switchroom 0.18.6 → 0.18.8

package/dist/host-control/main.js

@@ -21495,6 +21495,7 @@ var SessionContinuitySchema = exports_external.object({
   max_turns_in_briefing: exports_external.number().int().positive().optional().describe("Cap on recent user/assistant turn pairs fed to the summarizer."),
   resume_mode: exports_external.enum(["auto", "continue", "handoff", "none"]).optional().describe("How to resume the next session. 'handoff' (default as of #362) " + "never passes --continue; a fresh Claude starts each restart and " + "reads a briefing assembled from recent Telegram messages, Hindsight " + "recall, and today's daily memory file. 'auto' uses --continue when " + "the latest JSONL is smaller than resume_max_bytes, else falls back " + "to the handoff briefing. 'continue' always passes --continue. " + "'none' starts completely fresh every time."),
   resume_max_bytes: exports_external.number().int().positive().optional().describe("Byte threshold above which 'auto' mode falls back to handoff " + "instead of --continue. Default 2_000_000 (~2MB). Large transcripts " + "can blow out the context window even with prefix caching, and " + "--continue replay is known-fragile at scale."),
+  boot_resume: exports_external.enum(["always", "in-flight", "never"]).optional().describe("How the gateway auto-resumes a turn that was IN FLIGHT when the " + "agent restarted. 'in-flight' (default) resumes genuinely " + "interrupted work even after a deliberate/operator restart — a " + "sanctioned restart landing mid-turn no longer silently drops the " + "work. 'always' forces resume unconditionally (same as the " + "SWITCHROOM_BOOT_RESUME_ALWAYS=1 escape hatch). 'never' is the " + "quota-saving posture: don't auto-replay work across a clean " + "restart — but the user is STILL sent a passive notice of what was " + "in flight (silence is never used). Independent of the at-most-once " + "resume ledger and the bounded resume-chain loop-guard, which always " + "apply. Threaded to the gateway as SWITCHROOM_BOOT_RESUME."),
   session_retention_max_count: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): keep at most this many " + "newest session transcripts under .claude/projects; older ones " + "past both this count and the age bound are pruned by the Stop " + "hook. The newest sessions (and the handoff source) are always " + "kept. Default 20; set 0 to disable the count bound."),
   session_retention_max_age_days: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): prune session transcripts " + "older than this many days (a file is deleted only when it is BOTH " + "over the count bound and older than this). Default 30; set 0 to " + "disable the age bound.")
 }).optional();
@@ -28737,7 +28738,7 @@ import { existsSync as existsSync7, readFileSync as readFileSync5 } from "node:f
 import { dirname as dirname4, join as join6 } from "node:path";
 
 // src/build-info.ts
-var VERSION = "0.18.6";
+var VERSION = "0.18.8";
 
 // src/cli/resolve-version.ts
 function readPackageVersion() {

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

@@ -4212,6 +4212,7 @@ var init_schema = __esm(() => {
     max_turns_in_briefing: exports_external.number().int().positive().optional().describe("Cap on recent user/assistant turn pairs fed to the summarizer."),
     resume_mode: exports_external.enum(["auto", "continue", "handoff", "none"]).optional().describe("How to resume the next session. 'handoff' (default as of #362) " + "never passes --continue; a fresh Claude starts each restart and " + "reads a briefing assembled from recent Telegram messages, Hindsight " + "recall, and today's daily memory file. 'auto' uses --continue when " + "the latest JSONL is smaller than resume_max_bytes, else falls back " + "to the handoff briefing. 'continue' always passes --continue. " + "'none' starts completely fresh every time."),
     resume_max_bytes: exports_external.number().int().positive().optional().describe("Byte threshold above which 'auto' mode falls back to handoff " + "instead of --continue. Default 2_000_000 (~2MB). Large transcripts " + "can blow out the context window even with prefix caching, and " + "--continue replay is known-fragile at scale."),
+    boot_resume: exports_external.enum(["always", "in-flight", "never"]).optional().describe("How the gateway auto-resumes a turn that was IN FLIGHT when the " + "agent restarted. 'in-flight' (default) resumes genuinely " + "interrupted work even after a deliberate/operator restart — a " + "sanctioned restart landing mid-turn no longer silently drops the " + "work. 'always' forces resume unconditionally (same as the " + "SWITCHROOM_BOOT_RESUME_ALWAYS=1 escape hatch). 'never' is the " + "quota-saving posture: don't auto-replay work across a clean " + "restart — but the user is STILL sent a passive notice of what was " + "in flight (silence is never used). Independent of the at-most-once " + "resume ledger and the bounded resume-chain loop-guard, which always " + "apply. Threaded to the gateway as SWITCHROOM_BOOT_RESUME."),
     session_retention_max_count: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): keep at most this many " + "newest session transcripts under .claude/projects; older ones " + "past both this count and the age bound are pruned by the Stop " + "hook. The newest sessions (and the handoff source) are always " + "kept. Default 20; set 0 to disable the count bound."),
     session_retention_max_age_days: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): prune session transcripts " + "older than this many days (a file is deleted only when it is BOTH " + "over the count bound and older than this). Default 30; set 0 to " + "disable the age bound.")
   }).optional();

