perf(skills): list categories in system prompt

[haran2001]

Summary

• Replace the full per-skill system-prompt index with a category-level index and skill counts.
• Instruct agents to call skills_list(category=...) before loading matching skills with skill_view(...).
• Update prompt-builder tests for category-only rendering, filtering, and conditional skill visibility.

Test Plan

• python -m pytest tests/agent/test_prompt_builder.py -q -o 'addopts='

Closes #26786

[teknium1]

Thanks for targeting a meaningful fixed prompt-cost source. The full-index premise still exists on current main (agent/prompt_builder.py:1660-1667), but the proposed global category-only strategy conflicts with a later verified design decision.

Problems

• Current main intentionally keeps every skill name visible: agent/prompt_builder.py:1622-1628 and commit ee1a744ac document that models did not reliably use skills_list to rediscover omitted agent-created skills. The PR's skills_list(category=...) replacement at agent/prompt_builder.py:1177-1179 reintroduces that failure mode.
• Renaming the block at agent/prompt_builder.py:1199-1201 leaves agent/context_breakdown.py:15 and hermes_cli/prompt_size.py:21 unable to find and attribute the skills index.

Suggested changes

• Preserve name visibility and reduce description noise only, following the current names-only demotion contract and tests at tests/agent/test_prompt_builder.py:429-475.
• Retain the existing wrapper tag or update all block consumers and their tests.

This is an automated hermes-sweeper review.

[teknium1]

Current main intentionally does not rely on skills_list to rediscover omitted skills: ee1a744ac records a real failure where an agent-created skill disappeared from the index and was not rediscovered. Preserve skill names rather than making this category-only.

[teknium1]

This renamed wrapper is not consumed by the current prompt diagnostics: agent/context_breakdown.py:15 and hermes_cli/prompt_size.py:21 only match <available_skills>. Update those consumers and tests, or retain the existing tag.

[kshitijk4poor]

Closing on the design axis after checking the current skills-index implementation: main deliberately keeps every skill visible in the index at all times — the never-hide decision (ee1a744) demotes long categories to names-only under pressure but never removes entries, precisely because models do NOT reliably reach for skills_list to rediscover what the index stops showing them. This PR's categories-only index (with a 'call skills_list(category=...) to inspect' nudge) is the exact pattern that decision rejected: in practice the model skips the extra round-trip and misses skills. The category-count suffix rider is neutral-to-negative once full entries stay (extra tokens, no signal). Thank you — the token concern is real, and the compact-guidance angle (#72813) is the shape of it we can take.