Add per-turn and per-image model overrides by deusyu · Pull Request #29923 · NousResearch/hermes-agent

deusyu · 2026-05-21T16:39:25Z

Summary

Add /model --once support across CLI, TUI, and gateway, restoring the previous model/runtime after one assistant turn.
Add one-call provider and model overrides to image_generate.
Preserve existing session/global /model behavior while adding cleanup for pending one-turn overrides.

Tests

scripts/run_tests.sh tests/hermes_cli/test_model_switch_once_flags.py tests/hermes_cli/test_cli_model_once.py tests/hermes_cli/test_apply_model_switch_result_context.py tests/hermes_cli/test_model_switch_custom_providers.py tests/tools/test_image_generation_plugin_dispatch.py tests/tools/test_image_generation.py tests/gateway/test_model_command_custom_providers.py tests/gateway/test_model_switch_persistence.py tests/gateway/test_session_model_override_routing.py tests/test_tui_gateway_server.py -- -q
uv run --no-sync python -m pytest -q tests/hermes_cli/test_model_switch_once_flags.py tests/hermes_cli/test_cli_model_once.py tests/hermes_cli/test_apply_model_switch_result_context.py tests/hermes_cli/test_model_switch_custom_providers.py tests/tools/test_image_generation_plugin_dispatch.py tests/tools/test_image_generation.py tests/gateway/test_model_command_custom_providers.py tests/gateway/test_model_switch_persistence.py tests/gateway/test_session_model_override_routing.py tests/test_tui_gateway_server.py::test_config_set_model_once_keeps_env_and_records_restore tests/test_tui_gateway_server.py::test_config_set_model_once_requires_live_session tests/test_tui_gateway_server.py::test_config_set_model_session_switch_clears_pending_once_restore tests/test_tui_gateway_server.py::test_restore_agent_model_runtime_falls_back_to_switch_model tests/test_tui_gateway_server.py::test_config_set_model_syncs_inference_provider_env tests/test_tui_gateway_server.py::test_config_set_model_global_persists tests/test_tui_gateway_server.py::test_config_set_model_syncs_tui_provider_env
uv run --no-sync python -m pytest -q tests/hermes_cli/test_user_providers_model_switch.py tests/hermes_cli/test_model_provider_persistence.py tests/hermes_cli/test_model_switch_variant_tags.py tests/hermes_cli/test_model_switch_context_display.py tests/hermes_cli/test_anthropic_model_flow_stale_oauth.py tests/hermes_cli/test_custom_provider_model_switch.py tests/hermes_cli/test_codex_cli_model_picker.py tests/run_agent/test_switch_model_context.py tests/run_agent/test_switch_model_fallback_prune.py tests/tools/test_image_generation_env.py
python3.12 -m compileall hermes_cli/model_switch.py cli.py tui_gateway/server.py gateway/run.py tools/image_generation_tool.py tests/hermes_cli/test_model_switch_once_flags.py tests/hermes_cli/test_cli_model_once.py tests/gateway/test_model_command_custom_providers.py tests/gateway/test_model_switch_persistence.py tests/gateway/test_session_model_override_routing.py tests/tools/test_image_generation_plugin_dispatch.py tests/tools/test_image_generation.py tests/test_tui_gateway_server.py
git diff --check

lkz-de · 2026-07-02T08:30:41Z

I checked current main plus nearby image-gen work while validating a local patch here. #29923 already covers the core once-per-call provider / model override surface for image_generate, so I do not think a separate fresh PR to main for that broad feature is justified.

What still looks missing relative to current provider behavior is narrower:

plugins/image_gen/openrouter/__init__.py should thread an explicit per-call model into _resolve_model_chain() so OpenRouter/Nous routes actually honor the override at provider level rather than only at generic dispatch level.
the managed Krea fast-path should read the per-call model_override before falling back to configured model.
in direct FAL mode, importing the SDK may still need FAL_KEY hydrated from Hermes config when the user keeps secrets only in the Hermes env/config file.

Local verification I ran while checking the overlap:

pytest -q tests/tools/test_image_generation.py tests/tools/test_image_generation_image_to_image.py tests/plugins/image_gen/test_openrouter_compat_provider.py -o addopts= → 115 passed

So my recommendation is: keep #29923 as the main PR, and if helpful I can prepare a focused follow-up patch for those provider-specific gaps rather than open a duplicate broad override PR.

lkz-de · 2026-07-02T18:22:47Z

I reproduced a remaining tool-surface gap on current main while testing image editing.

Provider plugins can already honor per-call model selection, but image_generate still drops explicit provider / model args inside _handle_image_generate, so the request can silently route using configured defaults instead of the requested backend/model.