package/dist/vault/broker/server.js

@@ -4212,6 +4212,7 @@ var init_schema = __esm(() => {
     max_turns_in_briefing: exports_external.number().int().positive().optional().describe("Cap on recent user/assistant turn pairs fed to the summarizer."),
     resume_mode: exports_external.enum(["auto", "continue", "handoff", "none"]).optional().describe("How to resume the next session. 'handoff' (default as of #362) " + "never passes --continue; a fresh Claude starts each restart and " + "reads a briefing assembled from recent Telegram messages, Hindsight " + "recall, and today's daily memory file. 'auto' uses --continue when " + "the latest JSONL is smaller than resume_max_bytes, else falls back " + "to the handoff briefing. 'continue' always passes --continue. " + "'none' starts completely fresh every time."),
     resume_max_bytes: exports_external.number().int().positive().optional().describe("Byte threshold above which 'auto' mode falls back to handoff " + "instead of --continue. Default 2_000_000 (~2MB). Large transcripts " + "can blow out the context window even with prefix caching, and " + "--continue replay is known-fragile at scale."),
+    boot_resume: exports_external.enum(["always", "in-flight", "never"]).optional().describe("How the gateway auto-resumes a turn that was IN FLIGHT when the " + "agent restarted. 'in-flight' (default) resumes genuinely " + "interrupted work even after a deliberate/operator restart — a " + "sanctioned restart landing mid-turn no longer silently drops the " + "work. 'always' forces resume unconditionally (same as the " + "SWITCHROOM_BOOT_RESUME_ALWAYS=1 escape hatch). 'never' is the " + "quota-saving posture: don't auto-replay work across a clean " + "restart — but the user is STILL sent a passive notice of what was " + "in flight (silence is never used). Independent of the at-most-once " + "resume ledger and the bounded resume-chain loop-guard, which always " + "apply. Threaded to the gateway as SWITCHROOM_BOOT_RESUME."),
     session_retention_max_count: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): keep at most this many " + "newest session transcripts under .claude/projects; older ones " + "past both this count and the age bound are pruned by the Stop " + "hook. The newest sessions (and the handoff source) are always " + "kept. Default 20; set 0 to disable the count bound."),
     session_retention_max_age_days: exports_external.number().int().nonnegative().optional().describe("Session-JSONL retention (issue #2792): prune session transcripts " + "older than this many days (a file is deleted only when it is BOTH " + "over the count bound and older than this). Default 30; set 0 to " + "disable the age bound.")
   }).optional();

