chore(release): v1.1.4 — MCP annotations + tool description polish

[fengyat]

Why

Glama's score page flagged the same root cause across all four tools — no MCP annotations declared, so reviewer scoring penalizes the Behavior dimension (2-4 out of 5) for every tool. Sample reviewer comment from orcarouter_model_card:

No annotations are present, so the description bears full burden. It indicates a read-like operation (get info) but does not explicitly state it is non-destructive, mention authorization needs, rate limits, or potential error conditions.

This PR closes that gap. It also tightens descriptions for the other low-scoring dimensions (Usage Guidelines, Completeness, Parameters).

What changes

File	Change
src/tools/types.ts	Add ToolAnnotations interface + optional annotations field on ToolDefinition.
src/server.ts	Pass annotations through to the tools/list response (spread ...(t.annotations ? { annotations: t.annotations } : {})).
src/tools/chat.ts	Add annotations (readOnly:false, idempotent:false, openWorld:true) + describe models fallback chain interaction, isError:true surface, ORCAROUTER_API_KEY requirement.
src/tools/models_list.ts	Add annotations (readOnly:true, idempotent:true, openWorld:true) + enumerate returned fields + note filter composition.
src/tools/model_card.ts	Add annotations (read-only / idempotent / open-world) + explain when to use vs. models_list + note isError:true surface.
src/tools/providers_list.ts	Add annotations + enumerate returned fields + note zero-arg call shape.
test/server.test.ts	New case: tools/list response includes MCP annotations on every tool asserts the per-tool annotation map (4 tools × 5 fields).
package.json, server.json	Version bump 1.1.3 → 1.1.4.
CHANGELOG.md	v1.1.4 entry.

Per-tool annotations table

Tool	readOnly	destructive	idempotent	openWorld
orcarouter_chat	F	F	F	T
orcarouter_models_list	T	F	T	T
orcarouter_model_card	T	F	T	T
orcarouter_providers_list	T	F	T	T

Reasoning: catalog tools are read-only queries against the public OrcaRouter catalog API — same inputs give the same outputs (idempotent), and they reach an external service (open world). The chat tool reaches an external service too, but isn't read-only (it triggers an LLM completion, which the upstream provider may count, log, and bill), and isn't idempotent (same prompt produces different completions).

What it unlocks downstream

• Glama TDQS: Behavior dimension should jump from 2-4 to ~5 across all tools. Expected overall TDQS average: 4.2 → ~4.6. Already at grade A; this firms up the score.
• MCP clients with annotation-aware UI: Claude Desktop, Cursor, Windsurf, etc. can render the catalog tools as "safe to call without confirmation" and the chat tool as "calls an external LLM".
• Better LLM tool selection: agents that consume tools/list get richer signal about each tool's character.

Verification

• npm run typecheck — clean
• npm test — 116/116 (115 existing + 1 new annotations test)
• npm run build — 26.51 KB ESM bundle, clean
• server.json validates against the official MCP Registry schema

Release plan

After merge, push the version tag and the CI auto-publish workflow (introduced in #3) takes over:

git tag v1.1.4
git push origin v1.1.4
# CI: typecheck → test → build → npm publish → mcp-publisher publish

Once npm + MCP Registry are updated, Glama auto-syncs from GitHub (typically within a few hours), re-runs TDQS scoring, and the new annotations land in the public score.

[fengyat]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

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".