fix(agent): demote non-coding skill categories to names-only — never hide skills by OutThisLife · Pull Request #44342 · NousResearch/hermes-agent

OutThisLife · 2026-06-11T15:07:26Z

Problem

The coding posture (#43316) pruned non-coding skill categories from the system prompt's skill index under the default auto mode. The pruning was "discovery-only" in theory — skills_list still returned everything — but that premise failed in practice: models don't reach for skills_list to rediscover what the index stops showing them, and agent-created skills are the model's accumulated project memory (runbooks, learned pitfalls, operating rules). A skill in a pruned category silently vanished mid-project and the agent fell back to a stale sibling skill, with nothing telling the user why.

An earlier revision of this PR gated pruning behind the opt-in focus mode. That was the wrong fix in the other direction: someone opening a worktree doesn't know agent.coding_context exists, so the index-noise reduction would effectively never apply for anyone.

Fix: demote, don't hide

Non-coding categories are now demoted to names-only in the index, in every coding mode (auto/on/focus):

<available_skills>
  autonomous-ai-agents:
    - live-test-hermes: Live-test hermes builds
  gaming [names only]: allthemons10-ops, mc-backup
  github:
    - pr-review: Review pull requests
</available_skills>
(Categories marked [names only] are outside the current coding context, so their
descriptions are omitted — the skills work normally and load with skill_view(name) as usual.)

Every skill name stays in the prompt. Memory-anchored recall ("load <name>") works in all modes — nothing can silently vanish.
The noise reduction still ships by default. Descriptions (the bulk of the index tokens) are dropped for demoted categories.
Deny-list semantics unchanged. Unknown/custom categories and coding-adjacent ones (github, devops, data-science, research, …) keep full entries; the general (non-coding) posture demotes nothing.

API

Renamed to match the honest semantics:

RuntimeMode.hidden_skill_categories() → compact_skill_categories()
coding_hidden_skill_categories() → coding_compact_skill_categories()
build_skills_system_prompt(hidden_categories=…) → compact_categories=…

Tests

test_skill_demotion_scoped_to_coding_posture — demotion applies in auto/on/focus, never in the general posture; coding-adjacent categories exempt.
test_compact_categories_demoted_to_names_only — demoted skill names visible, descriptions dropped, footer note present.
test_compact_categories_demote_nested_and_miss_cache_separately — nested categories demote via their parent; compacted and full renders don't share a cache entry.

tests/agent/test_coding_context.py + tests/agent/test_prompt_builder.py: 177 passed.

github-actions · 2026-06-11T15:08:18Z

🔎 Lint report: `bb/posture-skill-gate` vs `origin/main`

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10717 on HEAD, 10717 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 5609 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

…hide skills Real-world failure with the original index pruning: under the default auto posture, an agent-created ops skill in a demoted category vanished from the prompt's skill index mid-project, and the agent silently fell back to a stale sibling skill instead. The "discovery-only" premise didn't hold — models do not reach for skills_list to rediscover what the index stops showing them, and agent-created skills are the model's accumulated project memory (runbooks, pitfalls, operating rules). Gating pruning behind the opt-in focus mode was the wrong fix too: users opening a worktree don't know the config exists, so the index-noise win would effectively never ship. Instead, the coding posture now DEMOTES non-coding categories rather than hiding them: each demoted category renders as a single names-only line ("gaming [names only]: allthemons10-ops, mc-backup") with a footer note explaining the omitted descriptions. Every skill name stays in the prompt, so memory-anchored recall ("load <name>") keeps working in every mode, while the description noise is still cut. Applies in auto/on/focus alike; the general posture demotes nothing. Deny-list semantics unchanged — unknown/custom categories and coding-adjacent ones keep full entries. API renamed to match the honest semantics: hidden_skill_categories → compact_skill_categories, build_skills_system_prompt(hidden_categories=) → compact_categories=.

…44387) The coding posture's names-only demotion of non-coding skill categories (#44342) applied under the default auto mode, silently changing the skill index for every user in a git repo. Index changes must be opt-in: demotion now only fires under agent.coding_context=focus, alongside the toolset collapse. auto/on leave the skill index untouched; focus semantics are unchanged (demoted, never hidden; deny-list keeps coding-adjacent and custom categories at full entries).

…hide skills (#44342) Real-world failure with the original index pruning: under the default auto posture, an agent-created ops skill in a demoted category vanished from the prompt's skill index mid-project, and the agent silently fell back to a stale sibling skill instead. The "discovery-only" premise didn't hold — models do not reach for skills_list to rediscover what the index stops showing them, and agent-created skills are the model's accumulated project memory (runbooks, pitfalls, operating rules). Gating pruning behind the opt-in focus mode was the wrong fix too: users opening a worktree don't know the config exists, so the index-noise win would effectively never ship. Instead, the coding posture now DEMOTES non-coding categories rather than hiding them: each demoted category renders as a single names-only line ("gaming [names only]: allthemons10-ops, mc-backup") with a footer note explaining the omitted descriptions. Every skill name stays in the prompt, so memory-anchored recall ("load <name>") keeps working in every mode, while the description noise is still cut. Applies in auto/on/focus alike; the general posture demotes nothing. Deny-list semantics unchanged — unknown/custom categories and coding-adjacent ones keep full entries. API renamed to match the honest semantics: hidden_skill_categories → compact_skill_categories, build_skills_system_prompt(hidden_categories=) → compact_categories=.

…44387) The coding posture's names-only demotion of non-coding skill categories (#44342) applied under the default auto mode, silently changing the skill index for every user in a git repo. Index changes must be opt-in: demotion now only fires under agent.coding_context=focus, alongside the toolset collapse. auto/on leave the skill index untouched; focus semantics are unchanged (demoted, never hidden; deny-list keeps coding-adjacent and custom categories at full entries).

…hide skills (NousResearch#44342) Real-world failure with the original index pruning: under the default auto posture, an agent-created ops skill in a demoted category vanished from the prompt's skill index mid-project, and the agent silently fell back to a stale sibling skill instead. The "discovery-only" premise didn't hold — models do not reach for skills_list to rediscover what the index stops showing them, and agent-created skills are the model's accumulated project memory (runbooks, pitfalls, operating rules). Gating pruning behind the opt-in focus mode was the wrong fix too: users opening a worktree don't know the config exists, so the index-noise win would effectively never ship. Instead, the coding posture now DEMOTES non-coding categories rather than hiding them: each demoted category renders as a single names-only line ("gaming [names only]: allthemons10-ops, mc-backup") with a footer note explaining the omitted descriptions. Every skill name stays in the prompt, so memory-anchored recall ("load <name>") keeps working in every mode, while the description noise is still cut. Applies in auto/on/focus alike; the general posture demotes nothing. Deny-list semantics unchanged — unknown/custom categories and coding-adjacent ones keep full entries. API renamed to match the honest semantics: hidden_skill_categories → compact_skill_categories, build_skills_system_prompt(hidden_categories=) → compact_categories=.

…ousResearch#44387) The coding posture's names-only demotion of non-coding skill categories (NousResearch#44342) applied under the default auto mode, silently changing the skill index for every user in a git repo. Index changes must be opt-in: demotion now only fires under agent.coding_context=focus, alongside the toolset collapse. auto/on leave the skill index untouched; focus semantics are unchanged (demoted, never hidden; deny-list keeps coding-adjacent and custom categories at full entries).