switchroom 0.18.6 → 0.18.7

package/profiles/_base/cron-session.sh.hbs

@@ -87,36 +87,75 @@ CRON_APPEND_PROMPT="You are the cheap background cron worker for {{name}}. You r
 # claude's project key matches the pre-seeded trust state (.claude-cron).
 cd "{{agentDir}}" || exit 1
 
-# LiteLLM routing for the cron session — mirrors start.sh's boot block.
+# LiteLLM routing for the cron session — ports start.sh's boot contract.
 # IMPORTANT: the vault key is provisioned under the BASE agent name ({{name}}),
 # NOT the cron identity ({{name}}-cron). Use the Handlebars template var so the
-# lookup is a compile-time literal — no fragile runtime suffix-stripping.
-# Attribution headers carry $SWITCHROOM_AGENT_NAME (= {{name}}-cron) so LiteLLM
-# can distinguish cron spend from main-session spend per agent.
-# FAIL-OPEN: missing key OR unreachable proxy → strip routing env, fall back to
-# direct OAuth. An outage must never take the cron session dark.
+# lookup is a compile-time literal — no fragile runtime suffix-stripping. The
+# attribution headers use the SAME compile-time-literal form ({{name}}-cron,
+# identical to $SWITCHROOM_AGENT_NAME here — line 69 exports it) so this block
+# refers to the agent one consistent way instead of mixing a template literal
+# for the key with a runtime env var for the tags. Cron spend is attributed to
+# {{name}}-cron so LiteLLM distinguishes it from main-session spend per agent.
+#
+# Two boot failure modes, handled DIFFERENTLY (ported 2026-07 from start.sh so
+# cron no longer fails OPEN on a transient blip and silently runs UNTRACKED
+# direct OAuth for the fire — the product invariant is all model traffic stays
+# proxy-routed for cost/token tracking):
+#   - MISSING KEY → fail-open: strip routing, fall back to direct OAuth, LOUD
+#     log. Without a key there is no metered route to reach — correct here.
+#   - PROXY UNREACHABLE with key in hand → do NOT strip. Keep routing pointed at
+#     the proxy AND export the auth header so the session authenticates the
+#     moment the proxy recovers (the socat forwarder reconnects per-connection),
+#     LOUD log. Never a silent untracked direct-OAuth fallback.
+# Retry budget is DELIBERATELY short — 3 attempts, worst-case ~25s (3×5s curl
+# timeouts + 2×5s sleeps) — vs start.sh's 120s: a cron fire is
+# latency-sensitive and short-lived, and unreachable now KEEPS routing anyway,
+# so a long boot-probe would only add dead latency for no gain.
 if [ -n "${SWITCHROOM_LITELLM:-}" ] && command -v switchroom >/dev/null 2>&1; then
   sr_ll_key="$(switchroom vault get "litellm/{{name}}/api-key" 2>/dev/null || true)"
   sr_ll_ok=""
+  sr_ll_unreachable=""
   if [ -z "$sr_ll_key" ]; then
-    echo "litellm: no virtual key for cron '{{name}}' — falling back to direct OAuth (no tracking/guardrail)" >&2
