Summary

• Problem: Issue Inbound system event rendered twice per turn (runtime-context message + reply-body prepend) on Slack #95323 reports that a Slack DM (or any channel message that drains a system event) delivers the inbound label — e.g. System: [2026-06-20 13:59:51] Slack DM from Alice — to the model twice in the same turn: once inside the hidden runtime-context custom message (Message A) and again inline in the model-facing user turn (Message B). Both occurrences carry the same timestamp, confirming a single event rendered at two sites.
• Root Cause: When before_prompt_build hooks (or queued next-turn injections) return prependContext or appendContext — as active-memory, memory-lancedb, and agent_turn_prepare hooks do — attempt.ts sets hasPromptBuildContext = true and passes promptForModelBeforeRuntimeContextSplit (the full effective prompt after hooks, which includes the system-event prefix already prepended to queuedBody) as the modelPrompt argument to resolveRuntimeContextPromptParts. The function correctly extracts the system-event prefix into hiddenRuntimeContext / runtimeContext (→ Message A via the hidden custom transcript message). However, modelPromptText is taken from modelPrompt.text, which still contains the prefix verbatim. Because modelPromptText !== transcriptPrompt, the function returns it as promptSubmission.modelPrompt, which buildCurrentInboundPrompt wraps with currentInboundContext to form Message B — event prefix intact.
• Fix: After computing hiddenRuntimeContext, strip it from modelPromptText using removeLastPromptOccurrence before returning. lastIndexOf targets the last occurrence so hook content that coincidentally contains the same literal is not mis-targeted; the ?? modelPromptText fallback makes the branch a safe no-op when the substring is absent. The guard hiddenRuntimeContext && modelPrompt is falsy in every pre-existing code path (either hiddenRuntimeContext resolves to "" when modelPrompt.text === extracted.text, or modelPrompt is undefined), so the new branch is unreachable for all existing callers.
• What changed:
  • src/agents/embedded-agent-runner/run/runtime-context-prompt.ts — compute returnModelPromptText by stripping hiddenRuntimeContext from modelPromptText when both are non-empty; use it in the final return.
  • src/agents/embedded-agent-runner/run/runtime-context-prompt.test.ts — two regression tests: prepend-context hook case and append-context hook case; both fail on main and pass after the fix.
  • src/agents/embedded-agent-runner/run/attempt.llm-boundary.test.ts — model-boundary regression: asserts that after the fix, when insertRuntimeContextMessageForPrompt + normalizeMessagesForLlmBoundary assemble the LLM-bound message array, the system-event label appears only in the runtime-context custom message (Message A) and not in the model-facing user turn (Message B).
• What did NOT change: Config surface unchanged. Plugin surface unchanged. Gateway protocol unchanged. The !prompt.trim() (runtime-event) and emptyTranscriptMode === "model-prompt" early-return branches are untouched.

Reproduction

1. Configure a Slack session with the active-memory or memory-lancedb plugin so before_prompt_build returns prependContext on a memory hit (or any plugin that queues next-turn context injections).
2. Send a Slack DM to the agent.
3. Inspect the model API request (provider payload or transcript).
4. Before this PR: the model-facing user turn contains System: [timestamp] Slack DM from <user> inline, before the actual message text; the same label also appears in the preceding hidden runtime-context custom message — two occurrences, same timestamp.
5. After this PR: the system-event label appears only in the hidden runtime-context message (Message A); the model-facing user turn (Message B) contains only the hook context and the user text.

Real behavior proof

Behavior addressed (#95323, Slack inbound system event double-render): when a before_prompt_build hook (active-memory, memory-lancedb, or any next-turn injection) adds prepend or append context, resolveRuntimeContextPromptParts now returns modelPrompt without the system-event prefix and places it exclusively in runtimeContext; the model receives the event exactly once per turn.

Real environment tested (Linux, Node 22 — Vitest against the production resolveRuntimeContextPromptParts and insertRuntimeContextMessageForPrompt + normalizeMessagesForLlmBoundary implementations with inputs matching the attempt.ts call site for the hook+system-event scenario): promptForRuntimeContextSplit as the pre-hook queuedBody (system-event prefix + user text), transcriptPrompt as transcriptCommandBody (user text only), modelPrompt as promptForModelBeforeRuntimeContextSplit (hook context + queuedBody); model-boundary assembly uses insertRuntimeContextMessageForPrompt with the extracted runtime-context message and the stripped model-facing user message.

Exact steps or command run after this patch: node scripts/run-vitest.mjs src/agents/embedded-agent-runner/run/runtime-context-prompt.test.ts src/agents/embedded-agent-runner/run/attempt.llm-boundary.test.ts extensions/slack/src/monitor/message-handler/prepare.test.ts src/auto-reply/reply/get-reply-run.media-only.test.ts.

Evidence after fix (Vitest output):

runtime-context-prompt.test.ts     Tests  23 passed (23)   (includes 2 new)
attempt.llm-boundary.test.ts       Tests  24 passed (24)   (includes 1 new)
prepare.test.ts                    Tests  85 passed (85)
get-reply-run.media-only.test.ts   Tests  82 passed (82)

Observed result after fix: in the new model-boundary test, contextMsg.content contains the system-event label (Message A); String(userTurn.content) does not (Message B). In the two new helper tests, resolveRuntimeContextPromptParts returns modelPrompt: "prependContext\n\nuser text" and runtimeContext: "System: [ts] Slack DM from Alice" — event absent from the model prompt, present only in runtime context.

What was not tested: a live Slack session with a real active-memory hit end-to-end. The production flow from drainFormattedSystemEvents → buildReplyPromptBodies → queuedBody → attempt.ts call site → resolveRuntimeContextPromptParts → insertRuntimeContextMessageForPrompt → normalizeMessagesForLlmBoundary is covered at each layer by the focused unit tests; a full gateway round-trip was not driven.

Repro confirmation: the two new helper tests exercise the exact production input shapes (promptForRuntimeContextSplit, transcriptPromptForRuntimeSplit, promptForModelBeforeRuntimeContextSplit) from attempt.ts:4193-4210; both fail on main without the stripping step and pass after it. The model-boundary test confirms the resulting message structure has the event label only once.

Risk / Mitigation

• Risk: removeLastPromptOccurrence could incorrectly target hook-injected content that happens to be identical to the system-event text. Mitigation: lastIndexOf removes the last occurrence; since the system event is prepended to queuedBody before hooks wrap it, the system-event text appears after (in string position) any hook prependContext — so the last occurrence is always in the queuedBody portion, not the hook portion.
• Risk: stripping could fire on a modelPromptText that does not actually contain the system-event label (e.g. a path where hooks rewrote the body). Mitigation: removeLastPromptOccurrence returns null when the substring is absent; the ?? modelPromptText fallback leaves returnModelPromptText unchanged, and the final return behaves identically to the pre-fix code.

Change Type (select all)

• feat
• fix
• improve
• refactor
• docs
• chore

Scope (select all touched areas)

• Agents / runner
• Tests

Linked Issue/PR

Fixes #95323