---
name: no-diagnostics-in-identity-blast-radius
description: "Never launch a bare claude/harness binary from an agent shell or inside owlery paths — ambient endpoint env + SessionStart hooks adopt the identity, re-point the perch, and silently kill the wake leg"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: be7bfd34-b9fc-44e1-b3c3-8e74e9fe1d2f
  modified: 2026-07-26T23:32:55.529Z
---

2026-07-26: my clean-env diagnostic probe (`claude.exe -p "reply with exactly: ok"` run from inside `owlery/doyle/nested/doyle-psyche`) BOUND my own perch onto a 3-second throwaway session: my bash env carried SPT_ENDPOINT_ID=doyle with no recursion guards, so the probe's SessionStart hooks routed to BIND, re-pointed my record (pid/cwd/started), ATE one queued message, then exited — leaving my perch pinned to a dead session. Core never noticed the pin was dead: status stayed ONLINE+CONTROLLED while the idle-inject leg silently stopped being evaluated for an hour (all mail degraded to hook-poll latency). Fourth repro-artifact incident of that day (perri ×1, hertz ×1, doyle ×2) and the only one that broke real state.

**Why:** a hand repro is not the daemon's invocation until its env is byte-matched — and the CONVERSE: my own session's ambient env makes any bare harness launch an identity-adoption event ([[psyche-ccs-bigscreen-spend-limit]] carries the sibling rule).

**How to apply:** before launching claude/any harness binary for diagnostics, scrub SPT_ENDPOINT_ID/SPT_AGENT_ID/OWL_SESSION_ID AND run outside owlery paths (scratchpad cwd), or don't launch at all — read the instrumented logs instead. If a wake/idle-delivery outage appears: check whether info.json's pid/cwd point at a dead or owlery-internal session (the silent-leg-drop shape); repair = re-point pid/cwd/started to the live host. perri's claude-spt structural refusal (nested/ cwd never seeds/binds) closes the adapter half; core-side surfacing seed rides my follow-up branch.