[aviggiano]

Follow-up to #58.

• clear backend-specific reasoning whenever --agent switches agents, including --agent plus --model
• preserve configured reasoning for model-only overrides
• forward Claude reasoning effort through the supported --effort flag
• align CLI and config documentation

Validation:

• pnpm -w format:check
• pnpm -w lint
• pnpm -w build
• pnpm -w test
• pnpm -w run ci
• independent Smithers review approved with no findings

Greptile Summary

This follow-up to #58 fixes two related gaps in agent-override handling: backend-specific reasoning is now always cleared whenever --agent switches to a different agent (including when --model is also provided), and Claude's reasoningEffort is now forwarded to the Claude Code CLI via --effort rather than silently dropped.

• applyDefaultProfileOverrides restructured so delete profile.reasoning fires unconditionally on any cross-agent switch, decoupled from the overrides.model === undefined guard that controls delete profile.model.
• createClaudeAgent wires options.reasoningEffort to extraArgs: [\"--effort\", value] on SmithersClaudeCodeAgent; the template is excluded from the package tsconfig.json (.tsx pattern), so extraArgs validity is only caught at generated-project compile time.
• New unit test in config.test.ts and updated integration assertion in runtime.test.ts cover both the agent-only and agent-plus-model override paths.

Confidence Score: 4/5

Safe to merge; the logic change is small and well-scoped, and all stated validation steps (build, lint, tests, CI) reportedly pass.

The reasoning-clear fix in applyDefaultProfileOverrides is straightforward and the test coverage is good, but the new test in config.test.ts omits a pre-condition assertion — it checks that reasoning is null after the override without first confirming it was set, so a future regression that removes the clearing logic would go undetected. The extraArgs forwarding in the Claude template is also untested against real smithers-orchestrator types (templates are excluded from the tsconfig), leaving that integration unverified at the package level.

packages/config/test/config.test.ts — the new test case should assert the pre-condition; packages/runtime/src/templates/smithers/agents/claude.tsx — extraArgs validity is not type-checked in this repo.

Important Files Changed

Filename	Overview
packages/config/src/model-profiles.ts	Separates model and reasoning clearing logic so reasoning is always deleted on any agent switch, while model is only deleted when no explicit replacement is provided.
packages/runtime/src/templates/smithers/agents/claude.tsx	Removes the 'accepted and ignored' comment for reasoningEffort and wires it to ClaudeCodeAgent via extraArgs: ['--effort', value]; template is not type-checked against smithers-orchestrator since .tsx files are excluded from the tsconfig, so extraArgs validity is only caught at user-project compile time.
packages/config/test/config.test.ts	Adds two new test cases for the agent-switch reasoning-clear fix; missing an explicit pre-condition assertion (that reasoning was non-null before the override).
packages/runtime/test/runtime.test.ts	Extends existing agent-switch test to assert reasoningEffort is null after switching, and adds a template text-match assertion for the new --effort extraArgs pattern.
ultrafuzz.toml	Updates the comment on the [models.claude] profile to reflect that ClaudeAgent now forwards reasoning via --effort.
docs/cli.md	Documentation updated to accurately describe new reasoning-clearing semantics for --agent and --model overrides.
docs/reference/cli.md	Reference docs updated to match; no blank line between the revised paragraph and the --max-concurrency description (pre-existing issue, not new).
docs/config.md	Updates the ClaudeAgent reasoning section to document --effort passthrough instead of documenting it as a no-op.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["CLI: --agent / --model flags"] --> B{applyDefaultProfileOverrides}
    B --> C{agent override provided AND differs from profile.agent?}
    C -- No --> F[Profile unchanged, reasoning preserved]
    C -- Yes --> D{model override also provided?}
    D -- No --> E["delete profile.model\ndelete profile.reasoning"]
    D -- Yes --> G["delete profile.reasoning\nprofile.model = overrides.model"]
    E --> H[profile.agent = overrides.agent]
    G --> H
    F --> I{model override provided?}
    I -- Yes --> J[profile.model = overrides.model]
    I -- No --> K[Done]
    H --> K
    J --> K
    K --> L[Task spec: reasoningEffort = profile.reasoning ?? null]
    L --> M{reasoningEffort null?}
    M -- Yes --> N[Agent created without effort arg]
    M -- No --> O["ClaudeAgent: extraArgs: ['--effort', value]\nCodexAgent: config.model_reasoning_effort = value"]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["CLI: --agent / --model flags"] --> B{applyDefaultProfileOverrides}
    B --> C{agent override provided AND differs from profile.agent?}
    C -- No --> F[Profile unchanged, reasoning preserved]
    C -- Yes --> D{model override also provided?}
    D -- No --> E["delete profile.model\ndelete profile.reasoning"]
    D -- Yes --> G["delete profile.reasoning\nprofile.model = overrides.model"]
    E --> H[profile.agent = overrides.agent]
    G --> H
    F --> I{model override provided?}
    I -- Yes --> J[profile.model = overrides.model]
    I -- No --> K[Done]
    H --> K
    J --> K
    K --> L[Task spec: reasoningEffort = profile.reasoning ?? null]
    L --> M{reasoningEffort null?}
    M -- Yes --> N[Agent created without effort arg]
    M -- No --> O["ClaudeAgent: extraArgs: ['--effort', value]\nCodexAgent: config.model_reasoning_effort = value"]

Loading

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
packages/config/test/config.test.ts:242-262
**Missing pre-condition assertion weakens the regression test**

The test verifies that `reasoning` is `null` after the override but never asserts that it was non-null beforehand. If the default profile in `ultrafuzz.toml` ever loses its `reasoning = "xhigh"` entry, the test would still pass while silently proving nothing — the clearing logic could be deleted and every expectation would continue to hold. Adding an assertion before the `applyDefaultProfileOverrides` call would make the pre-condition explicit and keep the test honest as an invariant check.

_{Reviews (1): Last reviewed commit: "Fix Claude reasoning overrides" | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

[greptile-apps]

Missing pre-condition assertion weakens the regression test

The test verifies that reasoning is null after the override but never asserts that it was non-null beforehand. If the default profile in ultrafuzz.toml ever loses its reasoning = "xhigh" entry, the test would still pass while silently proving nothing — the clearing logic could be deleted and every expectation would continue to hold. Adding an assertion before the applyDefaultProfileOverrides call would make the pre-condition explicit and keep the test honest as an invariant check.

Prompt To Fix With AI

This is a comment left during a code review.
Path: packages/config/test/config.test.ts
Line: 242-262

Comment:
**Missing pre-condition assertion weakens the regression test**

The test verifies that `reasoning` is `null` after the override but never asserts that it was non-null beforehand. If the default profile in `ultrafuzz.toml` ever loses its `reasoning = "xhigh"` entry, the test would still pass while silently proving nothing — the clearing logic could be deleted and every expectation would continue to hold. Adding an assertion before the `applyDefaultProfileOverrides` call would make the pre-condition explicit and keep the test honest as an invariant check.

How can I resolve this? If you propose a fix, please make it concise.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!