feat(claude-code-settings): add four undocumented settings keys

[johnjspiteri]

Adds four settings keys that the Claude Code CLI reads but this schema does not declare.

Verified against Claude Code 2.1.236. Each key is declared in the binary's own embedded zod settings schema with a .describe() string, and the descriptions below are taken from there rather than inferred from usage:

Key	Type	Source
promptSuggestionEnabled	boolean	promptSuggestionEnabled: z.boolean().optional().describe("When false, prompt suggestions are disabled. When absent or true, prompt suggestions are enabled.")
skipWorkflowUsageWarning	boolean	skipWorkflowUsageWarning: z.boolean().optional().describe("@internal Whether the user has accepted the multi-agent workflow usage warning. Until set, auto permission mode prompts before running a workflow.")
precomputeCompactionEnabled	boolean	precomputeCompactionEnabled: z.boolean().optional().describe("Precompute the compaction summary in the background before it is needed. Only applies when auto-compact is on.")
modelProposedGoals	enum	modelProposedGoals: z.enum(["auto", "alwaysAsk", "disabled"]).optional().catch(undefined).describe("@internal Controls the ProposeGoal tool ...")

None of the four appears on https://code.claude.com/docs/en/settings, so each description carries the UNDOCUMENTED. prefix per the Undocumented Features section of CONTRIBUTING.md.

A note on the two @internal keys: they are marked internal in the binary, but both are user-facing in practice. precomputeCompactionEnabled is exposed in /config under Experimental, and skipWorkflowUsageWarning is persisted to settings.json when a user accepts the workflow warning dialog — so both show up in real config files and currently validate as unknown keys.

I deliberately excluded a fifth candidate, copyOnSelect. It is read as copyOnSelect ?? true and appears in the /config UI, but I could not find a zod declaration for it in the settings schema, so I could not confirm it is a settings key rather than state held elsewhere. Happy to add it in a follow-up if a maintainer can confirm.

Validation:

• node ./cli.js check --schema-name=claude-code-settings.json passes (pre-checks and Ajv).
• npx prettier --check passes.

Context: the Claude Code settings documentation points to this schema as "the published JSON schema" while noting that it "can lag behind the newest CLI releases". I have opened anthropics/claude-code#94232 upstream asking Anthropic to publish a schema generated from the CLI's own validation source, which would make this kind of manual backfill unnecessary.

[github-actions]

Thanks for the PR!

This section of the codebase is owned by @domdomegg, @bogini, @sarahdeaton, and @ant-kurt - if they write a comment saying "LGTM" then it will be merged.

[johnjspiteri]

Pushed a follow-up commit adding test coverage for the four keys.

I had initially only run node ./cli.js check as described under How to edit an existing JSON Schema, which passed, and missed that CI also runs node ./cli.js coverage. That flagged three real gaps in my change: test completeness, enum coverage and default value coverage.

Now covered:

• All three modelProposedGoals values across basic-config.json (auto), enum-coverage.json (alwaysAsk) and complete-config.json (disabled).
• "sometimes" added to invalid-enum-values.json as a negative test.
• Each boolean tested at both values, so the non-default case is exercised rather than only the default.

node ./cli.js coverage --schema-name=claude-code-settings.json now reports 8 passed, 0 failed.

One note in case it is useful for other contributors: the two pre-existing prettier warnings under src/test/prettierrc/ are untouched by this PR.