fix: enforce long-term memory settlement on completion

[kjore]

Closes #789

Summary

• add a one-time completion gate for normal tasks reaching 15 turns
• start long-term memory evaluation before exit when settlement has not alreadybots begun
• prevent duplicate settlement after an explicit start_long_term_update call
• preserve exemptions for autonomous flows and configurations where the tool is unavailable
• align the Chinese and English threshold wording to turn >= 15
• add regression tests for the threshold and one-time gate behavior

This forces memory evaluation, not unconditional memory modification. The evaluation may still conclude that nothing should be stored.

Behavior

• turn < 15: exit normally
• turn >= 15 with prior settlement: exit normally
• turn >= 15 without prior settlement: run settlement once, then exit
• rejected calls before turn 10 do not consume the completion gate

The gate is scoped to one task. A new user request creates a new handler and resets the gate.

Testing

• pytest tests/test_long_term_update_gate.py -q
  • 5 passed
• full Windows test suite
  • 278 passed, 2 skipped
  • the skipped tests require symbolic-link privileges unavailable to the current Windows account
• python -m py_compile ga.py agentmain.py
• git diff --check

[Copilot]

🟡 Changes recommended

There are a couple of correctness/robustness issues in the updated logging flow and the new tests (determinism depends on filesystem/CWD) that should be fixed before merging.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR enforces the “15+ turns must trigger start_long_term_update on completion” requirement at runtime by adding a one-time completion gate that starts long-term memory evaluation before exiting long-running tasks (while preserving exemptions like autonomous flows and tool-unavailable configurations).

Changes:

• Add a per-task, one-time completion gate in GenericAgentHandler.do_no_tool() that triggers long-term memory evaluation when turn >= 15 and settlement hasn’t started yet.
• Track settlement gating state via long_term_update_pending, and prevent duplicate settlement after explicit do_start_long_term_update().
• Align EN/CN tool schema wording to “15 or more turns”, and add regression tests for threshold + one-time gate behavior.

File summaries

File	Description
ga.py	Adds the turn >= 15 completion gate and refactors long-term update initiation to prevent duplicate settlement.
agentmain.py	Initializes the handler’s gate based on tool availability and autonomous-flow detection.
assets/tools_schema.json	Aligns English description wording to “15 or more turns”.
assets/tools_schema_cn.json	Aligns Chinese description wording to “达到15轮…”.
tests/test_long_term_update_gate.py	Adds regression coverage for threshold behavior and one-time gating.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

The auto-gate path inlines the long-term-update “tool result” into next_prompt instead of passing it via the normal tool_results channel, which can change prompt structure/behavior across backends and should be corrected before merging.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 1
• Review effort level: Lite