fix: accept max reasoning effort

[SamuelZ12]

summary

• accept max as a valid reasoning_effort across Hermes parsing and command surfaces
• update reasoning menus/help/docs so max appears anywhere users choose a reasoning level
• add regression coverage for CLI parsing, gateway /reasoning max, and the reasoning effort picker order

root cause

The Anthropic adapter already supported Claude 4.7 output_config.effort = max, but the shared reasoning parser and command validators still only accepted up to xhigh. That made agent.reasoning_effort: max silently parse as invalid, so Hermes skipped sending adaptive reasoning fields even though downstream Anthropic code could handle them.

testing

• /Users/samuelzhang/.hermes/hermes-agent/venv/bin/python -m pytest tests/cli/test_reasoning_command.py tests/gateway/test_reasoning_command.py tests/agent/test_anthropic_adapter.py tests/hermes_cli/test_reasoning_effort_menu.py -q -o 'addopts='

notes

• verified locally after patch that Claude requests include thinking: {"type": "adaptive", "display": "summarized"} and output_config: {"effort": "max"} for claude-opus-4-7

[Copilot]

Pull request overview

This PR updates Hermes’ reasoning-effort surfaces to accept max end-to-end (shared parsing, CLI + gateway commands, menus, and user-facing docs), closing a gap where downstream adapters supported max but upstream validation rejected it.

Changes:

• Extend shared reasoning effort parsing/validation to include max and propagate it through CLI + gateway /reasoning flows.
• Update UX surfaces (CLI command registry, reasoning effort picker ordering, docs) to list max as a valid option.
• Add regression tests covering CLI parsing, gateway /reasoning max, and reasoning picker ordering.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
website/docs/user-guide/features/batch-processing.md	Documents --reasoning_effort including max.
website/docs/user-guide/configuration.md	Updates config docs to list max as a valid agent.reasoning_effort.
tests/hermes_cli/test_reasoning_effort_menu.py	Adds test ensuring menu ordering places max after xhigh.
tests/gateway/test_reasoning_command.py	Adds gateway regression test for /reasoning max persistence + cache update.
tests/cli/test_reasoning_command.py	Extends CLI parsing test matrix to include max.
skills/autonomous-ai-agents/hermes-agent/SKILL.md	Updates slash-command docs to include max.
hermes_constants.py	Adds max to VALID_REASONING_EFFORTS and parser docstring.
hermes_cli/main.py	Updates canonical reasoning effort ordering to include max.
hermes_cli/commands.py	Adds max to /reasoning command subcommands for help/completions.
gateway/run.py	Updates gateway /reasoning help + validation to accept max.
cli.py	Updates CLI /reasoning help + “valid levels” output to include max.
batch_runner.py	Extends batch-runner --reasoning_effort validation/docs to accept max.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

/reasoning help text now lists max, but the no-args status output later in this handler still prints Usage: /reasoning <none|minimal|low|medium|high|xhigh|show|hide> (missing max). Please update that usage hint too so all CLI help surfaces are consistent.

[Copilot]

The handler docstring/usage now includes max, but the no-args response still returns _Usage:_ /reasoning <none|minimal|low|medium|high|xhigh|show|hide>`` (missing max). Please update that returned usage string as well so gateway help output matches the accepted levels.

[counterposition]

The root cause analysis looks right: the Anthropic adapter already knows how to emit output_config.effort = max, and the shared parser/command surfaces were rejecting it too early.

A couple suggestions before merging to make this fully robust:

1. Avoid leaking Anthropic-specific max to other transports.

VALID_REASONING_EFFORTS makes max globally accepted, but some non-Anthropic paths appear to forward reasoning_config["effort"] directly. For example, the Codex/Responses transport only clamps minimal -> low, so a user with agent.reasoning_effort: max who switches to a GPT/Codex model may send {"effort": "max"} to a backend that likely only supports up to high/xhigh. The generic Chat Completions OpenRouter/Nous path similarly forwards the config as-is.

I’d suggest normalizing max at the provider/model boundary, the same way the code already does for GitHub Models and Anthropic 4.6 xhigh -> max. For non-Anthropic providers, either clamp max to the strongest supported effort or omit it unless the provider explicitly advertises support.

2. There are a few stale help strings after applying this on current main:

• cli.py no-arg /reasoning usage still shows <none|minimal|low|medium|high|xhigh|show|hide>.
• gateway/run.py no-arg /reasoning usage still shows <none|minimal|low|medium|high|xhigh|show|hide>.
• gateway/platforms/discord.py slash-command description still says “none, minimal, low, medium, high, or xhigh.”

3. It would be useful to add one regression test for the cross-provider behavior, e.g. reasoning_effort=max on a Codex/OpenAI Responses path should not produce an unsupported reasoning.effort=max unless that path explicitly supports it.

Overall the patch is pointed in the right direction; I’d just make the accepted user-level enum distinct from the provider-specific wire enum so model switches don’t turn a Claude-only maximum into invalid requests elsewhere.

[teknium1]

Thanks for identifying the upstream validation gap. The gateway defect is still present on current main: gateway/slash_commands.py:2732 accepts only through xhigh, while hermes_constants.py:794-819 already accepts max.

Problems

• The patch targets the pre-refactor handler in gateway/run.py; current gateway dispatch lives in gateway/slash_commands.py:2634, so this needs a conflict-aware salvage rather than applying the old hunk.
• Provider projection is still needed. agent/transports/codex.py:155-166 only clamps minimal → low, so accepting max at the command/config layer can forward literal max to Codex/Responses.
• The live CLI help now resides in hermes_cli/cli_commands_mixin.py:2469-2537 and still omits max.

Suggested changes

• Port the gateway fix to gateway/slash_commands.py, update the live CLI/registry surfaces, and add a Codex wire-level regression that clamps or omits unsupported max values.

Automated hermes-sweeper review.

[teknium1]

Accepting max globally also reaches non-Anthropic transports. Please add provider-boundary normalization: current Codex handling only maps minimal → low, so it otherwise forwards max as reasoning.effort (agent/transports/codex.py:155-166).

[counterposition]

Thanks for this contribution @SamuelZ12 — you correctly identified a real gap at the time: the Anthropic adapter already understood max, but the shared parser and command surfaces still rejected it.

This has since been implemented on main by #62650 (merged 2026-07-12), which added both max and a new ultra level across all the surfaces this PR touched — the shared validation constants, the gateway and CLI /reasoning commands, the batch runner, docs, and tests — plus provider-side handling so the new levels are mapped safely per backend.

Since there's nothing left for this branch to add, closing as implemented-on-main. That's a timing outcome, not a knock on the change — thanks again for the well-tested PR.

[teknium1]

Closing — fully superseded on current main. You were the first to propose max as a reasoning effort level (April 18), and the feature has since landed piecemeal: max and ultra were added to VALID_REASONING_EFFORTS in #62650 (GPT-5.6 launch), and the remaining discoverability gap (help/status strings across CLI, gateway, and all 16 locales) was closed in #65651. Every surface this PR touched — constants, menus, batch runner, command registry, docs — now carries the extended levels.

Thanks for being early on this one; sorry it sat as long as it did.