Problem

Fixes #10896.

When context compression fires mid-task, the agent can silently drop the user's current request and either stall or re-execute a previously completed task.

Root cause

_find_tail_cut_by_tokens finishes with a call to _align_boundary_backward, which walks the cut boundary backward to avoid splitting a tool_call/tool_result group. In pathological cases this slides cut_idx past the most recent user message, pushing it into the compressed (middle) region.

The LLM summariser faithfully records that user message in ## Pending User Asks, but SUMMARY_PREFIX instructs the receiving model:

"Respond ONLY to the latest user message that appears AFTER this summary."

With the user message now inside the summary rather than in the live tail, the model has no post-summary user message to act on. It either stalls or — worse — interprets the summary's ## Completed Actions as still-pending work and re-executes it.

Reproduction (from issue):

1. Task A completes (many tool calls, large context).
2. User issues Task B.
3. Compression fires while Task B is in its first tool-call cycle.
4. _align_boundary_backward pulls cut_idx before user: "Task B".
5. Agent resumes, ignores Task B, and re-executes Task A.

Fix

1. Structural fix — _find_tail_cut_by_tokens

Added two new private methods:

_find_last_user_message_idx(messages, head_end)
Scans backward from the end to find the last user-role message index.

_ensure_last_user_message_in_tail(messages, cut_idx, head_end)
If the last user message is already in the tail (>= cut_idx), does nothing. If it fell into the middle region, sets cut_idx = last_user_idx directly — a user message is a clean exchange boundary with no tool_call/tool_result pair straddling it, so calling _align_boundary_backward again would over-correct.

Called as the final step of _find_tail_cut_by_tokens, after the existing _align_boundary_backward call:

# Align to avoid splitting tool groups
cut_idx = self._align_boundary_backward(messages, cut_idx)

# Ensure the most recent user message is always in the tail so the
# active task is never lost to compression (fixes #10896).
cut_idx = self._ensure_last_user_message_in_tail(messages, cut_idx, head_end)

2. Summary-template defence-in-depth

Even with the structural fix in place, iterative re-compressions over very long sessions could still lose task context. Three template-level changes add a second layer of protection:

New ## Active Task section (top of template, highest priority):

## Active Task
[THE SINGLE MOST IMPORTANT FIELD. Copy the user's most recent request or
task assignment verbatim — the exact words they used. ...]

Updated SUMMARY_PREFIX — explicitly points the receiving model to ## Active Task so it knows where to find its current objective.

Updated iterative-update prompt — adds CRITICAL: Update "## Active Task" to the list of things to maintain across re-compressions.

Impact

• Zero behaviour change when the last user message already sits inside the token-budget tail (the common case — _ensure_last_user_message_in_tail returns immediately).
• Fixes the regression when _align_boundary_backward over-corrects by anchoring cut_idx to the last user message.
• The tail may grow slightly larger than the token budget when anchoring fires, but this is bounded (at most one additional user message) and preferred over silent task loss.