Skip to content

feat(skills): skills.auto_load pins skills into every new session (salvage #74060/#26840) - #92048

Merged
teknium1 merged 3 commits into
mainfrom
cursor-inspired/skills-auto-load
Sep 15, 2026
Merged

feat(skills): skills.auto_load pins skills into every new session (salvage #74060/#26840)#92048
teknium1 merged 3 commits into
mainfrom
cursor-inspired/skills-auto-load

Conversation

@teknium1

@teknium1 teknium1 commented Aug 22, 2026

Copy link
Copy Markdown
Collaborator

Summary

skills.auto_load in config.yaml now pins skills as fully loaded at the start of every new session — CLI, TUI, gateway, cron, and API agents all included — the "always-on skill" pattern Cursor shipped as Custom Modes in its Aug 19, 2026 changelog.

Salvages #74060 by @ctaylor86 (itself an authorship-preserving salvage of #26840 by @ArcherQAQ, whose feature commit leads the branch) onto current main, adapted to the backgrounded --skills preload that landed since.

skills:
  auto_load:
    - my-workflow
    - github-pr-workflow

Their mechanism vs ours

  • Cursor: pick a skill → "Use as Mode" → the skill stays pinned in the chat, keeping the agent on a playbook.
  • Hermes: profile-scoped config list; skills are resolved once when the agent's system prompt is first built and the exact rendered bytes are reused across model switches, compression, and static-prefix restoration — so prompt caching stays intact. Missing/disabled skills warn and are skipped. --ignore-rules / HERMES_IGNORE_RULES=1 suppresses auto-load with the rest of the auto-injected context. Explicit --skills requests dedupe against auto-load by canonical name.

Changes

  • agent/skill_commands.py: resolve_auto_load_skills() + build_auto_load_prompt() (purpose-built activation note; same disabled-skill gate and Curator usage bump as --skills)
  • agent/system_prompt.py: resolve-once injection in the shared prompt path (all agent surfaces), lifecycle-stable
  • agent/agent_init.py, hermes_cli/cli_agent_setup_mixin.py: per-agent resolve-once cache, CLI hands its pre-resolved result to the lazily built agent
  • cli.py: auto-load resolution after session-ID creation; adaptation to current main — the --skills payload load is now a background thread joined by finalize_preloaded_skills(), so the dedup set is passed into the background loader and the activated-skills display merges at finalize time (auto_load first)
  • hermes_cli/config_defaults.py: skills.auto_load: [] default
  • Docs: website/docs/user-guide/cli.md (new section from the original PR) + website/docs/user-guide/configuration.md

Validation

Check Result
tests/agent/test_skills_auto_load.py + tests/cli/test_cli_preloaded_skills.py 29 passed
Prompt-cache/system-prompt suites (test_system_prompt, test_prompt_builder, test_prompt_caching, test_prompt_cache_boundary, test_prompt_cache_scope) 183 passed
E2E (isolated HERMES_HOME, real skill files, real build_system_prompt_parts on a real AIAgent) injection ✓, missing-skill warning ✓, resolve-once byte stability after on-disk skill mutation ✓, HERMES_IGNORE_RULES suppression ✓
ruff on all touched files clean
scripts/audit_pr_attribution.py --fix all emails mapped

Closes #26840. Closes #74060.

Infographic

skills.auto_load — pinned skills every session

This was referenced Aug 22, 2026
Closed
Closed
@github-actions

github-actions Bot commented Aug 22, 2026 •
edited
Loading

Copy link
Copy Markdown

૮ >ﻌ< ა ci review

ran on 4b7188c — fix(skills): auto-load resolves under the agent's own home a

⚠️ Warnings

OSV vulnerability scan · View job

76 known vulnerabilities found in pinned dependencies.

How to fix:

Review the findings in the Security tab. Update the affected dependencies if a patched version is available.

debug info

CI timings

CI timings · View report · View job

Wall time 6m39s vs 6m9s (+8.1%). 13 job(s) slower, 1 faster, 1 unchanged.

  • OS-specific tests / Windows-only tests: +29.0s
  • Python lints / Windows footguns (blocking): +15.0s
  • Docs Site / docs-site-checks: +13.0s
  • Python tests / e2e: +12.0s
  • OS-specific tests / macOS-only tests: -10.0s

@alt-glitch alt-glitch added type/feature New feature or request comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint comp/cli CLI entry point, hermes_cli/, setup wizard tool/skills Skills system (list, view, manage) area/config Config system, migrations, profiles P3 Low — cosmetic, nice to have sweeper:risk-compatibility Sweeper risk: may break existing users, config, migrations, defaults, or upgrades sweeper:risk-caching Sweeper risk: may break/degrade prompt caching or cache-key stability (invariant) labels Aug 22, 2026
@Sora-bluesky

Sora-bluesky commented Aug 22, 2026

Copy link
Copy Markdown

Cross-reference: #68608 (open, CI green) also changes agent/agent_init.py and agent/system_prompt.py, for an unrelated mechanism: it gates the kanban worker protocol (KANBAN_GUIDANCE) on HERMES_KANBAN_TASK instead of kanban_show tool presence, at the session-static resolution in init_agent and at the fallback in build_system_prompt_parts. The hunks sit in different regions from the skills preload here, and a merge of the two heads is clean. I will rebase #68608 once this lands.

@Enough1122

Enough1122 commented Aug 22, 2026

Copy link
Copy Markdown

AI code review — automated review for reference, author can ignore or act on any point.

