switchroom 0.15.2 → 0.15.4

package/bin/turn-pacing-hook.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+# UserPromptSubmit hook that emits the per-turn pacing directive.
+#
+# Wired into the agent's .claude/settings.json hooks.UserPromptSubmit by
+# scaffold.ts when context_efficiency inject_on_change is enabled (the
+# default). Replaces the inline `printf` that previously re-injected the
+# static directive on every single turn, causing ~3 KB of redundant context
+# per turn (measured ~290 KB over a 26-prompt session).
+#
+# Environment variables (set in the hook command by scaffold.ts):
+#
+#   TURN_PACING_DIRECTIVE - The full directive text to inject. Scaffold.ts
+#                           embeds the directive in the hook command string.
+#   TURN_PACING_HASH      - sha256 of the directive text (hex). Used to
+#                           detect when the directive changed between
+#                           scaffold regenerations.
+#
+# Inject-on-change semantics: on every fire:
+#   1. Read session_id from stdin JSON.
+#   2. Check per-session state file under $TELEGRAM_STATE_DIR/.hook-state/.
+#   3. If session matches AND directive hash matches recorded state → emit
+#      nothing (the model already has the directive in context).
+#   4. Otherwise emit the directive and record the new state.
+#
+# Fail open: any error in state management falls back to emitting the full
+# directive so context is never silently dropped.
+#
+# State files: $TELEGRAM_STATE_DIR/.hook-state/turn-pacing.<session_id>
+# Each file contains: <directive_hash>\n
+# Stale state files (different session_id) are cleaned up on each new-
+# session write (keeps at most 5 files to avoid accumulation).
+
+set -u
+
+DIRECTIVE="${TURN_PACING_DIRECTIVE:-}"
+DIRECTIVE_HASH="${TURN_PACING_HASH:-}"
+
+# If directive or hash are missing, emit whatever we have and exit.
+if [ -z "$DIRECTIVE" ]; then
+  exit 0
+fi
+
+emit_full() {
+  printf '%s\n' "$DIRECTIVE"
+}
+
+# Read stdin JSON to extract session_id.
+SESSION_ID=""
+if ! [ -t 0 ]; then
+  STDIN_JSON=$(cat 2>/dev/null || true)
+  if command -v jq >/dev/null 2>&1; then
+    SESSION_ID=$(printf '%s' "$STDIN_JSON" | jq -r '.session_id // empty' 2>/dev/null || true)
+  else
+    SESSION_ID=$(printf '%s' "$STDIN_JSON" | grep -o '"session_id":"[^"]*"' | sed 's/"session_id":"//;s/"//' 2>/dev/null || true)
+  fi
+  # Sanitise: session_id must be alphanumeric + hyphens only.
+  if ! printf '%s' "${SESSION_ID:-}" | grep -qE '^[a-zA-Z0-9_-]{1,128}$' 2>/dev/null; then
+    SESSION_ID=""
+  fi
+fi
+
+# If we can't get a session_id, emit full (fail-open).
+if [ -z "$SESSION_ID" ]; then
+  emit_full
+  exit 0
+fi
+
+# If hash is missing, emit full (fail-open).
+if [ -z "$DIRECTIVE_HASH" ]; then
+  emit_full
+  exit 0
+fi
+
+# Locate state dir. Prefer $TELEGRAM_STATE_DIR (set by start.sh in the
+# container); fall back to $HOME/.claude/switchroom-hookcache for dev use.
+STATE_BASE="${TELEGRAM_STATE_DIR:-${HOME:-/tmp}/.claude/switchroom-hookcache}"
+HOOK_STATE_DIR="$STATE_BASE/.hook-state"
+
+# Attempt to create the state dir; if that fails emit full (fail-open).
+if ! mkdir -p "$HOOK_STATE_DIR" 2>/dev/null; then
+  emit_full
+  exit 0
+fi
+
+STATE_FILE="$HOOK_STATE_DIR/turn-pacing.$SESSION_ID"
+
+# Check if already emitted for this session with the same directive.
+if [ -f "$STATE_FILE" ]; then
+  RECORDED=$(head -1 "$STATE_FILE" 2>/dev/null || echo "")
+  if [ "$RECORDED" = "$DIRECTIVE_HASH" ]; then
+    # Already injected and content unchanged — suppress.
+    exit 0
+  fi
+fi
+
+# New session or directive changed — emit and record state.
+emit_full
+
+# Write state: hash on first line.
+printf '%s\n' "$DIRECTIVE_HASH" > "$STATE_FILE" 2>/dev/null || true
+
+# Clean up stale state files (different session_id) — keep at most 5 files.
+# shellcheck disable=SC2012
+OLD_COUNT=$(ls -1 "$HOOK_STATE_DIR"/turn-pacing.* 2>/dev/null | grep -cv "turn-pacing\.$SESSION_ID$" || true)
+if [ "$OLD_COUNT" -gt 5 ]; then
+  ls -1t "$HOOK_STATE_DIR"/turn-pacing.* 2>/dev/null \
+    | grep -v "turn-pacing\.$SESSION_ID$" \
+    | tail -n +6 \
+    | xargs rm -f 2>/dev/null || true
+fi
+
+exit 0

