switchroom 0.15.40 → 0.15.42

package/dist/cli/ui/index.html

@@ -122,6 +122,15 @@
 
     .meta-item span { color: var(--text); }
 
+    /* Context-headroom gauge (RFC context-headroom-surface). */
+    .ctx-gauge { display: inline-flex; align-items: center; gap: 0.4rem; }
+    .ctx-text { color: var(--text); font-variant-numeric: tabular-nums; }
+    .ctx-bar {
+      display: inline-block; width: 48px; height: 6px; border-radius: 3px;
+      background: var(--bg-dim, rgba(255,255,255,0.12)); overflow: hidden;
+    }
+    .ctx-fill { display: block; height: 100%; border-radius: 3px; }
+
     .scope-toggle {
       cursor: pointer;
       color: var(--text);
@@ -1659,6 +1668,27 @@
       return 'inactive';
     }
 
+    // Context-headroom gauge (RFC context-headroom-surface): occupancy/cap
+    // with a fill bar, amber when tight (≥80%). '—' when no snapshot yet.
+    function fmtTok(n) {
+      return n >= 1000 ? (n / 1000).toFixed(n >= 10000 ? 0 : 1) + 'k' : String(n);
+    }
+    function formatContext(ctx) {
+      if (!ctx || ctx.state === 'unknown' || typeof ctx.occupancy !== 'number') {
+        return '<span style="color:var(--text-dim)">—</span>';
+      }
+      if (ctx.cap == null || ctx.pct == null) {
+        return `<span>${fmtTok(ctx.occupancy)} <span style="color:var(--text-dim)">(no cap)</span></span>`;
+      }
+      const pct = Math.max(0, Math.min(1, ctx.pct));
+      const tight = ctx.state === 'tight';
+      const col = tight ? 'var(--amber, #d99)' : 'var(--green)';
+      return `<span class="ctx-gauge" title="${fmtTok(ctx.occupancy)} / ${fmtTok(ctx.cap)} tokens — ${Math.round(pct * 100)}% of cap${tight ? ' (tight — /compact near)' : ''}">`
+        + `<span class="ctx-text">${fmtTok(ctx.occupancy)}/${fmtTok(ctx.cap)} (${Math.round(pct * 100)}%)</span>`
+        + `<span class="ctx-bar"><span class="ctx-fill" style="width:${(pct * 100).toFixed(0)}%;background:${col}"></span></span>`
+        + `</span>`;
+    }
+
     function formatUptime(timestamp) {
       if (!timestamp) return '--';
       const d = new Date(timestamp);
@@ -1693,6 +1723,7 @@
             <div class="meta-item"><label>Uptime </label><span>${formatUptime(a.uptime)}</span></div>
             <div class="meta-item"><label>Mem </label><span>${a.memory || '--'}</span></div>
             <div class="meta-item"><label>Last turn </label><span>${a.lastTurnAt ? formatTimestamp(a.lastTurnAt) : '—'}</span></div>
+            <div class="meta-item meta-context"><label>Context </label>${formatContext(a.context)}</div>
             <div class="meta-item"><label>Profile </label><span>${escapeHtml(a.extends)}</span></div>
             <div class="meta-item"><label>Auth </label><span>${a.auth.authenticated ? '✓' : '✗'}</span></div>
             <div class="meta-item"><label>Account </label><span>${a.primaryAccount ? escapeHtml(a.primaryAccount) : '<span style="color:var(--text-dim)">default</span>'}</span></div>

package/dist/host-control/main.js

@@ -13811,6 +13811,8 @@ var AgentMemorySchema = exports_external.object({
     max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
     cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
     min_overlap: exports_external.number().min(0).max(1).optional().describe("Minimum Jaccard token overlap [0.0–1.0] between the user " + "prompt and a memory's text for the memory to be injected. " + "Drops low-relevance matches before the count cap so weak hits " + "don't fill the slot on real queries. 0.0 disables (default — " + "current behaviour). Try 0.10–0.20 to start; observe the " + "`overlap_dropped` field via `switchroom memory recall-log`."),
+    types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
+    skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
     topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
   }).optional().describe("Auto-recall tuning knobs")
 }).optional();
@@ -13847,7 +13849,8 @@ var SubagentSchema = exports_external.object({
 var SessionSchema = exports_external.object({
   max_idle: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '2h', '30m')").optional().describe("Start a fresh session if the previous one has been idle " + "longer than this duration. Examples: '2h', '30m', '7200s'."),
   max_turns: exports_external.number().int().positive().optional().describe("Start a fresh session if the previous one has more user " + "turns than this. Useful for preventing context bloat on " + "long-running agents."),
-  max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context.")
+  max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context."),
+  idle_clear_after: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '3h', '90m')").optional().describe("Auto-run /clear (wipe the working context) after the live " + "session has been idle this long. Defaults to '3h' when unset " + "(on by default); set '0s' to disable. Long-term memory lives " + "in Hindsight, so a clear loses only the in-session thread.")
 }).optional();
 var SessionContinuitySchema = exports_external.object({
   enabled: exports_external.boolean().optional().describe("Master switch for the session-handoff briefing (default true)."),

