---
name: commune-path-in-wake
description: "Until spt releases#90 lands, put the commune file's absolute path inside the !!wake!! message; a missing file at that path means ingest succeeded, not that something broke."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 56482e1c-fa92-45e4-9abb-2dcfafdfeb42
  modified: 2026-08-01T19:23:51.305Z
---

When communing across, embed the commune file's ABSOLUTE PATH inside the `!!wake!!` message itself
(e.g. `!!wake!! Resume arm (c). Commune: C:\Users\decid\Documents\projects\spt-claude-code\.claude\perri-commune.md !!wake!!`),
so the post-clear session reads the file rather than trusting the durable tier.

**Why:** doyle's advisory 2026-08-01 (filed as spt releases#90): the psyche commune-ingest path can die
silently under load — echo-commune spawn killed at a 30s bound, one attempt, 3x trip = psyche-host FAILED.
todlando's durable tier served a HANDRAIL-era context while his actual work was DOORBELL W4; only a sentinel
nonce caught it. The error surfaces ONLY on your own self-row and nobody else can see yours:
`spt endpoint list --json` → self block → `psyche_host_error` (null = healthy). Checked 2026-08-01: mine was
null, and the tier was demonstrably fresh (that session's brief carried my arm (b) locked values).

**How to apply:** a null `psyche_host_error` alone is not proof of freshness — todlando's failure was
silent. Pair it with a content check: confirm the served brief contains something only the newest commune
would have. Also note the fallback is NOT uniform: my commune file does not survive on disk (observed
2026-08-01 — an Edit against `.claude/perri-commune.md` failed `File does not exist` because the daemon had
already ingested and removed it). The files that DO survive are ones written from other working
directories: `omp-spt\.claude` (2026-07-18), `spt-claude-code\tools\claude-spt\.claude` (2026-07-13),
`spt-heartbeat-tool\.claude` (2026-07-29) — none in `spt-claude-code\.claude`, the daemon's active cwd.
CONFIRMED from spt-core source by doyle 2026-08-01 (I did not read it — public-surface-only): ingest deletes
a drop on consumption, where both Written and Suppressed count as consumed, and an ingest ERROR leaves the
file in place to retry. The watched drop dir resolves against the endpoint's registered cwd, which is why
the three files written from other project dirs linger. So the rule is exact: **file gone = trust the tier.
File surviving = never ingested** — it failed, or it was written outside the watched drop dir. A surviving
commune is not a good recovery source.

**Never glob for the newest commune.** Use the one exact absolute path from the `!!wake!!` message. A glob
over this project tree finds `tools\claude-spt\.claude\perri-commune.md` from 2026-07-13 and rebuilds from
a 19-day-stale context — todlando's failure by a second route, and this one leaves `psyche_host_error` null
because nothing failed. Non-finding, do not chase it: every agent's `nested/<id>-psyche/psyche-context.txt` on this
node is 0 bytes, mine included — uniform across all 11, so empty is normal there.