Gate Claude-only wizard prompts on agent_backend=claude

[dcellison]

Problem

The make config wizard prompts the operator for several Claude-binary-specific knobs unconditionally, even when the operator has selected goose as the agent backend. Goose has no equivalent of these settings, so the prompts at best confuse the operator and at worst cause them to set values that are silently ignored at runtime.

Verified via audit of src/kai/install.py and src/kai/pool.py:

Env var	Wizard line	Currently gated?	Backend(s) that consume it
CLAUDE_AUTOCOMPACT_PCT	install.py:546-557	No	Claude only (ClaudeCodeBackend autocompact env override)
CLAUDE_EFFORT_LEVEL	install.py:559-573	No	Claude only (ClaudeCodeBackend --effort flag)
CLAUDE_USER	install.py:388-394	Only by if advanced:	Claude only (sudo isolation; Goose has no sudo path)

The other CLAUDE_* env vars in the wizard (CLAUDE_TIMEOUT_SECONDS, CLAUDE_MAX_CONTEXT_WINDOW, CLAUDE_MAX_BUDGET_USD -> BUDGET_CEILING, DEFAULT_MODEL <- CLAUDE_MODEL) are correctly not gated on backend — both GooseBackend and ClaudeCodeBackend accept the corresponding kwargs in pool.py.

Proposed solution

Add a backend-conditional gate to each of the three prompts identified above. The pattern that works:

1. Initialize the variable to its dataclass default before the conditional, so the env-emission if value != default block downstream still works without raising NameError when goose is selected.
2. Wrap the _prompt(...) / _prompt_choice(...) call in if agent_backend == "claude":.
3. The matching env-emission block in the lower half of _cmd_config() already short-circuits on the default-value check (if claude_effort_level != "high"), so no separate emission gate is needed - but verify each one to be sure.

For CLAUDE_USER: keep the existing if advanced: gate AND add agent_backend == "claude" as an additional condition. The os_user field has no meaning for Goose.

Tests

Every fixture in tests/test_install.py currently passes "claude" as the agent backend. This audit needs at least one "goose" fixture per gated prompt that asserts the prompt is skipped (otherwise the gating regresses silently the next time someone touches the wizard). The simplest path is to extend _base_inputs with an agent_backend parameter (default "claude") and add a single new TestCmdConfig::test_goose_backend_skips_claude_only_prompts test that:

• Calls _base_inputs(memory_block=["false"], agent_backend="goose") (omitting the autocompact, effort, and CLAUDE_USER inputs)
• Runs _cmd_config()
• Asserts the resulting install.conf env contains none of the three keys

Out of scope

• CLAUDE_MAX_SESSION_HOURS and CLAUDE_IDLE_TIMEOUT: also Claude-only at the consumer (ClaudeCodeBackend-only kwargs in pool.py), but neither has a wizard prompt today (operators hand-edit env). Adding wizard prompts AND gating them is a separate piece of work; out of scope here.
• Re-validation that already-set CLAUDE_* values are removed from install.conf when an operator switches an existing config from claude to goose. Today the wizard preserves prior values across backend changes; whether that should change is a separate question.

Acceptance criteria

• Wizard does not prompt for CLAUDE_AUTOCOMPACT_PCT when agent_backend == "goose"
• Wizard does not prompt for CLAUDE_EFFORT_LEVEL when agent_backend == "goose"
• Wizard does not prompt for CLAUDE_USER when agent_backend == "goose" (regardless of advanced setting)
• When the operator selects goose, the resulting install.conf contains none of the three keys above
• When the operator selects claude, all three prompts continue to fire as today (no behavior regression)
• Existing tests/test_install.py fixtures (all currently use "claude") continue to pass without changes to their input lists
• At least one new test exercises the "goose" path and asserts the prompts are skipped

[dcellison]

Triage by Kai

Triage summary: The config wizard unconditionally prompts for three Claude-specific settings (CLAUDE_AUTOCOMPACT_PCT, CLAUDE_EFFORT_LEVEL, CLAUDE_USER) even when the operator selects the goose backend, where these settings are silently ignored. The fix is straightforward: gate each prompt on agent_backend == 'claude' and add a test covering the goose path.

Priority: medium
Labels applied: enhancement, good first issue