fix(agent): don't re-pin sessions to provider-default base_url on credential swap

[AIalliAI]

Fixes #44070

Root cause

Two compounding issues, both in the credential-pool recovery path:

1. _swap_credential discards the configured base_url override.

Pool entries seeded from env keys store only the provider's registry default endpoint (base_url = env_url or pconfig.inference_base_url in agent/credential_pool.py). Session resolution layers the user's model.base_url / <PROVIDER>_BASE_URL override on top (pool_url_is_default in hermes_cli/runtime_provider.py:371-382) — but that override only exists at resolution time; the pool entry on disk keeps the default URL.

AIAgent._swap_credential (run_agent.py) adopted the raw entry URL. So the first recoverable error (401/403/429) that routed through _recover_with_credential_pool silently re-pointed the session at the provider default — permanently, because the swap mutates _client_kwargs (which every per-request client is rebuilt from) and restore_primary_runtime is gated on _fallback_activated, which credential rotation doesn't set.

For the reporter's Xiaomi MiMo token-plan setup the sequence is:

1. Turn 1 resolves https://token-plan-cn.xiaomimimo.com/v1 (config override over the pool entry's default) → works, including tool calls.
2. Turn 2 replays the full history + tool schemas → bigger payload → trips the plan's rate limit (or an auth blip) → recover_with_credential_pool → try_refresh_current() / mark_exhausted_and_rotate() → _swap_credential(entry) → base_url flips to https://api.xiaomimimo.com/v1 (the entry's stored default).
3. The token-plan key can't route on the pay-per-use host → the Go backend answers 404 page not found → all 3 retries hit it → API call failed after 3 retries: HTTP 404: 404 page not found.
4. Every later turn reuses the mutated _client_kwargs → the session never recovers.

This also explains "works intermittently via CLI" — it works until the first 429/401 of a session, then that session is wedged.

(I probed both hosts: …/v1/chat/completions exists on both; wrong-path requests return openresty HTML 404s and bad keys return JSON 401s, so the plain Go-style 404 page not found body the reporter saw is consistent with an authenticated request the backend can't route — i.e. a plan key on the wrong host — not with a malformed path.)

2. switch_model() keeps the old provider's credential pool attached.

The fallback path got this exact fix in #33163 ("leaving it attached means downstream recovery calls _swap_credential with a primary entry which overwrites the agent's base_url back to the primary's endpoint — every fallback request then 404s against the wrong host"), but switch_model() never did. After switching to e.g. DeepSeek mid-conversation, the first recoverable error could swap the agent back onto the old provider's endpoint and credentials — matching the reported "switching models mid-conversation has no effect; the error persists".

Fix

• _swap_credential now resolves the URL via a new _pool_entry_swap_base_url helper: when the pool entry carries only the provider's registry-default endpoint and the session is running on a different (configured) URL, the session's URL is preserved; the key still rotates. Entries with genuinely per-credential endpoints (kimi/zai region resolution, custom pools seeded from config, Nous inference_base_url) keep winning, so multi-account pools with distinct endpoints are unaffected.
• switch_model() detaches a credential pool whose provider differs from the switch target, mirroring Fallback to OpenRouter retains primary's base_url — requests go to ChatGPT Codex with openrouter/auto model → HTTP 404 #33163 and the defensive guard in recover_with_credential_pool (Fallback provider 429 can exhaust primary provider credential pool #33088).

Tests

tests/run_agent/test_44070_credential_swap_base_url.py — 8 new tests covering: override preserved on default-URL entries, per-credential endpoints still adopted, missing-URL entries, no-override pass-through, unknown/custom providers, end-to-end _swap_credential, and cross-provider vs same-provider pool detach on switch_model.

Regression suites pass: test_fallback_credential_isolation.py, test_credential_pool_interrupt.py, test_codex_xai_oauth_recovery.py, test_credential_pool_routing.py, test_switch_model_*.py, test_primary_runtime_restore.py, test_run_agent.py (482 tests).

[liuhao1024]

Verification: reviewed diff — clean fix for credential-swap URL contamination.

The new _pool_entry_swap_base_url() method correctly solves the problem: pool entries seeded from env keys store only the provider's registry default endpoint, while session resolution layers the user's model.base_url override on top. Without this fix, a 401/429-triggered credential recovery adopts the raw entry URL and permanently pins the session to the wrong host.

Checked:

• _pool_entry_swap_base_url compares entry.base_url against PROVIDER_REGISTRY[inference_base_url] — when they match (registry default), the agent's current URL is preserved. Entries with genuinely per-credential endpoints (region resolution, custom pools) still win.
• switch_model now detaches the credential pool when switching to a different provider (mirrors the fallback-path fix from Fallback to OpenRouter retains primary's base_url — requests go to ChatGPT Codex with openrouter/auto model → HTTP 404 #33163). Same-provider switches keep the pool.
• Test coverage is thorough: registry-default entry preserves override, per-credential endpoint wins, missing URL keeps current, agent-on-default passes through, unknown provider adopts entry URL, end-to-end _swap_credential preserves override, cross-provider switch detaches pool, same-provider switch keeps pool.

No issues found.

[AIalliAI]

Requesting maintainer review — this is ready to land from my side. Standalone fork CI is pending first-run approval here; the rollup branch in #44061 carrying this session's batch is fully green on upstream CI (all test shards, typecheck, e2e).

[teknium1]

Thanks for tracing the resolved-runtime versus pooled-entry endpoint mismatch. The base-URL preservation portion remains relevant on current main: run_agent.py:4518-4543 still adopts and persists the raw pooled URL, while hermes_cli/runtime_provider.py:481-491 can replace a registry-default pool URL with configured model.base_url.

Problems

• The switch_model() pool-detach change is stale. Current main reloads the target provider pool at agent/agent_runtime_helpers.py:1894-1900; the added cross-provider test expecting None conflicts with the existing target-pool recovery contract (tests/run_agent/test_switch_model_pool_reload_52727.py:84-111).
• The new tests use a bare agent and mocks, so they do not exercise the runtime-resolution chain that produces the mismatch.

Suggested changes

• Salvage the _pool_entry_swap_base_url() / _swap_credential() change.
• Replace the detach expectation with the current target-pool reload behavior, and add a temp-HERMES_HOME integration case covering configured Xiaomi model.base_url plus rotation of a default-URL pool entry.

Automated hermes-sweeper review.

[teknium1]

Current main now reloads load_pool(new_provider) during a cross-provider switch (agent/agent_runtime_helpers.py:1894-1900) so the new provider retains 401/429 recovery. Please reconcile this detach-only behavior and its None expectation with the target-pool reload contract rather than replacing it.