What does this PR do?

Large openai-codex subscription requests can spend longer than the fixed TTFB cutoff in backend admission or prompt prefill before the first SSE event is emitted. Hermes treated that as a no-first-byte stream failure and reconnected, which caused long-context Codex turns to retry even though the backend could still complete successfully.

This updates the Codex Responses watchdog policy to distinguish three cases:

• small requests with no first SSE event still fail fast and reconnect
• large openai-codex requests wait for backend prefill instead of being killed by the first-byte watchdog
• streams that emit at least one SSE event and then go idle are handled by a separate event-idle watchdog

How this differs from recent Codex stream fixes

This complements the existing Codex stream fixes rather than replacing them:

• fix(agent): recover Codex Responses streams with null output #32963 recovered from response.completed.output = null crashes after stream events had already been collected.
• refactor(codex): drop SDK responses.stream() helper; consume events directly #33042 removed the SDK responses.stream() helper and consumes raw responses.create(stream=True) events directly, making that null-output parser failure structurally impossible.
• fix(codex): update silent-hang workaround hint + wire into TTFB watchdog #33133 improved the user-facing silent-hang workaround hint and wired that hint into the TTFB watchdog path.

Those fixes address stream consumption once events arrive, or improve the message shown when the TTFB watchdog kills a request. This PR addresses the separate pre-first-event case where hosted Codex accepts a large request but spends a long time in admission/prompt prefill before emitting the first SSE event.

Related Issue

Addresses the no-first-byte / large-context TTFB portion of #33075.
Related: #22986, #7069, #32963, #33042, #33133

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✅ Tests (adding or improving test coverage)

Changes Made

• Updated agent/chat_completion_helpers.py to detect the hosted openai-codex Responses backend in the streaming watchdog path.
• Scaled stale and idle stream watchdog thresholds by estimated request context size.
• Disabled strict no-first-byte TTFB reconnects for large hosted Codex requests by default, while preserving HERMES_CODEX_TTFB_STRICT=1 for strict behavior.
• Kept fast reconnect behavior for small no-byte stalls and capped stale TTFB env values for small hosted Codex requests.
• Added a post-first-event idle watchdog so real stream stalls are still retried.
• Expanded tests/agent/test_codex_ttfb_watchdog.py with regression coverage for small no-byte stalls, first-event-then-idle stalls, large-context prefill delay, and strict-mode behavior.

How to Test

1. scripts/run_tests.sh tests/agent/test_codex_ttfb_watchdog.py
2. scripts/run_tests.sh tests/agent/test_codex_ttfb_watchdog.py tests/agent/test_auxiliary_client.py tests/run_agent/test_run_agent_codex_responses.py
3. Manual verification on macOS: long openai-codex turns with large estimated input context no longer reconnect solely because the hosted backend takes longer than the small-request TTFB cutoff before the first SSE event.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run relevant tests and all tests pass
• I've added tests for my changes
• I've tested on my platform: macOS

Documentation & Housekeeping

• Documentation update N/A
• cli-config.yaml.example update N/A
• CONTRIBUTING.md or AGENTS.md update N/A
• Cross-platform impact considered: no file I/O, terminal, process management, or platform-specific behavior changed
• Tool descriptions/schemas update N/A

Screenshots / Logs

Before this fix, large hosted Codex requests could show repeated reconnects like:

No first byte from provider in 12s (codex stream, model: gpt-5.5). Reconnecting.

After this fix, large hosted Codex requests wait for backend prefill; small no-byte stalls and post-first-event idle stalls are still retried.