Concrete repro from local testing:

call _handle_image_generate({... "provider": "openai-codex", "model": "gpt-image-2-medium", "image_url": ...})
expected: dispatch through Codex edit
current main behavior: explicit override is not threaded through tool dispatch / managed Krea routing / direct FAL model resolution

I opened a focused follow-up with tests here:

Fix image_generate per-call provider/model overrides #57293 Fix image_generate per-call provider/model overrides

That patch:

exposes optional provider / model on the image_generate schema
threads them through _handle_image_generate
honors them in plugin dispatch, managed Krea routing, and direct FAL model resolution
adds focused tests for explicit provider/model dispatch

Verification on the follow-up branch:

targeted override tests pass
a local end-to-end _handle_image_generate(...) call with explicit provider=openai-codex, model=gpt-image-2-medium, and image_url=... successfully produced an edited image

teknium1

Thanks for carrying the one-turn restore state across CLI, gateway, and TUI surfaces. The current-main premise is real: cli.py:8045 still uses the legacy parser, while tools/image_generation_tool.py:1181 and :1513 keep image backend/model selection user-configured only.

Problems

In the TUI path, the new --once branch avoids session["model_override"] (tui_gateway/server.py:2885 in this diff), but _run_prompt_submit() calls _sync_agent_model_with_config() before the turn (tui_gateway/server.py:8945). That helper only skips synchronization when session.get("model_override") is present (:3022), so the intended override can be reverted before its first request.
In the gateway path, the temporary override is written before the existing session-store persistence call (gateway/slash_commands.py:1863 in this diff; current persistence is at :1879-1887). The new restore helper restores only the in-memory map, so a restart can rehydrate the temporary override.

Suggested changes

Guard TUI pre-turn config synchronization with pending once state and add a test that observes the model actually used for the first turn.
Avoid persisting a once override, or restore the persisted value alongside the in-memory value; test gateway restart after the once turn.
Coordinate the overlapping image lane with open #59815.

Automated hermes-sweeper review.

teknium1 · 2026-07-13T09:35:55Z

@@ -2832,7 +2885,7 @@ def _apply_model_switch(
    # contamination bug). agent.switch_model() above already mutated the right


_run_prompt_submit() calls _sync_agent_model_with_config() before the turn, and its only opt-out is session["model_override"]. Because this branch suppresses that key for --once, the configured model can replace the one-turn model before run_conversation; preserve a transient guard until the first turn finishes.

teknium1 · 2026-07-13T09:35:55Z

@@ -1850,6 +1863,14 @@ async def _finish_switch() -> str:
                "base_url": result.base_url,


This temporary override is still passed to the unconditional session-store write below. _restore_pending_one_turn_model_override() restores only the in-memory map, so a restart may rehydrate this one-turn choice permanently. Skip the write for --once, or restore the persisted override too.

lkz-de · 2026-07-14T00:29:38Z

I rechecked the current head (003b9acd) against current main, and teknium's two /model --once findings reproduce from my side too:

in tui_gateway/server.py, the --once path avoids session["model_override"], so the pre-turn _sync_agent_model_with_config() path can still clobber the one-turn override before first use in fresh/resumed/config-change cases
in gateway/slash_commands.py, the temporary override still goes through the existing set_model_override(...) write-through path, so a gateway restart can rehydrate what should have been a one-turn-only choice

Given the review's note to coordinate the image lane with #59815, the cleanest split now looks like:

keep Add per-turn and per-image model overrides #29923 focused on the /model --once lane
drop tools/image_generation_tool.py, plugins/image_gen/fal/__init__.py, and the image override tests from this PR
let fix(image_generate): honor per-call provider/model overrides and warn-fallback unknown FAL overrides #59815 carry the image_generate provider/model override lane

#59815 is mine — happy to rebase it around whatever lands here first if that helps keep the split low-friction.

Adds --once to /model across CLI, TUI, and gateway: switch model for the next turn only, restoring the previous model in a finally block so success, exception, and interrupt all revert. Parsing extends parse_model_flags_detailed(); resolve_persist_behavior() treats --once as a persistence opt-out; --global + --once is rejected. Salvaged from PR #29923 (image-generation lane split to #59815 per review; conflict resolution against current main by the maintainers).