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "switchroom",
   "//version": "NOT the release version — source of truth is the git tag, resolved by scripts/build.mjs:resolveVersion() (see CLAUDE.md > Standard release process). This field is stale by design and only the Layer-4 dev/non-tag fallback for build.mjs + src/cli/resolve-version.ts; do NOT bump it expecting a release to pick it up. npm-pack tarball naming needs a real version — do that as an UNCOMMITTED pack-time bump (see release step 6), never a committed one.",
-  "version": "0.18.6",
+  "version": "0.18.8",
   "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": {
@@ -24,9 +24,9 @@
     "build": "node scripts/build.mjs",
     "build:cli": "node scripts/build.mjs && bun build --compile --target=bun-linux-x64 --minify bin/switchroom.ts --outfile switchroom-linux-amd64",
     "pretest": "npm run build",
-    "test": "vitest run && bun test telegram-plugin/tests/history.test.ts telegram-plugin/tests/cross-turn-card-gate.test.ts telegram-plugin/tests/emission-authority-open-gate.test.ts telegram-plugin/tests/emission-authority-ping-gate.test.ts telegram-plugin/tests/emission-authority-card-drain-gate.test.ts telegram-plugin/tests/per-topic-current-turn.test.ts telegram-plugin/tests/history-reaper.test.ts telegram-plugin/tests/ipc-server-client.test.ts telegram-plugin/tests/ipc-server-race.test.ts telegram-plugin/tests/gateway-bridge.test.ts telegram-plugin/tests/gateway-startup-mutex.test.ts telegram-plugin/tests/gateway-clean-shutdown-marker.test.ts telegram-plugin/tests/boot-card-dedupe.test.ts telegram-plugin/tests/boot-card-reason.test.ts telegram-plugin/tests/progress-update.test.ts telegram-plugin/tests/quota-cache.test.ts telegram-plugin/tests/unhandled-rejection-policy.test.ts telegram-plugin/tests/registry-turns.test.ts telegram-plugin/registry/subagents.test.ts telegram-plugin/registry/subagents-bugs.test.ts telegram-plugin/tests/subagent-watcher-parent-turn-key.test.ts telegram-plugin/tests/subagent-nested-dispatch.test.ts telegram-plugin/tests/nested-worker-visibility-harness.test.ts telegram-plugin/tests/turns-writer.test.ts telegram-plugin/tests/resume-inbound-builder.test.ts telegram-plugin/registry/api-registry.test.ts telegram-plugin/registry/turns-schema.test.ts telegram-plugin/tests/idle-footer-wiring.test.ts telegram-plugin/tests/subagent-tracker-hooks.test.ts telegram-plugin/tests/gateway-update-placeholder-dispatch.test.ts telegram-plugin/tests/status-query-telemetry.test.ts telegram-plugin/tests/reaction-trigger.test.ts telegram-plugin/tests/reaction-trigger-flow.test.ts telegram-plugin/gateway/webhook-ingest-server.test.ts telegram-plugin/tests/skill-proposal-card.test.ts",
+    "test": "vitest run && bun test telegram-plugin/tests/history.test.ts telegram-plugin/tests/cross-turn-card-gate.test.ts telegram-plugin/tests/emission-authority-open-gate.test.ts telegram-plugin/tests/emission-authority-ping-gate.test.ts telegram-plugin/tests/emission-authority-card-drain-gate.test.ts telegram-plugin/tests/per-topic-current-turn.test.ts telegram-plugin/tests/history-reaper.test.ts telegram-plugin/tests/ipc-server-client.test.ts telegram-plugin/tests/ipc-server-race.test.ts telegram-plugin/tests/ipc-server-query-pending-permission.test.ts telegram-plugin/tests/gateway-bridge.test.ts telegram-plugin/tests/gateway-startup-mutex.test.ts telegram-plugin/tests/gateway-clean-shutdown-marker.test.ts telegram-plugin/tests/boot-card-dedupe.test.ts telegram-plugin/tests/boot-card-reason.test.ts telegram-plugin/tests/progress-update.test.ts telegram-plugin/tests/quota-cache.test.ts telegram-plugin/tests/unhandled-rejection-policy.test.ts telegram-plugin/tests/registry-turns.test.ts telegram-plugin/registry/subagents.test.ts telegram-plugin/registry/subagents-bugs.test.ts telegram-plugin/tests/subagent-watcher-parent-turn-key.test.ts telegram-plugin/tests/subagent-nested-dispatch.test.ts telegram-plugin/tests/nested-worker-visibility-harness.test.ts telegram-plugin/tests/turns-writer.test.ts telegram-plugin/tests/resume-inbound-builder.test.ts telegram-plugin/registry/api-registry.test.ts telegram-plugin/registry/turns-schema.test.ts telegram-plugin/tests/idle-footer-wiring.test.ts telegram-plugin/tests/subagent-tracker-hooks.test.ts telegram-plugin/tests/gateway-update-placeholder-dispatch.test.ts telegram-plugin/tests/status-query-telemetry.test.ts telegram-plugin/tests/reaction-trigger.test.ts telegram-plugin/tests/reaction-trigger-flow.test.ts telegram-plugin/gateway/webhook-ingest-server.test.ts telegram-plugin/tests/skill-proposal-card.test.ts",
     "test:vitest": "vitest run",
-    "test:bun": "bun test src/vault/grants.test.ts src/vault/grants-db.test.ts src/vault/write-grants.test.ts src/vault/broker/server-grants.test.ts src/vault/broker/server-write-grants.test.ts src/vault/broker/server-mint-grant-passphrase-attest.test.ts src/vault/broker/server-passphrase-attest.test.ts src/vault/broker/server-mint-grant-posture-attest.test.ts src/vault/broker/server-admin-only-keys.test.ts src/vault/broker/client-token.test.ts src/vault/broker/server-unlock.test.ts src/vault/broker/auto-unlock.test.ts src/vault/broker/drift-detection.test.ts tests/vault-broker-passphrase.test.ts src/cli/vault-get-broker.test.ts src/vault/resolver-via-broker.test.ts src/vault/broker/scope.test.ts src/vault/broker/server.test.ts src/litellm/provision-apply-e2e.test.ts src/drive/disconnect.test.ts src/drive/grants.test.ts src/drive/oauth.test.ts src/drive/onboarding.test.ts src/drive/reconciler.test.ts src/drive/vault-slots.test.ts src/drive/wrapper.test.ts src/vault/approvals/kernel.test.ts src/vault/approvals/approval-origin.test.ts src/vault/approvals/schema-idempotent.test.ts