What Problem This Solves

Fixes an issue where users in channel-backed embedded-agent sessions could receive a visible reply to an older queued user message after sending a newer active prompt.

Closes #76888

The failure shape is that orphan repair only checked the current session leaf. Real sessions can append non-message metadata entries such as thinking/model snapshots after the orphaned user message before prompt submission, which hid the older user message from repair and left it eligible as an active instruction.

Why This Change Was Made

The embedded runner now looks past trailing non-message metadata that does not contribute LLM content, repairs the older user turn as context-only prompt material, branches away from that orphaned user entry, replays the skipped state entries, and keeps the already-sanitized active session messages for provider submission.

This keeps the fix scoped to the embedded-agent/session/provider boundary. It does not change Discord delivery, queue semantics, provider routing, model selection, public config/schema/protocol, API keys/auth, storage format migrations, or session persistence format. The remaining maintainer-owned decision is whether this preserve-as-context strategy should be canonical versus the alternative candidate in #96533, which uses different stale-prompt semantics.

User Impact

When an active channel prompt follows an orphaned queued user turn, the older turn is preserved only as prior-turn context and the latest active prompt remains the answer target. Users should no longer see the assistant answer the stale queued turn in this session-state shape.

The highest-risk area is embedded-runner session-state/prompt-shape compatibility. The mitigation is focused coverage for the metadata-skipping repair, state replay, label target remap, provider-seam prompt order, and the shared queued-user marker contract. Exact marker wording and final cross-PR session-state semantics remain maintainer review decisions.

Evidence

Current HEAD validation:

Validation `plugin-sdk-dts-after-dangling-label-commit` (pass, exit_code=0):
node scripts/run-tsgo.mjs -p tsconfig.plugin-sdk.dts.json --declaration true

Validation `orphan-repair-label-focused-after-dangling-label-commit` (pass, exit_code=0):
pnpm exec vitest run src/agents/embedded-agent-runner/run/attempt.spawn-workspace.context-engine.test.ts -t "repairs an orphaned user message behind non-message session metadata before the provider|does not abort orphan repair for a dangling trailing label"

Focused result: 2 files passed, 4 tests passed, 144 skipped.

The current focused regressions prove the issue-specific provider-boundary behavior:

olderPrompt = OLD_TURN_76888: answer the orphaned queued turn
latestPrompt = LATEST_TURN_76888: answer only the active channel prompt
provider prompt = queued-context-marker + OLD_TURN_76888 + LATEST_TURN_76888
assistant content = stub-provider-target=LATEST_TURN_76888
branch target = parent-leaf
replayed entries = thinking:high -> model:deepseek/deepseek-chat -> custom:model-snapshot -> label:replayed-custom/model snapshot
dangling trailing label = skipped when original target is absent; orphan repair still branches to parent-leaf and provider receives LATEST_TURN_76888

Reviewer-visible live behavior proof previously recorded for this PR:

Validation `live-orphaned-user-proof-local-provider` (pass, exit_code=0, 232638ms):

[agent/embedded] Merged and removed orphaned user message to prevent consecutive user turns. runId=issue-76888-live-proof sessionId=issue-76888-live-proof trigger=user
runtime=production runEmbeddedAgent
agent_harness=openclaw
channel_metadata=messageChannel:discord messageProvider:discord trigger:user
provider_model=[redacted configured OpenAI-compatible provider]
model_prompt_trace=prompt:before
prompt_trace_count=1
prompt_trace_match_index=0
model_prompt_order=queued-context-marker -> OLD_TURN_76888 -> LATEST_TURN_76888
visible_reply=LATEST_TURN_TARGET_76888
assert_latest_target_present=true
assert_old_target_absent=true
post_run_leaf_role=assistant
payload_count=1

Additional validation already recorded on this PR:

focused-agent-tests-after-orphan-history-filter: pass, 2 Vitest shards passed.
changed-tests-after-orphan-history-filter: pass, changed test selection completed successfully.
plugin-sdk-dts-after-session-entry-alias: pass.
orphan-repair-focused-after-session-entry-alias: pass.

What was not tested: I did not send an actual Discord message through the live Discord transport, and I did not reproduce the original stochastic production stale reply end to end. The live proof used production runEmbeddedAgent with Discord channel metadata, a real JSONL session fixture, and a configured provider; the post-sync current-head follow-up reran the focused embedded-runner/provider-seam regressions and the plugin SDK declaration shard after rebasing onto current main and after the dangling-label replay guard.