelModal(){ el('model-modal').hidden=true; }
 async function saveModelSettings(){
   /* Build the PUT body from CURRENT field state: empty base_url/model/max
      tokens = clear (null server-side); backend '' clears the named-backend
-     key; api_key only rides when typed; enable_thinking only when toggled. */
+     key; api_key only rides when typed; enable_thinking only when toggled.
+     Managed (hosted) mode: backend/base_url/api_key never ride the body. */
   const body={};
-  body.backend=el('mm-backend').value;
-  const bu=el('mm-baseurl').value.trim(); body.base_url=bu?bu:null;
+  if(!M.managed){
+    body.backend=el('mm-backend').value;
+    const bu=el('mm-baseurl').value.trim(); body.base_url=bu?bu:null;
+    const ak=el('mm-apikey').value; if(ak) body.api_key=ak;
+  }
   const mo=el('mm-model').value.trim(); body.model=mo?mo:null;
-  const ak=el('mm-apikey').value; if(ak) body.api_key=ak;
   const mt=el('mm-maxtok').value.trim(); body.max_tokens=mt?Number(mt):null;
   if(mmThinkTouched) body.enable_thinking=el('mm-think').checked;
   try{
@@ -3754,6 +3780,31 @@ function agoLabel(iso){
 function buildCapture(){
   const host=el('cap-rows');
   host.innerHTML='';
+  /* #530 follow-up: the distill-inbox row — rendered FIRST because every
+     render branch below returns early (a call after them would only run in
+     the degrade). The absent-block guard keeps it hidden on a pre-#529
+     server; the warn threshold is always the payload's echo (#476
+     degrade-literal precedent), never a page literal. */
+  const qrow=el('cap-queue');
+  qrow.innerHTML='';
+  qrow.className='cap-row2 cap-qrow';
+  const q=D&&D.distill_queue?D.distill_queue:null;
+  qrow.hidden=!q;
+  if(q){
+    const thr=q.warn_age_hours||24;
+    const oldest=(typeof q.oldest_age_hours==='number')?q.oldest_age_hours:null;
+    const stale=oldest!==null&&oldest>thr;
+    if(stale) qrow.classList.add('stale');
+    qrow.title='Distill inbox — deliveries are stored durably at POST time and distilled by the scheduled nightly run before consolidation'
+      +(stale?' — the oldest item is '+oldest+' h old (warn age '+thr+' h): the scheduled drain is behind':'');
+    qrow.innerHTML='<span class="cap-dot"></span>'
+      +'<span class="hn">Distill inbox</span>'
+      +'<span class="cap-bar"><span class="fill'+(stale?' warn':'')+'" style="width:'+(oldest===null?0:Math.min(1,oldest/thr)*100)+'%"></span></span>'
+      +'<span class="st"><b>'+fmt(q.depth||0)+'</b> queued'
+      +(oldest===null?' · inbox empty':' · oldest '+(oldest>=48?(oldest/24).toFixed(1)+'d':oldest.toFixed(1)+'h'))
+      +(stale?'<span class="cap-stale" title="oldest queued item past the '+thr+' h warn age — the scheduled drain is behind">&#9888; stale</span>':'')
+      +'</span>';
+  }
   const sub=el('cap-sub');
   const ch=D&&D.capture_health?D.capture_health:null;
   /* fix round 7: the rolling-window view — the card shows the normal 30 day

