Summary

Keep a known model context length when Hermes only guessed the next context probe tier and that guessed tier is already below the prompt being recovered.

This is a narrower alternative to #14499. It follows the same recovery shape as #14743: if the provider error does not contain a trustworthy parseable context limit, do not mutate context_length to an untrusted lower value. Compress with the known window instead, or fail closed if compression cannot reduce the session.

User impact

This is a high-frequency Codex/GPT-5.4 usability issue for my setup. I hit this at least five times per day, and it significantly disrupts normal Hermes usage. Local session history shows the issue was reported repeatedly on 2026-04-23 and 2026-04-24, including after hermes update.

The observed failure mode from the attached evidence in #14499:

Provider: openai-codex Model: gpt-5.4
Endpoint: https://chatgpt.com/backend-api/codex
Context: 492 msgs, ~285,378 tokens
Context length exceeded - stepping down: 1,050,000 -> 128,000 tokens
gpt-5.4 270K/128K

Once Hermes believes the context is 128K, an otherwise recoverable long-session overflow turns into repeated compression/failure against an artificially tiny window.

What changed

• Added _should_keep_context_length_on_untrusted_probe(...).
• When parse_context_limit_from_error(...) returns no concrete limit and get_next_probe_tier(...) would shrink below the current prompt estimate, keep old_ctx.
• Preserve existing behavior when:
  • provider returned a concrete parseable limit;
  • guessed probe tier is still above current prompt estimate;
  • normal non-Codex unknown-provider probe-down applies.
• Added pure helper tests and a run_conversation regression test modeled after fix(agent): preserve MiniMax context length on delta-only overflow (salvage #9170) #14743.

Related work

• fix(agent): preserve MiniMax context length on delta-only overflow (salvage #9170) #14743: merged MiniMax fix that preserves known context length when provider overflow text is not a real context limit.
• [codex] Prevent long-context probe collapse #14499: broader previous attempt that changed global probe tiers. This PR is intentionally narrower.
• gpt-5.4 shows 32k context in Hermes instead of 1,050,000 #5173 / Fix gpt-5.4 context length resolution for Codex #5174 / fix(model_metadata): add gpt-5.x context lengths + guard against poisoned cache #5179: related GPT-5.4 context resolution/cache issues, but do not cover this untrusted probe shrink path.

Validation

./venv/bin/python -m pytest tests/test_ctx_halving_fix.py tests/run_agent/test_run_agent.py::TestRunConversation::test_untrusted_probe_below_prompt_keeps_known_context_length tests/run_agent/test_run_agent.py::TestRunConversation::test_minimax_delta_overflow_keeps_known_context_length tests/run_agent/test_run_agent.py::TestRunConversation::test_non_minimax_delta_overflow_still_probes_down -q
30 passed in 7.37s

./venv/bin/python -m py_compile run_agent.py
git diff --check