hermes research decision-grade Pattern A+B mixed — single trial Fleet-wide adoption arc · 2 sub-agents + main-thread synthesis · 4 deliverables

T1 — Pantheon Scheduling Stack — Fleet-Wide Capability Profile

Companion: T2-nous-hermes-fleet-profile.md (mirror axes for matrix join in T3). Source arcs: scheduled-tasks-fleet-protocol v0 (2026-05-20) + n8n-mastra-slack (2026-05-24). Lens: fleet-wide adoption candidate evaluation. Not viska-specific.

Stack inventory (executors covered)

Axis profiles

Scoring: 0 = not supported, 1 = barely, 2 = workable with friction, 3 = solid, 4 = strong, 5 = best-in-class.

Axis 1 — Triggering modes

Executor cron event manual conversational file-watch score note
launchd ✓ ✗ ✓ (launchctl start) ✗ ✗ (kqueue possible, unused) 2 StartCalendarInterval only. src: 07-launchd §plist
/loop ✓ ✗ ✓ in-session ✓ (Claude can author cron mid-turn) ✗ 2 Session-bound; conversational is unique
Routines ✓ ✓ (GitHub webhook, API trigger) ✓ (per-routine API POST) ✗ ✗ 4 Three triggers attachable + combinable. src: 02
CC Desktop ✓ ✗ ✓ via UI ✓ (self-modify via update_scheduled_task — FORBIDDEN in fleet) ✗ 2 Local only. src: 03
n8n ✓ (multi-rule per node) ✓ (webhooks, native triggers) ✓ (Execute Workflow) ✗ ✓ (Local File Trigger) 4 Broadest surface. src: T2 §1
Mastra Workflow ✗ native (in-process setInterval fragile) ✓ (HTTP) ✓ (POST endpoint) ✗ ✗ 2 Not a scheduler; downstream consumer. src: T4 §schedule-caveat
Slack ✗ ✓ (Events API, slash, interact) ✓ (slash command) ✓ (mention thread) ✗ 3 Trigger surface only, not executor

Axis 2 — Persistence + state durability

Executor restart reboot sleep state location score
launchd ✓ ✓ ✓ catches up on wake ~/Library/LaunchAgents/ plist 5
/loop ✗ ✗ ✗ (must be open session) session memory; restored on --resume if unexpired; 7d max 1
Routines ✓ ✓ n/a (cloud) claude.ai cloud per-account 5
CC Desktop ✓ ✓ ✗ sleep skips (catch-up: 1 most-recent missed) Desktop-internal store + SKILL.md on disk 3
n8n ✓ ✓ (process supervised) n/a (server) n8n Postgres executions table; 30d max (EXECUTIONS_DATA_MAX_AGE) 3 — silent-stall failure mode (CRIT, 2yr unresolved). src: T2 footgun #2
Mastra Workflow ✓ ✓ n/a @mastra/pg against Supabase Postgres 4
Slack ✓ (Slack-managed) n/a n/a Slack workspace storage 5 (as delivery surface)

Axis 3 — Tool surface

Executor tool surface available at runtime score
launchd shell ProgramArguments only; any executable on disk 2
/loop full session tools (MCP, plugins, subagents inherit) 5
Routines connectors configured per task; MCP via "Connectors" config; no permission prompts during run; allowlist network egress 4
CC Desktop config files + connectors; MCP servers per-task; permission mode per-task 4
n8n 400+ native nodes (14 NATIVE-T+A in fleet scope), Code node, HTTP Request, Execute Sub-workflow. src: T1-services-landscape 5
Mastra Workflow Mastra Tools (Zod-typed), sub-Agents, sub-Workflows, MCP servers, 11 OTel exporters. src: T4 §1,2 4
Slack Block Kit + chat.postMessage variants + thread_ts n/a (delivery, not invocation)

Axis 4 — Cost model

Executor cost score (cost-efficiency)
launchd $0 — OS-native 5
/loop rolled into Claude Code subscription 4
Routines subscription-tier daily run cap; one-offs exempt. src: 02 §limits 3
CC Desktop rolled into Claude Code subscription 4
n8n self-host on hostinger-vps (fixed VPS cost); Enterprise gates multi-main + External Secrets. src: T2 footgun #1, #7 4
Mastra Workflow self-host; LLM API pay-per-token for Agent step 3
Slack free workspace tier covers fleet volume; per-app-per-workspace rate budget 5

Axis 5 — Isolation / sandbox posture

Executor process network secret tenant-boundary score
launchd OS process none enforced env vars + Keychain per-user (LaunchAgents vs LaunchDaemons) 3
/loop inherits session session permissions session env session-scoped 2
Routines cloud session Trusted-default allowlist (403 host_not_allowed) claude.ai-stored OAuth per-operator-account (NOT per-fleet-agent — identity leakage). src: 02 §identity 3
CC Desktop local process; worktree-isolation toggle (off-by-default — fleet MUST force on) inherits machine inherits machine per-machine 3
n8n single-main process all workflows share; queue mode adds worker pool per-workflow HTTP egress (no allowlist) n8n DB encrypted-but-not-rotated; Enterprise External Secrets only per-workflow; no namespace isolation between mimir-n8n + pantheon-ops authorship 2 — multi-replica double-fire is a tenant-boundary breach. src: T2 footgun #1
Mastra Workflow one Node process per mastra-bridge instance mastra-bridge ACL via Tailscale token per-workflow env / Hades-deposited per-workflow + per-Agent (Tool constraint = sandbox) 4 — analyze-only Agent posture eliminates LLM-driven side-effect class
Slack n/a n/a bot-token per workspace (Hades) workspace-level (slack-boas vs slack-viska budgets independent). src: T5 §3 4