src/vault/broker/server-approvals.test.ts telegram-plugin/tests/boot-probes.test.ts telegram-plugin/tests/boot-version-string.test.ts telegram-plugin/tests/history.test.ts telegram-plugin/tests/cross-turn-card-gate.test.ts telegram-plugin/tests/emission-authority-open-gate.test.ts telegram-plugin/tests/emission-authority-ping-gate.test.ts telegram-plugin/tests/emission-authority-card-drain-gate.test.ts telegram-plugin/tests/per-topic-current-turn.test.ts telegram-plugin/tests/history-reaper.test.ts telegram-plugin/tests/ipc-server-client.test.ts telegram-plugin/tests/ipc-server-race.test.ts telegram-plugin/tests/gateway-bridge.test.ts telegram-plugin/tests/gateway-startup-mutex.test.ts telegram-plugin/tests/gateway-clean-shutdown-marker.test.ts telegram-plugin/tests/boot-card-dedupe.test.ts telegram-plugin/tests/boot-card-reason.test.ts telegram-plugin/tests/progress-update.test.ts telegram-plugin/tests/quota-cache.test.ts telegram-plugin/tests/silent-reply-guard.test.ts telegram-plugin/tests/unhandled-rejection-policy.test.ts telegram-plugin/tests/registry-turns.test.ts telegram-plugin/registry/subagents.test.ts telegram-plugin/registry/subagents-bugs.test.ts telegram-plugin/tests/subagent-watcher-parent-turn-key.test.ts telegram-plugin/tests/subagent-nested-dispatch.test.ts telegram-plugin/tests/nested-worker-visibility-harness.test.ts telegram-plugin/tests/turns-writer.test.ts telegram-plugin/tests/resume-inbound-builder.test.ts telegram-plugin/tests/subagent-tracker-hooks.test.ts telegram-plugin/tests/resolve-calling-subagent.test.ts telegram-plugin/tests/gateway-update-placeholder-dispatch.test.ts telegram-plugin/tests/status-query-telemetry.test.ts telegram-plugin/tests/reaction-trigger.test.ts telegram-plugin/tests/reaction-trigger-flow.test.ts telegram-plugin/tests/subagent-watcher-workflow-visibility.test.ts telegram-plugin/uat/load-env.test.ts telegram-plugin/uat/feed-matcher.test.ts telegram-plugin/uat/uat-driver.test.ts telegram-plugin/gateway/webhook-ingest-server.test.ts telegram-plugin/tests/skill-proposal-card.test.ts",
+    "test:bun": "bun test src/vault/grants.test.ts src/vault/grants-db.test.ts src/vault/write-grants.test.ts src/vault/broker/server-grants.test.ts src/vault/broker/server-write-grants.test.ts src/vault/broker/server-mint-grant-passphrase-attest.test.ts src/vault/broker/server-passphrase-attest.test.ts src/vault/broker/server-mint-grant-posture-attest.test.ts src/vault/broker/server-admin-only-keys.test.ts src/vault/broker/client-token.test.ts src/vault/broker/server-unlock.test.ts src/vault/broker/auto-unlock.test.ts src/vault/broker/drift-detection.test.ts tests/vault-broker-passphrase.test.ts src/cli/vault-get-broker.test.ts src/vault/resolver-via-broker.test.ts src/vault/broker/scope.test.ts src/vault/broker/server.test.ts src/litellm/provision-apply-e2e.test.ts src/drive/disconnect.test.ts src/drive/grants.test.ts src/drive/oauth.test.ts src/drive/onboarding.test.ts src/drive/reconciler.test.ts src/drive/vault-slots.test.ts src/drive/wrapper.test.ts src/vault/approvals/kernel.test.ts src/vault/approvals/approval-origin.test.ts src/vault/approvals/schema-idempotent.test.ts src/vault/broker/server-approvals.test.ts telegram-plugin/tests/boot-probes.test.ts telegram-plugin/tests/boot-version-string.test.ts telegram-plugin/tests/history.test.ts telegram-plugin/tests/cross-turn-card-gate.test.ts telegram-plugin/tests/emission-authority-open-gate.test.ts telegram-plugin/tests/emission-authority-ping-gate.test.ts telegram-plugin/tests/emission-authority-card-drain-gate.test.ts telegram-plugin/tests/per-topic-current-turn.test.ts telegram-plugin/tests/history-reaper.test.ts telegram-plugin/tests/ipc-server-client.test.ts telegram-plugin/tests/ipc-server-race.test.ts telegram-plugin/tests/ipc-server-query-pending-permission.test.ts telegram-plugin/tests/gateway-bridge.test.ts telegram-plugin/tests/gateway-startup-mutex.test.ts telegram-plugin/tests/gateway-clean-shutdown-marker.test.ts telegram-plugin/tests/boot-card-dedupe.test.ts telegram-plugin/tests/boot-card-reason.test.ts telegram-plugin/tests/progress-update.test.ts telegram-plugin/tests/quota-cache.test.ts telegram-plugin/tests/silent-reply-guard.test.ts telegram-plugin/tests/unhandled-rejection-policy.test.ts telegram-plugin/tests/registry-turns.test.ts telegram-plugin/registry/subagents.test.ts telegram-plugin/registry/subagents-bugs.test.ts telegram-plugin/tests/subagent-watcher-parent-turn-key.test.ts telegram-plugin/tests/subagent-nested-dispatch.test.ts telegram-plugin/tests/nested-worker-visibility-harness.test.ts telegram-plugin/tests/turns-writer.test.ts telegram-plugin/tests/resume-inbound-builder.test.ts telegram-plugin/tests/subagent-tracker-hooks.test.ts telegram-plugin/tests/resolve-calling-subagent.test.ts telegram-plugin/tests/gateway-update-placeholder-dispatch.test.ts telegram-plugin/tests/status-query-telemetry.test.ts telegram-plugin/tests/reaction-trigger.test.ts telegram-plugin/tests/reaction-trigger-flow.test.ts telegram-plugin/tests/subagent-watcher-workflow-visibility.test.ts telegram-plugin/uat/load-env.test.ts telegram-plugin/uat/feed-matcher.test.ts telegram-plugin/uat/uat-driver.test.ts telegram-plugin/gateway/webhook-ingest-server.test.ts telegram-plugin/tests/skill-proposal-card.test.ts",
     "test:watch": "vitest",
     "lint": "tsc --noEmit && node scripts/check-plugin-references.mjs && bash scripts/check-bot-api-wrapping.sh && node scripts/check-bun-test-imports.mjs && node scripts/check-no-pii-secrets.mjs && node scripts/check-vault-test-hermeticity.mjs && node scripts/check-no-broadcast-delivery.mjs && node scripts/check-stale-tool-descriptions.mjs && node scripts/check-web-subscription-honest.mjs",
     "lint:tsc": "tsc --noEmit",

