fix: dedup TTL expiry, compression provider fallback, TCP keepalive CLOSE-WAIT

[d3momonkey]

Summary

Fixes three bugs identified in issues #10306, #10314, and #10324.

---

Fix 1 — #10306: MessageDeduplicator TTL never expires on low-traffic instances

Files: gateway/platforms/wecom.py, gateway/platforms/dingtalk.py

The _is_duplicate() method only pruned expired entries when the cache exceeded DEDUP_MAX_SIZE (1000). On low-traffic instances the cache stays small, so message IDs accumulated forever and were permanently treated as duplicates.

Fix: Check TTL on every lookup (not just size-overflow). The prune condition now fires when msg_id in self._seen_messages OR when over max size, so expired entries are evicted before the membership test.

---

Fix 2 — #10314: Context compression falls back to OpenRouter when summary_provider=auto

File: agent/auxiliary_client.py

When compression.summary_provider=auto (default) and summary_model is empty (default), _resolve_task_provider_model() returned ("auto", None, ...) which fed into _resolve_auto(). That chain picks OpenRouter first and uses google/gemini-3-flash-preview — a model that 404s on non-OpenRouter providers (DashScope, custom endpoints, etc.). This caused a 600s cooldown loop, making the agent appear frozen.

Fix: Added _read_main_provider() and wired it into _resolve_task_provider_model(): when task is compression and the result is still auto with no model, read the main configured provider/model directly and return it, skipping the OpenRouter fallback chain.

---

Fix 3 — #10324: Agent hangs indefinitely on CLOSE-WAIT when provider drops connection

File: run_agent.py

When a custom provider (e.g. LiteLLM proxy) drops a connection mid-stream, the socket enters TCP CLOSE-WAIT. Because httpx uses epoll_wait for async I/O and a CLOSE-WAIT socket with no buffered data does not become readable, the 900s timeout never starts and the process hangs forever requiring kill -9.

Fix: Add TCP keepalive socket options to the httpx transport in _create_openai_client():

• SO_KEEPALIVE=1 — enable keepalives
• TCP_KEEPIDLE=30 — start probes after 30s idle
• TCP_KEEPINTVL=10 — probe every 10s
• TCP_KEEPCNT=3 — give up after 3 failures

Worst-case detection window: ~60s instead of infinite. try/except fallback handles macOS/Windows where TCP_KEEPIDLE may be unavailable.

---

Testing

• tests/gateway/ — all passing (pre-existing failures unchanged)
• tests/tools/ — all passing (pre-existing failures unchanged)
• No new test failures introduced

[d3momonkey]

Update: Added two more fixes to this branch:

• [Bug]: Gateway fails to start with --config flag: json.load() used instead of yaml.safe_load() #10216 — Gateway --config flag now correctly parses YAML files. Was using json.load() on a YAML config, instantly crashing with JSONDecodeError.
• When Hermes calls the MCP tool, if the returned result is in Chinese, it will consume three times as many tokens. This is because the ensure_ascii parameter in json.dumps within mcp_tool.py is set to True. #10234 — MCP tool now passes ensure_ascii=False to all 19 json.dumps() calls. Previously Chinese/CJK characters were escaped to \uXXXX sequences, inflating token costs ~3-4x for non-ASCII content.

Also investigated #10174 (on_memory_write bridge missing in sequential path) — the bridge described in the ticket does not exist anywhere in the codebase (neither in _invoke_tool nor anywhere else). Both paths are already symmetric. Closing that one as a false report.

[d3momonkey]

Update 2: Four more fixes added to the branch:

• [Bug]: hermes update --check not recognized — documented but not implemented #10318 — hermes update --check now works. Runs git fetch + git rev-list to count commits behind, prints status without pulling.
• Bug: _load_skill_payload rebuilds wrong skill_dir for external skills (from skills.external_dirs) #10313 — External skill paths fixed. skill_view() now returns absolute skill_dir in JSON; _load_skill_payload() uses it directly instead of blindly prepending SKILLS_DIR.
• fix(cli): load_cli_config() overwrites gateway's TERMINAL_CWD from MESSAGING_CWD #10225 — load_cli_config() no longer clobbers TERMINAL_CWD set by the gateway from MESSAGING_CWD. Existing value is preserved when terminal.cwd is '.'" or 'auto'`.
• Bug: auxiliary_client._client_cache accumulates AsyncOpenAI clients indefinitely, causing fd exhaustion in long-running gateway #10200 — _client_cache now has a hard cap of 32 entries with LRU eviction. Oldest async clients are force-closed when the limit is hit, preventing fd exhaustion in long-running gateways.