-  elif command -v curl >/dev/null 2>&1 && [ -n "$ANTHROPIC_BASE_URL" ] \
-       && ! curl -fsS -m 5 -o /dev/null "${SWITCHROOM_LITELLM_BASE:-${ANTHROPIC_BASE_URL%/anthropic}}/health/liveliness" 2>/dev/null; then
-    echo "litellm: proxy unreachable at ${SWITCHROOM_LITELLM_BASE:-$ANTHROPIC_BASE_URL} — falling back to direct OAuth (no tracking/guardrail this session)" >&2
+    echo "litellm(cron): no virtual key for '{{name}}' — falling back to direct OAuth (untracked, unguarded)" >&2
+  elif command -v curl >/dev/null 2>&1 && [ -n "$ANTHROPIC_BASE_URL" ]; then
+    # Bounded retry probe — short budget for cron latency (3 attempts,
+    # worst-case ~25s: 3×5s curl timeouts + 2×5s sleeps).
+    sr_ll_url="${SWITCHROOM_LITELLM_BASE:-${ANTHROPIC_BASE_URL%/anthropic}}/health/liveliness"
+    sr_ll_try=0
+    sr_ll_up=""
+    while :; do
+      if curl -fsS -m 5 -o /dev/null "$sr_ll_url" 2>/dev/null; then sr_ll_up="1"; break; fi
+      sr_ll_try=$(( sr_ll_try + 1 ))
+      [ "$sr_ll_try" -ge 3 ] && break
+      sleep 5
+    done
+    if [ -z "$sr_ll_up" ]; then
+      # Proxy unreachable but the key is present. Do NOT strip: keep routing so
+      # the cron session self-heals once the proxy is back.
+      sr_ll_unreachable="1"
+      echo "litellm(cron): proxy unreachable at ${SWITCHROOM_LITELLM_BASE:-$ANTHROPIC_BASE_URL} after 3 retries — routing LEFT IN PLACE so it self-heals; NOT falling back to untracked direct Anthropic OAuth." >&2
+    else
+      sr_ll_ok="1"
+    fi
+    unset sr_ll_url sr_ll_try sr_ll_up
   else
     sr_ll_ok="1"
   fi
-  if [ -n "$sr_ll_ok" ]; then
+  if [ -n "$sr_ll_ok" ] || [ -n "$sr_ll_unreachable" ]; then
+    # Key fetched successfully — export the auth header REGARDLESS of the probe
+    # outcome so the virtual key reaches the process env. On unreachable this is
+    # what lets the session authenticate the moment the proxy recovers instead
+    # of 401-ing (the same self-heal fix as start.sh's inner/outer blocks).
     export ANTHROPIC_CUSTOM_HEADERS="x-litellm-api-key: Bearer $sr_ll_key
-x-litellm-customer-id: $SWITCHROOM_AGENT_NAME
-x-litellm-tags: agent:$SWITCHROOM_AGENT_NAME,profile:${SWITCHROOM_AGENT_PROFILE:-default}"
+x-litellm-customer-id: {{name}}-cron
+x-litellm-tags: agent:{{name}}-cron,profile:${SWITCHROOM_AGENT_PROFILE:-default}"
     export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
   else
-    # Fail-open: drop every routing var so the claude CLI talks to Anthropic
-    # directly on its OAuth credential (subscription path), unproxied.
+    # Missing-key fail-open: drop every routing var so the claude CLI talks to
+    # Anthropic directly on its OAuth credential (subscription path), unproxied.
     unset ANTHROPIC_BASE_URL ANTHROPIC_SMALL_FAST_MODEL SWITCHROOM_LITELLM SWITCHROOM_LITELLM_BASE
   fi
-  unset sr_ll_key sr_ll_ok
+  unset sr_ll_key sr_ll_ok sr_ll_unreachable
 fi
 
