…#17344)

Bug
---
Reporter traced a 6-session compression chain in which every child
session carried the *identical* original first user request \u2014 as if
no progress had been made across the entire chain. After resume (and
on new sessions opened post-compression), the model would re-execute
the original first task instead of continuing from the handoff
summary's '## Active Task' section.

Root cause
----------
ContextCompressor protects 'protect_first_n=3' messages at the head
(default: [system, user1, assistant1]). On every compaction cycle
that head is preserved verbatim. So after compaction N the message
list looks like:

  [system + compaction note, user1 (ORIGINAL), assistant1, summary,
   ...tail..., latest_user]

The summary tells the model 'resume from ## Active Task', but
'user1 (ORIGINAL)' is sitting right next to it as a still-prominent
user-role message. Most models latch onto the first plausible
unanswered request and re-execute it \u2014 the structured summary
guidance loses against direct attention on a user message.

After 6 compactions, that same user1 has been re-anchored 6 times.
The chain never makes progress because every cycle's first task is
the original one.

Fix
---
On the *second and subsequent* compactions \u2014 detected by checking
whether messages[0] (the system prompt) already carries the compaction
note we appended last time \u2014 shrink protect_first_n to 1 for that
call. The original [user1, assistant1] then flow into the summariser
pool alongside everything else, and the structured ## Active Task
section becomes the sole steering signal as designed.

The shrink is per-call; self.protect_first_n is left untouched so
fresh sessions continue to use the configured default.

Detection signal
----------------
Reuses the existing compaction note that compress() already writes
to the system prompt on first compaction. A new
_COMPRESSION_NOTE_SENTINEL constant captures a stable substring of
that note so that PR NousResearch#17301 (which is also editing the note text)
will not break detection \u2014 the sentinel matches both the current
note and NousResearch#17301's expanded version. The new helper
ContextCompressor._is_recompaction(messages) does the lookup with
no I/O, returns False on malformed input, and handles multimodal
system content via the existing _content_text_for_contains() helper.

Why not just strengthen SUMMARY_PREFIX
--------------------------------------
SUMMARY_PREFIX already says 'Respond ONLY to the latest user message
that appears AFTER this summary.' Stronger prose helps marginally
but cannot compete with structural attention on a head-preserved
user message. The reporter explicitly noted: 'the model responds as
if the session had just started.' That's an architectural problem,
not a wording problem; this PR fixes the architecture.

Why not also fix the 'parent_session_id NULL' observation
---------------------------------------------------------
The reporter raised that as a related concern. It's covered by the
existing NousResearch#15000 / resolve_resume_session_id() chain-walk and the
end_session()/create_session() hand-off in run_agent.py:8888 which
explicitly passes parent_session_id=old_session_id. If the reporter
sees NULL it's likely a separate DB-write failure path \u2014 worth its
own bug. This PR stays focused on the message-level fix that
actually unbreaks the user-visible 'restarts first task' behaviour.

Coordination with PR NousResearch#17301 / issue NousResearch#17251
------------------------------------------
PR NousResearch#17301 (open, by HiddenPuppy) addresses a sibling problem:
SUMMARY_PREFIX over-applies 'background reference' framing to the
agent's persistent memory and skills. Both are real, both stem from
the compaction handoff being too easily misinterpreted by the
model, but they're orthogonal fixes \u2014 NousResearch#17301 carves out exceptions
inside SUMMARY_PREFIX text, this PR shrinks protect_first_n on
recompaction. They compose cleanly; merge order doesn't matter.

Tests
-----
- TestIsRecompaction (6 cases): sentinel detection \u2014 fresh prompt,
  prompt-with-note, multimodal content, empty messages, non-system
  first message, garbage content. Helper never raises.
- TestRecompactionShrinksProtectFirstN (5 cases): first compaction
  preserves first exchange in head; recompaction demotes it to the
  summariser pool; latest user message survives in tail; configured
  protect_first_n attribute is not mutated; protect_first_n=1
  configurations are no-ops on recompaction.
- TestRecompactionMinForCompressGate (1 case): _min_for_compress
  early-return uses the effective (post-shrink) head count, so
  small recompacted sessions can still compress.

99 passed across the four compression test files (87 pre-existing +
12 new) on Python 3.12, no regressions.