| name | slack-relayfile-readdown-debugging |
| description | Use when inbound Slack content (channel messages, thread replies, or DMs) doesn't read down to local disk, or an integration-event preview says "Message unavailable" though the bytes may be on disk. Covers reading mount sync state, write-only-vs-mirror mounts, the bare-vs-suffixed channel-id drop, the preview two-shape gap, mount wedges, and the relayfile#262 backfill trap. |
Debugging Slack read-down failures (relayfile mount ↔ cloud fs/events)
You are diagnosing why inbound Slack content isn't materializing under the local mount (~/.agentworkforce/pear/relayfile/workspaces/<ws>/slack/..., mirrored to .integrations/slack/...). Localize to a layer, then prove it with the mount log / state file.
First read the mount state
For <mount>/.relay/:
mount.log launch line: sync=mirror|write-only mode=poll. sync=write-only ⇒ zero read-down (push only). sync=mirror ⇒ reads down + pushes. Steady mount sync cycle completed ≈ 30s healthy (or ~5min if websocket fell back to poll). Repeated mount sync cycle failed: context deadline exceeded with 0 completed = a wedge. restart fast-path: … skipping bootstrap full pull ⇒ no full pull this launch.state.json: syncMode, eventsCursor/lastEventAt (null/null on a write-only mount that never read down), files{}, pendingWriteback.
Symptom → root cause → fix
- Top-level msgs and/or DMs don't read down, but THREAD replies do → those mounts are
sync=write-only while threads is mirror. Root: pear integration-mounts.ts mountSpecsFor made the dual-purpose command roots write-only with no read-down leg. Fix: set all Slack mounts to syncMode:'mirror' (mirror is a strict superset of write-only) — pear#170; needs a Pear rebuild+restart.
- ⚠️ relayfile#262 trap: flipping write-only→mirror does NOT backfill already-missed messages. A write-only mount sets
BootstrapComplete=true with no pull; on the mirror restart the fast-path seeds the cursor to the tip and skips the bootstrap full pull → only NEW messages read down. Backfill = clearState the mount (delete .relay/) before relaunch — but only if its outbound command dirs are clean (else re-dispatch risk).
- Thread replies specifically don't read down (incremental); only a full pull recovers them → bare-vs-suffixed: cloud's fs/events feed emitted the BARE channel id (
/slack/channels/<id>/threads/...) while the mount root + fs/tree are SUFFIXED (<id>__<slug>). The Go daemon's isUnderRemoteRoot (syncer.go) is a pure literal-prefix match → bare-rooted events silently continue-dropped (no fetch/404/error). Fix: cloud emits the suffixed path at the fs/events source (cloud#2010) — deploys server-side, no Pear restart. Same class as the bridge fix pear#155/cloud#1995, one layer over.
- File IS on disk with content, but the preview says "unavailable" → a PREVIEW gap, not read-down. Two causes: (a) preview reader was SDK-read-only with no local-disk fallback + didn't translate bare→suffixed (pear#155/#163); (b) two record shapes — older/flat records have top-level
text; newer wrap under payload (payload.text). A reader checking only one shape renders "unavailable" though bytes are present. Read-down ≠ preview: confirm by reading the file (handle BOTH shapes).
- A mount wedges (
context deadline exceeded, 0 completed, no recovery) → pear#160: health poll force-restarted only on AUTH failures + read the wrong state path + timeout unset. Fix: generous RELAYFILE_MOUNT_TIMEOUT + a non-auth wedge / stalled-revision detector (queueForcedRestart(path,…,{clearState:true})). A wedged-but-alive mount child is NOT auto-respawned by electron — only an app restart respawns it.
Fix-to-layer matching (don't cross wires)
- bare-vs-suffixed fs/events, record paths, DM D→U mapping = cloud, deploys without a Pear restart.
- mount
syncMode / topology / health = pear, needs a rebuild+restart.
isUnderRemoteRoot literal-prefix, write-only-skips-pull, restart fast-path = relayfile Go daemon (api.relayfile.dev is Cloud's TS worker, not this Go server).
