…1481)

## What was broken

Slack agents in prd receive messages but never reply — no LLM request is
ever made. Observed live on 2026-06-10 in `iterate` project stream
`/agents/slack/c08r1smtzgd/ts-1781124999-011519`:

- the slack-agent processor rendered the webhook into a triggering
`agent/input-added` at **offset 9** (20:56:46.2)
- the agent processor's `subscription-configured` event landed at
**offset 15** (20:56:46.7) — the AGENT DO wake hook appends it after D1
reads and workspace setup, so slack-agent reliably wins this race on a
cold thread
- the host anchors side effects at the subscription-configured offset
(`stream-processor-host.ts`), so the input at offset 9 was reduced as
historical replay and its scheduling side effect was skipped: no
`llm-request-scheduled`, no `llm-request-requested`, no `openai-ws`
activity, no reply — and nothing ever retriggers it
- visible fingerprint in the stream: capability-noted renders exist only
for offsets above the anchor (18–23), none for 8–9

The anchor mechanism is correct for re-attach (don't re-fire historical
LLM requests), but it shipped in #1402 without anything making the
*first* message of a new thread durable. Regression from #1402, same
symptom as #1372 but a different mechanism. **Every first message of
every new prod Slack thread is dropped.**

## The fix

Make the trigger a durable obligation in reduced state instead of a
fire-and-forget side effect:

- **`AgentState.pendingTriggerOffset`** — set by a triggering
`input-added`, cleared by `llm-request-scheduled` /
`llm-request-requested` / `llm-request-queued`. If it survives in
reduced state, the scheduling side effect never ran.
- **`subscriber-connected` reconciliation recovers it** (the presence
fact always lands above the anchor, so this handler always runs live):
schedule a request when idle, append the queued fact when a request is
in flight (never interrupts in-flight work). Appends are keyed off the
trigger event exactly like the live path
(`agent/llm-request-scheduled@<offset>`), so raced duplicates dedup in
the stream.
- **Gated on `pendingTriggerOffset <= sideEffectsAfterOffset`** so
recovery fires only for anchor-skipped triggers and never races the live
`input-added` handler. Crash/restart cases above the anchor remain owned
by the existing scheduled-phase reconciliation.
- `StreamProcessor.processEvent` args now expose
`sideEffectsAfterOffset` (the batch-level hook already had it); the core
processor's inline path passes 0 (inline appends are always live).

The scheduled phase needs no queued fact on recovery: its handoff
rebuilds the request body from full committed history, which already
includes the skipped trigger.

## Verification

- Unit tests replay the prod stream shape: trigger below anchor +
subscriber-connected above → exactly one `llm-request-scheduled@9`;
non-triggering inputs don't recover; in-flight requests get a queued
fact; live triggers aren't double-scheduled.
- New token-gated e2e (`schedules and completes an LLM request for a
plain routed Slack message`) drives a real Slack root message + routed
webhook through webhook → input → scheduled → requested →
completed(success) against a live deployment.
- `pnpm typecheck && pnpm lint && pnpm format && pnpm test` all green.
- E2E run against the preview deployment with the real Slack bot token:
results to follow in a comment.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **High Risk**
> Changes core agent LLM scheduling and subscriber-connected
reconciliation on a production outage path; incorrect gating could
double-schedule or miss triggers on every new Slack thread.
> 
> **Overview**
> Fixes **first-message silence** on new Slack thread streams when a
triggering `input-added` lands **before** the agent subscription is
configured: the host’s side-effect anchor replays that input into state
but skips scheduling, so no LLM turn ever starts.
> 
> **Agent processor** now records **`pendingTriggerOffset`** in reduced
state for triggering inputs and clears it when a durable
schedule/request/queue fact exists. On **`subscriber-connected`**, when
that offset is at or below the anchor, it **recovers** the missed
obligation—`llm-request-scheduled` when idle (same idempotency key as
the live path) or **`llm-request-queued`** when a request is already in
flight—without double-scheduling live triggers above the anchor.
**`#appendLlmRequestScheduled`** arms the debounce timer with the
**committed** `requestId` after idempotent dedup so raced recovery paths
don’t wedge the handoff.
> 
> **Streams**: `processEvent` receives **`sideEffectsAfterOffset`** so
reconcilers can detect anchor-skipped side effects; the core inline path
passes **`0`** (always live).
> 
> **Verification**: new unit coverage for anchor-skip recovery, deduped
schedule, queue-when-busy, and no recovery for non-triggering inputs;
token-gated e2e asserts routed Slack webhook → scheduled → requested →
completed(success).
> 
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
eb3a7ac. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

<!-- CLOUDFLARE_PREVIEW -->
## Environment Config Lease
<!-- CLOUDFLARE_PREVIEW_STATE -->
<!--
{
  "apps": {
    "os": {
      "appDisplayName": "OS",
      "appSlug": "os",
      "status": "deployed",
      "updatedAt": "2026-06-10T21:58:59.268Z",
      "headSha": "eb3a7ac0bb17f468c1d5490f0b6951bfe612374e",
      "message": null,
      "publicUrl": "https://os.iterate-preview-4.com",
"runUrl": "https://github.com/iterate/iterate/actions/runs/27308869802",
      "shortSha": "eb3a7ac"
    }
  },
  "environmentConfigLease": {
    "dopplerConfig": "preview_4",
    "leasedUntil": 1781132168766,
    "leaseId": "29fbdda0-4a62-44f1-8b9f-ebe4adac552c",
    "slug": "preview-4",
    "type": "environment-config-lease"
  }
}
-->
<!-- /CLOUDFLARE_PREVIEW_STATE -->
Lease: `preview-4`
Doppler config: `preview_4`
Type: `environment-config-lease`
Leased until: 2026-06-10T22:56:08.766Z

### OS
Status: deployed
Commit: `eb3a7ac`
Preview: https://os.iterate-preview-4.com
[Workflow
run](https://github.com/iterate/iterate/actions/runs/27308869802)
Updated: 2026-06-10T21:58:59.268Z
<!-- /CLOUDFLARE_PREVIEW -->

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>