refactor: delegate model resolution to Claude Code

[mychaelconnolly]

Sources and design precedent

• OpenAI codex-plugin-cc stable companion passes arbitrary model names to Codex and keeps only narrow convenience aliasing.
• OpenAI stable app-server wrapper leaves host-resolved model and effort values unset when not requested.
• Anthropic Claude Code model configuration assigns alias versions, account defaults, managed model restrictions, provider pins, and extended-context eligibility to Claude Code configuration.
• Anthropic programmatic usage documents supported headless invocation and resolved session metadata.
• OpenAI PR #471 and PR #454 explore dynamic catalogs but remain unmerged; they are research context, not stable precedent.

OpenAI alignment: host CLI owns model resolution

This PR makes the stable OpenAI boundary front-and-center: the wrapper owns orchestration and a small user-facing default policy, while the host CLI owns evolving model aliases, provider routing, account eligibility, and discovery.

Like OpenAI's reciprocal plugin, cc-plugin-codex now forwards arbitrary model aliases and full/provider-specific IDs directly to its host CLI. It removes stale version pins and does not force [1m] behavior that depends on account and provider configuration.

Anthropic alignment

Claude Code /model remains the authoritative account/provider-aware picker. Managed model restrictions, provider-specific names, alias versions, and extended-context eligibility stay under Claude Code or administrator control instead of being duplicated in this plugin.

Stack dependency

Depends on #76. This PR contains that Fable-support commit plus model-delegation and uniform-effort commits. Review only this PR's additional work with the stack comparison. After #76 merges, this PR's upstream diff should reduce to those follow-up changes.

What changed

• Removed fixed Opus, Sonnet, and Haiku model mappings.
• Trimmed and forwarded every explicit model value without rewriting it.
• Standardized implicit effort for friendly aliases: Fable, Opus, Sonnet, and Haiku each default to high.
• Kept opus as the model used when --model is omitted.
• Kept explicit --effort authoritative.
• Stopped guessing effort for full or provider-specific model IDs.
• Documented Claude Code /model as the authoritative discovery mechanism.
• Updated unit, integration, E2E, and skill-contract expectations for pass-through routing and the uniform alias effort policy.

Intentional local policy

OpenAI's stable companion leaves omitted model and effort values to Codex. This plugin intentionally keeps a small compatibility policy: omitted model selects the friendly opus alias, and every friendly alias defaults to high effort. Alignment covers host-owned model resolution; it does not require identical product defaults.

Full and provider-specific IDs remain pure pass-through values with no inferred effort because their supported capabilities can vary by provider and configuration.

User impact

Claude Code can advance aliases without a plugin release. Users get one predictable high effort default across Fable, Opus, Sonnet, and Haiku, while explicit effort settings still win. Users can inspect current model choices with /model and pass any listed alias or full ID to the plugin.

Validation

• npm run check — 504 unit tests, 43 integration tests, and 22 E2E tests passed.
• Added integration coverage proving all four friendly aliases reach Claude Code with --effort high.
• PR feat: add Fable using host-owned model routing #76's live no-tools Fable smoke remains valid with pass-through routing.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fc64a6dae5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Keep built-in aliases case-insensitive

For users who type a documented family alias with capitalization (for example $cc:review --model Opus or --model Sonnet), this now forwards Opus/Sonnet literally, while Claude Code documents the aliases as lowercase (opus, sonnet, haiku, fable) and warns that --model values are not prechecked, so mistyped values fail on the first request (checked https://code.claude.com/docs/en/model-config). The old alias lookup was case-insensitive, so these invocations previously worked; normalize only recognized built-in aliases before returning so provider-specific IDs remain untouched.

Useful? React with 👍 / 👎.