 # Create the cron tmux session DETACHED (-d), not in attach mode. This is a

package/profiles/_base/start.sh.hbs

@@ -46,6 +46,17 @@ if [ "$SWITCHROOM_RUNTIME" = "docker" ] && [ -z "$SWITCHROOM_DOCKER_TMUX_INNER"
   # compose env and the inner-pass export. Pinned by the scaffold-order test.
   export SWITCHROOM_AGENT_NAME="{{name}}"
 
+  # Boot-resume policy (session_continuity.boot_resume; default 'in-flight').
+  # The gateway's boot-resume block reads SWITCHROOM_BOOT_RESUME to decide
+  # whether a genuinely in-flight turn is auto-resumed across a deliberate
+  # restart ('in-flight' — default), always ('always'), or downgraded to a
+  # passive notice ('never'). MUST be exported here, before the gateway fork
+  # below, or the daemon never sees it and every agent defaults silently — same
+  # start.sh env-fork landmine as the channels.telegram knobs. The later
+  # SWITCHROOM_RESUME_MODE block is set AFTER the fork (it drives the inner
+  # `claude` --continue path, not the gateway), so this needs its own hoist.
+  export SWITCHROOM_BOOT_RESUME="{{#if bootResumeMode}}{{{bootResumeMode}}}{{else}}in-flight{{/if}}"
+
   # Gateway-consumed env MUST be exported HERE, before the gateway fork
   # below. The gateway daemon reads channels.telegram.* knobs (and any
   # agent env) from process.env at startup — e.g. SWITCHROOM_TG_STREAM_
@@ -120,17 +131,19 @@ if [ "$SWITCHROOM_RUNTIME" = "docker" ] && [ -z "$SWITCHROOM_DOCKER_TMUX_INNER"
     else
       sr_ll_ok="1"
     fi
-    if [ -n "$sr_ll_ok" ]; then
+    if [ -n "$sr_ll_ok" ] || [ -n "$sr_ll_unreachable" ]; then
+      # Key fetched successfully — export the proxy auth header REGARDLESS of the
+      # reachability probe outcome. The header is what carries the virtual key
+      # into the process env; without it the gateway's sr-* discovery
+      # (gateway.ts discoverSrModels) is dead and, once the proxy recovers,
+      # every request 401s until a manual restart. Since routing is LEFT IN
+      # PLACE on unreachable (below/above), the header must be too so the
+      # session actually self-heals when litellm returns. The probe result gates
+      # only log wording + inner-pass _LITELLM_OK, never the header.
       export ANTHROPIC_CUSTOM_HEADERS="x-litellm-api-key: Bearer $sr_ll_key
 x-litellm-customer-id: $SWITCHROOM_AGENT_NAME
 x-litellm-tags: agent:$SWITCHROOM_AGENT_NAME,profile:${SWITCHROOM_AGENT_PROFILE:-default}"
       export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
-    elif [ -n "$sr_ll_unreachable" ]; then
-      # Proxy unreachable at boot: routing intentionally LEFT IN PLACE (loud
-      # warning already emitted) so it survives into the inner pass and
-      # self-heals once litellm returns. NO unset — the permanent strip here
-      # was what made the inner self-heal fix inert on docker.
-      :
     else
       # Missing-key fail-open: gateway talks direct OAuth, unproxied.
       unset ANTHROPIC_BASE_URL ANTHROPIC_SMALL_FAST_MODEL SWITCHROOM_LITELLM SWITCHROOM_LITELLM_BASE
@@ -1016,8 +1029,15 @@ if [ -n "${SWITCHROOM_LITELLM:-}" ] && command -v switchroom >/dev/null 2>&1; th
   else
     sr_ll_ok="1"
   fi
-  if [ -n "$sr_ll_ok" ]; then
-    _LITELLM_OK="1"
+  if [ -n "$sr_ll_ok" ] || [ -n "$sr_ll_unreachable" ]; then
+    # Key fetched successfully — export the proxy auth header REGARDLESS of the
+    # reachability probe outcome. Routing is LEFT IN PLACE on unreachable so the
+    # session self-heals when litellm returns; the header MUST survive with it,
+    # otherwise every request 401s once the proxy recovers (the "self-heals"
+    # promise below was inert without this) and gateway sr-* discovery stays
+    # dead. The probe outcome gates ONLY the _LITELLM_OK flag (sr-* override
+    # drop) and the log wording above — never the header.
+    #
     # Newline-separated Name: Value pairs (claude CLI ANTHROPIC_CUSTOM_HEADERS format).
     # Tags: agent:<name> for per-agent spend tracking; profile:<profile> for
     # fleet-level cost breakdown by role. Per-turn tags (cron vs telegram) are
@@ -1031,13 +1051,12 @@ x-litellm-tags: agent:$SWITCHROOM_AGENT_NAME,profile:${SWITCHROOM_AGENT_PROFILE:
     # models configured in the proxy appear in the /model picker and can be
     # selected via --model. Without this, the CLI only knows its bundled list.
     export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
-  elif [ -n "$sr_ll_unreachable" ]; then
-    # Proxy unreachable at boot: routing intentionally LEFT IN PLACE (loud
-    # warning already emitted above) so it self-heals when litellm returns.
-    # _LITELLM_OK stays empty so any sr-* session model override is still
-    # dropped below (claude would 4xx an unknown model against a down endpoint).
-    # NO unset here — that permanent strip was the bug this branch fixes.
-    :
+    if [ -n "$sr_ll_ok" ]; then
+      # Probe succeeded: proxy is live, so honor any sr-* session model override
+      # below. On unreachable, _LITELLM_OK stays empty so the override is dropped
+      # (claude would 4xx an unknown model against a down endpoint).
+      _LITELLM_OK="1"
+    fi
   else
     # Missing key fail-open: drop every routing var so the claude CLI talks to
     # Anthropic directly on its OAuth credential (subscription path), unproxied.