package/bin/workspace-dynamic-hook.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 # UserPromptSubmit hook for dynamic workspace bootstrap (MEMORY.md, daily
-# notes, HEARTBEAT.md).
+# notes).
 #
 # Wired into the agent's .claude/settings.json hooks.UserPromptSubmit by
 # scaffold.ts. On every inbound user prompt, this script re-renders the
@@ -9,17 +9,27 @@
 #
 # Configuration is via env vars (set at start.sh time):
 #
-#   SWITCHROOM_AGENT_NAME - The agent name (required, set in start.sh)
+#   SWITCHROOM_AGENT_NAME       - The agent name (required, set in start.sh)
+#   SWITCHROOM_INJECT_ON_CHANGE - When "1", enables inject-on-change
+#                                 semantics: content is only emitted when the
+#                                 session_id changes or file content changes.
+#                                 When "0" or unset (legacy mode), content is
+#                                 emitted every turn as before (full emit).
+#                                 Default in new agents is "1" (set by
+#                                 scaffold.ts when inject_on_change is true,
+#                                 the default).
 #
 # Failure modes (all silent — workspace injection must never block the turn):
 #   - switchroom CLI missing  → exit 0 with no output
 #   - workspace dir missing   → exit 0 with no output
 #   - workspace render fails  → exit 0 with no output
 #   - empty result set        → exit 0 with no output
+#   - state dir error         → emit full content (fail-open)
 
 set -u
 
 AGENT_NAME="${SWITCHROOM_AGENT_NAME:-}"
+INJECT_ON_CHANGE="${SWITCHROOM_INJECT_ON_CHANGE:-0}"
 
 if [ -z "$AGENT_NAME" ]; then
   exit 0
@@ -29,6 +39,26 @@ if ! command -v switchroom >/dev/null 2>&1; then
   exit 0
 fi
 