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
@@ -967,8 +980,8 @@ fi
 # per-session cost/issue tracking is handled out-of-band by correlating
 # LiteLLM's request log against switchroom's turn ledger, never by mutating the
 # claude protocol.
-# _LITELLM_OK is read by the session-model-override block below: an sr-*
-# (LiteLLM/OpenRouter) session override must be DROPPED when the proxy is
+# _LITELLM_OK is read by the session-model resolution block below: an sr-*
+# (LiteLLM/OpenRouter) session override must NOT be applied when the proxy is
 # unreachable at boot, or claude would 4xx an unknown model against Anthropic.
 # Empty by default (proxy not confirmed live); set to "1" only on a successful
 # liveliness probe. Deliberately NOT unset with the sr_ll_* scratch vars.
@@ -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.
@@ -1046,54 +1065,131 @@ x-litellm-tags: agent:$SWITCHROOM_AGENT_NAME,profile:${SWITCHROOM_AGENT_PROFILE:
   unset sr_ll_key sr_ll_ok sr_ll_unreachable
 fi
 
-# --- Session-only model override (carrier: .session-model-override) ---
+# --- Session model resolution (durable .session-model + .relaunch-model-intent) ---
+#
+# Contract: reference/rfcs/session-model-stickiness.md. A positively-
+# confirmed `/model X` switch persists as `{{agentDir}}/.session-model`
+# (one-line JSON written by the gateway). Whether THIS boot applies it is
+# decided by the one-shot `.relaunch-model-intent` file, stamped by the
+# gateway immediately before every switchroom-managed bounce:
 #
-# The Telegram `/model <sr-*>` switch can't use claude's in-REPL `/model`
-# picker for non-Anthropic (LiteLLM/OpenRouter `sr-*`) models — the CLI's
-# native picker rejects unknown ids ("Model not found"). Instead the gateway
-# writes the chosen model token to `.session-model-override` and gracefully
-# restarts the agent; THIS boot picks it up and launches `claude --model
-# <token>` directly (no picker validation at launch — LiteLLM routes it).
+#   fresh (<10 min by embedded ts) "keep" intent  → apply the override
+#   anything else (revert / absent / stale / corrupt) → delete the override,
+#     boot the configured default, and (when an override was actually
+#     removed) write a `.session-model-alert` notice naming why.
+#
+# Default is REVERT (operator decision 2026-07): a crash, OOM, raw
+# `docker restart`, host reboot, or deploy writes no intent and therefore
+# reverts to the yaml model. Watchdog/recovery bounces run gateway code and
+# stamp "keep" before SIGTERM, so a confirmed switch survives them.
 #
-# One-shot + session-only: the carrier is consumed (rm) on this one boot
-# regardless of validity, so the next restart reverts to the configured
-# default `{{{modelQ}}}`. Mirrors the `.force-fresh-session` marker's
-# gateway↔start.sh round-trip. See telegram-plugin/gateway/model-command.ts.
 # NB {{{modelQ}}} is already shell-single-quoted by the scaffold (it renders as
 # a quoted token, e.g. 'claude-sonnet-5'), so it is assigned BARE here — never
 # inside additional double quotes, which would embed the literal quote chars in
 # the value and break `claude --model`.
 _EFFECTIVE_MODEL={{{modelQ}}}
+# Record the RESOLVED configured default (raw, unquoted) every boot, before
+# override resolution. The gateway copies this into the override's
+# `configuredDefaultAtWrite`, so both sides of the invalidation compare below
+# come from the same resolver. Overwrite, not consumed.
+printf '%s\n' "$_EFFECTIVE_MODEL" > "{{agentDir}}/.configured-default-model" 2>/dev/null || true
+
+# Migration shim (one release, RFC §7): a leftover one-shot
+# `.session-model-override` carrier means an OLD gateway wrote it immediately
+# before this very bounce — it is the newest user intent and WINS over any
+# `.session-model` on disk. Convert (overwrite) + consume, and apply THIS
+# boot (the carrier's presence is itself the keep signal).
+_sm_keep=""
 if [ -f "{{agentDir}}/.session-model-override" ]; then
-  _override="$(tr -d '[:space:]' < "{{agentDir}}/.session-model-override" 2>/dev/null || true)"
+  _mig="$(tr -d '[:space:]' < "{{agentDir}}/.session-model-override" 2>/dev/null || true)"