package/dist/calibration.d.ts

@@ -416,3 +416,34 @@ export declare function resolveOllamaFlushEvery(): number;
  *  HICORTEX_OLLAMA_FLUSH_WAIT_MS wins; invalid warns and keeps
  *  OLLAMA_FLUSH_WAIT_MS. */
 export declare function resolveOllamaFlushWaitMs(): number;
+/**
+ * Re-check interval (ms) while the drain's yield signal reports the LLM
+ * endpoint busy (#529 owner ruling 05.10.2026: interactive use always
+ * yields). The drain waits one interval between probe re-checks instead of
+ * polling in a tight loop; 30 s is short enough that the run resumes within
+ * half a minute of the owner going idle, long enough not to hammer the
+ * signal endpoint while a long interactive session runs.
+ */
+export declare const DRAIN_YIELD_POLL_MS = 30000;
+/**
+ * Per-item cap (ms) on yield waits (#529). The run deadline is the primary
+ * bound; this cap is the fail-open backstop so a misconfigured signal that
+ * always answers "busy" cannot stall one item past half an hour — the item
+ * proceeds and the next between-items checks decide again. 30 min matches
+ * the order of a long local-model generation, the longest legitimate busy.
+ */
+export declare const DRAIN_YIELD_WAIT_CAP_MS: number;
+/**
+ * Timeout (ms) for ONE yield-probe GET (#529). Deliberately short: the probe
+ * sits between drain items, and an endpoint that answers busy-signals at all
+ * answers them fast — anything slower reads as unreachable (fail-open).
+ */
+export declare const DRAIN_YIELD_PROBE_TIMEOUT_MS = 3000;
+/**
+ * Oldest-item age (hours) above which the distill inbox warns in
+ * `hicortex status` and on the dashboard (#529). 24 h ≈ two missed scheduled
+ * runs on the default 2×/day consolidation grid — items older than that are
+ * not "waiting for tonight" anymore, they are stuck, and the operator should
+ * look at the drain's last outcome.
+ */
+export declare const DRAIN_QUEUE_WARN_AGE_HOURS = 24;