+# ---------------------------------------------------------------------------
+# Inject-on-change: read session_id from stdin JSON (when enabled)
+# ---------------------------------------------------------------------------
+SESSION_ID=""
+if [ "$INJECT_ON_CHANGE" = "1" ]; then
+  # Read stdin (non-blocking: if no stdin supplied in tests, just skip).
+  if ! [ -t 0 ]; then
+    STDIN_JSON=$(cat 2>/dev/null || true)
+    if command -v jq >/dev/null 2>&1; then
+      SESSION_ID=$(printf '%s' "$STDIN_JSON" | jq -r '.session_id // empty' 2>/dev/null || true)
+    else
+      SESSION_ID=$(printf '%s' "$STDIN_JSON" | grep -o '"session_id":"[^"]*"' | sed 's/"session_id":"//;s/"//' 2>/dev/null || true)
+    fi
+    # Sanitise session_id — must be alphanumeric + hyphens only.
+    if ! printf '%s' "${SESSION_ID:-}" | grep -qE '^[a-zA-Z0-9_-]{1,128}$' 2>/dev/null; then
+      SESSION_ID=""
+    fi
+  fi
+fi
+
 # Cache directory shared with the post-render dedupe sidecar below. We
 # need it earlier than the original code so the mtime-based fast-skip
 # can read its body file before invoking the (~800ms) switchroom CLI.
@@ -46,15 +76,15 @@ BODY_FILE="$CACHE_DIR/workspace-dynamic.${CACHE_DATE}.body"
 # Mtime-fast-skip: if BODY_FILE exists AND is newer than every workspace
 # source the renderer reads, we can emit the cached body and skip the
 # ~800ms `switchroom workspace render` invocation entirely. The renderer
-# reads MEMORY.md, HEARTBEAT.md, today's daily, yesterday's daily — see
+# reads MEMORY.md, today's daily, yesterday's daily — see
 # `loadDynamicBootstrapFiles` in src/agents/workspace.ts.
 #
 # Skip semantics: a source file that doesn't exist contributes a "very
 # old" mtime (epoch-0 via stat fallback), which never invalidates the
 # cache. A source file that's been updated since BODY_FILE's mtime
 # triggers a fresh render. Forensics measured this fast-path saving
-# ~825ms on the common case (chat turns where MEMORY/HEARTBEAT haven't
-# changed since the last turn).
+# ~825ms on the common case (chat turns where MEMORY hasn't changed
+# since the last turn).
 #
 # Resolve the agent's workspace dir. Switchroom uses
 # `~/.switchroom/agents/<name>/workspace/` by default. We avoid invoking
@@ -75,7 +105,7 @@ if [ -f "$BODY_FILE" ]; then
   # missing), emit the cache and exit.
   body_mtime=$(stat -c '%Y' "$BODY_FILE" 2>/dev/null || echo 0)
   newest_src_mtime=0
-  for src in "$WS_DIR/MEMORY.md" "$WS_DIR/HEARTBEAT.md" "$TODAY_FILE" "$YESTERDAY_FILE"; do
+  for src in "$WS_DIR/MEMORY.md" "$TODAY_FILE" "$YESTERDAY_FILE"; do
     if [ -f "$src" ]; then
       src_mtime=$(stat -c '%Y' "$src" 2>/dev/null || echo 0)
       if [ "$src_mtime" -gt "$newest_src_mtime" ]; then
@@ -84,18 +114,38 @@ if [ -f "$BODY_FILE" ]; then
     fi
   done
   if [ "$body_mtime" -gt "$newest_src_mtime" ]; then
-    # Fast path: every source is older than the cached body. Emit and
-    # skip the renderer entirely. Saves ~800ms cold-start of the
-    # switchroom CLI plus the actual render work.
+    # Fast path: every source is older than the cached body.
+    # In inject-on-change mode, also check the session-state file — if the
+    # session_id matches the last-emitted session AND the hash matches, we
+    # can suppress entirely (model already has this content in context).
+    if [ "$INJECT_ON_CHANGE" = "1" ] && [ -n "$SESSION_ID" ]; then
+      SESSION_STATE_DIR="${TELEGRAM_STATE_DIR:-${HOME:-/tmp}/.claude/switchroom-hookcache}/.hook-state"
+      SESSION_STATE_FILE="$SESSION_STATE_DIR/ws-dynamic.$SESSION_ID"
+      if [ -f "$SESSION_STATE_FILE" ]; then
+        RECORDED_HASH=$(head -1 "$SESSION_STATE_FILE" 2>/dev/null || echo "")
+        # Read hash from body cache to compare with session state.
+        CACHED_HASH=$(head -1 "$CACHE_FILE" 2>/dev/null || echo "")
+        if [ -n "$CACHED_HASH" ] && [ "$CACHED_HASH" = "$RECORDED_HASH" ]; then
+          # Same session, same content — suppress injection.
+          exit 0
+        fi
+      fi
+      # Different session or hash changed — emit, then update session state.
+      mkdir -p "$SESSION_STATE_DIR" 2>/dev/null || true
+      CACHED_HASH_FOR_STATE=$(head -1 "$CACHE_FILE" 2>/dev/null || echo "")
+      if [ -n "$CACHED_HASH_FOR_STATE" ]; then
+        printf '%s\n%s\n' "$CACHED_HASH_FOR_STATE" "$SESSION_ID" > "$SESSION_STATE_FILE" 2>/dev/null || true
+      fi
+    fi
     cat "$BODY_FILE"
     exit 0
   fi
 fi
 
-# Render the dynamic workspace files (MEMORY.md, today/yesterday daily,
-# HEARTBEAT.md). The render command exits 0 and returns empty string if the
-# workspace doesn't exist or all dynamic files are missing/empty, so no
-# special-casing needed here.
+# Render the dynamic workspace files (MEMORY.md, today/yesterday daily).
+# The render command exits 0 and returns empty string if the workspace
+# doesn't exist or all dynamic files are missing/empty, so no special-casing
+# needed here.
 #
 # --warning-mode off: truncation warnings go to the stable render (where they
 # can surface during scaffold/reconcile), not the per-turn path where they'd
@@ -107,8 +157,8 @@ fi
 WS_DYNAMIC=$(timeout 3 switchroom workspace render "$AGENT_NAME" --dynamic --warning-mode off 2>/dev/null || true)
 
