feat(reasoning): expose "max" effort level for Anthropic 4.6/4.7 adaptive thinking - #25401
Closed
Vic563 wants to merge 1 commit into
Closed
feat(reasoning): expose "max" effort level for Anthropic 4.6/4.7 adaptive thinking#25401Vic563 wants to merge 1 commit into
Vic563 wants to merge 1 commit into
Conversation
…tive thinking
The Anthropic adapter has supported "max" since Opus 4.7 launched (it's
also the strongest level Opus/Sonnet 4.6 accept), and ADAPTIVE_EFFORT_MAP
already routes it through to output_config.effort. But "max" was never
listed in VALID_REASONING_EFFORTS, so it failed parse_reasoning_effort()
validation everywhere upstream — config.yaml loaders, /reasoning slash
command, gateway, batch_runner — and users had to lie via "xhigh" + a 4.6
model to actually request it.
This makes "max" a first-class reasoning level alongside minimal/low/medium
/high/xhigh:
- Add "max" to VALID_REASONING_EFFORTS in hermes_constants.py and
document it as the Anthropic 4.7+ adaptive-thinking ceiling
- Update CLI help, gateway _load_reasoning_config, batch_runner help
+ valid_efforts, autocomplete subcommands, setup wizard canonical
order, and all 16 locale files (status + unknown_arg messages)
- Update bundled hermes-agent skill + corresponding website doc page
- Update website/docs/user-guide/configuration.md reference
Provider compatibility:
- Anthropic adapter (4.6+): pass-through via existing ADAPTIVE_EFFORT_MAP;
4.6/4.7 both accept "max" natively
- LM Studio: added "max" to _LM_VALID_EFFORTS; the existing allowed_options
clamp will downgrade if a model doesn't expose it
- Gemini 3 Flash/Pro: "max" maps to the same "high" thinking level as
"xhigh" (matches the existing xhigh treatment)
- Codex Responses (OpenAI/xAI): clamp "max" → "high" via _effort_clamp,
matching the existing minimal → low pattern. Same clamp applied in
auxiliary_client.py for consistency
Tests:
- test_hermes_constants.py: documented set guard now requires "max";
new test_max_is_a_valid_level covers normalization
- test_codex_transport.py: new test_max_effort_clamped verifies the
Codex backend never sees a literal "max"
- test_anthropic_adapter.py: pre-existing tests already cover max
on 4.6/4.7
360 reasoning/anthropic/codex/chat-completions tests pass. The 10
TestResolveAnthropicToken failures are pre-existing on origin/main
(env-var leakage, unrelated to this change).
Contributor
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
Adds a new "max" reasoning effort level to the existing set (none|minimal|low|medium|high|xhigh), wiring it through the constants registry, CLI/gateway commands, provider transports, localization strings, and documentation.
Changes:
- Add
"max"toVALID_REASONING_EFFORTSand accept it in parsing/validation paths. - Map
"max"appropriately in provider adapters (Codex/auxiliary clamp to"high"; Gemini maps to"high"; LM Studio accepts). - Update CLI/gateway help text, slash-command subcommands, and all locale strings; add tests.
Reviewed changes
Copilot reviewed 31 out of 31 changed files in this pull request and generated 6 comments.