package/dist/calibration.js

@@ -25,7 +25,7 @@
  */
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.STAGE_BELIEF_STRENGTH = exports.STAGE_FADING_STRENGTH = exports.STAGE_FADING_DAYS = exports.WEAK_PRIMARY_FLOOR = exports.CORRECTION_REWRITE_MIN_CONFIDENCE = exports.CORRECTION_MIN_SIMILARITY = exports.DEDUP_AUTO_MERGE_THRESHOLD = exports.IMPORTANCE_SETTLE_WINDOW_DAYS = exports.IMPORTANCE_CEILING = exports.BM25_WEIGHT_DOMAIN = exports.BM25_WEIGHT_PROJECT = exports.BM25_WEIGHT_BODY = exports.BOTH_CHANNEL_BOOST = exports.RRF_VECTOR_WEIGHT = exports.RRF_FTS_WEIGHT = exports.RRF_COMPOSITE_WEIGHT = exports.RRF_K = exports.SCOPE_AFFINITY_WEIGHT = exports.SUPERSEDED_DEMOTION = exports.RECENCY_HEAD_DAYS = exports.RECENCY_HEAD_WEIGHT = exports.RECENCY_HOURLY_DECAY = exports.SCORE_RECENCY_WEIGHT = exports.CONNECTIONS_SATURATION_DEGREE = exports.SCORE_CONNECTIONS_WEIGHT = exports.SCORE_STRENGTH_WEIGHT = exports.SCORE_SIMILARITY_WEIGHT = exports.VOLATILE_GATE_MAX_CHARS = exports.VOLATILE_STATUS_FILTER = exports.MEMORY_PRECISION_PROMPT_EXCERPT_CHARS = exports.MEMORY_PRECISION_DIVERGENCE_MIN_SHOWN = exports.MEMORY_PRECISION_REDUNDANT_ABOVE = exports.MEMORY_PRECISION_WINDOW_DAYS = exports.RECALL_USES_AXIS_MAX = exports.RECALL_USES_NORMAL_MAX = exports.RECALL_USES_LOW_MAX = exports.RECALL_RESHOW_TURNS = exports.NOVELTY_FLOOR_SLOTS = exports.RECALL_TITLE_CHARS = exports.RECALL_MIN_PROMPT_CHARS = exports.RECALL_MAX_ITEMS = exports.RECALL_MIN_SIMILARITY = exports.SESSION_INTENT_WEIGHT = exports.COLD_EXPOSURE_SLOTS = exports.RECENT_WINDOW_DAYS = exports.RECENT_LIMIT = exports.SEARCH_LIMIT = exports.PROMOTION_STRENGTH_FLOOR = exports.PROMOTION_RATE = exports.DECAY_HALF_LIFE_DAYS = void 0;
-exports.DIAGNOSTIC_ENV_TIER = exports.OLLAMA_FLUSH_WAIT_MS = exports.OLLAMA_FLUSH_EVERY = exports.NUM_CTX = exports.DEFAULT_FIRST_RUN_LOOKBACK_DAYS = exports.ENRICH_STRENGTH_DELTA = exports.STAGE_TRUTH_STRENGTH = void 0;
+exports.DRAIN_QUEUE_WARN_AGE_HOURS = exports.DRAIN_YIELD_PROBE_TIMEOUT_MS = exports.DRAIN_YIELD_WAIT_CAP_MS = exports.DRAIN_YIELD_POLL_MS = exports.DIAGNOSTIC_ENV_TIER = exports.OLLAMA_FLUSH_WAIT_MS = exports.OLLAMA_FLUSH_EVERY = exports.NUM_CTX = exports.DEFAULT_FIRST_RUN_LOOKBACK_DAYS = exports.ENRICH_STRENGTH_DELTA = exports.STAGE_TRUTH_STRENGTH = void 0;
 exports.resolveNumCtx = resolveNumCtx;
 exports.resolveOllamaFlushEvery = resolveOllamaFlushEvery;
 exports.resolveOllamaFlushWaitMs = resolveOllamaFlushWaitMs;
@@ -514,3 +514,40 @@ function resolveOllamaFlushWaitMs() {
     console.warn(`[hicortex] env HICORTEX_OLLAMA_FLUSH_WAIT_MS=${JSON.stringify(raw)} is not a positive finite number — using default ${exports.OLLAMA_FLUSH_WAIT_MS}.`);
     return exports.OLLAMA_FLUSH_WAIT_MS;
 }
+// ---------------------------------------------------------------------------
+// Distill inbox / drain family (#529) — delivery-compute separation. Delivery
+// (cheap, must never lose data) stores segments durably; compute (expensive,
+// scheduled) happens only inside the nightly's drain stage. These constants
+// shape the drain's interactive-yield seam and the inbox's visibility signal.
+// ---------------------------------------------------------------------------
+/**
+ * Re-check interval (ms) while the drain's yield signal reports the LLM
+ * endpoint busy (#529 owner ruling 05.10.2026: interactive use always
+ * yields). The drain waits one interval between probe re-checks instead of
+ * polling in a tight loop; 30 s is short enough that the run resumes within
+ * half a minute of the owner going idle, long enough not to hammer the
+ * signal endpoint while a long interactive session runs.
+ */
+exports.DRAIN_YIELD_POLL_MS = 30_000;
+/**
+ * Per-item cap (ms) on yield waits (#529). The run deadline is the primary
+ * bound; this cap is the fail-open backstop so a misconfigured signal that
+ * always answers "busy" cannot stall one item past half an hour — the item
+ * proceeds and the next between-items checks decide again. 30 min matches
+ * the order of a long local-model generation, the longest legitimate busy.
+ */
+exports.DRAIN_YIELD_WAIT_CAP_MS = 30 * 60_000;
+/**
+ * Timeout (ms) for ONE yield-probe GET (#529). Deliberately short: the probe
+ * sits between drain items, and an endpoint that answers busy-signals at all
+ * answers them fast — anything slower reads as unreachable (fail-open).
+ */
+exports.DRAIN_YIELD_PROBE_TIMEOUT_MS = 3_000;
+/**
+ * Oldest-item age (hours) above which the distill inbox warns in
+ * `hicortex status` and on the dashboard (#529). 24 h ≈ two missed scheduled
+ * runs on the default 2×/day consolidation grid — items older than that are
+ * not "waiting for tonight" anymore, they are stuck, and the operator should
+ * look at the drain's last outcome.
+ */
+exports.DRAIN_QUEUE_WARN_AGE_HOURS = 24;

