Share runtime skill rendering and selection

[jif-oai]

Why

Host and extension skills currently use different rendering and selection helpers. Before one step snapshot can own every source, those helpers need to understand the same neutral catalog entries without losing the mature host behavior.

Doing this separately keeps the dynamic cutover small and reviewable.

What changed

• Generalize the existing bounded skills renderer to source-neutral catalog entries.
• Preserve host path aliases, scope ordering, description shrinking, omission warnings, and the existing render report/metrics.
• Move exact-locator and plain-name selection onto the shared catalog model.
• Share the bounded skill-instruction/read result used by the existing injection path.
• Adapt the current extension path without changing when skills refresh or who owns model context yet.
• Use the mature fair allocator for extension catalogs too, so one oversized description cannot consume the whole list.
• Keep catalog order deterministic and enforce an 8 KB byte cap even for multibyte text.
• Ignore a plain $name when it also names a connector; an exact skill locator still works.

Example

The renderer now accepts entries such as:

host         deploy -> (file: /repo/.codex/skills/deploy/SKILL.md)
executor     deploy -> (environment resource: skill://demo@1/skills/deploy/SKILL.md)
orchestrator search -> (orchestrator resource: skill://orchestrator/search/SKILL.md)

It still applies one hard metadata budget, shortens descriptions before dropping skill names, and keeps host aliases when they save space. Absolute paths and aliases are evaluated independently, so a long absolute path cannot hide a usable alias.

This PR does not change refresh timing or ownership. It does make the existing extension catalog use the same bounded rendering rules as host skills. The next PR uses these shared primitives to remove the parallel owners and refresh the complete snapshot per model step.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 70e6930d9b

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]

Expose activation snapshots to connector consumers

When a thread is initialized with SelectedCapabilityActivation, this path only publishes the computed connectors into ConnectorSnapshotState; the existing runtime still reads ConnectorSnapshot from ExtensionDataInit in core/src/mcp.rs (connector_snapshot_for_thread and the selected-MCP routing check). In that activation mode the selected connector declarations are therefore invisible to tool construction and hosted-app routing even though they were resolved here, so selected plugin apps will be omitted or routed as standalone MCP servers until the active snapshot is stored in the type consumers actually read (or those consumers are switched to ConnectorSnapshotState).

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Refresh the Bazel lockfile for new Rust deps

This patch adds new Rust dependencies/features in Cargo.toml but only commits Cargo.lock; Bazel's crate universe will still be based on the old MODULE.bazel.lock, so the Bazel lock drift check will fail until just bazel-lock-update is run and the resulting lockfile diff is included.

AGENTS.md reference: AGENTS.md:L37-L39

Useful? React with 👍 / 👎.

[jif-oai]

Folded into #29960. The skill source, rendering, selection, and per-step World State changes are now one coherent PR with one runtime boundary.