[kshitijk4poor]

Open bot tiles no longer cold-start or pin pooled backends from the background, so interactive bot opens stop timing out on "waiting for a free local slot" (#103375; salvage of #103399 by @kyssta-exe).

Fixes #103375. Supersedes #103399 (cherry-picked to preserve authorship — the original commit carried a hermes-pr@nousresearch.com identity that is not the contributor's, so it is re-authored to @kyssta-exe's GitHub noreply).

What changes

• Passive read intent (from fix(desktop): prevent background tile reconcile from starving pool slots (#103375) #103399): HermesApiRequest.passive, getLatestSessionMessages(id, scope, { passive }); reconcileTileTranscripts reads every tile passively.
• Passive is decided inside the resolvers, not the dispatchers. ensureBackend / ensureRegistryBackend take { passive }: an existing pool entry is served without refreshing lastActiveAt; no entry → throw before the eviction/spawn step. The original PR recomputed the pool key in the two REST dispatchers and missed forced-local (conn:local::<profile>) and primary-routed profiles, which would have made every bot-tile reconcile throw against a warm backend (silently — the reconcile catch is bare) and regressed background deliveries into open bot chats ([Bug]: Open Bot Chat stops receiving messages — no on-screen append/re-key path (background delivery AND mid-conversation model switch) #93942).
• Keepalive sweep touches only open sockets. touchSecondaryGateways pinged every wantOpen secondary every 60 s, including ones stuck in reconnect backoff. A tile pins its owner scope wantOpen for its lifetime, so an unreachable backend stayed keepalive-fresh forever and neither LRU eviction nor the idle reaper could free its slot.
• Bounded stalled redials. reconnectSecondary fail-stopped only on removed-connection / deleted-profile errors; a pool-slot wait timeout fell through to an unbounded backoff loop that re-issued a background spawn per tick per tile — the repeating same-order "waiting for a free local slot … timed out … will retry" cycle in the issue's log. After 3 consecutive stalled dials (slot-wait timeout or the 20 s dial timeout) the scope parks (wantOpen=false); fast failures (a restarting gateway) keep the ordinary unbounded backoff. Only a successful open refills the budget (an interleaved fast failure can't). Every explicit open of the scope — the viewed scope's Reconnect / request retry via ensureActiveGatewayOpen, and the online/focus/wake nudge (reconnectSecondaryGateways) — re-arms it with a fresh budget.
• Passive reads stay out of backendDialClaims in the registry dispatcher: they never dial, and an interactive open coalescing onto an in-flight passive read would otherwise inherit its "no warm backend" rejection.

Root cause

Tile-pinned owner scopes are wantOpen for the tile's lifetime; every automatic path that dials them (reconnect backoff, keepalive touch, REST reads) had spawn/slot-holding side effects with no bound, so N tiles > pool size saturated the pool permanently.

Validation

Check	Result
tsc -p tsconfig.electron.json / tsc -p .	clean
eslint on touched files	0 errors
use-background-sync.test.ts (3 failing on #103399)	34/34 pass (assert passive intent)
backend-dial-claim.test.ts (1 failing on #103399), registry-primary-profile-scope.test.ts	14/14, 3/3 (both are source-grep tests pinned to exact call text; strings updated)
gateway-connection-lifecycle.test.ts (+6 invariant tests)	22/22; each guard test goes red when its guard is removed (touch on closed socket / infinite stall budget / budget refilled by a fast failure / no re-arm on the active scope / no re-arm on the wake nudge); the fast-failure test pins that ECONNREFUSED never parks
Full desktop vitest (--project electron + --project ui)	2149 + 7498 pass; 2 failures (managed-ssh-update, use-prompt-actions/utils) reproduce on unmodified origin/main — pre-existing
hermes-desktop-app-dev/scripts/main-boot-smoke.sh on the final branch	main process boots
Live multi-profile repro	not run — no 20-profile desktop install available here; mechanism traced against the issue's log

Not changed: resumeTile's non-passive transcript prefetch — it shares the main-process dial claim with the session.resume dial for the same scope, adds no spawn, and a cold tile needs it to paint (omit_messages resume).

Also relevant: #103401 (same symptom family), #105264, #105390 (open, priority threading — complementary).

[alt-glitch]

This was generated by AI during triage.

Supersedes #103399 (salvage with authorship preserved). Related: #103375 (issue), #102828 (pool cap), #93942 (background delivery regression the original approach would have re-introduced).