fix(context): thread custom_providers to all context-length resolution call sites

[TurgutKural]

Problem

custom_providers[].models.<id>.context_length overrides were only honored at agent startup (agent_init) and /model switch (model_switch). Several sibling call paths called get_model_context_length() without passing custom_providers, causing them to fall through to endpoint probing and the 256K/131K hardcoded defaults — even when the user had an explicit per-model override configured.

Observed symptom (log spam on every session start with a custom provider model):

INFO agent.model_metadata: Could not detect context length for model 'qwen3.8-max-preview' at https://… — defaulting to 256,000 tokens (probe-down).
INFO agent.model_metadata: Using hardcoded context length 131,072 for model 'qwen3.8-max-preview' (custom endpoint, catalog match on 'qwen')

This fires twice per startup because both the built-in compressor and the context engine plugin resolve context length independently.

Affected call sites

Call site	File	Impact
ContextCompressor._resolve_context_length	agent/context_compressor.py	Deferred first-access probe ignores per-model override
_candidate_context_window	agent/auxiliary_client.py	Fallback chain screening uses wrong context window
_trim_messages_for_reference	agent/moa_loop.py	MoA reference trimming uses wrong window
get_model_info	hermes_cli/web_server.py	WebUI shows wrong auto-detected context

Fix

Thread custom_providers through each path:

• ContextCompressor: new custom_providers constructor parameter, passed from agent_init (where _custom_providers is already resolved). Used in _resolve_context_length().
• auxiliary_client: _candidate_context_window lazily loads custom_providers from config via load_config_readonly() + get_compatible_custom_providers() (matching the existing lazy-config pattern in the file).
• moa_loop: new _load_custom_providers() helper with the same lazy pattern, called from _trim_messages_for_reference.
• web_server: get_model_info extracts custom_providers from the config it already loads.

All changes are backward-compatible: custom_providers defaults to None and every config load is wrapped in try/except so failures fall through to the existing probe chain.

Tests

20 new tests in tests/agent/test_custom_provider_context_threading.py:

• ContextCompressor (5): override honored, fallthrough without override, config_context_length precedence, instance storage, default None
• auxiliary_client (3): override honored, graceful config failure, empty model
• moa_loop (3): helper returns list, helper returns None on failure, trim respects 1M context
• web_server (1): custom_providers kwarg passed to resolver
• get_custom_provider_context_length (8): extended edge-case coverage (list-format models, zero/negative/string values, first-match-wins, non-dict model cfg)

All 90 existing related tests pass without modification:

• tests/hermes_cli/test_custom_provider_context_length.py (6)
• tests/agent/test_model_metadata.py (42)
• tests/agent/test_moa_context_max_tokens.py (2)
• tests/run_agent/test_switch_model_context.py (6)
• tests/run_agent/test_invalid_context_length_warning.py (3)
• tests/hermes_cli/test_model_switch_context_display.py (3)

Closes the sibling-path gap left by #15779.

[teknium1]

Thanks for tracing the missing propagation paths. The core premise is verified on current main: agent/model_metadata.py:2343-2350 only applies a per-model custom-provider override when both the route and custom_providers are supplied; the current auxiliary, MoA, and WebUI callers omit that list.

Problems

• The claimed all-call-sites coverage is incomplete. model_tools.py:621-641 resolves an active custom endpoint, then calls get_model_context_length() without custom_providers; the tool-search context gate therefore retains the same fallback/default behavior.
• The inactive-agent /context fallback at gateway/slash_commands.py:785-791 passes only model_name, with no route identity or compatible custom-provider list, so it cannot satisfy the resolver's override preconditions.

Suggested changes

• Extend the shared propagation to model_tools._resolve_active_context_length() and add a behavioral test using a temporary configured custom provider.
• Either resolve the configured route/providers for the no-resident-agent /context fallback or reuse a helper that already does so, then test that branch.

This is an automated hermes-sweeper review.

[TurgutKural]

Thanks — both suggested changes are now implemented in e043661470 (on top of today's rebase onto aec331899e):

1. model_tools._resolve_active_context_length() — now loads the compatible custom_providers list from config and passes it to get_model_context_length(), so the tool-search context gate honors per-model context_length overrides instead of falling back to generic metadata for custom endpoints.

2. Inactive-agent /context fallback (gateway/slash_commands.py) — the fallback now resolves the configured model.provider route identity and the compatible custom_providers list before calling the resolver, instead of passing only model_name. This satisfies the resolver's override preconditions on the no-resident-agent path.

Regression coverage added in tests/agent/test_custom_provider_context_threading.py:

• behavioral capture test asserting get_model_context_length() receives custom_providers + route provider from the tool-search gate (and that the gate degrades to 0 on config-load failure),
• behavioral capture test for the /context fallback (custom_providers + provider + model kwargs), plus a fail-open test when config loading raises.

Local validation on the rebased worktree: tests/agent/test_custom_provider_context_threading.py (24 passed) and tests/test_model_tools.py (24 passed); py_compile clean on all three touched files.

[andrexibiza]

Coordination from the web_server.py god-file decomposition (epic #78647): the custom-endpoints family has been extracted into web_routers/custom_endpoints.py in #79127 (slice R3-C1). Verified LOW overlap with this PR — adjacent regions, disjoint code. Both compose; no conflict expected. Interlocked: #79127 and the full slice wave (#79123-#79129).

[TurgutKural]

Rebased onto current upstream/main (395c70d). The three red slices on the previous head (2/12, 6/12: test_session_api.py; 12/12: test_goal_continuation_drain.py + test_goal_max_turns_config.py + test_transcription_tools.py) were the Aug-18 gateway timing races and a transcription idle-timeout flake, de-flaked upstream in a75d1b5 / c70e152 / 5633764 — none reproduce on the rebased head; all 12 test slices and all required checks are green.

The two sweeper review items remain implemented from the previous cycle: model_tools._resolve_active_context_length() passes the configured custom_providers list to get_model_context_length(), and the inactive-agent /context fallback in gateway/slash_commands.py resolves route identity + compatible providers before calling the resolver. Regression coverage in tests/agent/test_custom_provider_context_threading.py.