PR #17349: fix(compressor): shrink protect_first_n on recompaction (#17344)

• Repository: NousResearch/hermes-agent
• Author: Sanjays2402
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/17349

Description (problem / solution / changelog)

Closes #17344.

Bug

Reporter traced a 6-session compression chain in which every child session carried the identical original first user request — as if no progress had been made. After resume (or on new sessions opened post-compression), the model re-executes the original first task instead of continuing from the handoff summary's ## Active Task.

Root cause

ContextCompressor protects protect_first_n=3 messages at the head — [system, user1, assistant1]. On every cycle that head is preserved verbatim:

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

SUMMARY_PREFIX says "resume from ## Active Task," but user1 (ORIGINAL) is sitting right next to it as a still-prominent user-role message. The model latches onto the first plausible unanswered request and re-executes it — structured summary prose loses against direct attention on a user message. After 6 cycles that same user1 has been re-anchored 6 times.

Fix

On the second and subsequent compactions — detected by checking whether messages[0] (the system prompt) already carries the compaction note we appended last time — shrink protect_first_n to 1 for that call. The original [user1, assistant1] then flow into the summariser pool, 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.

effective_protect_first_n = self.protect_first_n
if self._is_recompaction(messages) and self.protect_first_n > 1:
    effective_protect_first_n = 1

Detection signal

Reuses the existing compaction note already written to the system prompt on first compaction. A new _COMPRESSION_NOTE_SENTINEL constant captures a stable substring ("earlier conversation turns have been compacted into a handoff summary") so PR #17301 — which expands the note text — will not break detection. 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?

The 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.

Coordination with PR #17301 / #17251

PR #17301 (open, by @HiddenPuppy) addresses a sibling problem: SUMMARY_PREFIX over-applies "background reference" framing to memory and skills. Both fixes stem from the same root concern (compaction handoff misinterpreted by the model) but are orthogonal — #17301 carves out exceptions inside SUMMARY_PREFIX text; this PR shrinks protect_first_n on recompaction. They compose cleanly; merge order doesn't matter.

Out of scope

The reporter also flagged parent_session_id = NULL observations on chained sessions. That's a separate DB-write concern — run_agent.py:8891 explicitly passes parent_session_id=old_session_id and resolve_resume_session_id (#15000) handles chain-walking. If NULL is observed it's likely a different write-path failure and deserves its own bug. This PR stays focused on the message-level fix that unbreaks the user-visible "restarts first task" behaviour.

Tests

TestIsRecompaction (6 cases) — sentinel detection edge cases:

• test_fresh_system_prompt_is_not_recompaction
• test_system_prompt_with_compaction_note_is_recompaction
• test_empty_messages_safe
• test_non_system_first_message_is_not_recompaction
• test_multimodal_system_content_is_inspected
• test_garbage_content_does_not_raise

TestRecompactionShrinksProtectFirstN (5 cases) — behavioural:

• test_first_compaction_preserves_first_exchange_in_head (control)
• test_recompaction_demotes_first_exchange_to_summary (the bug)
• test_recompaction_preserves_latest_user_message_in_tail
• test_recompaction_keeps_protect_first_n_attribute_unchanged
• test_protect_first_n_one_no_op_for_recompaction

TestRecompactionMinForCompressGate (1 case) — _min_for_compress early-return uses the effective (post-shrink) head count.

$ python -m pytest tests/agent/test_context_compressor.py \
                   tests/agent/test_context_compressor_recompaction.py \
                   tests/run_agent/test_compression_boundary_hook.py \
                   tests/run_agent/test_compression_persistence.py \
                   tests/run_agent/test_413_compression.py -q
99 passed in 8.26s

87 pre-existing + 12 new, zero regressions.

Changed files

• agent/context_compressor.py (modified, +48/-2)
• tests/agent/test_context_compressor_recompaction.py (added, +234/-0)