 # Empty render → emit nothing AND do NOT cache the empty body. Caching an
-# empty body would re-emit empty forever even after MEMORY/HEARTBEAT come
-# back online, defeating the whole purpose of the hook.
+# empty body would re-emit empty forever even after MEMORY comes back online,
+# defeating the whole purpose of the hook.
 if [ -z "$WS_DYNAMIC" ]; then
   exit 0
 fi
@@ -126,7 +176,44 @@ if [ -f "$CACHE_FILE" ]; then
   OLD_HASH=$(head -1 "$CACHE_FILE" 2>/dev/null || echo "")
 fi
 
+# Session-state helper (shared by both the hash-match and hash-changed paths).
+_ws_record_session_state() {
+  local hash="$1" sid="$2"
+  if [ -n "$hash" ] && [ -n "$sid" ]; then
+    local sdir="${TELEGRAM_STATE_DIR:-${HOME:-/tmp}/.claude/switchroom-hookcache}/.hook-state"
+    [ -d "$sdir" ] || mkdir -p "$sdir" 2>/dev/null || true
+    if [ -d "$sdir" ]; then
+      printf '%s\n%s\n' "$hash" "$sid" > "$sdir/ws-dynamic.$sid" 2>/dev/null || true
+      # Keep at most 5 ws-dynamic state files to bound growth.
+      # shellcheck disable=SC2012
+      local old_count
+      old_count=$(ls -1 "$sdir"/ws-dynamic.* 2>/dev/null | grep -cv "ws-dynamic\.$sid$" || true)
+      if [ "$old_count" -gt 5 ]; then
+        ls -1t "$sdir"/ws-dynamic.* 2>/dev/null \
+          | grep -v "ws-dynamic\.$sid$" \
+          | tail -n +6 \
+          | xargs rm -f 2>/dev/null || true
+      fi
+    fi
+  fi
+}
+
 if [ -n "$NEW_HASH" ] && [ "$NEW_HASH" = "$OLD_HASH" ] && [ -f "$BODY_FILE" ]; then
+  # Content unchanged since last render. In inject-on-change mode, check if
+  # we already injected this content in the current session — if so, suppress.
+  if [ "$INJECT_ON_CHANGE" = "1" ] && [ -n "$SESSION_ID" ]; then
+    SESSION_STATE_DIR="${TELEGRAM_STATE_DIR:-${HOME:-/tmp}/.claude/switchroom-hookcache}/.hook-state"
+    SESSION_STATE_FILE="$SESSION_STATE_DIR/ws-dynamic.$SESSION_ID"
+    if [ -f "$SESSION_STATE_FILE" ]; then
+      RECORDED_HASH=$(head -1 "$SESSION_STATE_FILE" 2>/dev/null || echo "")
+      if [ "$RECORDED_HASH" = "$NEW_HASH" ]; then
+        # Same session, same content — suppress injection.
+        exit 0
+      fi
+    fi
+    # New session or session state mismatch — emit and record.
+    _ws_record_session_state "$NEW_HASH" "$SESSION_ID"
+  fi
   cat "$BODY_FILE"
 else
   # Refresh sidecar: write hash + body, then echo body. Touch the body
@@ -135,6 +222,9 @@ else
   if [ -n "$NEW_HASH" ]; then
     printf '%s\n' "$NEW_HASH" > "$CACHE_FILE" 2>/dev/null || true
     printf '%s\n' "$WS_DYNAMIC" > "$BODY_FILE" 2>/dev/null || true
+    if [ "$INJECT_ON_CHANGE" = "1" ] && [ -n "$SESSION_ID" ]; then
+      _ws_record_session_state "$NEW_HASH" "$SESSION_ID"
+    fi
   fi
   printf '%s\n' "$WS_DYNAMIC"
 fi

package/bin/workspace-stable-hook.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 # UserPromptSubmit hook for stable workspace content (AGENTS.md, SOUL.md,
-# USER.md, IDENTITY.md, TOOLS.md, HEARTBEAT.md).
+# USER.md, IDENTITY.md, TOOLS.md).
 #
 # Used only when channels.telegram.hotReloadStable is true. In that mode,
 # the stable workspace render moves from a session-start bake into
@@ -34,7 +34,7 @@ if ! command -v switchroom >/dev/null 2>&1; then
 fi
 