package/dist/capture.d.ts

@@ -90,6 +90,13 @@ export interface PostResult {
         completion: number;
         total: number;
     };
+    /**
+     * #529: the 201 confirmed durable QUEUEING, not distillation — a queue-mode
+     * server stores the segment and the nightly's drain distills it in the
+     * scheduled run. Contract-wise identical to any other 201 (the cursor
+     * advances); the capture loop only swaps its log line on it.
+     */
+    queued?: boolean;
 }
 export type PostFn = (body: DistillBody) => Promise<PostResult>;
 export interface CaptureOptions {

package/dist/capture.js

@@ -301,7 +301,16 @@ async function captureBatches(batches, opts) {
                 }
                 if (advancesBoundary)
                     lastConfirmedEnd = seg.segEnd;
-                console.log(`[hicortex]     → ${result.distilled ?? 0} memories (segment ${body.segment_id})`);
+                // #529: a queue-mode 201 means "durably stored, distilled on the
+                // server's next scheduled run" — same confirmation, different log
+                // line, so a nightly log never reads "0 memories" as a failure when
+                // the server simply queued the segment.
+                if (result.queued) {
+                    console.log(`[hicortex]     Queued on server (segment ${body.segment_id}) — distilled on the next scheduled run`);
+                }
+                else {
+                    console.log(`[hicortex]     → ${result.distilled ?? 0} memories (segment ${body.segment_id})`);
+                }
                 for (const d of result.dropped ?? []) {
                     // #489: the response array carries BOTH gate kinds (substance +
                     // volatility) — the server's own log names the exact gate; this

package/dist/dashboard.d.ts

@@ -268,6 +268,20 @@ export interface DashboardData {
      * pre-#476 recall-card rendering when the block is absent.
      */
     memory_precision: MemoryPrecision;
+    /**
+     * #529 — the distill inbox: live queue depth + oldest-item age (the
+     * drain-side health signal; delivery-side health is capture_health
+     * above). ALWAYS present; `oldest_age_hours` null when empty. The warn
+     * threshold is ECHOED like every threshold (the page renders its own
+     * stale badge from it, never hardcodes 24). Live view only — no snapshot
+     * history (depth is ~0 at snapshot time by construction). A pre-#529
+     * server omits the block; the page guards.
+     */
+    distill_queue: {
+        depth: number;
+        oldest_age_hours: number | null;
+        warn_age_hours: number;
+    };
     digest: {
         date: string | null;
         run_at: string | null;
@@ -667,15 +681,19 @@ export interface DashboardModelSettings {
     /** Constant true — the daemon resolves config at boot, so every write
      *  lands on restart. The modal footnotes it. */
     applies_on_restart: true;
+    /** Present ONLY in hosted mode — the console uses it to disable the managed inputs (#514). */
+    managed?: true;
 }
 /**
  * The pure GET handler: echo the CONFIG values raw (null when unset — the UI
  * shows defaults as placeholders, so this never resolves them) + the runtime
- * provider from the daemon's in-memory llmConfig.
+ * provider from the daemon's in-memory llmConfig. `hostedMode` (#514) only
+ * adds `managed: true` to the body — self-hosted responses never carry the
+ * key at all (byte-identical wire).
  */
 export declare function handleDashboardModelGet(config: Record<string, unknown> | null | undefined, llmConfig: {
     provider: string;
-} | null): {
+} | null, hostedMode?: boolean): {
     status: 200;
     body: DashboardModelSettings;
 };
@@ -684,10 +702,14 @@ export declare function handleDashboardModelGet(config: Record<string, unknown>
  * injected writer (which THROWS on a malformed config — the adapter maps that
  * to a 500 and the file stays untouched), answer with the fresh GET shape
  * built from the post-write config. `null` REMOVES a config key.
+ *
+ * `hostedMode` (#514): the managed fields (see HOSTED_MANAGED_MODEL_KEYS) are
+ * rejected BEFORE any validation or persistence — a refused PUT writes
+ * nothing, not even the allowed keys that rode the same body.
  */
 export declare function handleDashboardModelPut(body: unknown, persist: (updates: Record<string, unknown>) => Record<string, unknown>, getLlmConfig: () => {
     provider: string;
-} | null): {
+} | null, hostedMode?: boolean): {
     status: number;
     body: DashboardModelSettings | {
         error: string;
@@ -696,21 +718,24 @@ export declare function handleDashboardModelPut(body: unknown, persist: (updates
 /**
  * Express adapter for GET /dashboard/model. Bearer-only by construction
  * (mounted after createAuthMiddleware — NO exemption, unlike the page shells:
- * this carries install config). Failures surface as a 500 {error}.
+ * this carries install config). Failures surface as a 500 {error}. The
+ * boot-resolved `hostedMode` flag (#514) threads straight through to the pure
+ * handler — it is a boot constant, not a per-request value.
  */
 export declare function dashboardModelGetHandler(getConfig: () => Record<string, unknown> | null | undefined, getLlmConfig: () => {
     provider: string;
-} | null): express.RequestHandler;
+} | null, hostedMode?: boolean): express.RequestHandler;
 /**
  * Express adapter for PUT /dashboard/model. Same auth posture as the GET.
  * The injected persist closure owns the config path (the server passes
  * init.ts persistConfigUpdates over stateDir/config.json); its load/persist
  * failures (malformed config, unwritable file) map to a 500 {error} with the
- * file left untouched — never a silent partial write.
+ * file left untouched — never a silent partial write. `hostedMode` (#514)
+ * threads to the pure handler's managed-field gate.
  */
 export declare function dashboardModelPutHandler(persist: (updates: Record<string, unknown>) => Record<string, unknown>, getLlmConfig: () => {
     provider: string;
-} | null): express.RequestHandler;
+} | null, hostedMode?: boolean): express.RequestHandler;
 /** The PUT's body: {machine?: string|null, harness: string, paused: boolean}. */
 export declare function handleDashboardCapturePausePut(body: unknown, setPause: (machine: string, harness: string, paused: boolean) => string | null): {
     status: number;

