Various bugs and fixes I discovered

[lucid trout]

📋 Here's a number of fixes I had to implement in order to work around real issues I ran into. I am providing a relatively slim list (sorted by severity) in the hopes that you may find it useful - those were all genuine problems that caused real issues in my workflow and needed fixing, not just personal flavour. Some of it has already been fixed and removed from the list, about status of the rest I'm unsure. If you require more context to any of the points, let me know and I'll happily provide.

Gateway cached-agent invalidation fix (1/2)
Adds context/compression settings to the gateway agent cache signature, so changing model.context_length or compression config rebuilds the cached agent instead of keeping stale state.
File: gateway/run.py

Gateway hygiene compression hardening
Makes the hard message limit configurable via compression.hygiene_hard_message_limit.
Fixes duplicate model= kwarg when hygiene compression spins up a compression agent.
Files: gateway/run.py, tests/gateway/test_session_hygiene.py

Background process notifications no longer create fake agent turns
notify_on_complete and watch_patterns now send ephemeral chat notifications instead of injecting synthetic messages into the agent transcript.
This prevents transcript bloat and weird pairing/authorization side effects.
Files: gateway/run.py, tools/terminal_tool.py, tools/process_registry.py, tests.

Checkpoint behavior change
Only one filesystem checkpoint per turn total, even across multiple directories.
Files: tools/checkpoint_manager.py, test.

Manual GPT-5.5 + 272k context cap
Adds gpt-5.5 to Codex model list.
Forces GPT-5.4 / GPT-5.5 family context to 272,000, overriding larger provider metadata.
Files: agent/model_metadata.py, hermes_cli/codex_models.py

Codex auxiliary reasoning support
Makes auxiliary Codex calls translate extra_body.reasoning into Responses API top-level reasoning.
Maps minimal → low.
Adds reasoning.encrypted_content include.
Files: agent/auxiliary_client.py, tests.

Codex auth fallback hardening
If _read_codex_tokens() fails, reads auth.json directly for the OpenAI Codex OAuth access token.
File: agent/auxiliary_client.py

[heady vale]

Thanks, this is useful. I checked these against current main.

The GPT-5.5 / Codex 272k context-cap piece is already covered now.

The watch-pattern spam side and checkpoint duplicate-entry cleanup are partially covered, but not the exact behavior you described. Current main still injects background completion/watch notifications as synthetic internal message events, and checkpoints are still deduped per directory per turn, not one checkpoint total per turn.

The gateway cached-agent signature still does not appear to include the context/compression settings you called out. The hygiene hard message limit is still hardcoded, not configurable. I also do not see the auxiliary Codex reasoning translation or the direct Codex auth fallback landed in the current code path.

So yes, several of these still look like valid fixes. If you already have local patches, the most useful next step is to open PRs for the remaining items separately so they can be reviewed without getting tangled together. I’m going to keep this thread open as captured engineering feedback rather than close it as resolved.