Greptile Summary

This PR successfully adds independent /effort control for Anthropic Claude 4.6 models, decoupling effort configuration from the existing /think toggle. The implementation is clean, well-tested, and follows existing code patterns.

Key Changes:

• New EffortLevel type and normalization functions in src/auto-reply/thinking.ts (following the same pattern as ThinkLevel)
• /effort command registered with /e alias in the commands registry
• Effort level stored in session state and displayed in /status output
• output_config.effort injected into Anthropic API payloads via createAnthropicEffortWrapper() in extra-params.ts
• Model-aware validation: max effort restricted to Opus 4.6 models only
• 96 new test lines covering effort normalization, model support detection, and directive parsing
• Documentation updated in docs/providers/anthropic.md

Implementation Quality:

• Properly scoped to anthropic provider only (line 753 in extra-params.ts)
• OpenRouter continues using thinking level for reasoning.effort (backward compatible)
• Comprehensive test coverage for all new functions
• Session persistence properly implemented
• Follows existing directive handling patterns

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The implementation is exceptionally clean and well-architected. It follows established patterns in the codebase, includes comprehensive test coverage (96 new test lines), and properly validates edge cases like model-specific effort levels. The changes are properly scoped to only affect Anthropic provider, maintaining backward compatibility with OpenRouter and other providers. All code paths are accounted for from command registration through session persistence to API invocation.
• No files require special attention

_{Last reviewed commit: b2d5df6}

Good catch @chatgpt-codex-connector — fixed in 920a35b. Added hasEffortDirective: false and effortLevel: undefined to the unauthorized sender gate in get-reply-directives.ts, consistent with the other command directives.

To use Codex here, create a Codex account and connect to github.

Second catch fixed in 5d77d6f — added hasEffortDirective to the hasInlineDirective guard in get-reply-directives.ts (line 237). Now /effort in mixed chat content triggers the directive-only check and gets cleared, consistent with /think and other inline directives.

All three Codex review comments addressed and resolved:

1. P1 — Auth bypass (920a35b): hasEffortDirective: false + effortLevel: undefined added to unauthorized sender gate
2. P1 — Mixed content (5d77d6f): hasEffortDirective added to hasInlineDirective guard
3. P2 — Stale max effort (80ab716): resolvedEffortLevel now clamps max to high when model doesn't support it (e.g. session stored /effort max on Opus then switched to Sonnet)

All checks pass: pnpm check && pnpm build ✅

All 6 Codex review threads now addressed and resolved:

Round 1 (fixed earlier):

1. P1 — Auth bypass: effort directives stripped for unauthorized senders ✅
2. P1 — Mixed content: effort added to inline directive guard ✅
3. P2 — Stale max on model switch: clamped in directive resolution ✅

Round 2 (fixed in b587a67):
4. P1 — Propagation: false positive — effortLevel was already wired through the full chain ✅
5. P2 — Max effort on fallback: extra-params.ts now clamps max→high for non-Opus models ✅
6. P1 — Gate to 4.6 models: new supportsEffort() function checks both provider AND model family before injecting effort ✅

Additional improvements in this round:

• Added EFFORT_MODEL_PREFIXES for all Claude 4.6 family (opus-4-6, sonnet-4-6, sonnet-4-5)
• Added supportsEffort() export with 4 new test cases (34 total thinking tests)
• Effort injection docs added to docs/tools/slash-commands.md (fixes doc test)

All tests pass (72/72), pnpm check && pnpm build clean.

PR Summary

What: Independent /effort control for Anthropic Claude 4.6 models (Opus and Sonnet), decoupled from the existing /think toggle.

Why: Anthropic Claude 4.6 models support effort as an independent parameter from thinking. Currently OpenClaw conflates them via /think. With effort separated, users can run thinking: off + effort: high (the #1 non-reasoning config on Artificial Analysis benchmarks per #9837).

How:

• New /effort slash command (/e alias) with levels: off, low, medium, high, max (max = Opus only)
• supportsEffort() / supportsMaxEffort() model capability gates (Claude 4.6 family only)
• Effort wired through the full chain: directives → session state → embedded runs → followup queue → status display → gateway APIs → Swift protocol
• Falls back gracefully: persists on unsupported models (no-op at runtime, consistent with /think behavior), clamps max → high on non-Opus

CI note: The 3 failing checks are pre-existing on main:

• bun test: nodes-utils.test.ts assertion (Bun object iteration order differs from Node)
• windows test (x2): plugins/loader.test.ts + config/includes.test.ts

All Linux Node checks pass. Single squashed commit, 42 files changed.

Refs: #9837

@GodsBoy are you running this as a local patch for yourself at the moment?

Yes — running it as a local patch on my production OpenClaw instance. Been using it daily for a few weeks now. The main use case for me is running sub-agents with thinking disabled but effort high (nothink mode) for faster non-reasoning tasks, while keeping effort: high with thinking enabled for complex reasoning sessions. The two controls are genuinely independent in the API and it makes a real difference in practice.

Would love a second pair of eyes on it if you're interested in testing. Happy to answer any questions about the implementation.

Thanks for flagging — but the 3.1 change doesn't cover this PR.

The 3.1 changelog entry ("set adaptive as the default thinking level for Anthropic Claude 4.6 models") only changes the default thinking level — equivalent to auto-enabling /think for 4.6 models. That's the thinking parameter.

This PR is about the effort parameter — a completely separate API field that controls reasoning depth independently of whether thinking blocks are active at all. The word "effort" doesn't appear in the 3.1 changelog and the feature isn't in the codebase.

The use case this unlocks that 3.1 doesn't address:

• Thinking off + effort high — the "nothink" configuration. No thinking token overhead, but maximum reasoning budget for computation. This scored #1 out of 60 non-reasoning models on the Artificial Analysis Intelligence Index in TTFT-adjusted benchmarks.
• Thinking adaptive + effort low — reduced cost for lighter tasks where full effort isn't needed.
• Any mix of the two controls independently — which is what Anthropic's API actually exposes.

I've been running this as a local patch on production for several weeks. The two controls are genuinely independent at the API level and the current /think toggle cannot replicate them. Happy to rebase on latest main if that helps move this forward.