fix(agent): persist complete model runtime state on /model switch

[zccyman]

Summary

Both gateway /model and CLI /model --global now persist the complete runtime tuple to config.yaml:

model.default
model.provider
model.base_url      # NEW: was missing or conditionally written
model.api_mode      # NEW: was never persisted

Before

Gateway /model (gateway/run.py):

model_cfg["default"] = result.new_model
model_cfg["provider"] = result.target_provider
if result.base_url:                    # ← only writes when truthy
    model_cfg["base_url"] = result.base_url
# api_mode: never persisted

CLI /model --global (cli.py, 2 locations):

save_config_value("model.default", result.new_model)
if result.provider_changed:
    save_config_value("model.provider", result.target_provider)
# base_url: never persisted
# api_mode: never persisted

After

Gateway — unconditional with empty-string clearing:

model_cfg["base_url"] = result.base_url or ""
model_cfg["api_mode"] = result.api_mode or ""

CLI — two additional save_config_value calls:

save_config_value("model.base_url", result.base_url or "")
save_config_value("model.api_mode", result.api_mode or "")

Impact

Fixes the class of bugs where switching models leaves stale endpoint/protocol state:

Scenario	Before	After
Switch from custom → built-in provider	Old base_url survives	Cleared to ""
Switch from anthropic_messages → openai	Old api_mode survives	Cleared to ""
Switch to a provider with explicit base_url	Not persisted (CLI)	Persisted correctly
Restart gateway after /model switch	May use stale endpoint	Uses correct state

Testing

• 5 new tests covering all persistence paths
• 29 existing model-switch tests passing (0 regressions)

Files changed

File	Change
gateway/run.py	Persist base_url + api_mode unconditionally
cli.py	Add save_config_value for base_url + api_mode (2 locations)
tests/gateway/test_model_switch_config_persistence.py	New: 5 tests

Closes #25107
Closes #25106
See also #25105 (kimi-for-coding alias normalization — separate issue)

[teknium1]

Thanks for identifying the remaining CLI persistence gap. The CLI premise is still present on current main: cli.py:7945-7948 and cli.py:8257-8260 persist only model.default and conditionally model.provider.

Problems

• The gateway hunk targets a path that has since moved. Live picker and typed persistence are now in gateway/slash_commands.py:1666-1673 and gateway/slash_commands.py:1916-1923; the PR is currently conflicting.
• The proposed unconditional gateway writes conflict with the later endpoint-clearing contract. Both live paths call clear_model_endpoint_credentials(..., clear_base_url=True), which removes api_mode and base_url (hermes_cli/config.py:4566-4569), added by f6deabca0d8010d485d4d3a8744d0ceb797eacb9.
• tests/gateway/test_model_switch_config_persistence.py simulates the intended assignments instead of exercising a production handler, so it cannot regress the actual CLI or gateway paths.

Suggested changes

• Salvage the CLI fix at both live branches, and port gateway behavior to both current gateway/slash_commands.py sites while preserving non-custom clearing semantics.
• Replace simulation tests with temp-HERMES_HOME tests driving both CLI paths and the gateway typed/picker paths; tests/gateway/test_model_picker_persist.py:15-18 is the existing callback-driven pattern.

This is an automated hermes-sweeper review.

[teknium1]

Thanks for identifying this persistence-gap class early — you were 2 months ahead of the fixes that eventually landed. Closing as superseded on current main, in three pieces:

• CLI base_url/api_mode sync on global switch → 398634e ([Bug]: CLI --global model switch does not persist/clear model.base_url and model.api_mode #25106 fix by @joaomarcos, Jul 8), which also clears stale values
• Gateway persistence → the handler moved to gateway/slash_commands.py, which now writes the full tuple (default/provider/base_url/api_mode) with the endpoint-credential-clearing contract from f6deabc
• Unconditional model.provider write on CLI --global → fix(cli): persist provider on global model switch #69818 (merged today)

Nothing from this branch remains unlanded, and the gateway hunk targets a path that has since moved (per the earlier sweeper review). Appreciate the report and the diagnosis — the bug class was real and is now fully covered with regression tests on all paths.