Field notes

Re-engage via in-session suppression (hook-only), not Slack backfill

Date: 2026-06-16 Status: shipped Supersedes: the re-engage backfill approach (reconstructing the gap from Slack on re-@mention), which was abandoned before merge.

Problem#

When a channel thread disengages (a user @mentions a colleague), the gateway gate dropped every later plain message — they never reached the Claude session. On re-@mention the agent resumed blind to everything said in the gap.

The first attempt rebuilt that gap on re-engage by re-fetching the thread from Slack and injecting a synthetic preamble. It worked but was heavy: a Slack round-trip, name resolution, recency caps, and a truncation story — all to reconstruct context the session could simply have kept in the first place.

Better framing: keep the gap in the session#

Don't drop and later reconstruct — let disengaged messages flow into the session so the transcript stays continuously populated, and just stop the model from responding while disengaged. On re-engage there is nothing to backfill: the context is already there, as real conversation history.

The mechanism is a UserPromptSubmit hook keyed on the thread's engaged flag.

The load-bearing detail: `continue:false`, NOT `decision:"block"`#

A throwaway spike against the pinned SDK (@anthropic-ai/claude-agent-sdk 0.3.173) settled which hook return achieves "persist but don't process". The hook fires before any model call, so no real auth is needed to observe it:

hook return	model runs?	user prompt persisted to transcript?
`decision:"block"`	no	no — discarded before the enqueue/persist step (no jsonl written at all)
(no hook)	yes	yes — `{"type":"user",…}` written even when the model call itself fails
`continue:false`	no (no `assistant` event)	yes — `{"type":"user",…}` lands in the transcript jsonl

So decision:"block" is the wrong primitive despite what the type docs imply (UserPromptSubmitHookSpecificOutput.suppressOriginalPrompt governs the ephemeral block message shown in-the-moment, not durable persistence). continue:false enqueues/persists the user turn, then halts before generation — exactly "session stays populated, model never processes it", at ~0 output tokens. On re-engage the next turn resumes with the gap already in history.

Implementation#

src/agent/manager.ts
- disengagedHookDecision(row) — pure, exported for tests: returns {continue:false, suppressOutput:true, …} when row.engaged === 0, else {continue:true}.
- A UserPromptSubmit hook calls it (reading the row live via Sessions.findById, not closed over — a re-@mention takes effect on the very next message) and increments disengaged_suppressed_total.
src/gateway/core/gateway.ts
- The engagement gate no longer drops disengaged messages. When a thread has a session, both the disengaging colleague-mention and subsequent plain replies are routed into the session via handleMessage(args, {suppress:true}).
- suppress skips all Slack-visible feedback (no 👀, no "thinking…", no ✅ on done) and swaps the reply directive for a "recorded for context only — do NOT reply" note. A colleague-mention in a thread with no session is still dropped (never spin one up for an unrelated thread).
src/metrics.ts — slaude_disengaged_suppressed_total.

Tests#

tests/agent/disengage-hook.test.ts — the pure decision (engaged / disengaged / no-row), and an explicit guard that it never emits decision:"block".
tests/gateway/core/gateway-seam.test.ts — rewrote the engage/disengage durability block to the new contract: disengaged messages are recorded but suppressed (classified by envelope), never processed; no-session colleague-mention is dropped; disengage survives a restart; re-@mention re-engages.

Note#

The user-facing guarantee is unchanged from the durability work — slaude does not reply after you turn to a colleague. What changed is the mechanism: the gap is retained as genuine session history instead of being dropped and later rebuilt from Slack.