This salvage does the two hard things right. First, cache safety by construction: auto-load resolves once per agent lifecycle and the exact rendered bytes are reused across model switches, compression, and static-prefix restoration (_auto_load_skills_resolved + seeded result), with HERMES_IGNORE_RULES captured at first resolution so the prompt stays stable even if env changes mid-session — the system prompt remains byte-stable for the conversation's life, which is the invariant that matters most here. Second, contributor credit: authorship-preserving salvage from #74060/#26840 with the contributors email mapping included.

The CLI adaptation to the backgrounded --skills preload is also correct: pre-resolve after session creation (so ${HERMES_SESSION_ID} substitution sees the real ID), seed the agent cache so the shared prompt path doesn't re-read config, pass the canonical-name exclusion set into the background loader, and merge display names auto_load-first at finalize time. E2E validation against a real agent with real skill files, including byte stability after on-disk mutation, is the verification standard the repo asks for on context-affecting paths.

Minor points:

  1. First-turn latency on non-CLI surfaces: gateway/TUI/cron agents resolve lazily inside the first build_system_prompt_parts, so each configured skill's tree walk (~0.5s per the description for large libraries) lands inside the first turn rather than a startup phase. Probably imperceptible for 1–3 skills; worth a doc line so users with long lists know where the cost goes, and a natural follow-up would be warm-preloading where a background phase already exists (as the CLI does).
  2. Missing/disabled entries warn once per agent and skip — good failure direction; consider including whether a similarly-named skill exists in the warning to shorten the typo-debugging loop.
  3. Config edits take effect next session (resolve happens at first build) — consistent with the repo's deferred-invalidation default; no action needed, just confirming that's intended behavior worth stating in configuration.md (the new section covers usage; a one-line "picks up on next session" note would help).

ArcherQAQ and others added 2 commits September 14, 2026 19:03
@ArcherQAQ @teknium1
feat(skills): skills.auto_load pins skills into every new session's p…
498c2fa 
…rompt

`skills.auto_load: [name, ...]` in config.yaml renders the listed skills
as fully loaded skill blocks in the system prompt of every new agent —
CLI, TUI, gateway, cron and API — the persistent counterpart of `-s`.
Missing or operator-disabled names are warned about and skipped; a
config typo never blocks session start.

Re-implementation of #26840 by @ArcherQAQ (via #74060) against current
main — the original patch targeted the pre-decomposition cli.py /
system_prompt.py god files; design and diagnosis preserved. The loader
reuses `_load_skill_blocks` (same disabled gate and Curator usage bump
as `-s`) instead of a parallel loop.
@ctaylor86 @teknium1
fix(skills): auto-load resolves once per agent and dedupes against -s
756ff37 
The system prompt must stay byte-stable for the life of a conversation:
`_auto_load_skills_result` is seeded in `_SESSION_STATE` and filled on
the FIRST prompt build only (HERMES_IGNORE_RULES captured then too), so
model switches, compression and static-prefix restoration reuse the
exact rendered bytes rather than re-reading config or skill files.

CLI: auto_load renders in the existing background `--skills` preload
thread (real session id for ${HERMES_SESSION_ID}), `-s` names dedupe
against the auto-loaded canonical names via
`build_preloaded_skills_prompt(excluded_loaded_names=)`, the activated
skills line shows auto_load first, and the lazily built agent is seeded
with the pre-resolved bytes. `--ignore-rules` skips auto-load with the
rest of the auto-injected context.

Re-implementation of #74060 by @ctaylor86 against current main.
@teknium1
teknium1 force-pushed the cursor-inspired/skills-auto-load branch from 42f27c7 to 756ff37 Compare September 15, 2026 02:06
@teknium1
fix(skills): auto-load resolves under the agent's own home and skips …
4b7188c 
…internal forks

Why: build_auto_load_prompt read config via ambient load_config_readonly()
and looked skills up under the ambient SKILLS_DIR. Gateway bot threads lose
the HERMES_HOME ContextVar, so a bot profile's pinned skills came from the
launch profile — the docs promise profile scoping. _auto_load_parts now
passes home_override=_agent_home(agent) and build_auto_load_prompt binds it
for config, disabled-list and <home>/skills lookup, the same seam
_skills_prompt uses via skills_dir_override.

_auto_load_parts was unconditional and injected pinned SKILL.md bytes into
delegate children, curator/background_review forks and gateway hygiene
agents; it now mirrors _skills_prompt's gate (nothing without the skills
toolset) and returns [] when skip_context_files is set.

cli.py's HERMES_IGNORE_RULES check used == "1" while system_prompt used
is_truthy_value; both use is_truthy_value now.

Tests stay at 4: the build test asserts the home-scoped resolution, the
ignore-rules test also covers the subagent / no-skills-toolset gates.
@teknium1
teknium1 merged commit 286e723 into main Sep 15, 2026
37 checks passed
@teknium1
teknium1 deleted the cursor-inspired/skills-auto-load branch September 15, 2026 10:34
@s0xn1ck s0xn1ck mentioned this pull request Sep 19, 2026
Open
This was referenced Sep 28, 2026
Closed
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

area/config Config system, migrations, profiles comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint comp/cli CLI entry point, hermes_cli/, setup wizard P3 Low — cosmetic, nice to have sweeper:risk-caching Sweeper risk: may break/degrade prompt caching or cache-key stability (invariant) sweeper:risk-compatibility Sweeper risk: may break existing users, config, migrations, defaults, or upgrades tool/skills Skills system (list, view, manage) type/feature New feature or request

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

6 participants

@teknium1 @Sora-bluesky @Enough1122 @alt-glitch @ArcherQAQ @ctaylor86