 # Render the stable workspace files (AGENTS.md, SOUL.md, IDENTITY.md, USER.md,
-# TOOLS.md, HEARTBEAT.md). The render command exits 0 and returns empty string
+# TOOLS.md). The render command exits 0 and returns empty string
 # if the workspace doesn't exist or all stable files are missing/empty, so no
 # special-casing needed here.
 #

package/dist/agent-scheduler/index.js

@@ -11090,7 +11090,8 @@ var TelegramChannelSchema = exports_external.object({
   stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI — " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events — stable " + "order, per-tool status emojis, fires only on semantic transitions."),
   stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
   clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message — Reading X, Searching the web for Y, …) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) — no post-then-delete. Per-agent " + "override; cascades defaults → profile → agent (per-key)."),
-  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes — suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
   orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
   cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
   deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),

package/dist/auth-broker/index.js

@@ -11090,7 +11090,8 @@ var TelegramChannelSchema = exports_external.object({
   stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI — " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events — stable " + "order, per-tool status emojis, fires only on semantic transitions."),
   stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
   clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message — Reading X, Searching the web for Y, …) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) — no post-then-delete. Per-agent " + "override; cascades defaults → profile → agent (per-key)."),
-  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes — suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
   orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
   cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
   deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),
@@ -12363,6 +12364,47 @@ function resolveConsumerProbeIntervalMs(env) {
   return DEFAULT_CONSUMER_PROBE_INTERVAL_MS;
 }
 
