Sweep every effective ordinary-skill mount and operator preference #629

New issue

Open

opened 2026-07-23 08:12:25 +00:00 by coilysiren · 0 comments

coilysiren commented

2026-07-23 08:12:25 +00:00

Owner

Parent and boundaries

Parent: AOS #655
Uses the admission and placement doctrine from AOS #646.
Inbox #249 remains the comprehensive skill-type study. This sweep categorizes current material and does not close that study.

Goal

Sweep every effective ordinary-skill mount, not only the AOS source folder. Preserve recurring operator preferences such as language and framework choices while removing model-native knowledge, misplaced role depth, volatile facts, and repo-only context from the universal candidate catalog.

Snapshot to investigate

AOS currently has 76 ordinary SKILL.md sources.
Other AOS substrate repos contribute additional repo-local sources.
agentic-os-kai contributes the private operator overlay.
AOSH currently contributes 53 hardware, machine, model, home, leaderboard, and repo sources through aggregate_hardware.
Harnesses and plugins can add their own eager frontmatter.
Repo-pointer skills can escape their owner repo through aggregation.

These counts are a dated starting point. The implementation derives the live inventory from canonical manifests and mount configuration.

Required inventory

Derive AOS substrate roots from the committed substrate manifest.
Include every ordinary source that a composed session can actually discover from the public AOS layer, private operator layer, current repo, harness/plugin layer, and generated mount surface.
Treat AOSH as a separate provider, not as an AOS substrate repo.
Identify and remove the aggregate_hardware path that places the AOSH catalog into unrelated sessions. Hardware, machine, model, and leaderboard material becomes repo-local or explicitly retrieved/gated.
Detect repo-pointer skills that are globally aggregated instead of staying repo-local.

Decision for every source

Record one evidence-backed outcome:

keep as ordinary cross-role discovery
move to person context because the preference matters in nearly every session
move to role-scoped COMPOSED.md
keep only as repo-local context or a generated repo-pointer
replace volatile facts with runtime retrieval
move durable explanation to repository docs
replace a mechanical invariant with code or a validator
merge with an overlapping capability
delete as model-native, stale, generic, or unproven

The preference lane explicitly captures positive and negative operator choices for each language and tool family, including framework choices, disliked stacks, and the reason or boundary behind the preference.

Acceptance

Produce a machine-readable inventory with provider repo, source path, mount path, audience, trigger, eager frontmatter bytes/tokens, decision, reason, owner, and freshness event.
Apply the decision across all effective roots, not only the 76 AOS sources.
Ordinary placement remains only where every role benefits from candidate discovery.
Baseline language, framework, HTML, and accessibility knowledge does not become a skill without a local or measured specialist delta.
Smaller-model scaffolding is gated to the model, role, or repo that needs it.
AOSH skills no longer auto-load through agentic-os-kai or the Codex catalog.
Before-and-after reports show source count and eager frontmatter cost by provider and harness.
Validators and docs reflect the final placement rules.
The sweep cross-references any new candidate type back to inbox #249 without closing it.

## Parent and boundaries * Parent: AOS #655 * Uses the admission and placement doctrine from AOS #646. * Inbox #249 remains the comprehensive skill-type study. This sweep categorizes current material and does not close that study. ## Goal Sweep every effective ordinary-skill mount, not only the AOS source folder. Preserve recurring operator preferences such as language and framework choices while removing model-native knowledge, misplaced role depth, volatile facts, and repo-only context from the universal candidate catalog. ## Snapshot to investigate * AOS currently has 76 ordinary SKILL.md sources. * Other AOS substrate repos contribute additional repo-local sources. * agentic-os-kai contributes the private operator overlay. * AOSH currently contributes 53 hardware, machine, model, home, leaderboard, and repo sources through aggregate_hardware. * Harnesses and plugins can add their own eager frontmatter. * Repo-pointer skills can escape their owner repo through aggregation. These counts are a dated starting point. The implementation derives the live inventory from canonical manifests and mount configuration. ## Required inventory * Derive AOS substrate roots from the committed substrate manifest. * Include every ordinary source that a composed session can actually discover from the public AOS layer, private operator layer, current repo, harness/plugin layer, and generated mount surface. * Treat AOSH as a separate provider, not as an AOS substrate repo. * Identify and remove the aggregate_hardware path that places the AOSH catalog into unrelated sessions. Hardware, machine, model, and leaderboard material becomes repo-local or explicitly retrieved/gated. * Detect repo-pointer skills that are globally aggregated instead of staying repo-local. ## Decision for every source Record one evidence-backed outcome: * keep as ordinary cross-role discovery * move to person context because the preference matters in nearly every session * move to role-scoped COMPOSED.md * keep only as repo-local context or a generated repo-pointer * replace volatile facts with runtime retrieval * move durable explanation to repository docs * replace a mechanical invariant with code or a validator * merge with an overlapping capability * delete as model-native, stale, generic, or unproven The preference lane explicitly captures positive and negative operator choices for each language and tool family, including framework choices, disliked stacks, and the reason or boundary behind the preference. ## Acceptance * Produce a machine-readable inventory with provider repo, source path, mount path, audience, trigger, eager frontmatter bytes/tokens, decision, reason, owner, and freshness event. * Apply the decision across all effective roots, not only the 76 AOS sources. * Ordinary placement remains only where every role benefits from candidate discovery. * Baseline language, framework, HTML, and accessibility knowledge does not become a skill without a local or measured specialist delta. * Smaller-model scaffolding is gated to the model, role, or repo that needs it. * AOSH skills no longer auto-load through agentic-os-kai or the Codex catalog. * Before-and-after reports show source count and eager frontmatter cost by provider and harness. * Validators and docs reflect the final placement rules. * The sweep cross-references any new candidate type back to inbox #249 without closing it.

coilyco-ops referenced this issue

2026-07-23 17:54:12 +00:00

[epic] Bound composed-agent context across selected roles and harnesses #655

coilyco-ops changed title from ~~do a sweep on personal prefs for each language / tool category~~ to Sweep every effective ordinary-skill mount and operator preference

2026-07-23 17:55:02 +00:00

coilyco-ops referenced this issue

2026-07-23 17:55:59 +00:00

Enforce generated repo-pointer presence for every eligible managed repo #658

coilyco-ops referenced this issue

2026-07-24 00:40:27 +00:00

Measure Goose startup context before the ordinary-skill refactor #656

coilyco-ops added the

autonomy

async-consult

priority

labels