   rm -f "{{agentDir}}/.session-model-override"
-  # Shape gate — kept aligned with MODEL_ARG_RE in telegram-plugin/gateway/model-command.ts.
-  # `/` is allowed for OpenRouter-style `sr-vendor/model` ids; it is not a shell
-  # metachar inside the double-quoted `claude --model "$_EFFECTIVE_MODEL"` usage below.
-  if printf '%s' "$_override" | grep -Eq '^[A-Za-z0-9][]A-Za-z0-9._[/-]{0,99}$'; then
-    _EFFECTIVE_MODEL="$_override"
-    # LiteLLM-down guard for an sr-* OVERRIDE only (operator decision 2026-07):
-    # with the proxy unreachable at boot there IS a Claude configured default to
-    # fall back to, so DROP the override rather than boot on an unreachable sr-*
-    # model that would 4xx against Anthropic — log LOUDLY and write an alert
-    # sentinel the gateway turns into a Telegram message. (The passthrough→router
-    # repoint for a LIVE sr-* — override OR configured default — is done once,
-    # post-resolution, below.)
-    case "$_override" in
+  if printf '%s' "$_mig" | grep -Eq '^[A-Za-z0-9][]A-Za-z0-9._[/-]{0,99}$'; then
+    printf '{"model":"%s","configuredDefaultAtWrite":"%s","ts":%s}\n' "$_mig" "$_EFFECTIVE_MODEL" "$(( $(date +%s) * 1000 ))" > "{{agentDir}}/.session-model" 2>/dev/null || true
+    _sm_keep="1"
+    echo "session-model: migrated legacy one-shot carrier '$_mig' to durable .session-model (applying this boot)" >&2
+  else
+    echo "session-model: ignoring malformed legacy .session-model-override (failed shape gate)" >&2
+  fi
+  unset _mig
+fi
+
+# One-shot intent: consume unconditionally, honor only a fresh "keep".
+# Freshness clock is the EMBEDDED ts (ms), never file mtime — same clock the
+# gateway writes with.
+_sm_reason=""
+if [ -f "{{agentDir}}/.relaunch-model-intent" ]; then
+  _int_raw="$(cat "{{agentDir}}/.relaunch-model-intent" 2>/dev/null || true)"
+  rm -f "{{agentDir}}/.relaunch-model-intent"
+  _int="$(printf '%s' "$_int_raw" | sed -n 's/.*"intent"[[:space:]]*:[[:space:]]*"\([a-z]*\)".*/\1/p')"
+  _int_ts="$(printf '%s' "$_int_raw" | sed -n 's/.*"ts"[[:space:]]*:[[:space:]]*\([0-9]\{1,\}\).*/\1/p')"
+  _sm_reason="$(printf '%s' "$_int_raw" | sed -n 's/.*"reason"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')"
+  if [ "$_int" = "keep" ] && [ -n "$_int_ts" ] && [ $(( $(date +%s) * 1000 - _int_ts )) -lt 600000 ]; then
+    _sm_keep="1"
+  fi
+  unset _int_raw _int _int_ts
+fi
+
+if [ -f "{{agentDir}}/.session-model" ]; then
+  _smf="$(cat "{{agentDir}}/.session-model" 2>/dev/null || true)"
+  _sm_model="$(printf '%s' "$_smf" | sed -n 's/.*"model"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')"
+  _sm_cfg="$(printf '%s' "$_smf" | sed -n 's/.*"configuredDefaultAtWrite"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')"
+  _sm_ts="$(printf '%s' "$_smf" | sed -n 's/.*"ts"[[:space:]]*:[[:space:]]*\([0-9]\{1,\}\).*/\1/p')"
+  if [ -z "$_sm_keep" ]; then
+    # No fresh keep intent → deliberate restart, crash, external container
+    # restart, or deploy. Revert to the configured default and say so.
+    rm -f "{{agentDir}}/.session-model"
+    _sm_why="${_sm_reason:-no keep intent — crash, external container restart, or deploy}"
+    echo "session-model: reverting to configured default '$_EFFECTIVE_MODEL' — $_sm_why (session override '$_sm_model' cleared)" >&2
+    printf 'Session model override `%s` was cleared — this relaunch reverted to the configured default `%s` (%s). Re-issue /model %s to switch back.\n' "$_sm_model" "$_EFFECTIVE_MODEL" "$_sm_why" "$_sm_model" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+    unset _sm_why
+  # Shape gate — kept BYTE-IDENTICAL with MODEL_ARG_RE in
+  # telegram-plugin/gateway/model-command.ts. `/` is allowed for
+  # OpenRouter-style `sr-vendor/model` ids; it is not a shell metachar inside
+  # the double-quoted `claude --model "$_EFFECTIVE_MODEL"` usage below.
+  # The wc -c/-l check rejects EMPTY and MULTILINE values first: the sed
+  # extraction above emits one match per matching LINE, and `grep -Eq`
+  # passes if ANY line matches — a multiline value must never reach
+  # `claude --model` (parity with parseSessionModel's single-string check).
