fix(agent): scope model.context_length override to config default model

[TheTom]

Closes #62152

Problem

The explicit model.context_length in config.yaml is written for the config's default model, but init_agent applied it to every agent build unconditionally. A new session created with a per-session model override (desktop/mobile picker choosing a different provider/model) inherited the override and reported the wrong context window: e.g. a Codex gpt-5.6 session (real window 372k) showing a local default model's 131072, with the compressor threshold firing accordingly.

The in-session /model switch path already guards against exactly this by clearing _config_context_length on swap (agent_runtime_helpers.py). The build path had no equivalent, so switching mid-session gave the correct window while picking the same model for a new chat did not.

Fix

New pure helper _scope_config_context_length_to_default_model in agent/agent_init.py, wired in right after the existing int-parse of the override:

• reads the config default the same way hermes_cli/web_server.py::get_model_info does (default, falling back to name)
• if the config names a default model and agent.model differs, returns None so get_model_context_length() auto-detects the real window for the actual model
• conservative otherwise: non-dict model config, no named default, or empty agent model all keep current behavior

Per-model context_length entries under custom_providers are untouched and still resolve after the scoped override falls through to None (that block already keys on model + base_url).

Tests

tests/run_agent/test_config_context_length_model_scoping.py (new, 11 tests):

• integration (real AIAgent, mirroring test_invalid_context_length_warning.py): override applies when the session model matches the config default; override does not leak onto a session with a different model (auto-detection wins); plain-string model config unchanged
• unit coverage of the helper: match/mismatch, name fallback, whitespace, missing default, missing agent model, non-dict config

tests/run_agent/test_invalid_context_length_warning.py: two tests incidentally built the agent with a model that mismatched the config default while asserting the override applied; they now pass the matching model explicitly, with a comment noting the scoping constraint.

python3 -m pytest tests/run_agent/test_config_context_length_model_scoping.py tests/run_agent/test_invalid_context_length_warning.py passes 16/16 on this branch. ruff check clean on the touched files.

[alt-glitch]

This was generated by AI during triage.

Competing-PR cluster for #62152 (model.context_length override leaking onto a session running a different model). This PR and #62124 both scope the override in agent/agent_init.py via the same core mechanism (drop the override when the config's default model differs from the active model), but frame different triggers: this PR targets the per-session model-override / picker leak; #62124 targets the CLI --model/skill/fallback mismatch and additionally adds a provider-prefix normalizer, a visible mismatch warning, and a companion anti-thrashing guard (#62125). #62124 is the earlier and broader entrant. Flagging so a maintainer can pick the canonical one rather than merging both.

[teknium1]

Thanks for isolating the initialization leak; current main still reads the override unconditionally at agent/agent_init.py:1632-1639, so the underlying bug is real.

Problems

• agent/agent_init.py:280 compares raw identifiers, but current main normalizes agent.model for non-aggregator providers at agent/agent_init.py:464-471. A valid configured default such as zai/glm-4.6 can therefore compare unequal to normalized glm-4.6 and lose its intended override.
• The gateway hygiene path remains unscoped: gateway/run.py:11050-11056 reads the config override before resolving session runtime at gateway/run.py:11079-11087, then passes it to context resolution at gateway/run.py:11093-11099. This can still size pre-agent compression for an overridden session with the default model's window.

Suggested changes

• Normalize the configured identifier with the same provider-aware normalization used for agent.model, with a prefixed-model regression test.
• Scope the gateway hygiene override after session-runtime resolution and cover that path with a regression test.

Automated hermes-sweeper review.

[teknium1]

agent.model has already been normalized for non-aggregator providers in the init path, while _default_model remains raw config text. Normalize the config value with the same provider-aware normalizer before comparing; otherwise a valid value such as zai/glm-4.6 mismatches normalized glm-4.6 and drops the intended override.

[TheTom]

Both findings addressed, thanks for the precise pointers.

1. Provider-prefix normalization: _scope_config_context_length_to_default_model now takes the provider and runs the configured default through the same normalize_model_for_provider path used on agent.model (same aggregator-provider exclusion), so default: zai/glm-4.6 matches the normalized glm-4.6 and keeps its override. Regression tests cover the prefixed-default keep case, the genuinely-different-model drop case, and the aggregator-provider passthrough.

2. Gateway hygiene path: gateway/run.py now re-scopes _hyg_config_context_length after _resolve_session_agent_runtime, using the resolved session model and provider, before context resolution consumes it. New tests/gateway/test_session_hygiene.py covers both directions; the leak test fails without the fix.

Also rebased onto current main. On the competing-PR note from triage: #62124 covers the CLI/skill/fallback trigger in agent_init only; this PR now additionally covers the gateway pre-agent hygiene sizing, which neither PR previously did. Happy to converge with #62124 in whichever direction maintainers prefer.

[kshitijk4poor]

Closing as superseded — the core mechanism here (scope a global model.context_length to the configured default model/route, drop it when the session's actual model or route differs) landed on main on July 21–22 via #69114 (fix(compression): prevent stale-budget retry loops, commit 377244f) and commit 63dd651 (fix(providers): scope route-owned runtime settings):

• agent/agent_init.py now drops the pin on _model_mismatch or _route_mismatch (normalized model comparison + route-URL comparison, including custom_providers base_url resolution)
• gateway/run.py's message path applies the same scoping (_msg_model != _msg_configured_model check + should_clear_context_pin_async route check)

This PR was opened before that landed and correctly identified the bug (#62152) and the fix direction — thank you for the detailed analysis and tests; the normalization-aware default-model comparison you implemented is exactly the shape that ended up on main.