package/dist/dashboard.js

@@ -46,6 +46,7 @@ const config_read_js_1 = require("./config-read.js");
 const consolidate_js_1 = require("./consolidate.js");
 const capture_health_js_1 = require("./capture-health.js");
 const recall_precision_js_1 = require("./recall-precision.js");
+const distill_queue_js_1 = require("./distill-queue.js");
 const capture_pause_js_1 = require("./capture-pause.js");
 const state_js_1 = require("./state.js");
 const retrieval_js_1 = require("./retrieval.js");
@@ -585,6 +586,13 @@ function handleDashboardData(db, query, config) {
                 shown_sum: live.adoption?.shown_sum ?? 0,
                 used_sum: live.adoption?.used_sum ?? 0,
             }),
+            // #529: live inbox state — depth + oldest age + the echoed warn
+            // threshold. Live only (no series): depth at snapshot time is ~0 by
+            // construction, so history would be a flat line.
+            distill_queue: {
+                ...(0, distill_queue_js_1.readQueueStats)(db),
+                warn_age_hours: calibration_js_1.DRAIN_QUEUE_WARN_AGE_HOURS,
+            },
             // #423 phase 3: pauses + presence in one block — one source of truth
             // for the rail's dots, toggles and the capture card's PAUSED badges.
             fleet: {
@@ -1005,9 +1013,11 @@ function configString(v) {
 /**
  * The pure GET handler: echo the CONFIG values raw (null when unset — the UI
  * shows defaults as placeholders, so this never resolves them) + the runtime
- * provider from the daemon's in-memory llmConfig.
+ * provider from the daemon's in-memory llmConfig. `hostedMode` (#514) only
+ * adds `managed: true` to the body — self-hosted responses never carry the
+ * key at all (byte-identical wire).
  */
-function handleDashboardModelGet(config, llmConfig) {
+function handleDashboardModelGet(config, llmConfig, hostedMode = false) {
     const cfg = config ?? {};
     return {
         status: 200,
@@ -1022,6 +1032,7 @@ function handleDashboardModelGet(config, llmConfig) {
             enable_thinking: typeof cfg.enableThinking === "boolean" ? cfg.enableThinking : null,
             api_key_set: Boolean(configString(cfg.llmApiKey)),
             applies_on_restart: true,
+            ...(hostedMode ? { managed: true } : {}),
         },
     };
 }
@@ -1034,6 +1045,12 @@ const MODEL_PUT_KEYS = {
     max_tokens: "maxTokens",
     enable_thinking: "enableThinking",
 };
+/** #514 — hosted mode: these model-settings fields are controlled by the
+ *  hosting service (it owns the endpoint and the credentials), so the PUT
+ *  rejects them wholesale. Key PRESENCE in the body is the trigger — the
+ *  value is never inspected (null or nested garbage is a change attempt like
+ *  any other). Self-hosted installs never hit this gate. */
+const HOSTED_MANAGED_MODEL_KEYS = ["backend", "base_url", "api_key"];
 /** The backends init ever writes (absence of llmBackend + baseUrl+apiKey =
  *  the openai-compat path). "" clears (the modal's "auto" option). */
 const MODEL_BACKEND_VALUES = new Set(["", "ollama", "claude-cli"]);
@@ -1042,12 +1059,32 @@ const MODEL_BACKEND_VALUES = new Set(["", "ollama", "claude-cli"]);
  * injected writer (which THROWS on a malformed config — the adapter maps that
  * to a 500 and the file stays untouched), answer with the fresh GET shape
  * built from the post-write config. `null` REMOVES a config key.
+ *
+ * `hostedMode` (#514): the managed fields (see HOSTED_MANAGED_MODEL_KEYS) are
+ * rejected BEFORE any validation or persistence — a refused PUT writes
+ * nothing, not even the allowed keys that rode the same body.
  */