Fixes the two review defects that kept PR #29923 open, plus docs: - gateway: exclude --once from the session-store write-through. The once-override lived only in memory before, but the write-through persisted it, so a gateway restart before the finally-restore rehydrated a supposedly one-turn model permanently. - TUI: skip _sync_agent_model_with_config while a one-turn restore is pending. The once-model is deliberately not pinned as a session model_override, so the config sync saw a model mismatch and clobbered the once-override back to the config model before the turn ran. - tests: real _handle_model_command drive asserting --once never touches set_model_override while --session still does; restore-pop idempotency. - docs: /model --once in configuring-models.md with an honest prompt-cache cost note (one-shot switch breaks the cached prefix twice; wins for short sessions and cheap-to-expensive escalation).

teknium1 · 2026-07-18T21:22:21Z

Merged via PR #67113 — your commit was cherry-picked onto current main with your authorship preserved in git history. Thanks for a well-built feature: the ModelFlagParseResult dataclass refactor with the legacy 5-tuple shim was exactly the right compatibility approach, and the restore-in-finally pattern across all three frontends held up cleanly through review.

Two adjustments landed on top in the salvage:

the gateway lane no longer writes the once-override through to the session store (a restart mid-turn used to rehydrate it permanently)
the TUI skips the pre-turn config-model sync while a once-override is active (it used to clobber the override before the turn ran)

The per-call image_generate override lane stayed out per the review discussion — that half remains tracked in #59815. Docs for --once are live on the configuring-models page.

Adds --once to /model across CLI, TUI, and gateway: switch model for the next turn only, restoring the previous model in a finally block so success, exception, and interrupt all revert. Parsing extends parse_model_flags_detailed(); resolve_persist_behavior() treats --once as a persistence opt-out; --global + --once is rejected. Salvaged from PR NousResearch#29923 (image-generation lane split to NousResearch#59815 per review; conflict resolution against current main by the maintainers).

Fixes the two review defects that kept PR NousResearch#29923 open, plus docs: - gateway: exclude --once from the session-store write-through. The once-override lived only in memory before, but the write-through persisted it, so a gateway restart before the finally-restore rehydrated a supposedly one-turn model permanently. - TUI: skip _sync_agent_model_with_config while a one-turn restore is pending. The once-model is deliberately not pinned as a session model_override, so the config sync saw a model mismatch and clobbered the once-override back to the config model before the turn ran. - tests: real _handle_model_command drive asserting --once never touches set_model_override while --session still does; restore-pop idempotency. - docs: /model --once in configuring-models.md with an honest prompt-cache cost note (one-shot switch breaks the cached prefix twice; wins for short sessions and cheap-to-expensive escalation).

Add per-turn and per-image model overrides

4f2e857

Merge upstream main into model overrides PR

90223d8

y0shua1ee mentioned this pull request May 29, 2026

feat: add image generation fallback controls #32320

Open

19 tasks

Merge upstream main into model overrides PR

cf7733f

github-actions Bot mentioned this pull request Jul 3, 2026

🦞 OpenClaw 生态日报 2026-07-04 huajiao1998/big_model_radar#15

Open

lkz-de mentioned this pull request Jul 6, 2026

fix(image_generate): honor per-call provider/model overrides and warn-fallback unknown FAL overrides #59815

Closed

Merge upstream main into model overrides PR

003b9ac

teknium1 reviewed Jul 13, 2026

View reviewed changes

teknium1 mentioned this pull request Jul 18, 2026

feat: /model --once one-turn model override (salvages #29923) #67113

Merged

teknium1 closed this Jul 18, 2026

pierrenode mentioned this pull request Aug 18, 2026

fix(gateway): thread session overrides through Bot Chat capability rebuild #88826

Open

3 tasks

Navigation Menu

Uh oh!

Add per-turn and per-image model overrides - #29923

Add per-turn and per-image model overrides#29923
deusyu wants to merge 4 commits into
NousResearch:mainfrom
deusyu:codex-model-overrides-29914

deusyu commented May 21, 2026

Uh oh!

lkz-de commented Jul 2, 2026 •

edited

Loading

Uh oh!

lkz-de commented Jul 2, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 13, 2026

Uh oh!

teknium1 Jul 13, 2026

Uh oh!

lkz-de commented Jul 14, 2026

Uh oh!

teknium1 commented Jul 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		@@ -2832,7 +2885,7 @@ def _apply_model_switch(
		# contamination bug). agent.switch_model() above already mutated the right

		@@ -1850,6 +1863,14 @@ async def _finish_switch() -> str:
		"base_url": result.base_url,

Uh oh!

Conversation

deusyu commented May 21, 2026

Summary

Tests

Uh oh!

lkz-de commented Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lkz-de commented Jul 2, 2026

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 13, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 13, 2026

Choose a reason for hiding this comment

Uh oh!

lkz-de commented Jul 14, 2026

Uh oh!

teknium1 commented Jul 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

lkz-de commented Jul 2, 2026 •

edited

Loading