…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.

After context compression, the model can respond to stale context instead
of the user's latest message when large tool outputs consume the entire
tail token budget. This is because `_find_tail_cut_by_tokens()` uses a
pure token-budget walk with a hard minimum of 3 messages — which may only
protect tool results, not the user's actual request.

This fix adds a **user-message anchor** to `_find_tail_cut_by_tokens()`:

1. Find the index of the most recent user message
2. Set it as a floor for the tail boundary — the token-budget walk can
   include more messages (if budget allows) but never fewer than everything
   from the last user message onward
3. After the walk, enforce `cut_idx <= user_anchor` so the user message
   is never summarized away regardless of token pressure

**Scenario fixed** (reproduced from real user report):
- User sends "analyze commits after 10b0633"
- Assistant runs 3 large tool calls (git log, git diff, notion read)
- Each tool output is ~40KB → tail token budget exhausted
- OLD behavior: min_tail=3 protects only [tool, assistant, tool] —
  user message gets summarized → model responds about stale context
- NEW behavior: user-message anchor forces tail to start at or before
  the user message → model sees the actual request

This complements the v0.9.0 SUMMARY_PREFIX rewrite (NousResearch#8107) which addressed
the model interpreting summaries as active instructions. Together, they
fix the two root causes of post-compaction incoherent responses (NousResearch#7133).

Refs: NousResearch#7133, NousResearch#8107

…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.

…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.

…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.

…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.

…tions (NousResearch#8107)

After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.

Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.

Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):

1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
   this summary — respond ONLY to the latest user message AFTER it'

2. Summarizer preamble (shared by both prompts) adds:
   - 'Do NOT respond to any questions' (from OpenCode's approach)
   - 'Different assistant' framing (from Codex) to create psychological
     distance between summary content and active conversation

3. New summary sections:
   - '## Resolved Questions' — tracks already-answered questions with
     their answers, preventing re-answering (from Claude Code's
     'Pending user asks' pattern)
   - '## Pending User Asks' — explicitly marks unanswered questions
   - '## Remaining Work' replaces '## Next Steps' — passive framing
     avoids reading as active instructions

4. merge-summary-into-tail path now inserts a clear separator:
   '--- END OF CONTEXT SUMMARY — respond to the message below ---'

5. Iterative update prompt now instructs: 'Move answered questions to
   Resolved Questions' to maintain the resolved/pending distinction
   across multiple compactions.