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

New issue

Closed

opened 2026-07-23 17:54:12 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

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

Member

Goal

Make composed-agent startup context deterministic, role-bounded, and affordable before and after the ordinary-skill refactor. Use one representative open-source harness for the structural baseline, then expand measurement only where the result would change a concrete decision.

Ownership

AOSH owns Kai’s model-opaque role-intent-harness selections and backend routing.
AOS owns the public-safe projection, context measurement engine, skill admission rules, and substrate definition.
agentic-os-kai owns Kai’s private mcporter inventory and harness projection policy.
Infrastructure owns host and fleet convergence.
Each product repo owns its unconditional AGENTS.md doctrine and repo-local pointer skill.

No shipped AOS runtime fetches upward from AOSH. The hardware, machine, model, and leaderboard skill catalog must not auto-load into unrelated sessions.

Existing anchors

AOS #650 syncs the confirmed AOSH role-intent-harness board into an AOS-owned artifact. It no longer blocks the representative Goose baseline.
AOS #656 measures one fixed ops / operational-decision / goose lane before and after the skill refactor.
AOS #629 sweeps the ordinary skill catalog and personal preferences.
AOS #646 defines the admission and placement rubric.
AOS #653 measures behavioral instruction adherence after structural delivery is proven.
Inbox #249 remains the comprehensive skill-type study. Easy exemplars and this context campaign do not close it.

Workstreams

Measure the representative Goose lane with explicit eager-versus-lazy provenance before and after the ordinary-skill sweep.
Sweep every effective ordinary-skill mount, including private, substrate, plugin, and hardware aggregation roots.
Inventory effective AGENTS.md cascades across substrate and managed product repos, then clip duplicated product doctrine.
Make generated repo-pointer skills a deterministic repo-local presence contract.
Natively mount the high-value tailnet Forgejo MCP where issue work happens while leaving the rest of the MCP inventory lazy unless explicitly selected.
Re-run the Goose structural baseline before interpreting the behavioral results in AOS #653.

Done condition

The fixed Goose lane has deterministic before-and-after startup-context reports using the same role, intent, repository, and CWD.
Reports distinguish global doctrine, repo AGENTS.md cascade, role context, ordinary and composed skill frontmatter, plugins, and eager MCP schemas.
AOSH hardware skills no longer enter every catalog.
Managed product repos carry only repo-specific unconditional doctrine.
Repo-pointer presence and drift are enforced from canonical repo manifests.
Forgejo issue tools are available without AWS SSO detours on the selected harness surfaces.
The campaign records before-and-after eager bytes and tokens and leaves no untracked fleet mutations.
A broader harness matrix remains deferred unless the Goose result identifies a decision that another harness must resolve.

Ticket graph and execution order

AOS #656 establishes the fixed Goose baseline from existing committed composition wiring. AOS #650 can proceed independently.
AOS #629 sweeps every effective ordinary-skill mount and removes AOSH global aggregation.
AOS #657 inventories real AGENTS.md cascades. Infrastructure #607 clips the managed product fleet from that evidence.
AOS #658 enforces repo-pointer presence and repo-local scope. Infrastructure #608 performs the fleet rollout.
agentic-os-kai #748 defines selected native MCP exposure with Forgejo first. Infrastructure #609 converges it across hosts and warded agents.
AOS #656 remeasures the fixed Goose lane after the reductions and selected MCP addition.
AOS #653 evaluates behavioral adherence only after the structural payload is known.

The comprehensive skill-type study remains inbox #249 throughout this campaign.

## Goal Make composed-agent startup context deterministic, role-bounded, and affordable before and after the ordinary-skill refactor. Use one representative open-source harness for the structural baseline, then expand measurement only where the result would change a concrete decision. ## Ownership * AOSH owns Kai’s model-opaque role-intent-harness selections and backend routing. * AOS owns the public-safe projection, context measurement engine, skill admission rules, and substrate definition. * agentic-os-kai owns Kai’s private mcporter inventory and harness projection policy. * Infrastructure owns host and fleet convergence. * Each product repo owns its unconditional `AGENTS.md` doctrine and repo-local pointer skill. No shipped AOS runtime fetches upward from AOSH. The hardware, machine, model, and leaderboard skill catalog must not auto-load into unrelated sessions. ## Existing anchors * AOS #650 syncs the confirmed AOSH role-intent-harness board into an AOS-owned artifact. It no longer blocks the representative Goose baseline. * AOS #656 measures one fixed `ops / operational-decision / goose` lane before and after the skill refactor. * AOS #629 sweeps the ordinary skill catalog and personal preferences. * AOS #646 defines the admission and placement rubric. * AOS #653 measures behavioral instruction adherence after structural delivery is proven. * Inbox #249 remains the comprehensive skill-type study. Easy exemplars and this context campaign do not close it. ## Workstreams * Measure the representative Goose lane with explicit eager-versus-lazy provenance before and after the ordinary-skill sweep. * Sweep every effective ordinary-skill mount, including private, substrate, plugin, and hardware aggregation roots. * Inventory effective `AGENTS.md` cascades across substrate and managed product repos, then clip duplicated product doctrine. * Make generated repo-pointer skills a deterministic repo-local presence contract. * Natively mount the high-value tailnet Forgejo MCP where issue work happens while leaving the rest of the MCP inventory lazy unless explicitly selected. * Re-run the Goose structural baseline before interpreting the behavioral results in AOS #653. ## Done condition * The fixed Goose lane has deterministic before-and-after startup-context reports using the same role, intent, repository, and CWD. * Reports distinguish global doctrine, repo `AGENTS.md` cascade, role context, ordinary and composed skill frontmatter, plugins, and eager MCP schemas. * AOSH hardware skills no longer enter every catalog. * Managed product repos carry only repo-specific unconditional doctrine. * Repo-pointer presence and drift are enforced from canonical repo manifests. * Forgejo issue tools are available without AWS SSO detours on the selected harness surfaces. * The campaign records before-and-after eager bytes and tokens and leaves no untracked fleet mutations. * A broader harness matrix remains deferred unless the Goose result identifies a decision that another harness must resolve. ## Ticket graph and execution order 1. AOS #656 establishes the fixed Goose baseline from existing committed composition wiring. AOS #650 can proceed independently. 2. AOS #629 sweeps every effective ordinary-skill mount and removes AOSH global aggregation. 3. AOS #657 inventories real `AGENTS.md` cascades. Infrastructure #607 clips the managed product fleet from that evidence. 4. AOS #658 enforces repo-pointer presence and repo-local scope. Infrastructure #608 performs the fleet rollout. 5. agentic-os-kai #748 defines selected native MCP exposure with Forgejo first. Infrastructure #609 converges it across hosts and warded agents. 6. AOS #656 remeasures the fixed Goose lane after the reductions and selected MCP addition. 7. AOS #653 evaluates behavioral adherence only after the structural payload is known. The comprehensive skill-type study remains inbox #249 throughout this campaign.

coilyco-ops referenced this issue

2026-07-23 17:54:35 +00:00

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

coilyco-ops referenced this issue

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

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

coilyco-ops referenced this issue

2026-07-23 17:55:18 +00:00

Inventory effective AGENTS.md cascades across substrate and product repos #657

coilyco-ops referenced this issue

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

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