-function handleDashboardModelPut(body, persist, getLlmConfig) {
+function handleDashboardModelPut(body, persist, getLlmConfig, hostedMode = false) {
     if (body === null || typeof body !== "object" || Array.isArray(body)) {
         return { status: 400, body: { error: "Body must be a JSON object of model settings" } };
     }
     const input = body;
+    // #514 hosted-mode gate: the hosting service owns the endpoint and the
+    // credentials, so a tenant PUT may not touch them. Key PRESENCE is the
+    // trigger; the rejection precedes validation AND persistence, so nothing
+    // is written on a refused PUT (atomicity for mixed allowed+blocked bodies).
+    if (hostedMode) {
+        const attempted = HOSTED_MANAGED_MODEL_KEYS.filter((k) => k in input);
+        if (attempted.length > 0) {
+            const named = attempted.map((k) => `'${k}'`).join(", ");
+            return {
+                status: 403,
+                body: {
+                    error: `${named} ${attempted.length === 1 ? "is" : "are"} managed by the hosting service and cannot be changed here (managed fields: ${HOSTED_MANAGED_MODEL_KEYS.join(", ")})`,
+                },
+            };
+        }
+    }
     const updates = {};
     for (const [key, value] of Object.entries(input)) {
         const configKey = MODEL_PUT_KEYS[key];
@@ -1129,17 +1166,19 @@ function handleDashboardModelPut(body, persist, getLlmConfig) {
     // persist THROWS on a malformed config (strict load) — propagated to the
     // adapter → 500, file untouched. On success it returns the fresh config.
     const fresh = persist(updates);
-    return handleDashboardModelGet(fresh, getLlmConfig());
+    return handleDashboardModelGet(fresh, getLlmConfig(), hostedMode);
 }
 /**
  * Express adapter for GET /dashboard/model. Bearer-only by construction
  * (mounted after createAuthMiddleware — NO exemption, unlike the page shells:
- * this carries install config). Failures surface as a 500 {error}.
+ * this carries install config). Failures surface as a 500 {error}. The
+ * boot-resolved `hostedMode` flag (#514) threads straight through to the pure
+ * handler — it is a boot constant, not a per-request value.
  */
-function dashboardModelGetHandler(getConfig, getLlmConfig) {
+function dashboardModelGetHandler(getConfig, getLlmConfig, hostedMode = false) {
     return (_req, res) => {
         try {
-            const { status, body } = handleDashboardModelGet(getConfig(), getLlmConfig());
+            const { status, body } = handleDashboardModelGet(getConfig(), getLlmConfig(), hostedMode);
             res.status(status).json(body);
         }
         catch (err) {
@@ -1152,12 +1191,13 @@ function dashboardModelGetHandler(getConfig, getLlmConfig) {
  * The injected persist closure owns the config path (the server passes
  * init.ts persistConfigUpdates over stateDir/config.json); its load/persist
  * failures (malformed config, unwritable file) map to a 500 {error} with the
- * file left untouched — never a silent partial write.
+ * file left untouched — never a silent partial write. `hostedMode` (#514)
+ * threads to the pure handler's managed-field gate.
  */
-function dashboardModelPutHandler(persist, getLlmConfig) {
+function dashboardModelPutHandler(persist, getLlmConfig, hostedMode = false) {
     return (req, res) => {
         try {
-            const { status, body } = handleDashboardModelPut(req.body ?? null, persist, getLlmConfig);
+            const { status, body } = handleDashboardModelPut(req.body ?? null, persist, getLlmConfig, hostedMode);
             res.status(status).json(body);
         }
         catch (err) {