+// src/auth/broker/account-eligibility.ts
+var WALL_PCT = 99.5;
+var HEALTHY_CLEAR_PCT = 80;
+var SNAPSHOT_STALE_AGE_MS = 24 * 60 * 60 * 1000;
+function snapshotFresh(s, now, maxAgeMs = SNAPSHOT_STALE_AGE_MS) {
+  return !!s && now - s.capturedAt <= maxAgeMs && s.capturedAt <= now + 60000;
+}
+function snapshotWalled(s) {
+  return s.fiveHourUtilizationPct >= WALL_PCT || s.sevenDayUtilizationPct >= WALL_PCT;
+}
+function snapshotClearlyHealthy(s) {
+  return s.fiveHourUtilizationPct < HEALTHY_CLEAR_PCT && s.sevenDayUtilizationPct < HEALTHY_CLEAR_PCT;
+}
+function isAccountBlocked(opts) {
+  const { mark, snapshot, now } = opts;
+  if (snapshotFresh(snapshot, now)) {
+    const markedAt = mark?.marked_at ?? 0;
+    if (snapshot.capturedAt >= markedAt) {
+      return snapshotWalled(snapshot);
+    }
+  }
+  return mark !== undefined && mark.exhausted_until > now;
+}
+function snapshotShouldClearMark(snapshot, mark, now) {
+  if (!mark)
+    return false;
+  if (!snapshotFresh(snapshot, now))
+    return false;
+  if (snapshot.capturedAt < (mark.marked_at ?? 0))
+    return false;
+  return snapshotClearlyHealthy(snapshot);
+}
+function clampMarkExpiry(opts) {
+  const { proposedUntil, now, shortMs, snapshot } = opts;
+  const shortCeil = now + shortMs;
+  if (proposedUntil <= shortCeil)
+    return proposedUntil;
+  const liveContradictsWeeklyWall = snapshotFresh(snapshot, now) && snapshot.sevenDayUtilizationPct < WALL_PCT;
+  return liveContradictsWeeklyWall ? shortCeil : proposedUntil;
+}
+
 // src/util/atomic.ts
 import { randomBytes } from "node:crypto";
 import { closeSync, constants, fsyncSync, openSync, renameSync, rmSync, writeSync } from "node:fs";
@@ -14030,8 +14072,11 @@ class AuthBroker {
     return this.accountWithFailover(account);
   }
   isAccountExhausted(account) {
-    const q = this.quota[account];
-    return q !== undefined && q.exhausted_until > this.now();
+    return isAccountBlocked({
+      mark: this.quota[account],
+      snapshot: this.lastQuotaCache[account],
+      now: this.now()
+    });
   }
   accountWithFailover(account) {
     if (!account || !this.isAccountExhausted(account))
@@ -14067,7 +14112,7 @@ class AuthBroker {
       const creds = readAccountCredentials(label, this.home);
       const meta = readAccountMeta(label, this.home);
       const q = this.quota[label];
-      const exhausted = q !== undefined && q.exhausted_until > this.now();
+      const exhausted = this.isAccountExhausted(label);
       const lq = this.lastQuotaCache[label];
       return {
         label,
@@ -14142,7 +14187,7 @@ class AuthBroker {
   cacheQuotaSnapshot(label, result) {
     if (!result.ok)
       return;
-    this.lastQuotaCache[label] = {
+    const snapshot = {
       fiveHourUtilizationPct: result.data.fiveHourUtilizationPct,
       sevenDayUtilizationPct: result.data.sevenDayUtilizationPct,
       fiveHourResetAt: result.data.fiveHourResetAt?.toISOString() ?? null,
@@ -14152,6 +14197,13 @@ class AuthBroker {
       overageDisabledReason: result.data.overageDisabledReason,
       capturedAt: this.now()
     };
+    this.lastQuotaCache[label] = snapshot;
+    if (snapshotShouldClearMark(snapshot, this.quota[label], this.now())) {
+      delete this.quota[label];
+      this.persistQuota();
+      process.stdout.write(`auth-broker: live probe shows ${label} healthy (5h=${snapshot.fiveHourUtilizationPct}% 7d=${snapshot.sevenDayUtilizationPct}%) — cleared stale exhaustion mark
+`);
+    }
   }
   async fleetQuotaProbeTick() {
     for (const label of listAccounts(this.home)) {
@@ -14183,14 +14235,21 @@ class AuthBroker {
         this.logErr(`consumer-quota-probe ${label}: ${err.message}`);
         continue;
       }
+      this.cacheQuotaSnapshot(label, result);
       const decision = quotaIndicatesExhaustion(result);
       if (!decision.exhausted)
         continue;
-      const exhaustedUntil = decision.until ?? this.now() + MARK_EXHAUSTED_DEFAULT_MS;
+      const now = this.now();
+      const exhaustedUntil = clampMarkExpiry({
+        proposedUntil: decision.until ?? now + MARK_EXHAUSTED_DEFAULT_MS,
+        now,
+        shortMs: MARK_EXHAUSTED_DEFAULT_MS,
+        snapshot: this.lastQuotaCache[label]
+      });
       const existing = this.quota[label]?.exhausted_until;
       if (existing !== undefined && existing >= exhaustedUntil)
         continue;
-      this.quota[label] = { exhausted_until: exhaustedUntil };
+      this.quota[label] = { exhausted_until: exhaustedUntil, marked_at: now };
       this.persistQuota();
       this.audit({ op: "mark-exhausted", identity: { kind: "operator" }, account: label, ok: true });
       process.stdout.write(`auth-broker: consumer-quota-sensor marked ${label} exhausted until ${new Date(exhaustedUntil).toISOString()} — consumer(s) fail over
@@ -14224,8 +14283,14 @@ class AuthBroker {
       socket.write(encodeError(id, "ACCOUNT_NOT_FOUND", "no active account configured"));
       return;
     }
-    const exhaustedUntil = until ?? this.now() + MARK_EXHAUSTED_DEFAULT_MS;
-    this.quota[account] = { exhausted_until: exhaustedUntil };
+    const now = this.now();
+    const exhaustedUntil = clampMarkExpiry({
+      proposedUntil: until ?? now + MARK_EXHAUSTED_DEFAULT_MS,
+      now,
+      shortMs: MARK_EXHAUSTED_DEFAULT_MS,
+      snapshot: this.lastQuotaCache[account]
+    });
+    this.quota[account] = { exhausted_until: exhaustedUntil, marked_at: now };
     this.persistQuota();
     const rolled = this.fanoutFailoverFor(account);
     const rolledTo = this.nextHealthyAccount(account, this.config.auth?.fallback_order ?? []);
@@ -14752,9 +14817,7 @@ class AuthBroker {
       const cand = order[(start + i) % order.length];
       if (!cand)
         continue;
-      const q = this.quota[cand];
-      const exhausted = q !== undefined && q.exhausted_until > this.now();
-      if (!exhausted && accountExists(cand, this.home))
+      if (!this.isAccountExhausted(cand) && accountExists(cand, this.home))
         return cand;
     }
     return null;

package/dist/cli/notion-write-pretool.mjs

@@ -11838,7 +11838,8 @@ var TelegramChannelSchema = exports_external.object({
   stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI \u2014 " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events \u2014 stable " + "order, per-tool status emojis, fires only on semantic transitions."),
   stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
   clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message \u2014 Reading X, Searching the web for Y, \u2026) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) \u2014 no post-then-delete. Per-agent " + "override; cascades defaults \u2192 profile \u2192 agent (per-key)."),
-  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes \u2014 suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
   orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
   cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
   deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),