Add 'maximum' reasoning level for codex xhigh support

[mariusvniekerk]

Summary

• Adds a new ReasoningMaximum level ("maximum") that maps to codex's "xhigh" reasoning effort, enabling access to codex's highest reasoning tier
• Accepts "maximum", "max", and "xhigh" as aliases in both ParseReasoningLevel and NormalizeReasoning
• For agents without an xhigh equivalent (droid, kilo, pi), maximum maps to their highest available level (same as thorough)
• Updates CLI flag help text across all 7 commands and config TOML comments to include the new level

Closes #386

🤖 Generated with Claude Code

[wesm]

@mariusvniekerk want to add support for high/max claude efforts in this too?

[roborev-ci]

roborev: Combined Review (17b78a7)

Verdict: The PR successfully introduces the maximum reasoning level but contains a medium severity behavioral regression in workflow-specific configuration resolution.

Medium

• Workflow-specific agent/model overrides do not follow the new maximum level:
  ResolveAgentForWorkflow / ResolveModelForWorkflow now receive level=="maximum", but the lookup layer only knows fast, standard, and thorough. In practice, --reasoning maximum will skip review_agent_thorough, review_model_thorough,
  fix_agent_thorough, etc., and fall back to the generic workflow/default values instead. That is a silent regression for users who route “deep” reviews to a specific agent/model via *_thorough config, and it can defeat the intended Codex xhigh path. Suggested fix: either add explicit
  *_maximum config fields everywhere, or have workflow lookup fall back from maximum to thorough before dropping to generic defaults.
  • config.go

config.go

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (4b08cc4)

Verdict: Introduces a new maximum reasoning level, but requires fixes for missing compatibility checks and incomplete configuration plumbing.

Medium

• internal/agent/claude.go
  The new buildArgs() path now emits --effort for fast, thorough, and maximum, but there is no
  compatibility check for that flag. review defaults to thorough, so any older Claude Code install that lacks --effort will start failing even on the default path. The file already probes --dangerously-skip-permissions; --effort needs the same treatment, or the code should degrade cleanly when unsupported.

• internal/config/config.go
  The commit teaches config normalization/comments about maximum, but the workflow-specific agent/model plumbing still only has *_fast, *_standard , and *_thorough fields/cases. That means once reasoning resolves to maximum, ResolveAgentForWorkflow / ResolveModelForWorkflow skip all level-specific overrides and silently fall back to the generic agent/model. Either add *_maximum config slots and switch cases, or explicitly map
  maximum to thorough in those resolvers if that is the intended behavior. This also needs config-level tests; the new tests only cover parsing and direct agent flag mapping.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (0a177d7)

Verdict: All agents agree the code is clean, with no Medium, High, or Critical issues found.

The changes safely introduce medium and maximum reasoning levels across the CLI and configuration, including specific mappings for Claude and Codex. No security
or high-severity functional issues were identified.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[mariusvniekerk]

@wesm Added support for the levels in claude too. Also added an explicit medium level.