package/dist/db.js

@@ -763,6 +763,51 @@ const MIGRATIONS = [
       `);
         },
     },
+    {
+        version: 24,
+        name: "distill_queue",
+        up: (db) => {
+            // #529 — the durable distill inbox. POST /distill in queue mode stores
+            // the REDACTED segment here and answers 201 immediately (no LLM call);
+            // the nightly drain distills rows oldest-first and deletes each row in
+            // the same transaction as its memory inserts. One table in the shared
+            // SQLite DB — one store, same backup and tooling as the corpus (owner
+            // decision; not files). The row carries every /distill wire field needed
+            // to re-distill later (attribution, project, session_date, privacy) so
+            // the drain reproduces the sync path's insert exactly.
+            //
+            // UNIQUE(session_id, segment_id) is the delivery-time idempotency key: a
+            // queued-but-undistilled segment re-POST answers 200 skipped (the
+            // handler's prechecks consult this table), and a failed drain attempt
+            // never duplicates on retry. Legacy no-segment posts normalize
+            // segment_id ''. session_id stays NULL for posts that carry no session
+            // key — SQLite unique indexes treat NULLs as distinct, so those rows
+            // (which have no dedup key by construction, same as the sync path's
+            // sourceSession-undefined inserts) each enqueue separately.
+            // Idempotent: IF NOT EXISTS (the v18/v21 pattern — plain CREATE).
+            db.exec(`
+        CREATE TABLE IF NOT EXISTS distill_queue (
+          id INTEGER PRIMARY KEY AUTOINCREMENT,
+          session_id TEXT,
+          segment_id TEXT NOT NULL DEFAULT '',
+          source_agent TEXT,
+          source_agent_id TEXT,
+          source_domain TEXT,
+          source_machine TEXT,
+          project TEXT,
+          session_date TEXT,
+          privacy TEXT,
+          text TEXT NOT NULL,
+          arrived_at TEXT NOT NULL
+        )
+      `);
+            db.exec(`
+        CREATE UNIQUE INDEX IF NOT EXISTS idx_distill_queue_session_segment
+          ON distill_queue(session_id, segment_id)
+      `);
+            db.exec("CREATE INDEX IF NOT EXISTS idx_distill_queue_arrived ON distill_queue(arrived_at)");
+        },
+    },
 ];
 /**
  * Run all pending migrations against the database.

package/dist/dedup.js

@@ -670,8 +670,8 @@ async function runDedup(options = {}) {
             }
             for (const c of plan.mismatchSkipped) {
                 // #206 decision 2: project_mismatch is the only skip reason left on
-                // this rail (the source_agent rail was removed) — the bedrock dry run
-                // sizes the project rail alone off this line.
+                // this rail (the source_agent rail was removed) — the reference dry
+                // run sizes the project rail alone off this line.
                 console.log(`[hicortex]   SKIPPED (project_mismatch): ${c.memberIds.map((id) => id.slice(0, 8)).join(", ")}`);
             }
             // #393 guard-C: listed for review like the mismatch clusters — a