+  elif [ "$(printf '%s' "$_sm_model" | wc -c)" -eq 0 ] || [ "$(printf '%s' "$_sm_model" | wc -l)" -ne 0 ] || ! printf '%s' "$_sm_model" | grep -Eq '^[A-Za-z0-9][]A-Za-z0-9._[/-]{0,99}$'; then
+    rm -f "{{agentDir}}/.session-model"
+    echo "session-model: ignoring malformed .session-model (failed shape gate) — using configured default '$_EFFECTIVE_MODEL'" >&2
+  elif [ "$_sm_cfg" != "$_EFFECTIVE_MODEL" ]; then
+    # switchroom.yaml `model:` changed since the switch → the override is
+    # against a default that no longer exists. Invalidate + announce.
+    rm -f "{{agentDir}}/.session-model"
+    echo "session-model: configured default changed ('$_sm_cfg' → '$_EFFECTIVE_MODEL') — clearing session override '$_sm_model'" >&2
+    printf 'The configured default model changed (`%s` → `%s`), so your session override to `%s` was cleared — the agent booted on the new configured default. Re-issue /model %s if you still want it.\n' "$_sm_cfg" "$_EFFECTIVE_MODEL" "$_sm_model" "$_sm_model" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+  elif [ -n "$_sm_ts" ] && [ $(( $(date +%s) * 1000 - _sm_ts )) -gt 604800000 ]; then
+    # 7-day staleness bound (belt-and-braces vs version-rollback resurrection
+    # and long-forgotten overrides).
+    rm -f "{{agentDir}}/.session-model"
+    echo "session-model: session override '$_sm_model' is older than 7 days — expiring it, using configured default '$_EFFECTIVE_MODEL'" >&2
+    printf 'Session model override `%s` expired (older than 7 days) — the agent booted on its configured default `%s`. Re-issue /model %s if you still want it.\n' "$_sm_model" "$_EFFECTIVE_MODEL" "$_sm_model" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+  else
+    case "$_sm_model" in
       sr-*)
         if [ -z "$_LITELLM_OK" ]; then
-          _EFFECTIVE_MODEL={{{modelQ}}}
-          echo "session-model: LiteLLM proxy unreachable at boot — DROPPING sr-* override '$_override', booting on configured default '$_EFFECTIVE_MODEL' instead (would otherwise 4xx against Anthropic)" >&2
-          echo "session-model: the requested sr-* model is UNAVAILABLE this session; re-issue /model $_override once LiteLLM is reachable" >&2
-          printf 'LiteLLM proxy was unreachable at boot, so the session-only model switch to `%s` was dropped — the agent booted on its configured default `%s` instead. Re-issue /model %s once LiteLLM is reachable.\n' "$_override" "$_EFFECTIVE_MODEL" "$_override" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+          # sr-* + LiteLLM unreachable: booting the override would 4xx against
+          # Anthropic. Boot the configured default but RETAIN the durable file
+          # — it re-applies on the next keep relaunch once the proxy is back.
+          echo "session-model: LiteLLM proxy unreachable at boot — NOT applying sr-* override '$_sm_model' this boot; booting configured default '$_EFFECTIVE_MODEL' (override retained, re-applies next relaunch)" >&2
+          printf 'LiteLLM proxy was unreachable at boot, so the session model override `%s` was not applied — the agent booted on its configured default `%s`. The override is retained and will re-apply on the next switchroom-managed relaunch once LiteLLM is reachable; /model default drops it.\n' "$_sm_model" "$_EFFECTIVE_MODEL" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+        else
+          _EFFECTIVE_MODEL="$_sm_model"
         fi
         ;;
+      *)
+        _EFFECTIVE_MODEL="$_sm_model"
+        ;;
     esac
-  else
-    echo "session-model: ignoring malformed .session-model-override (failed shape gate) — using configured default '$_EFFECTIVE_MODEL'" >&2
+    if [ "$_EFFECTIVE_MODEL" = "$_sm_model" ]; then
+      echo "session-model: keeping session override '$_sm_model' across this switchroom-managed relaunch${_sm_reason:+ ($_sm_reason)}" >&2
+      printf 'Session model override `%s` kept across this relaunch%s. It reverts on /restart, agent restart, crash, or an external container restart; /model default clears it now.\n' "$_sm_model" "${_sm_reason:+ ($_sm_reason)}" > "{{agentDir}}/.session-model-alert" 2>/dev/null || true
+    fi
   fi
-  unset _override
+  unset _smf _sm_model _sm_cfg _sm_ts
 fi
+unset _sm_keep _sm_reason
 