package/dist/vault/approvals/kernel-server.js

@@ -11419,6 +11419,8 @@ var init_schema = __esm(() => {
       max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       min_overlap: exports_external.number().min(0).max(1).optional().describe("Minimum Jaccard token overlap [0.0–1.0] between the user " + "prompt and a memory's text for the memory to be injected. " + "Drops low-relevance matches before the count cap so weak hits " + "don't fill the slot on real queries. 0.0 disables (default — " + "current behaviour). Try 0.10–0.20 to start; observe the " + "`overlap_dropped` field via `switchroom memory recall-log`."),
+      types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
+      skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
       topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
     }).optional().describe("Auto-recall tuning knobs")
   }).optional();
@@ -11455,7 +11457,8 @@ var init_schema = __esm(() => {
   SessionSchema = exports_external.object({
     max_idle: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '2h', '30m')").optional().describe("Start a fresh session if the previous one has been idle " + "longer than this duration. Examples: '2h', '30m', '7200s'."),
     max_turns: exports_external.number().int().positive().optional().describe("Start a fresh session if the previous one has more user " + "turns than this. Useful for preventing context bloat on " + "long-running agents."),
-    max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context.")
+    max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context."),
+    idle_clear_after: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '3h', '90m')").optional().describe("Auto-run /clear (wipe the working context) after the live " + "session has been idle this long. Defaults to '3h' when unset " + "(on by default); set '0s' to disable. Long-term memory lives " + "in Hindsight, so a clear loses only the in-session thread.")
   }).optional();
   SessionContinuitySchema = exports_external.object({
     enabled: exports_external.boolean().optional().describe("Master switch for the session-handoff briefing (default true)."),

