fix(cliproxyapi): rewrite mcp_ references in system prompt + tool descriptions

[NomenAK]

Summary

Follow-up to #2165 (Anthropic-shape body routing and ^mcp_[^_] tool-name rewrite). The original rewrite renamed the tools but left every textual reference to those names unchanged in the system prompt and tool descriptions. The model then reads a prompt that says "use `mcp_call`" while the tool catalog only contains `Mcp_call` — it cannot match the name, gives up on tool-calling, and falls back to plain-text responses.

Anthropic-SDK clients (Capy, claude-cli) render plain text outside of designated message tools as a reasoning trace, so the response visibly lands in the "Thought" panel rather than as a normal assistant message. The fix here closes that loop.

Related Issues

• Closes #
• Follow-up to fix(cliproxyapi): Anthropic-shape body routing and gate compatibility #2165 (this PR was hinted at as a TODO in fix(cliproxyapi): Anthropic-shape body routing and gate compatibility #2165 — "The reverse map below is propagated to chatCore via body._toolNameMap" — but the in-text rewrite was not part of that PR's scope.)

Motivation

Captured live artifact (Capy /v1/messages → OmniRoute → cliproxyapi → Anthropic, 2026-05-12) :

```
// providerRequest (post-#2165 rewrite)
tools: [
{ name: "Mcp_query", description: "Discover MCP tools via mcp_query. Use before mcp_call." },
{ name: "Mcp_call", description: "..." }
]
system: [{ text: "ALWAYS use mcp_query first to discover available tools, THEN use mcp_call." }]
```

```
// providerResponse — Claude does NOT use any tool, emits plain text
content: [{ type: "text", text: "Test received. What would you like me to do?" }]
stop_reason: "end_turn"
```

Same body shape with this PR applied :

```
// providerRequest (post-#2165 + this PR)
tools: [
{ name: "Mcp_query", description: "Discover MCP tools via Mcp_query. Use before Mcp_call." },
{ name: "Mcp_call", description: "..." }
]
system: [{ text: "ALWAYS use Mcp_query first... THEN use Mcp_call." }]
```

```
// providerResponse — Claude correctly uses the tool
content: [{ type: "tool_use", id: "...", name: "Mcp_query", input: {...} }]
stop_reason: "tool_use"
```

(Response-side tool name is restored to the client's original namespace by the existing reverse-map mechanism — Capy sees `mcp_query` back in `tool_use` blocks, no client-visible change.)

Changes

`open-sse/executors/cliproxyapi.ts` :

1. Add `MCP_NAME_REF_RE = /\bmcp_(?=[^_\\s])/g` — the same shape as `MCP_RESERVED_PREFIX_RE` (the existing tool-name detector) but in a global-replace form with a left word-boundary and a right negative-lookahead.
2. In `applyMcpToolNameRewrite`, after collecting the reverse map, apply the regex to :
   • Top-level `system` blocks (both string and array-of-text-blocks forms — Anthropic supports both).
   • Each `tools[*].description`.
3. Skip `messages[*].content` text blocks — those may carry user-supplied text we shouldn't mutate (covered by the "does not mutate message content" test).
4. The rewrite is gated on `reverseMap.size > 0` — bodies without any `mcp_*` tools incur no scan.

Regex shape — edge cases handled (covered by tests)

• `mcp__experimental` → preserved (double underscore : `[^_]` lookahead fails)
• `my_mcp_helper` → preserved (left `\b` doesn't match between two word chars)
• `mcp_ ` (with trailing space) → preserved (`[^_\\s]` lookahead fails)
• Markdown : `(mcp_call)` → rewritten to `(Mcp_call)` (intentional ; `\b` matches the `(` → `m` boundary)

Validation

• `npm run lint` (clean)
• `npm run test:unit` — `tests/unit/cliproxyapi-executor.test.ts` — 43/43 pass (6 new tests added)
• Live smoke against real Capy /v1/messages → CPA → Anthropic : tool-calling restored, response renders as message (not Thought)

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces logic to ensure consistency between tool names and their textual references in system prompts and tool descriptions. It adds a regex-based transformation that converts 'mcp_' prefixes to 'Mcp_' within the system prompt and tool descriptions when MCP tools are being used, preventing potential tool-calling failures due to naming mismatches. Comprehensive unit tests were included to verify the transformation logic, handle various system prompt formats, and ensure that user-supplied message content is preserved. I have no feedback to provide as there were no review comments.

[NomenAK]

Closing after 2026-05-12 empirical bisect.

This PR added two transformations to align model perception of mcp_* tool names:

1. Tool name rewrite (mcp_X → Mcp_X on tool defs, tool_use blocks, tool_choice) — gate-dodge for Anthropic's reserved ^mcp_[^_] namespace.
2. Prose rewrite (same regex applied to system prompt text + tool descriptions) — to keep the model's textual references consistent with the rewritten tool catalog.

Bisect against live Anthropic via CPA cloak (Claude Opus 4.7, full Capy body with 24 tools incl. mcp_query, 5-turn stress test):

• With prose rewrite OFF (name rewrite still ON) → Claude calls the rewritten tools correctly, 5/5 turns. The prose rewrite was modifying client content with no measurable benefit.
• With ALL rewrite OFF (OMNIROUTE_DISABLE_MCP_REWRITE=1) → still HTTP 200 + tool_use(message_user). The ^mcp_ gate apparently does not fire on the current Anthropic+CPA path on this body, though 2026-05-11 char-by-char probe had previously confirmed it. We're keeping the name rewrite (cheap, no harm, gate may resume) but dropping the prose pass.

Fork-side: NomenAK/OmniRoute@cleanup/remove-over-engineered-strips contains the name-only rewrite via commit 3166060 ("drop mcp_ prose rewrite, keep name-only rewrite"). If upstream wants the name-only rewrite as a clean standalone PR, happy to open one — let me know.