[bompus]

Adds an opt-in skill-routing modifier (off by default, nothing changes for existing setups): at the start of each task the agent checks what the user is doing and picks the one available skill that best fits, keeps it for follow-ups and small edits, picks again on a new task or changed goal, and uses none (without saying so) when none fits.

Problem. With many skills installed, agents often finish a task without calling one that fits, even when the request matches the skill's description. Hosts load only each skill's description up front, so the model has to notice the match on its own.

Evidence (Claude Code 2.1.318, claude plugin eval, a 60-skill catalog, only the Skill tool allowed, 3 runs per case, tool-call graders on the raw trace, no model judge):

• 18 existing routing cases with no other rules: skill calls on the should-fire prompts rose from 30/42 to 42/42 on Sonnet and from 13/42 to 42/42 on Haiku. Every miss without the block was "no skill called", never a wrong skill.
• 10 new, uncoached prompts with a full set of user rules in both arms (6 that fit a skill, 4 that fit none): Sonnet 13/18 to 18/18, Haiku 11/18 to 14/18 on the five bullets in this change. The four no-skill prompts called no skill in every arm (12/12 each). A rerun of the six-bullet version gave the same counts as its first run.
• Without the bullet that asks the agent to name the skill it used, Sonnet matched the six-bullet version and Haiku was within one trial.

Limits. Small: 3 runs and six fit-a-skill prompts in the second test; the first test's cases were written for their skills and were not held out. Only the first message of a task was tested, so the "keep the same skill for follow-ups" bullet has no measurement. The block was supplied as appended system text.

Why a modifier. Some users prefer to rely on skill descriptions alone, and the cost is a little extra reasoning at the start of each task (about 0.1 to 0.5 extra turns per trial here). It composes after Session purpose.

Changes: rules/modifiers/skill-routing.md, a README modifier row, a CHANGELOG entry under Unreleased. bun test test/ (974 pass, 0 fail) and bun run check pass locally.

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 1 finding(s).

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 1 finding(s).

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 0 finding(s).

Review 3: only high findings and above are posted from here on.

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 0 finding(s).

Review 4: only high findings and above are posted from here on.