 # sr-* passthrough→router repoint — ONE post-resolution gate covering BOTH the
 # /model override path AND the configured-default path (`model: sr-*` in

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/skills/switchroom-runtime/SKILL.md

@@ -178,4 +178,6 @@ This is **the persistent-shell wedge.** Claude Code keeps a single `bash` subpro
 
 **Triggering causes to avoid.** The wedge most often follows: (a) a long `npm test` / `bun test` run, (b) any command that was `!`-interrupted mid-flight, (c) heredoc-style commands the shell's stdin couldn't fully consume. Prevention: dispatch heavy test suites to a worker sub-agent (so the wedge dies with the worker) rather than running them in your own session, and use `run_in_background: true` for long jobs.
 
+**Mid-flight responsiveness — never block your turn on a long foreground watch.** A blocking foreground command (`gh pr checks --watch`, `sleep`-and-poll loops, a long `docker build`) pins your whole turn: a user message arriving mid-watch waits behind it, sometimes for minutes. The gateway posts a deterministic "⏳ Queued — currently inside `<tool>`" ack on your behalf (#2995), but the ack is a mitigation, not a license. Run long watches as background tasks (`run_in_background: true`, then poll `BashOutput` between other work) so mid-flight questions get real answers in seconds.
+
 A sentinel file at `$TELEGRAM_STATE_DIR/wedge-detected.json` records the most recent wedge detection. Operators can `cat` it for forensic timestamps; you don't normally need to read it yourself.

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,