Verification Result

After restarting hermes-gateway, I re-tested the LCM/context compression runtime behavior for the active gpt-5.4 session.

Expected

• context_length = 524288
• threshold_tokens = 393216

Runtime Check

Using lcm_status(), the active session reports:

• context_length: 524288
• threshold_tokens: 393216

This confirms the runtime values are applied correctly, not just present in config.

Real Compression Test

I then executed a direct ContextCompressor.compress() test in the Hermes project virtualenv with:

• model: gpt-5.4
• provider: yunfeiplus
• threshold_percent = 0.75
• config_context_length = 524288
• current_tokens = threshold_tokens + 5000

Observed result:

• context_length=524288
• threshold_tokens=393216
• input_messages=39
• output_messages=33
• summary_count=1
• summary_prefix_ok=True

A valid compaction summary was generated with the expected prefix:

• [CONTEXT COMPACTION — REFERENCE ONLY]

Conclusion

Verification passed.

Confirmed:

1. LCM context_length is now 524288
2. threshold_tokens is now 393216
3. Real compression is functioning correctly when the token count exceeds threshold
4. This is a runtime-confirmed result, not only a static config change

Overlaps with #12316 — both unify context-length resolution across runtime and /model display. Also related to #12380, #14382, and #8785.

Thanks for the careful work on this, @leavrcn! The context-length alignment bug you identified was independently fixed on main after this PR was filed.

This is an automated hermes-sweeper review.

• Commit 125de0205 (fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K — fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K #15844) landed on main ~April 25 and addresses the same display/runtime divergence:
  • New hermes_cli.config.get_custom_provider_context_length() helper — single source of truth for per-model overrides
  • resolve_display_context_length() in hermes_cli/model_switch.py gained a custom_providers= kwarg (line 530)
  • AIAgent.switch_model() in run_agent.py re-reads custom_providers from live config on every /model switch (line 2149)
  • Two gateway/run.py call sites wired up (lines 5529-5530, 5676-5677)
  • Regression tests added: tests/hermes_cli/test_model_switch_context_display.py and tests/hermes_cli/test_custom_provider_context_length.py

Your test_model_switch_context_alignment.py (384 lines) isn't on main yet — if it covers additional edge cases (negative values, bool context_length, provider-name collisions), maintainers may want to cherry-pick those tests separately.