Router accepts mismatched subagent model tags and strips user text

[MicroMilo]

Summary

The router subagent tag parser accepts malformed tags whose opening and closing tag families do not match. For example, <ccr-subagent-model>...</pilotdeck-subagent-model> is currently treated as a valid control tag.

This can silently turn ordinary user-provided text into router control input: detectSubagent marks the request as a subagent request and extracts a model hint, while stripSubagentTagFromMessages removes the malformed text from the user message.

Code path

• src/router/scenario/subagentDetector.ts:3 defines SUBAGENT_TAG_PATTERN with independent alternations for the opening and closing tag families.
• src/router/scenario/subagentDetector.ts:17 uses the regex in detectSubagent and sets taggedInUserMessage, modelHint, and isSubagent.
• src/router/scenario/subagentDetector.ts:53 uses the same regex in stripSubagentTagFromMessages and removes the matched text.
• src/router/scenario/decideScenario.ts:26 consumes the detection result for router scenario selection.
• src/router/RouterRuntime.ts:510 enables subagent tag stripping when the decision is considered a subagent request.

Steps to reproduce

Tested against local HEAD dbb2b416.

From the repository root:

node --import tsx --input-type=module <<'NODE'
import {
  detectSubagent,
  stripSubagentTagFromMessages,
} from './src/router/scenario/subagentDetector.ts';

const messages = [
  {
    role: 'user',
    content: [
      {
        type: 'text',
        text: '<ccr-subagent-model>openai/gpt-4o</pilotdeck-subagent-model> explain this markup',
      },
    ],
  },
];

const tools = [{ name: 'agent', inputSchema: {} }];

console.log(JSON.stringify(detectSubagent(messages, tools, true)));
console.log(JSON.stringify(stripSubagentTagFromMessages(messages)[0].content[0].text));
NODE

Current output:

{"isSubagent":true,"modelHint":"openai/gpt-4o","missingAgentTool":false,"taggedInUserMessage":true}
" explain this markup"

Expected behavior

Only same-family tag pairs should be accepted as router control tags, for example:

<pilotdeck-subagent-model>...</pilotdeck-subagent-model>
<ccr-subagent-model>...</ccr-subagent-model>

A cross-family pair should be ignored by detectSubagent and preserved by stripSubagentTagFromMessages:

<ccr-subagent-model>...</pilotdeck-subagent-model>

Actual behavior

The cross-family pair is accepted as valid. This causes:

• detectSubagent to return isSubagent: true and taggedInUserMessage: true for malformed user text;
• modelHint to be extracted from the malformed tag body;
• stripSubagentTagFromMessages to delete text that was not a valid control tag.

Existing coverage

I searched existing issues for subagentDetector, subagent-model, cross-family, and malformed subagent. The only nearby matches were subagent configuration issues (#247 and #248), which do not cover this tag parser behavior.

Suggested fix

Bind the closing tag family to the opening tag family, for example with a backreference:

const SUBAGENT_TAG_PATTERN =
  /<(pilotdeck|ccr)-subagent-model>([\s\S]+?)<\/\1-subagent-model>/i;

With that shape, the model hint capture would move from match[1] to match[2].

Suggested tests

• same-family pilotdeck tags are detected and stripped;
• same-family ccr tags are detected and stripped;
• cross-family tags are not detected as subagent control tags;
• cross-family tags are preserved by stripSubagentTagFromMessages;
• ordinary user text containing malformed XML-like snippets is not silently removed.

---

Submitted with Codex.