package/dist/vault/broker/server.js

@@ -11419,6 +11419,8 @@ var init_schema = __esm(() => {
       max_memories: exports_external.number().int().min(0).optional().describe("Cap on the number of memories injected into the prompt by " + "auto-recall, regardless of token budget. Plugin default is 12. " + "0 disables the cap (all memories Hindsight returns are injected)."),
       cache_ttl_secs: exports_external.number().int().min(0).optional().describe("Per-session recall cache TTL in seconds. When > 0, identical " + "(prompt, bank) within the same session reuse the cached recall " + "result instead of round-tripping to Hindsight. 0 disables. " + "Default is 600 (10 min) for switchroom-managed agents."),
       min_overlap: exports_external.number().min(0).max(1).optional().describe("Minimum Jaccard token overlap [0.0–1.0] between the user " + "prompt and a memory's text for the memory to be injected. " + "Drops low-relevance matches before the count cap so weak hits " + "don't fill the slot on real queries. 0.0 disables (default — " + "current behaviour). Try 0.10–0.20 to start; observe the " + "`overlap_dropped` field via `switchroom memory recall-log`."),
+      types: exports_external.array(exports_external.string()).optional().describe("Hindsight fact types to recall. Switchroom default is " + '["world", "experience", "observation"] — the synthesized ' + "`observation` tier is on by default. Set to " + '["world", "experience"] to opt out of observation-backed ' + "recall for this agent (or fleet-wide under defaults)."),
+      skip_trivial: exports_external.boolean().optional().describe("Skip recall on plausibly-stateless trivial turns (time/date/" + "greeting). Switchroom default true — saves the recall arm + " + "injected tokens on turns that never need memory, guarded so it " + "never skips a turn that references user/project/session state. " + "Set false to always run recall."),
       topic_filter_mode: exports_external.enum(["soft-preamble", "hard-filter"]).optional().describe("Supergroup-mode cross-topic memory behaviour. Default " + "(unset) → soft-preamble: recall returns memories from all " + "topics, and a 'Current topic: …' preamble tells the model " + "to self-scope. hard-filter: drop any recalled memory whose " + "metadata.thread_id differs from the active inbound's topic. " + "Flip to hard-filter when the recall_log shows binding " + "failures (model surfacing the right memory but applying " + "it to the wrong topic).")
     }).optional().describe("Auto-recall tuning knobs")
   }).optional();
@@ -11455,7 +11457,8 @@ var init_schema = __esm(() => {
   SessionSchema = exports_external.object({
     max_idle: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '2h', '30m')").optional().describe("Start a fresh session if the previous one has been idle " + "longer than this duration. Examples: '2h', '30m', '7200s'."),
     max_turns: exports_external.number().int().positive().optional().describe("Start a fresh session if the previous one has more user " + "turns than this. Useful for preventing context bloat on " + "long-running agents."),
-    max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context.")
+    max_context_tokens: exports_external.number().int().positive().optional().describe("Proactively run /compact when the live context window " + "occupancy (latest assistant turn input + cache-read + " + "cache-creation tokens) reaches this many tokens. Opt-in: " + "unset means rely on Claude Code's native auto-compaction. " + "Useful on large-window models (e.g. 1M Opus) to hold a " + "deliberately lean working context."),
+    idle_clear_after: exports_external.string().regex(/^\d+[smh]$/, "Duration must be a number followed by s, m, or h (e.g. '3h', '90m')").optional().describe("Auto-run /clear (wipe the working context) after the live " + "session has been idle this long. Defaults to '3h' when unset " + "(on by default); set '0s' to disable. Long-term memory lives " + "in Hindsight, so a clear loses only the in-session thread.")
   }).optional();
   SessionContinuitySchema = exports_external.object({
     enabled: exports_external.boolean().optional().describe("Master switch for the session-handoff briefing (default true)."),
@@ -16355,6 +16358,9 @@ class VaultBroker {
     this.passphrase = passphrase;
     this._setReadinessSentinel(true);
   }
+  reload(config) {
+    this.config = config;
+  }
   _setReadinessSentinel(ready) {
     const p = process.env.SWITCHROOM_VAULT_BROKER_READY_PATH;
     if (!p || p.length === 0)
@@ -16425,6 +16431,9 @@ class VaultBroker {
   _getSecretsRef() {
     return this.secrets;
   }
+  _getConfigRef() {
+    return this.config;
+  }
   bindAgentSocket(socketPath) {
     const abs = resolve6(socketPath);
     const agentName = socketPathToAgent(abs);
@@ -17856,6 +17865,19 @@ async function main() {
   }
   const broker = new VaultBroker;
   registerShutdownHandlers(broker);
+  process.on("SIGHUP", () => {
+    (async () => {
+      try {
+        const { loadConfig: loadConfig2 } = await Promise.resolve().then(() => (init_loader(), exports_loader));
+        broker.reload(loadConfig2(configPath));
+        process.stdout.write(`vault-broker: SIGHUP reload — config refreshed
+`);
+      } catch (err) {
+        process.stderr.write(`vault-broker: SIGHUP reload failed (keeping previous config): ${err.message}
+`);
+      }
+    })();
+  });
   if (perAgentTargets.length > 0) {
     await broker.start(legacySocketPath, configPath, vaultPath);
     process.stdout.write(`vault-broker: legacy socket listening on ${legacySocketPath}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "switchroom",
-  "version": "0.15.40",
+  "version": "0.15.42",
   "description": "Run Claude Code 24/7 on your Claude Pro/Max subscription over Telegram. Open-source alternative to OpenClaw and NanoClaw — no API keys.",
   "type": "module",
   "bin": {

package/profiles/_base/start.sh.hbs

@@ -332,6 +332,21 @@ export HINDSIGHT_RECALL_CACHE_TTL_SECS={{hindsightRecallCacheTtlSecs}}
 {{#if (isNumber hindsightRecallMinOverlap)}}
 export HINDSIGHT_RECALL_MIN_OVERLAP={{hindsightRecallMinOverlap}}
 {{/if}}
+# Recall fact types (memory.recall.types cascade). Switchroom default is
+# world,experience,observation (the synthesized `observation` tier is ON
+# by default, set in the plugin settings.json). Export only when the
+# operator OPTED OUT via switchroom.yaml — comma-separated; the env value
+# wins over settings.json. e.g. ["world","experience"] drops
+# observation-backed recall for this agent.
+{{#if hindsightRecallTypes}}
+export HINDSIGHT_RECALL_TYPES="{{hindsightRecallTypes}}"
+{{/if}}
+# Trivial-turn recall skip (memory.recall.skip_trivial cascade). On by
+# default (plugin settings.json). Export only when the operator overrode
+# it; set false to always run recall.
+{{#if hindsightRecallSkipTrivial}}
+export HINDSIGHT_RECALL_SKIP_TRIVIAL={{hindsightRecallSkipTrivial}}
+{{/if}}
 # PR6 — supergroup-mode topic tagging. JSON map of {alias: thread_id}
 # parsed by retain.py + recall.py to (a) stamp chat_id/thread_id/topic_alias
 # into retained memory metadata and (b) emit a "Current topic: …" preamble