Running total: 9 bugs fixed across 9 files in this PR.

[alt-glitch]

Fix 3 (TCP keepalive CLOSE-WAIT) overlaps with already-merged #10933 and #11277 which addressed #10324. Fix 1 (dedup TTL) overlaps with closed #10403. Verify which sub-fixes are still needed.

[teknium1]

Thanks for the thorough multi-fix PR, @d3momonkey! After an automated hermes-sweeper review against current main, all 9 fixes bundled here have already been independently merged.

Evidence by fix:

• [Bug]: MessageDeduplicator TTL never expires unless cache exceeds max_size #10306 (dedup TTL) — MessageDeduplicator.is_duplicate() in gateway/platforms/helpers.py:46 already checks TTL on every lookup. Merged in PR fix: enforce TTL in MessageDeduplicator + use yaml for gateway --config #10509 (sha 2edbf1556).
• [Bug]: Context compression falls back to OpenRouter default model instead of main provider when summary_provider=auto #10314 (compression fallback) — _read_main_provider() exists and is wired into _resolve_task_provider_model() in agent/auxiliary_client.py:1038. Issue closed 2026-04-16.
• Agent hangs indefinitely on CLOSE-WAIT when custom provider drops connection mid-response #10324 (TCP keepalive CLOSE-WAIT) — SO_KEEPALIVE/TCP_KEEPIDLE/TCP_KEEPINTVL/TCP_KEEPCNT present in run_agent.py. Merged via PR fix: enable TCP keepalives to detect dead provider connections #10933 (sha 12b109b66) and re-merged after revert in fix: enable TCP keepalives to detect dead provider connections (#10324) #11277 (sha 8c478983e). This overlap was also flagged by @alt-glitch.
• [Bug]: Gateway fails to start with --config flag: json.load() used instead of yaml.safe_load() #10216 (gateway --config yaml) — gateway/run.py:111 uses yaml.safe_load(). Merged in PR fix: enforce TTL in MessageDeduplicator + use yaml for gateway --config #10509.
• When Hermes calls the MCP tool, if the returned result is in Chinese, it will consume three times as many tokens. This is because the ensure_ascii parameter in json.dumps within mcp_tool.py is set to True. #10234 (MCP ensure_ascii) — All 19 json.dumps() calls in tools/mcp_tool.py already pass ensure_ascii=False. Merged in PR fix: add ensure_ascii=False to all MCP json.dumps calls #10512 (sha 861efe274).
• [Bug]: hermes update --check not recognized — documented but not implemented #10318 (update --check) — _cmd_update_check() implemented at hermes_cli/main.py:5880. Merged in sha dc5e02ea7.
• Bug: _load_skill_payload rebuilds wrong skill_dir for external skills (from skills.external_dirs) #10313 (external skill paths) — tools/skills_tool.py:1338 returns absolute skill_dir; agent/skill_commands.py:61-66 uses it directly. Merged in PR fix: use absolute skill_dir for external skills #10587 (sha 4fdcae6c9).
• fix(cli): load_cli_config() overwrites gateway's TERMINAL_CWD from MESSAGING_CWD #10225 (TERMINAL_CWD clobber) — Guard at cli.py:520-531 preserves an existing gateway-set TERMINAL_CWD. A broader CWD refactor also merged in sha 3c42064ef.
• Bug: auxiliary_client._client_cache accumulates AsyncOpenAI clients indefinitely, causing fd exhaustion in long-running gateway #10200 (client cache fd leak) — _CLIENT_CACHE_MAX_SIZE = 64 with LRU eviction at agent/auxiliary_client.py:2355. Merged in PR fix: bound auxiliary client cache to prevent fd exhaustion in long-running gateways #10470 (sha 6391b4677).

This is an automated hermes-sweeper review. Closing as implemented on main.