Axis 6 — Observability

Executor logs traces rule-18 evidence integration score
launchd StandardOutPath/StandardErrorPath keys per-plist (BOTH required or silent loss) none application-emitted only 2
/loop session transcript transcript only none 1
Routines claude.ai run list — "green ≠ success" built-in anti-pattern. src: 02 §observability-gap none hostile by default — rule-18 violation surface 1
CC Desktop Desktop session list none none 2
n8n executions table, /api/v1/executions REST, /metrics self-host only (NOT cloud). src: T2 footgun #13 n/a application-emitted via collect-evidence terminal step 3
Mastra Workflow per-step LLM/Tool/Workflow spans via 11 first-party OTel exporters (Langfuse, Datadog, etc.). src: T4 §5.3 ✓ first-class deterministic writeRun step persists verification_evidence JSONB; rule-18 CHECK constraint enforces at DB layer. src: REPORT §7.3 5
Slack per-message ts = delivery proof; permalink audit n/a slack_message_ts in verification_evidence JSONB. src: REPORT §7.1 4

Axis 7 — Multi-tenant boundary

Executor boundary score
launchd per-machine, per-user. K20N0S is shared executor — agents distinguished only by plist Label namespace. 2
/loop session = tenant; no cross-tenant possible 4
Routines per-operator-account. ONE account holds all fleet routines — collapses agent identity into operator. src: 02 §identity 1
CC Desktop per-machine; per-task permission mode 3
n8n per-workflow; mimir-n8n + pantheon-ops author same instance — no namespace; double-fire on multi-replica DOSes shared Postgres + downstream APIs. src: T2 footgun #1 2
Mastra Workflow per-workflow; per-Agent Tool-constrained; can deploy per-tenant mastra-bridge instances 4
Slack per-workspace per-app rate budget; slack-boas noise cannot DOS slack-viska. src: T5 §3 / REPORT §6.3 5

Axis 8 — Failure modes + auto-pause

Executor retry auto-pause silent-stall risk score
launchd KeepAlive semantics; no retry-on-fail backoff knob ✗ (manual) silent skip if StandardErrorPath missing 2
/loop no retry semantics ✗ session-close = loss 1
Routines platform retry opaque ✗ per-account daily cap may silently drop 1
CC Desktop per-task; no backoff knob ✗ sleep-through = skip (1 catch-up max) 2
n8n Retry-On-Fail (no backoff knob); Error Workflow pattern. src: T2 §3, footgun #12 ✗ native (Inngest-style auto-pause-after-N not built-in) CRIT silent-stall (2yr unresolved). src: T2 footgun #2 2
Mastra Workflow per-step retry + workflow-level error branch ✗ native (built fleet-side via PRD §3 ladder) low — observable via OTel spans 4
Slack platform handles n/a rate-limit storm visible via 429 / ratelimited 4

Fleet-applied compensations (designed-in, not executor-native): PRD v0 §3 ladder = 1 failure log → 2 handoff → 3 paused_by_failure=true + athena decision-request → Cerberus P-SCHED-5..8 probes. src: PRD §3, REPORT §8.2

Axis 9 — Learning / feedback loop

Executor reflection step learns from outcome rows score
launchd ✗ ✗ 0
/loop conversational session memory only 2
Routines ✗ ✗ 0
CC Desktop session generates SKILL.md body edits; update_scheduled_task MCP allows self-rescheduling (FORBIDDEN in fleet). src: 03 §self-modifying 1 (fleet-disabled) 1
n8n ✗ ✗ 0
Mastra Workflow analyze step LLM produces reflection_text stored alongside evidence in fleet_schedule_runs; downstream consumers can mine the corpus. Memory primitive (@mastra/memory) available but not wired in fleet posture. src: T4 §1, REPORT §1.1 3
Slack thread analysis (Shape A) lands on same ts parent for compounding context. src: REPORT §6.2 2

Stack-level gap: no feedback loop closes today. reflection_text is written but no executor reads it back to adjust cadence, retry policy, or evidence-shape expectations. Adoption candidate axis.

Axis 10 — Cross-machine + sleep/wake survival

Executor wake catchup cross-machine needs always-on score
launchd ✓ wake-catchup primary advantage over cron. src: 07 ✗ per-host (executor_machine field disambiguates). src: PRD §5 K20N0S no-sleep configured 4
/loop ✗ ✗ session must be open 0
Routines n/a (cloud) ✓ — API trigger reachable from any Tailscale host. src: 02 §triggers ✗ 5
CC Desktop 1 catch-up of most-recent missed; "Keep computer awake" required; closing lid still sleeps. src: 03 §missed-run ✗ per-machine machine must run 2
n8n ✗ no missed-run replay; silent skip. src: T2 §1 failure-modes ✓ (hostinger-vps) hostinger-vps always-on 3
Mastra Workflow n/a (consumes from n8n trigger) localhost:3002 currently K20N0S-only; Tailscale-exposure recommended. src: REPORT §11 Q3 K20N0S 3
Slack n/a ✓ cloud n/a 5

Cross-stack gaps surfaced

Cross-stack strengths

Architecture invariants worth preserving in any adoption decision