package/profiles/default/CLAUDE.md.hbs

@@ -105,7 +105,7 @@ By default, every restart starts a **fresh `claude` session** — the in-flight
   - **`resume_interrupted`** (operator restart / SIGTERM / crash): pick the work back up and carry it to completion. Briefly tell the user you're resuming and roughly how long ago it was interrupted — then just do it. Do NOT ask whether to resume.
   - **`resume_watchdog_timeout`** (hang-watchdog killed it after no progress): do NOT silently resume — it may hang the same way. Tell the user plainly that your last turn was killed after N minutes of no progress, roughly what it was doing, and ask whether to retry or take a different angle. Report only the honest cause; don't invent a deeper root cause.
   The one-shot `SWITCHROOM_PENDING_*` env vars are passive forensic context for the wake-audit / "why did you restart" protocols — not the resume trigger.
-- **`.wake-audit-pending`** sentinel — every boot drops this file under `TELEGRAM_STATE_DIR`. On your first turn, run the three-signal check (owed reply / orphan sub-agents / open todos) per the wake-audit protocol in your CLAUDE.md, then `rm -f` the sentinel.
+- **`.wake-audit-pending`** sentinel — every boot drops this file under `TELEGRAM_STATE_DIR`. On your first turn, run the three-signal check (owed reply / orphan sub-agents / open todos) per the wake-audit protocol in the `switchroom-runtime` skill (`skills/switchroom-runtime/SKILL.md`), then `rm -f` the sentinel.
 
 A config-summary greeting card is sent automatically by the SessionStart hook — you don't need to announce yourself. If your context feels thin (after compaction or any fresh session), proactively recall from Hindsight before proceeding.
 

package/telegram-plugin/dist/bridge/bridge.js

@@ -23132,6 +23132,20 @@ function createToolLabelSidecar(opts) {
   };
 }
 
+// model-label.ts
+function isModelSentinel(model) {
+  if (typeof model !== "string")
+    return true;
+  const m = model.trim();
+  if (m.length === 0)
+    return true;
+  if (m.startsWith("<"))
+    return true;
+  if (!/^[A-Za-z0-9][A-Za-z0-9._/-]*$/.test(m))
+    return true;
+  return false;
+}
+
 // session-tail.ts
 function isMultiAgentEnabled(env = process.env) {
   return env.PROGRESS_CARD_MULTI_AGENT !== "0";
@@ -23263,6 +23277,10 @@ function projectTranscriptLine(line) {
     if (!Array.isArray(content))
       return [];
     const events = [];
+    const mainModel = message?.model;
+    if (typeof mainModel === "string" && !isModelSentinel(mainModel)) {
+      events.push({ kind: "model", model: mainModel });
+    }
     const textEvents = projectAssistantTextBlocks(content, (text, blockIndex, lastInMessage) => ({ kind: "text", text, blockIndex, lastInMessage }));
     content.forEach((c, i) => {
       const ct = c.type;
@@ -23365,6 +23383,10 @@ function projectSubagentLine(line, agentId, state) {
     if (!Array.isArray(content))
       return [];
     const events = [];
+    const subModel = message?.model;
+    if (typeof subModel === "string" && !isModelSentinel(subModel)) {
+      events.push({ kind: "sub_agent_model", agentId, model: subModel });
+    }
     const textEvents = projectAssistantTextBlocks(content, (text, blockIndex, lastInMessage) => ({
       kind: "sub_agent_text",
       agentId,