Allow opting out of model invocation for installed skills (large monorepo skill-budget pain)

[echarrod]

Problem

Since v0.8.4 (PR #2099, closing #2098), specify init --ai claude writes disable-model-invocation: false on every generated SKILL.md, and the toggle was removed as "redundant" because the new default opts skills into auto-invocation. There is no opt-out flag.

This is a real problem in large monorepos that already register many domain skills. Claude Code reserves a small fraction of context for the skill-listing budget; when that budget is exhausted, lower-priority skills get truncated and Claude loses visibility on them entirely. Adding the 11 spec-kit skills as model-invocable pushes us over the limit and silently drops other skills we use far more often.

Two specific pain points:

1. Skill-listing budget exhaustion. With ~50 skills already in play (CI helpers, MR workflows, service tooling, etc.), the spec-kit 11 push us past the default skillListingBudgetFraction. The cheapest fix is to make spec-kit skills user-invocable only, so they cost nothing in the listing.
2. Over-triggering on generic prompts. Several spec-kit skill descriptions (speckit-implement, speckit-tasks, speckit-plan) are generic enough that Claude auto-invokes them on prompts that have nothing to do with SDD — e.g. someone asking to "implement" or "plan" something unrelated.

Workaround

We've patched our internally bundled copy to set disable-model-invocation: true on all 11 skills. But every fresh specify init reverts this, so we either lose the fix on upgrade or have to maintain a long-running patch.

Proposal

Add a --no-model-invocation flag to specify init (and persist to the init-options file) that emits disable-model-invocation: true in generated SKILL.md frontmatter. Default behaviour stays as-is (model-invocable). Opt-in users get manual /speckit.* invocation only.

Alternatively, a per-skill override in a config file would also solve it.

Happy to send a PR if the direction is welcome.

[mnriem]

Thanks for the detailed write-up — the skill-listing budget exhaustion problem is real and we'd welcome a PR for this.

One scope adjustment: rather than a top-level --no-model-invocation flag on specify init, this should be delivered as a Claude-specific integration option via --integration-options:

specify init my-project --integration claude --integration-options="--no-model-invocation"

The extension point is ClaudeIntegration.options() — same pattern Codex and Copilot use for their --skills flag. Implementation would be:

1. Add a --no-model-invocation option in ClaudeIntegration.options()
2. Thread the parsed flag through setup() → post_process_skill_content() to flip the hardcoded "false" to "true"
3. Update tests in tests/integrations/test_integration_claude.py

Default behavior stays as-is (model-invocable). The flag should also persist through init-options so specify integration update claude respects it.

Happy to answer questions on the integration internals if you hit any snags.