feat(doctor): deeper config validation — walk-up, overlays, unknown keys

[emal-avala]

Summary

Extends /doctor with deeper config validation. Catches the class of "my config silently did nothing" problems that #[serde(default)] tolerance otherwise hides.

New checks

Check	Category	Fires when
config:project_file via walk-up	Pass/Warn/Fail	Project settings found via ancestor walk, reports parse failures
config:project_local_file	Pass/Warn/Fail	.agent/settings.local.toml overlay detected, reports parse failures
Unknown key detection (all three config files)	Warn	Top-level section not in the known allow-list — catches typos like [permisions]
env:api_key_precedence	Warn	Multiple provider API-key env vars set; tells user which ones are being ignored

Known allow-list: api, ui, permissions, features, security, sandbox, mcp_servers, hooks, subagents, allowed_project_dirs.

Example session

Before: user has a sub-crate cwd in a monorepo and a typo'd [permisions] section in the repo-root config — /doctor either misses the file entirely or silently passes.

After:

  ✓ Project config: /repo/.agent/settings.toml
  ! Project config /repo/.agent/settings.toml has unknown key(s): permisions
  ✓ Project local overrides: /repo/.agent/settings.local.toml
  ! Multiple API-key env vars set: OPENAI_API_KEY, ANTHROPIC_API_KEY. Resolver uses the first — others ignored.

Why this matters

#[serde(default)] on the root config struct silently drops unknown fields — the right call at load time (forwards-compat with newer tools writing a shared file) but it converts config typos into total-silent no-ops. The allow-list guard test fails the moment the loader grows a new section without an accompanying /doctor entry, so this stays in sync.

Tests (5 new)

• Known sections produce an empty unknown-key list
• Typo section flagged correctly
• Unknowns sorted for stable output
• Allow-list completeness guard: fails the moment the loader grows a section that isn't taught here
• Non-table TOML root is a no-op

Full services::diagnostics suite: 21 pass, 0 fail. Clippy clean.

Test plan

• cargo test -p agent-code-lib --lib services::diagnostics
• cargo clippy --workspace --tests --no-deps -- -D warnings
• cargo fmt --all --check

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.