Show a summary per file
| File | Description |
|---|---|
| hermes_constants.py | Adds "max" to valid efforts; expanded docstring. |
| hermes_cli/main.py | Adds "max" to canonical order in prompt selection. |
| hermes_cli/commands.py | Adds "max" subcommand for /reasoning. |
| gateway/run.py | Accepts "max" and updates docstring. |
| cli.py | Updates /reasoning help/error text to include "max". |
| batch_runner.py | Adds "max" to valid efforts list and docstring. |
| agent/transports/codex.py | Clamps "max" → "high" for Codex backend. |
| agent/transports/chat_completions.py | Maps "max" for Gemini Flash/Pro. |
| agent/lmstudio_reasoning.py | Adds "max" to LM Studio's accepted set. |
| agent/auxiliary_client.py | Auxiliary client clamps "max" → "high". |
| locales/*.yaml (16 files) | Updates /reasoning usage and unknown-arg strings to include "max". |
| skills/.../SKILL.md, website/docs/... | Doc updates listing "max". |
| tests/test_hermes_constants.py | Adds test_max_is_a_valid_level; updates documented-levels set. |
| tests/agent/transports/test_codex_transport.py | Adds test_max_effort_clamped. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+56
to
57
| if effort not in {"minimal", "low", "medium", "high", "xhigh", "max"}: | ||
| effort = "medium" |
Comment on lines
+89
to
90
| _effort_clamp = {"minimal": "low", "max": "high"} | ||
| reasoning_effort = _effort_clamp.get(reasoning_effort, reasoning_effort) |
Comment on lines
71
to
73
| thinking_config["thinkingLevel"] = ( | ||
| "high" if effort in {"high", "xhigh"} else "low" | ||
| "high" if effort in {"high", "xhigh", "max"} else "low" | ||
| ) |
| rather than 400ing the request.""" | ||
| messages = [{"role": "user", "content": "Hi"}] | ||
| kw = transport.build_kwargs( | ||
| model="gpt-5.4", messages=messages, tools=[], |
Comment on lines
+202
to
+204
| treat it as their own ceiling — for example, OpenRouter/OpenAI-style | ||
| `reasoning.effort` consumers should map "max" to their highest supported | ||
| level. |
| @pytest.mark.parametrize( | ||
| "value", | ||
| ["bogus", "very-high", "max", "0", "off", "true", "default"], | ||
| ["bogus", "very-high", "0", "off", "true", "default"], |
Author
|
Closing — change is being kept as a local customization rather than upstreamed. Branch remains on the fork (Vic563/hermes-agent feat/reasoning-max-alias) for personal use. |
Author
|
Reopening — keeping the upstream contribution open as originally intended. |
alt-glitch
added
type/feature
Collaborator
This was referenced May 29, 2026
19 tasks
Contributor
|
Thanks for the thorough cross-provider analysis and tests. This is an automated hermes-sweeper review; the requested behavior is now implemented on main.
The member note identifying the earlier duplicate effort is consistent with this now-landed implementation. |
teknium1
added
the
sweeper:implemented-on-main
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The Anthropic adapter has supported
"max"since Opus 4.7 launched (it is also the strongest level Opus/Sonnet 4.6 accept), andADAPTIVE_EFFORT_MAPalready routes it through tooutput_config.effort. But"max"was never listed inVALID_REASONING_EFFORTS, so it failedparse_reasoning_effort()validation everywhere upstream — config.yaml loaders,/reasoningslash command, gateway,batch_runner— and users had to lie via"xhigh"+ a 4.6 model to actually request it.This makes
"max"a first-class reasoning level alongsideminimal/low/medium/high/xhigh.Changes
"max"toVALID_REASONING_EFFORTSinhermes_constants.pyand document it as the Anthropic 4.7+ adaptive-thinking ceiling_load_reasoning_config,batch_runnerhelp +valid_efforts, autocomplete subcommands, setup wizardcanonical_order, and all 16 locale files (status+unknown_argmessages)hermes-agentskill + corresponding website doc pagewebsite/docs/user-guide/configuration.mdreferenceProvider compatibility
ADAPTIVE_EFFORT_MAP; 4.6/4.7 both accept"max"natively"max"to_LM_VALID_EFFORTS; the existingallowed_optionsclamp will downgrade if a model does not expose it"max"maps to the same"high"thinking level as"xhigh"(matches the existing xhigh treatment)"max"→"high"via_effort_clamp, matching the existingminimal→lowpattern. Same clamp applied inauxiliary_client.pyfor consistencyTests
test_hermes_constants.py: documented-set guard now requires"max"; newtest_max_is_a_valid_levelcovers normalization (MAX,Max, etc.)test_codex_transport.py: newtest_max_effort_clampedverifies the Codex backend never sees a literal"max"test_anthropic_adapter.py: pre-existing tests already covermaxon 4.6/4.7360 reasoning/anthropic/codex/chat-completions tests pass. The 10
TestResolveAnthropicTokenfailures observed during local testing are pre-existing onorigin/main(env-var leakage in test isolation, unrelated to this change).