feat: stopPatterns option — explicit stop lines count as awaiting the user

[ChakraFusion]

Problem

Some agent setups end every turn with an explicit protocol line, e.g. STOP: NEEDS_DECISION <question> or STOP: BLOCKED <reason>, and keep the todo for that decision open until the user answers. The open-todos nudge then fires a few seconds later, and the agent tends to answer its own question instead of waiting.

In one such setup, 63 of 73 deliberate stop lines had open todos, so the plugin would have nudged after almost every one of them.

Change

A new opt-in option, stopPatterns (array of regex strings, default []):

• When the newest assistant message's visible text matches a pattern, hasPendingUserInput() returns true.
• Every idle nudge then stands down, exactly as it already does for an open question tool call. That covers idle and periodic open-todos nudges, the tool-text scan, action-intent callbacks and the task_complete visible nudge.
• A newer user message clears the gate as usual.
• Reasoning parts are ignored, and invalid regexes are skipped.
• With the default (no patterns), behaviour is unchanged.

Example:

"stopPatterns": ["STOP:\\s*(BLOCKED|NEEDS_DECISION)\\b"]

The two helpers are module-internal (not exported), so the bundled dist still exports only plugin-shaped values (the existing regression test stays green).

Tests

src/index.stop-patterns.test.ts:

• a turn ending in a stop line with open todos is not nudged;
• the same turn is nudged without the option (unchanged default);
• a silent turn end is still nudged with the option set;
• a stop line inside reasoning does not hold back the nudge;
• an invalid pattern is skipped.

Full suite: 599 pass, 0 fail (bun test, bun 1.2.23). README: new row in the options table, and a third gate under "User-input awareness".

🤖 Generated with Claude Code

[Mte90]

We have already Done patterns but not stop and it makes sense