Integration-event steer loss: optimistic dedup claim suppresses re-drive of lost steers (relay#1056 §5)

[kjgbot]

Summary

Integration-event injection can permanently drop a message when the steer→agent (PTY) hop loses it after the bridge has already claimed the logical-dedup key. Every subsequent re-delivery of the same logical message is then dedup-suppressed for the full SLACK_RECORD_REPLAY_TTL_MS (~1h), so the agent never receives it and no delivery-failure is surfaced. This has dropped real human (Khaliq/Will) Slack messages.

This is the "consistency bolt" described in relay#1056 §5 (delivery/read ack + re-drive).

Confirmed trace (slack-comms, 2026-06-07)

Dropped message ts 1780847052 ("read #1056 and tell me your thoughts"):

15:44:29  received
15:44:40  injecting   recipients=['slack-comms']   <-- claims the logical dedup key
15:44:47+ skipped duplicate path   (×9)            <-- every re-delivery suppressed

The bridge logged injecting (claimed the key) but the steer never reached the agent's conversation. No delivery-failure was logged. Contrast: the immediately-following answers message injected fine — so this is intermittent, tied to the agent PTY being busy/unreachable at inject time.

Why this is NOT the content-collapse path (and why #145 won't fix it)

• Failure mode (A): on main (pre-Slack integration event fixes + #125 workspace-key wiring #132) a targeted read returns no content → contentHash undefined → claimSlackLogicalInjection() returns false at the if (!contentHash ...) return false guard → collapses to once-per-key. This is what pear#145 fixes.
• Failure mode (B) — THIS issue: content is available, injecting is logged (key claimed), but the steer is lost on the agent hop. A later identical re-delivery hits existing.contentHashes.has(contentHash) → return false and is suppressed. DO-NOT-MERGE: Split #132 Slack integration event fix #145 does not help: with content defined, an unchanged re-delivery of a lost steer is still indistinguishable from a true duplicate without a delivery signal.

Fix direction

Make the dedup claim contingent on delivery, not on the inject attempt:

• Commit the logical/contentHash claim only once the steer is confirmed delivered to the agent (pear already has a broker delivery-confirmation signal — it was deliberately not awaited for system/guidance messages; reuse it here for integration-event injects).
• If delivery is not confirmed within a bounded window, release the provisional claim so a re-delivery re-drives, instead of suppressing for the full replay TTL.
• Align ack semantics with relay#1056 §5 / Will's answers: support both delivered and read acks, harness-driver-owned.

Spans pear (integration-event-bridge.ts claim/inject logic) + the broker/harness-driver delivery-confirmation signal (relay side).

Interim mitigation (active)

slack-comms treats ~/.agentworkforce/pear/integration-events.log as source of truth and OOB-reads any received event lacking a matching delivered injecting. This is load-bearing until the re-drive fix lands; it is manual and must not be the permanent answer.

Notes

• Sibling/parent: relay#1056 (Fleet Delivery RFC), §5 consistency bolt.
• File touched by the in-flight pear#132 split (pear#145) — stack the fix on the canonical bridge state to avoid churn.